COMPUTER-IMPLEMENTED SYSTEMS AND METHODS FOR TIME SERIES EXPLORATION

Abstract

Systems and methods are provided for analyzing unstructured time stamped data. A distribution of time-stamped data is analyzed to identify a plurality of potential time series data hierarchies for structuring the data. An analysis of a potential time series data hierarchy may be performed. The analysis of the potential time series data hierarchies may include determining an optimal time series frequency and a data sufficiency metric for each of the potential time series data hierarchies. One of the potential time series data hierarchies may be selected based on a comparison of the data sufficiency metrics. Multiple time series may be derived in a single-read pass according to the selected time series data hierarchy. A time series forecast corresponding to at least one of the derived time series may be generated.

Description

TECHNICAL FIELD

This document relates generally to time series analysis, and more particularly to structuring unstructured time series data into a hierarchical structure.

BACKGROUND

Many organizations collect large amounts of transactional and time series data related to activities, such as time stamped data associated with physical processes, such as product manufacturing or product sales. These large data sets may come in a variety of forms and often originate in an unstructured form that may include only a collection of data records having data values and accompanying time stamps.

Organizations often wish to perform different types of time series analysis on their collected data sets. However, certain time series analysis operators (e.g., a predictive data model for forecasting product demand) may be configured to operate on hierarchically organized time series data. Because an organization's unstructured time stamped data sets are not properly configured, the desired time series analysis operators are not able to properly operate on the organization's unstructured data sets.

SUMMARY

In accordance with the teachings herein, systems and methods are provided for analyzing unstructured time stamped data of a physical process in order to generate structured hierarchical data for a hierarchical time series analysis application. A plurality of time series analysis functions are selected from a functions repository. Distributions of time stamped unstructured data are analyzed to identify a plurality of potential hierarchical structures for the unstructured data with respect to the selected time series analysis functions. Different recommendations for the potential hierarchical structures for each of the selected time series analysis functions are provided, where the selected time series analysis functions affect what types of recommendations are to be provided, and the unstructured data is structured into a hierarchical structure according to one or more of the recommended hierarchical structures, where the structured hierarchical data is provided to an application for analysis using one or more of the selected time series analysis functions.

As another example, a system for analyzing unstructured time stamped data of a physical process in order to generate structured hierarchical data for a hierarchical time series analysis application includes one or more processors and one or more computer-readable storage mediums containing instructions configured to cause the one or more processors to perform operations. In those operations, a plurality of time series analysis functions are selected from a functions repository. Distributions of time stamped unstructured data are analyzed to identify a plurality of potential hierarchical structures for the unstructured data with respect to the selected time series analysis functions. Different recommendations for the potential hierarchical structures for each of the selected time series analysis functions are provided, where the selected time series analysis functions affect what types of recommendations are to be provided, and the unstructured data is structured into a hierarchical structure according to one or more of the recommended hierarchical structures, where the structured hierarchical data is provided to an application.

As a further example, a computer program product for analyzing unstructured time stamped data of a physical process in order to generate structured hierarchical data for a hierarchical time series analysis application, tangibly embodied in a machine-readable non-transitory storage medium, includes instructions configured to cause a data processing system to perform a method. In the method, a plurality of time series analysis functions are selected from a functions repository. Distributions of time stamped unstructured data are analyzed to identify a plurality of potential hierarchical structures for the unstructured data with respect to the selected time series analysis functions. Different recommendations for the potential hierarchical structures for each of the selected time series analysis functions are provided, where the selected time series analysis functions affect what types of recommendations are to be provided, and the unstructured data is structured into a hierarchical structure according to one or more of the recommended hierarchical structures, where the structured hierarchical data is provided to an application.

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 is a block diagram depicting a computer-implemented time series exploration system.

FIG. 2 is a block diagram depicting a time series exploration system configured to perform a method of analyzing unstructured hierarchical data for a hierarchical time series analysis application.

FIG. 3 is a block diagram depicting data structuring recommendation functionality.

FIG. 4 is a block diagram depicting selection of different recommended potential hierarchical structures based on associated time series analysis functions.

FIG. 5 is a flow diagram depicting performing a hierarchical analysis of the potential hierarchical structures.

FIG. 6 is a block diagram depicting a data structuring graphical user interface (GUI) for incorporating human judgment into a data structuring operation.

FIG. 7 is a block diagram depicting a wizard implementation of a data structuring GUI.

FIG. 8 is a block diagram depicting a data structuring GUI providing multiple data structuring process flows to a user for comparison.

FIG. 9 is a block diagram depicting a structuring of unstructured time stamped data in one pass through the data.

FIG. 10 is a block diagram depicting example SAS® procedures that can be combined to implement a method of analyzing unstructured time stamped data.

FIG. 11 depicts a block diagram depicting a time series explorer desktop architecture built on a SAS Time Series Explorer Engine.

FIG. 12 depicts a block diagram depicting a time series explorer enterprise architecture built on a SAS Time Series Explorer Engine.

FIGS. 13-19 depict example graphical interfaces for viewing and interacting with unstructured time stamped data, structured time series data, and analysis results.

FIG. 20 depicts an example internal representation of the panel series data.

FIG. 21 depicts reading/writing of the panel series data.

FIG. 22 depicts an example internal representation of the attribute data.

FIG. 23 depicts reading/writing of the attribute data.

FIG. 24 depicts an internal representation of derived attribute data.

FIG. 25 depicts reading/writing of derived attribute data.

FIGS. 26A, 26B, and 26C depict example systems for use in implementing a time series exploration system.

DETAILED DESCRIPTION

FIG. 1 is a block diagram depicting a computer-implemented time series exploration system. A time series exploration system 102 facilitates the analysis of unstructured time stamped data, such as data related to a physical process, in order to generate structured hierarchical time series data for a hierarchical time series application. For example, the time series exploration system 102 may receive unstructured (e.g., raw time stamped) data from a variety of sources, such as product manufacturing or product sales databases (e.g., a database containing individual data records identifying details of individual product sales that includes a date and time of each of the sales). The unstructured data may be presented to the time series exploration system 102 in different forms such as a flat file or a conglomerate of data records having data values and accompanying time stamps. The time series exploration system 102 can be used to analyze the unstructured data in a variety of ways to determine the best way to hierarchically structure that data, such that the hierarchically structured data is tailored to a type of further analysis that a user wishes to perform on the data. For example, the unstructured time stamped data may be aggregated by a selected time period (e.g., into daily time period units) to generate time series data and structured hierarchically according to one or more dimensions (attributes, variables). Data may be stored in a hierarchical data structure, such as a MOLAP database, or may be stored in another tabular form, such as in a flat-hierarchy form.

The time series exploration system 102 can facilitate interactive exploration of unstructured time series data. The system 102 can enable interactive structuring of the time series data from multiple hierarchical and frequency perspectives. The unstructured data can be interactively queried or subset using hierarchical queries, graphical queries, filtering queries, or manual selection. Given a target series, the unstructured data can be interactively searched for similar series or cluster panels of series. After acquiring time series data of interest from the unstructured data, the time series data can be analyzed using statistical time series analysis techniques for univariate (e.g., autocorrelation operations, decomposition analysis operations), panel, and multivariate time series data. After determining patterns in selected time series data, the time series data can be exported for subsequent analysis, such as forecasting, econometric analysis, pricing analysis, risk analysis, time series mining, as well as others.

Users 104 can interact with a time series exploration system 102 in a variety of ways. For example, FIG. 1 depicts at an environment wherein users 104 can interact with a time series exploration system 104 hosted on one or more servers 106 through a network 108. The time series exploration system 102 may analyze unstructured time stamped data of a physical process to generate structured hierarchical data for a hierarchical time series analysis application. The time series exploration system 102 may perform such analysis by accessing data, such as time series analysis functions and unstructured time stamped data, from a data store 110 that is responsive to the one or more servers 106.

FIG. 2 is a block diagram depicting a time series exploration system configured to perform a method of analyzing unstructured hierarchical data for a hierarchical time series analysis application. The time series exploration system 202 receives a selection of one or more time series analysis functions 204, such as time series analysis functions 204 that are customizable by a user that are stored in a function repository 206. The time series exploration system 202 accesses unstructured time-stamped data 208 and analyzes distributions of the unstructured time stamped data 208 to identify a plurality of potential hierarchical structures for the unstructured data with respect to the selected time series analysis functions (e.g., a selected function utilizes data according to item type and regions, and the system suggests a hierarchy including item type and region attributes (dimensions) as levels). The time series exploration system uses those potential hierarchical structures to provide different recommendations of which potential hierarchical structures are best suited according to selected time series analysis functions 204. The unstructured data 208 is structured into a hierarchical structure according to one or more of the recommended hierarchical structures (e.g., an automatically selected potential hierarchical structure, a potential hierarchical structure selected by a user) to form structured time series data 210. Such structured time series data 210 can be explored and otherwise manipulated by a time series exploration system 202, such as via data hierarchy drill down exploration capabilities, clustering operations, or search operations, such as a faceted search where data is explored across one or multiple hierarchies by applying multiple filters across multiple dimensions, where such filters can be added or removed dynamically based on user manipulation of a GUI. The structured time series data 210 is provided to a hierarchical time series analysis application 212 for analysis using one or more of the selected time series analysis functions 204 to generate analysis results 214. Such results 214 may also be in a hierarchical form such that drill down and other data exploration operations may be performed on the analysis results 214.

FIG. 3 is a block diagram depicting data structuring recommendation functionality. A time series exploration system 302 receives unstructured time stamped data 304 to process as well as a selection of time series analysis functions 306 (e.g., from a function repository 308) to be applied to the unstructured time stamped data 304. The time series exploration system 302 analyzes the unstructured time stamped data 304 to provide recommendations as to how the unstructured time stamped data should be structured to result in a best application of the time series analysis functions. For example, the data structuring recommendations functionality 310 may perform certain data distribution, time domain frequency analysis, and time series data mining on the unstructured time stamped data 304 to provide a recommendation of a hierarchical structure and data aggregation frequency for structuring the data for analysis by a time series analysis function. Other data analysis techniques may be used by the data structuring recommendations functionality 310, such as cluster analysis (e.g., a proposed cluster structure is provided via a graphical interface, and a statistical analysis is performed on structured data that is structured according to the selected cluster structure).

Based on the recommendations made by the data structuring recommendations functionality, the unstructured time stamped data 304 is structured to form structured time series data 312. For example, the recommendation for a particular time series analysis function and set of unstructured time stamped data may dictate that the unstructured time stamped data be divided into a number of levels along multiple dimensions (e.g., unstructured time stamped data representing sales of products for a company may be structured into a product level and a region level). The recommendation may identify a segmentation of the time series data, where such a segmentation recommendations provides one or more options for dividing the data based on a criteria, such as a user defined criteria or a criteria based upon statistical analysis results. The recommendation may further identify an aggregation frequency (e.g., unstructured time stamped data may be aggregated at a monthly time period). The structured time series data 312 is then provided to a hierarchical time series analysis application 314, where a selected time series analysis 306 function is applied to the structured time series data 312 to generate analysis results 316 (e.g., a time slice analysis display of the structured data or analysis results).

FIG. 4 is a block diagram depicting selection of different recommended potential hierarchical structures based on associated time series analysis functions. A time series exploration system 402 receives unstructured time stamped data 404 as well as a number of time series analysis functions 406 to be performed on the unstructured time stamped data. Data structuring recommendations functionality 408 provides recommendations for structures for the unstructured time series data 404 and may also provide candidate aggregation frequencies. As indicated at 410, the time series analysis functions 406 that are selected can have an effect on the recommendations made by the data structuring recommendations functionality 408. For example, the data structuring recommendations functionality 408 may recommend that the unstructured time stamped data 404 be structured in a first hierarchy based on first dimensions at a first aggregation time period because such a structuring will enable a first time series analysis function to operate optimally (e.g., that first function will provide results in a fastest time, with a least number of memory accesses, with a least number of processing cycles). When the data structuring recommendations functionality 408 considers a second time series analysis function, the recommendations functionality 408 may recommend a second, different hierarchical structure and a second, different aggregation time period for the unstructured time stamped data 404 to benefit processing by the second time series analysis function.

Upon selection of a hierarchical structure and an aggregation frequency for a particular time series analysis function, the time series exploration system 402 structures the unstructured time stamped data 404 accordingly, to generate structured time series data 412. The structured time series data 412 is provided to a hierarchical time series analysis application 414 that applies the particular time series analysis function to the structured time series data 412 to generate analysis results 416.

FIG. 5 is a flow diagram depicting automatically performing a hierarchical analysis of the potential hierarchical structures for use with a particular time series analysis function. Performing the hierarchical analysis for a potential hierarchical structure 502 includes aggregating the unstructured data according to the potential hierarchical structure and according to a plurality of candidate frequencies 504. An optimal frequency for the potential hierarchical structure is determined at 506 by analyzing at 508 a distribution of data across the aggregations from 504 for each candidate frequency to determine a candidate frequency data sufficiency metric 510. The analysis at 508 is repeated for each of the candidate frequencies to generate a plurality of candidate frequency data sufficiency metrics. An optimal frequency for the potential hierarchical structure is selected at 512 based on the sufficiency metrics 510.

The data sufficiency metric 510 that is associated with the selected optimal frequency is used to determine a data sufficiency metric the potential hierarchical structure at 514. Thus, the data sufficiency metric of the best candidate frequency may be imputed to the potential hierarchical structure or otherwise used to calculate a sufficiency metric for the potential hierarchical structure, as the potential hierarchical structure will utilize the optimal frequency in downstream processing and comparison. The performing of the hierarchical analysis 502 to identify an optimal frequency for subsequent potential hierarchical structures is repeated, as indicated at 516. Once all of the potential hierarchical structures have been analyzed, a data structure includes an identification of the potential hierarchical structures, an optimal frequency associated with each of the potential hierarchical structures, and a data sufficiency metric associated with each of the potential hierarchical structures.

At 518, one of the potential hierarchical structures is selected as the selected hierarchical structure for the particular time series analysis function based on the data sufficiency metrics for the potential hierarchical structures. The selected hierarchical structure 520 and the associated selected aggregation frequency 522 can then be used to structure the unstructured data for use with the particular time series analysis function.

The structured time series data can be utilized by a time series analysis function in a variety of ways. For example, all or a portion of the structured time series data may be provided as an input to a predictive data model of the time series analysis function to generate forecasts of future events (e.g., sales of a product, profits for a company, costs for a project, the likelihood that an account has been compromised by fraudulent activity). In other examples, more advanced procedures may be performed. For example, the time series analysis may be used to segment the time series data. For instance, the structured hierarchical data may be compared to a sample time series of interest to identify a portion of the structured hierarchical data that is similar to the time series of interest. That identified similar portion of the structured hierarchical data may be extracted, and the time series analysis function operates on the extracted similar portion.

In another example, the structured hierarchical data is analyzed to identify a characteristic of the structured hierarchical data (e.g., a seasonal pattern, a trend pattern, a growth pattern, a delay pattern). A data model is selected for a selected time series analysis function based on the identified characteristic. The selected time series analysis function may then be performed using the selected data model. In a different example, the selected time series analysis function may perform a transformation or reduction on the structured hierarchical data and provide a visualization of the transformed or reduced data. In a further example, analyzing the distributions of the time-stamped unstructured data may include applying a user defined test or a business objective test to the unstructured time stamped data.

Structuring unstructured time series data can be performed automatically (e.g., a computer system determines a hierarchical structure and aggregation frequency based on a set of unstructured time series data and an identified time series analysis function). Additionally, the process of structuring of the unstructured time series data may incorporate human judgment (e.g., structured judgment) at certain points or throughout. FIG. 6 is a block diagram depicting a data structuring graphical user interface (GUI) for incorporating human judgment into a data structuring operation. A time series exploration system 602 receives unstructured time stamped data 604 and a plurality of time series analysis functions 606. Based on time series analysis functions 606 that are selected, data structuring recommendations functionality 608 provides recommendations for ways to structure the unstructured time stamped data 604 for analysis by the time series analysis functions 606. A data structuring GUI 610 provides an interface for a user to provide input to the process. For example, the user may be provided with a number of recommendations for ways to structure the unstructured data 604 for analysis by a particular time series analysis function 606. The recommendations may include a metric that indicates how well each of the recommended structuring strategies is expected to perform when analyzed by the particular time series analysis function 606. The user can select one of the recommendations via the data structuring GUI 610, and the time series exploration system 602 structures the unstructured time stamped data 604 accordingly to produce structured time series data 612. The structured time series data 612 is provided to a hierarchical time series analysis application 614 that executes the particular time series analysis function 606 to generate analysis results 616.

FIG. 7 is a block diagram depicting a wizard implementation of a data structuring GUI. A time series exploration system 702 accesses unstructured time stamped data 704 and one or more selected time series analysis functions 706. For example, a user may specify a location of unstructured time stamped data 704 and a selection of a time series analysis function 706 to be executed using the unstructured time stamped data 704. Data structuring recommendations functionality 708 may provide recommendations for potential hierarchical structures and/or aggregation time periods for the unstructured time stamped data 704 that might provide best results (e.g., fastest, most efficient) for the particular time series analysis function 706 identified by the user. A user may interact with the time series exploration system 702 via a data structuring GUI 710 to provide human judgment input into the selection of a hierarchical structure to be applied to the unstructured time stamped data 704 to generate the structured time series data 712 that is provided to a hierarchical time series analysis application 714 to generate analysis results 716 based on an execution of the selected time series analysis function 706. Additionally, the data structuring GUI 710 can facilitate a user specifying the structuring of the unstructured time stamped data 704 entirely manually, without recommendation from the data structuring recommendations functionality 708.

The data structuring GUI 710 may be formatted in a variety of ways. For example, the data structuring GUI 710 may be provided to a user in a wizard form, where the user is provided options for selection in a stepwise fashion 718. In one example, a user is provided a number of potential hierarchical structures for the unstructured time stamped data 704 from which to choose as a first step. In a second step 722, the user may be provided with a number of options for a data aggregation time period for the hierarchical structure selected at 720. Other steps 724 may provide displays for selecting additional options for generating the structured time series data 712.

FIG. 8 is a block diagram depicting a data structuring GUI providing multiple data structuring process flows to a user for comparison. A time series exploration system 802 receives unstructured time series data 804 and unstructured time stamped data 806 and provides recommendations 808 for structuring the unstructured time stamped data 804. A data structuring GUI 810 provides an interface for a user to provide input into the process of generating the structured time series data 812 that is provided to a hierarchical time series analysis application 814 to generate analysis results 816 based on an execution of a time series analysis function 806.

In the example of FIG. 8, the data structuring GUI 810 provides a plurality of data structuring flows 818, 820, 822 (e.g., wizard displays) for allowing a user to enter selections regarding structuring of the unstructured time stamped data 804. The data structuring flows 818, 820, 822 may be presented to the user serially or in parallel (e.g., in different windows). The user's selections in each of the data structuring flows 818, 820, 822 are tracked at 824 and stored in a data structure at 826 to allow a user to move among the different structuring approaches 818, 820, 822 without losing the user's place. Thus, the user can make certain selections (e.g., a first hierarchical structure) in a first data structuring flow 818 and see results of that decision (e.g., data distributions, data sufficiency metrics) and can make similar decisions (e.g., a second hierarchical structure) in a second data structuring flow 820 and see results of that decision. The user can switch between the results or compare metrics of the results to make a decision on a best course of action, as enabled by the tracking data 824 stored in the data structuring GUI data structure 826.

As an example, a computer-implemented method of using graphical user interfaces to analyze unstructured time stamped data of a physical process in order to generate structured hierarchical data for a hierarchical forecasting application may include a step of providing a first series of user display screens that are displayed through one or more graphical user interfaces, where the first series of user display screens are configured to be displayed in a step-wise manner so that a user can specify a first approach through a series of predetermined steps on how the unstructured data is to be structured. The information the user has specified in the first series of screens and where in the first series of user display screens the user is located is storing in a tracking data structure. A second series of user display screens are provided that are displayed through one or more graphical user interfaces, where the second series of user display screens are configured to be displayed in a step-wise manner so that the user can specify a second approach through the series of predetermined steps on how the unstructured data is to be structured. The information the user has specified in the second series of screens and where in the second series of user display screens the user is located is storing in the tracking data structure. Tracking data that is stored in the tracking data structure is used to facilitate the user going back and forth between the first and second series of user display screens without losing information or place in either the first or second user display screens, and the unstructured data is structured into a hierarchical structure based upon information provided by the user through the first or second series of user display screens, where the structured hierarchical data is provided to an application for analysis using one or more time series analysis functions.

FIG. 9 is a block diagram depicting a structuring of unstructured time stamped data in one pass through the data. A time series exploration system 902 receives unstructured time stamped data 904 and a selection of one or more time series analysis functions 906 to execute on the unstructured data 904. The time series exploration system 902 may provide recommendations for structuring the data at 908, and a user may provide input into the data structuring process at 910. The unstructured data 904 is formatted into structured time series data 912 and provided to a hierarchical time series analysis application at 914, where a selected time series analysis function 906 is executed to generate analysis results 916.

Functionality for operating on the unstructured data in a single pass 918 can provide the capability to perform all structuring, desired computations, output, and visualizations in a single pass through the data. Each candidate structure runs in a separate thread. Such functionality 918 can be advantageous, because multiple read accesses to a database, memory, or other storage device can be costly and inefficient. In one example, a computer-implemented method of analyzing unstructured time stamped data of a physical process through one-pass includes a step of analyzing a distribution of time-stamped unstructured data to identify a plurality of potential hierarchical structures for the unstructured data. A hierarchical analysis of the potential hierarchical structures is performed to determine an optimal frequency and a data sufficiency metric for the potential hierarchical structures. One of the potential hierarchical structures is selected as a selected hierarchical structure based on the data sufficiency metrics. The unstructured data is structured according to the selected hierarchical structure and the optimal frequency associated with the selected hierarchical structure, where the structuring of the unstructured data is performed via a single pass though the unstructured data. The identified statistical analysis of the physical process is then performed using the structured data.

FIG. 10 is a block diagram depicting example SAS® procedures that can be combined to implement a method of analyzing unstructured time stamped data. In the example of FIG. 10, a SAS Time Series Explorer Engine (TSXEngine or PROC TIMEDATA) 1002 is utilized. Similar to the High Performance Forecasting Engine (HPFENGINE) for Forecast Server, the TSXENGINE 1002 provides large-scale processing and analysis of time-stamped data (e.g., serial or parallel processing). TSXENGINE provides both built-in capabilities and user-defined capabilities for extensibility. TSXENGINE can utilize one pass through a data set to create all needed computations. Because many time series related computations are input/output (I/O) bound, this capability can provide a performance improvement. Using the TSXENGINE can provide testability, maintainability, and supportability, where all numerical components can be performed in batch, where testing and support tools (e.g., SAS testing/support tools) can be utilized.

Given an unstructured time-stamped data set 1004, a data specification 1006 applies both a hierarchical and time frequency structure to form a structured time series data set. The TSXENGINE 1002 forms a hierarchical time series data set at particular time frequency. Multiple structures can be applied for desired comparisons, each running in a separate thread.

The data specification 1006 can be specified in SAS code (batch). The data specification API 1008 processes the interactively provided user information and generates the SAS code to structure the time series data 1004. The data specification API 1008 also allows the user to manage the various structures interactively.

Because there are many ways to analyze time series data, user-defined time series functions can be created using the FCMP procedure 1010 (PROC FCMP or the FCMP Function Editor) and stored in the function repository 1012. A function specification 1014 is used to describe the contents of the function repository 1012 and maps the functions to the input data set 1004 variables which allow for re-use. These functions allow for the transformation or the reduction of time series data. Transformations are useful for discovery patterns in the time series data by transforming the original time series 1004 into a more coherent form. Reductions summarize the time series data (dimension reductions) to a smaller number of statistics which are useful for parametric queries and time series ranking Additionally, functions (transformations, reductions, etc.) can receive multiple inputs and provide multiple outputs.

The function specification 1014 can be specified in SAS code (batch). The function specification API 1016 processes the interactively provided user information and generates the SAS code to create and map the user-defined functions. The function specification API 1016 allows the user to manage the functions interactively.

Because there are many possible computational details that may be useful for time series exploration, the output specification 1018 describes the requested output and form for persistent storage. The output specification 1018 can be specified in SAS code (batch). The output specification API 1020 processes the interactively provided user information and generates the need SAS code to produce the requested output. The output specification API 1020 allows the user to manage the outputs interactively.

Because there are many possible visualizations that may be useful for time series exploration, the results specification 1022 describes the requested tabular and graphical output for visualization. The results specification 1022 can be specified in SAS code (batch). The results specification API 1024 processes the interactively provided user information and generates the need SAS code to produce the requested output. The results specification API 1024 allows the user to manage the outputs interactively.

Given the data specification 1006, the function specification 1012, the output specification 1018, and the results specification 1022, the TSXENGINE 1002 reads the unstructured time-stamped data set 1004, structures the data set with respect to the specified hierarchy and time frequency to form a hierarchical time series, computes the transformations and reductions with respect user-specified functions, outputs the desired information in files, and visualizes the desire information in tabular and graphical form.

The entire process can be specified in SAS code (batch). The time series exploration API processes the interactively provided user information and generates the need SAS code to execute the entire process. The system depicted in FIG. 10 may process a batch of data in one pass through the data. Time-stamped data set can be very large, and multiple reads and write are not scalable. Thus, the TSXENGINE 1002 allows for one pass through the data set for all desired computations, output, and visualization. The depicted system is flexible and extensible. The user can define any time series function (transformations, reductions, etc.) and specify the variable mapping for re-use. Additionally, functions (transformations, reductions, etc.) can receive multiple inputs and provide multiple outputs. The system can provide coordinated batch and interactive management. The user can interactively manage all aspects of the time series exploration process. The system can also provide coordinated batch and interactive execution. The SAS code allows for batch use for scalability. The APIs allow for interactive use. Both can be coordinated to allow for the same results. The system can further provide coordinated batch and interactive persistence. A time series exploration API allows for the persistence of the analyses for subsequent post processing of the results. Further, the system can provide parallelization, where each set of time series is processed separately on separate computational threads.

FIG. 11 depicts a block diagram depicting a time series explorer desktop architecture built on a SAS Time Series Explorer Engine. Results from a TSXENGINE 1102 may be provided using a TSX API (e.g., Java Based). A desktop architecture allows for testability, maintainability, and supportability because all code generation can be performed in batch using JUnit test tools. Additionally, the desktop architecture can provide a simpler development and testing environment for the TSX API and TSX Client. The desktop architecture allows for integration with other desktop clients (e.g., SAS Display Manager, Desktop Enterprise Miner, JMP Pro).

FIG. 12 depicts a block diagram depicting a time series explorer enterprise architecture built on a SAS Time Series Explorer Engine. Results from a TSXENGINE 1202 may be provided using a TSX API (e.g., Java Based). The enterprise architecture allows for integration with Enterprise Solutions (e.g., promotion, migration, security, etc.). The enterprise architecture allows for integration with other enterprise clients (e.g., SAS as a Solution, SAS OnDemand, (Enterprise) Enterprise Miner, EG/AMO).

Structured time series data and analysis results, as well as unstructured time stamped data, can be displayed and manipulated by a user in many ways. FIGS. 13-19 depict example graphical interfaces for viewing and interacting with unstructured time stamped data, structured time series data, and analysis results. FIG. 13 is a graphical interface depicting a distribution analysis of unstructured time-stamped data. Such an interface can be provided to a user as part of a data structuring GUI. The interface displayed in FIG. 13 aids a user in exploring different potential hierarchical structures for a data set and metrics associated with those potential structures. FIG. 14 depicts a graphical interface displaying a hierarchical analysis of structured data. Hierarchical analysis helps a user determine whether structured data is adequate, such as for a desired time series analysis function. FIG. 15 is a graphical interface displaying a large scale visualization of time series data where large amounts of data are available to explore.

FIG. 16 is a graphical interface depicting a univariate time series statistical analysis of a structured set of data. Such an analysis can be used to discover patterns (e.g., seasonal patterns, trend patterns) in a structured time series automatically or via user input. FIG. 17 depicts a graphical interface showing panel and multivariate time series statistical analysis. Such an interface can be used to identify patterns in many time series. FIG. 18 depicts a graphical interface for time series clustering and searching. Clustering and searching operations can be used as part of an operation to identify similar time series. FIG. 19 depicts a graphical interface that provides a time slice analysis for new product diffusion analysis. A time slice analysis can be used for new product and end-of-life forecasting.

Certain algorithms can be utilized in implementing a time series exploration system. The following description provides certain notations related to an example time series exploration system.

Series Index

Let N represents the number of series recorded in the time series data set (or sample of the time series data set) and let i=1, . . . , N represent the series index. Typically, the series index is implicitly defined by the by groups associated with the data set under investigation.

Time Index

Let tε{t_i^b, (t_i^b+1), . . . , (t_i^e−1), t_i^e} represent the time index where t_i^b and t_i^e represent the beginning and ending time index for the i^th series, respectively. The time index is an ordered set of contiguous integers representing time periods associated with equally spaced time intervals. In some cases, the beginning and/or ending time index coincide, sometimes they do not. The time index may be implicitly defined by the time ID variable values associated with the data set under investigation.

Season Index

Let s□{s_i^b, . . . , s_i^e} represent the season index where s_i^b band s_i^e represent the beginning and ending season index for the i^th series, respectively. The season index may have a particular range of values, sε{1, . . . , S}, where S is the seasonality or length of the seasonal cycle. In some cases, the beginning and/or ending season index coincide, sometimes they do not. The season index may be implicitly defined by the time ID variable values and the Time Interval.

Cycle Index

Let l=1, . . . , L, represent the cycle index (or life-cycle index) and L_i=(t_i^e+1−t_i^b) represent the cycle length for the i^th series. The cycle index maps to the time index as follows: l=(t+1−t_i^b) and L_i=(t_i^e+1−t_i^b)). The cycle index represents the number of periods since introduction and ignores timing other than order. The cycle index may be implicitly defined by the starting and ending time ID variable values for each series.

Let L^P≦max_i (L_i) be the panel cycle length under investigation. Sometimes the panel cycle length is important, sometimes it is not. The analysts may limit the panel cycle length, L^P, under consideration, that is subset the data; or the analyst may choose products whose panel cycle length lies within a certain range.

Time Series

Let y_i,l represent the dependent time series values (or the series to be analyzed) where tε{t_i^b, . . . , t_i^e} to is the time index for the i^th dependent series and where i=1, . . . , N. Let

${\overrightarrow{y}}_i={\left\{y_{i,t}\right\}}_{t=t_i^b}^{t_i^e}$

represent the dependent time series vector for i^th dependent series. Let {right arrow over (Y)}⁽ⁱ⁾={{right arrow over (y)}_i}_i=1^N represent the vector time series for all of the dependent time series.

Let {right arrow over (x)}_i,t represent the independent time series vector that can help analyze the dependent series, y_i,t. Let {right arrow over (x)}_i,t={x_i,k,t}_k=1^K where k=1, . . . , K indexes the independent variables and K represents the number of independent variables. Let

${\stackrel{->}{X}}_i={\left\{{\stackrel{->}{x}}_{i,t}\right\}}_{t=t_i^b}^{t_i^e}$

represent the independent time series matrix for i^th dependent series. Let X^(t)={{right arrow over (X)}_i}_i=1^N represent matrix time series for all of the independent time series.

Together, (y_i,t,{right arrow over (x)}_i,t) represent the multiple time series data for the i^th dependent series. Together, (Y^(t),X^(t)) represent the panel time series data for all series (or a vector of multiple time series data).

Cycle Series

Each historical dependent time series, y_i,t, can be viewed as a cycle series (or life-cycle series) when the time and cycle indices are mapped: y_i,t=y_i,l where l=(t+1−t_i^b). Let {right arrow over (y)}_i={y_i,l}_l=1^Li represent the cycle series vector for i^th series. Let Y^(l)={{right arrow over (y)}_i}_i=1^N represent cycle series panel for all of the series. The time series values are identical to the cycle series values except for indexing (subscript).

Each independent time series vector can be indexed by the cycle index: {right arrow over (x)}_i,t={right arrow over (x)}_i,l where l=(t+1−t_i^b). Similarly {right arrow over (X)}_i={{right arrow over (x)}_i,l}_l=1^Li represents the independent time series matrix for i^th dependent series and X^(l)={{right arrow over (X)}_i}_i=1^N represents the matrix time series for all of the independent time series.

Together, (y_i,l,{right arrow over (x)}_i,l) represent the multiple cycle series data for the i^th dependent series. Together, (Y^(l), X^(l)) represent the panel cycle series data for all series (or a vector of multiple cycle series data).

Reduced Data

Given the panel time series data, (Y^(t),X^(t), reduce each multiple time series, (y_i,t,{right arrow over (x)}_i,t); to a reduced vector, {right arrow over (r)}_i={r_i,m}_m=1^M, of uniform length, M. Alternatively, given the panel cycle series data, (Y^(l),X^(l)), reduce each multiple cycle series, (y_i,l,{right arrow over (x)}_i,l), to a reduced data vector, {right arrow over (r)}_i={r_i,m}_m=1^M, of uniform length, M.

For example, features extracted from the i^th multiple time series, (y_i,t,{right arrow over (x)}_i,t). The features may be the seasonal indices where M is the seasonality, or the features may be the cross-correlation analysis results where M is the number of time lags.

The resulting reduced data matrix, R={{right arrow over (r)}_i}_i=1^N has uniform dimension (N×M). Uniform dimensions (coordinate form) are needed for many data mining techniques, such as computing distance measures and clustering data.

Similarity Matrix

Given the panel time series data, (Y^(t),X^(t)), compare each multiple time series, (y_i,t,{right arrow over (x)}_i,t), using similarity measures. Alternatively, given the panel cycle series data, (Y^(l),X^(l)); compare each multiple cycle series, (y_i,l,{right arrow over (x)}_i,l), using a similarity measures.

Let s_i,j=Sim({right arrow over (y)}_i,{right arrow over (y)}_j) represent the similarity measure between the i^th and j^th series. Let {right arrow over (s)}_i={s_i,j}_j=1^N represent the similarity vector of uniform length, N, for the i^th series.

The resulting similarity matrix, S={{right arrow over (s)}_i}_i=1^N has uniform dimension (N×N). Uniform dimensions (coordinate form) are needed for many data mining techniques, such as computing distance measures and clustering data.

Panel Properties Matrix

Given the panel time series data, (Y^(t),X^(t)) compute the reduce data matrix, R={{right arrow over (r)}_i}_i=1^N, and/or the similarity matrix, S={{right arrow over (s)}_i}_i=1^N. Alternatively, given the panel cycle series data, (Y^(l),X^(l)) compute the reduce data matrix, R={{right arrow over (r)}_i}_i=1^N, and/or the similarity matrix, S={{right arrow over (s)}_i}_i=1^N.

A panel properties matrix can be formed by merging the rows of the reduce data matrix and the similarity matrix.

Let P=, S) represent the panel properties matrix of uniform dimension (N×(M+N)). Let {right arrow over (p)}i=({right arrow over (r)}_i,{right arrow over (s)}_i) represent the panel properties vector for the i^th series of uniform dimension (1×(M+N)).

Distance Measures

Given the panel properties vectors, {right arrow over (p)}_i{p_i,j}_j=1^M+N, of uniform length, M+N, let d_i,j=D({right arrow over (p)}_i,{right arrow over (p)}_j) represent the distance between the panel properties vectors associated with i^th and j^th series where D( ) represents the distance measure. Let {right arrow over (d)}_i={d_i,j}_j=1^N be the distance vector associated with the i^th series. Let D={{right arrow over (d)}_i}_i=1^N be the distance matrix associated with all of the series.

Distance measures do not depend on time/season/cycle index nor do they depend on the reduction dimension, M. The dimensions of the distance matrix are (N×N).

If the distance between the Panel Properties Vectors is known, {right arrow over (p)}_i these distances can be used as a surrogate for the distances between the Panel Series Vectors, (y_i,t). In other words, {right arrow over (p)}_i is close {right arrow over (p)}_j to; then (y_i,t) is close to (y_j,t).

Attribute Index

Let K represents the number of attributes recorded in the attribute data and let k=1, . . . , K represent the attribute index.

For example, K could represent the number of attributes associated with the products for sale in the marketplace and k could represent the k^th attribute of the products.

There may be many attributes associated with a given time series. Some or all of the attributes may be useful in the analysis. In the following discussion, the attributes index, k=1, . . . , K, may represent all of the attributes or those attributes that are deemed important by the analyst.

Typically, the number of attribute variables is implicitly defined by the number of selected attributes.

Attribute Data

Let a_i,k represent the attribute data value for k^th attribute associated with i^th series. The attribute data values are categorical (ordinal, nominal) and continuous (interval, ratio). Let {right arrow over (a)}_i={a_i,k}_k=1^K represent the attribute data vector for the i^th series where i=1, . . . , N. Let A={{right arrow over (a)}_i}_i=1^N be the set of all possible attribute data vectors. Let A_k={a_i,k}_i=1^N be the set of attribute values for the k^th attribute for all the series.

For example, a_i,k could represent consumer demographic, product distribution, price level, test market information, or other information for the i^th product.

Analyzing the (discrete or continuous) distribution of an attribute variable values, A_k={a_i,k}_i=1^N, can be useful for new product forecasting in determining the attribute values used to select the pool of candidate products to be used in the analysis. In general, a representative pool of candidate products that are similar to the new product is desired; however, a pool that is too large or too small is often undesirable. A large pool may be undesirable because the pool may not be homogeneous in nature. A small pool may be undesirable because it may not capture all of the potential properties and/or variation.

Let A={{right arrow over (a)}_i}_i=1^N represent the attribute data set. In the following discussion, the attributes data set, A, may represent all of the attributes or those attributes that are deemed important to the analyses.

The attributes may not depend on the time/season/cycle index. In other words, they are time invariant. The analyst may choose from the set of the attributes and their attribute values for consideration. Sometimes the product attributes are textual in nature (product descriptions, sales brochures, and other textual formats). Text mining techniques may be used to extract the attribute information into formats usable for statistical analysis. Sometimes the product attributes are visual in nature (pictures, drawings, and other visual formats). This information may be difficult to use in statistical analysis but may be useful for judgmental analysis.

Derived Attribute Index

Let J represents the number of derived attributes computed from the time series data and let j=1, . . . , J represent the derived attribute index.

For example, J could represent the number of derived attributes associated with the historical time series data and j could represent the j^th derived attribute.

There may be many derived attributes associated with the historical time series data set. Some or all of the derived attributes may be useful in the analysis. In the following discussion, the derived attributes index, j=1, . . . , J, may represent all of the derived attributes or those derived attributes that are deemed important by the analyst.

Typically, the number of derived attribute variables is implicitly defined by the number of selected derived attributes.

Derived Attribute Data

Let g_i,j represent the derived attribute data value for j^th derived attribute associated with i^th series. The attribute data values are categorical (interval, ordinal, nominal). Let {right arrow over (g)}_i={g_i,j}_j=1^J represent the derived attribute data vector for the i^th series where i=1, . . . , N. Let G={{right arrow over (g)}_i}_i=1^N be the set of all possible derived attribute data vectors. Let G_j={g_i,j}_i=1^N be the set of attribute values for the j^th derived attribute for all the series.

For example, g_i,j could represent a discrete-valued cluster assignment, continuous-valued price elasticity, continuous-valued similarity measure, or other information for the i^th series.

Analyzing the (discrete or continuous) distribution of an derived attribute variable values, G_j={g_i,j}_i=1^N, is useful for new product forecasting in determining the derived attribute values used to select the pool of candidate products to be used in the analysis. In general, a representative pool of candidate products that are similar to the new product is desired; however, a pool that is too large or too small is often undesirable. A large pool may be undesirable because the pool may not be homogeneous in nature. A small pool may be undesirable because it may not capture all of the potential properties and/or variation.

Let G={{right arrow over (g)}_i}_i=1^N represent the derived attribute data set. In the following discussion, the derived attributes data set, G, may represent all of the derived attributes or those derived attributes that are deemed important to the analyses. The derived attributes may not depend on the time/season/cycle index. In other words, they may be time invariant. However, the means by which they are computed may depend on time. The analyst may choose from the set of the derived attributes and their derived attribute values for consideration.

Certain computations may be made by a time series exploration system. The following describes certain of those computations. For example, given a panel series data set, the series can be summarized to better understand the series global properties.

Univariate Time Series Descriptive Statistics

Given a time series, y_i,t, or cycle series, y_i,l, summarizes the time series using descriptive statistics. Typically, the descriptive statistics are vector-to-scalar data reductions and have the form: α_i=UnivariateDescrtiveStatistic({right arrow over (y)}_i)

For example:

• • Start, start_i, starting time ID value
  • End, end_i, ending time ID value
  • StartObs, startobs_i, starting observation
  • EndObs, endobs_i, ending observation
  • NObs, nobs n, number of observations
  • NMiss, nmiss_i, number of missing values
  • N, n_i, number of nonmissing values
  • Sum,

${\mathrm{sum}}_i=\sum _{t=t_i^b}^{t_i^e}y_{i,t}=\sum _{l=1}^{L_i}y_{i,l},$

missing values are ignored in the summation

• • Mean,

${\mu }_i=\frac{{\mathrm{sum}}_i}{\left(L_i-{\mathrm{nmiss}}_i\right)},$

• • StdDev,

${\sigma }_i=\sqrt{\frac{1}{\left(L_i-{\mathrm{nmiss}}_i-1\right)}\sum _{t=t_i^b}^{t_i^e}{\left(y_{i,t}-{\mu }_i\right)}^2}=\sqrt{\frac{1}{\left(L_i-{\mathrm{nmiss}}_i-1\right)}\sum _{l=1}^{L_i}{\left(y_{i,l}-{\mu }_i\right)}^2},$

missing values are ignored in the summation

• • Minimum,

$m_i=\begin{array}{c}\min \\ t\end{array}\left(y_{i,t}\right)=\begin{array}{c}\min \\ l\end{array}\left(y_{i,l}\right),$

missing values are ignored in the minimization

Maximum,

$M_i=\begin{array}{c}\max \\ t\end{array}\left(y_{i,t}\right)=\begin{array}{c}\max \\ l\end{array}\left(y_{i,l}\right),$

missing values are ignored in the maximization

• • Range, R_i=M_i−m_i

Time series descriptive statistics can be computed for each independent time series vector.

Vector Series Descriptive Statistics

Given a panel time series, Y^(t), or panel cycle series, Y^(t)=Y⁽ⁱ⁾, summarize the panel series using descriptive statistics. Essentially, the vector series descriptive statistics summarize the univariate descriptive statistics. Typically, the vector descriptive statistics are matrix-to-scalar data reductions and have the form: α=VectorDescriptiveStatistic(Y^(t))

Following are some examples:

• • Start, start, starting time ID value
  • End, end, ending time ID value
  • StartObs, startobs, starting observation
  • EndObs, endobs, ending observation
  • NObs, nobs, number of observations
  • NMiss, nmiss, number of missing values
  • N, n, number of nonmissing values

$\mathrm{Minimum},m=\begin{array}{c}\min \\ i\end{array}\left(m_i\right),$

missing values are ignored in the minimization

$\mathrm{Maximum},M=\begin{array}{c}\max \\ i\end{array}\left(M_i\right),$

missing values are ignored in the maximization

• • Range, R=M−m

Likewise, vector series descriptive statistics can be computed for each independent time series vector.

Certain transformations may be performed by a time series exploration system. The following describes certain example time series transformations.

Given a panel series data set, the series can be transformed to another series which permits a greater understanding of the series properties over time.

Univariate Time Series Transformations

Given a time series, y_i,t or cycle series, y_i,l, univariately transform the time series using a univariate time series transformation. Typically, univariate transformations are vector-to-vector (or series-to-series) operations and have the form: {right arrow over (z)}_i=UnivariateTransform({right arrow over (y)}_i)

Following are some examples:

• • Scale, {right arrow over (z)}_i=scale({right arrow over (y)}_i), scale the series from zero to one
  • CumSum, {right arrow over (z)}_i=cusum({right arrow over (y)}_i), cumulatively sum the series
  • Log, {right arrow over (z)}_i=log({right arrow over (y)}_i), series should be strictly positive
  • Square Root, {right arrow over (z)}_i=√{square root over ({right arrow over (y)}_i, series should be strictly positive
  • Simple Difference, z_i,t=(y_i,t−y_i,(t-1))
  • Seasonal Difference, z_i,t=(y_i,t−y_i,(t-S)), series should be seasonal
  • Seasonal Adjustment, z_t=SeasonalAdjusment({right arrow over (y)}_i)
  • Singular Spectrum, z_t=SSA({right arrow over (y)}_i)
  • Several transformations can be performed in sequence (e.g., a log simple difference).Transformations help analyze and explore the time series.

Multiple Time Series Transformations

Given a dependent time series, y_i,t, or cycle series, y_i,l, and an independent time series, x_i,t, multivariately transform the time series using a multiple time series transformation. Typically, multiple time series transforms are matrix-to-vector operations and have the form:

{right arrow over (z)}_i=MultipleTransforms({right arrow over (y)}_i,{right arrow over (x)}_i)For example:

• • Adjustment, {right arrow over (z)}_i=Adjustment({right arrow over (y)}_i,{right arrow over (x)}_i)Several multivariate transformations can be performed in sequence.

Vector Series Transformations

Given a panel time series, y_i,t, or panel cycle series, y_i,t=y_i,l, multivariately transform the panel series using a vector series transformation. Typically, the vector transformations are matrix-to-matrix (panel-to-panel) operations and have the form: Z=VectorTransform(Y)

Many vector transformations are just univariate transformations applied to each series individually. For each series index

{right arrow over (z)} _i=UnivariateTransform({right arrow over (y)}_i)i=1, . . . , N

Some vector transformations are applied to a vector series jointly.For example:

• • Standardization Z=(Ω⁻¹)′YQ⁻¹ Ω=cov(Y, Y)

Certain time series data reduction operations may be performed by a time series exploration system. Data mining techniques include clustering, classification, decision trees, and others. These analytical techniques are applied to large data sets whose observation vectors are relatively small in dimension when compared to the length of a transaction series or time series.

In order to effectively apply these data mining techniques to a large number of series, the dimension of each series can be reduced to a small number of statistics that capture their descriptive properties. Various transactional and time series analysis techniques (possibly in combination) can be used to capture these descriptive properties for each time series.

Many transactional and time series databases store the data in longitudinal form, whereas many data mining software packages utilize the data in coordinate form. Dimension reduction extracts important features of the longitudinal dimension of the series and stores the reduced sequence in coordinate form of fixed dimension. Assume that there are N series with lengths {T₁, . . . , T_N}.

In longitudinal form, each variable (or column) represents a single series, and each variable observation (or row) represents the series value recorded at a particular time. Notice that the length of each series, T_i, can vary.

{right arrow over (y)} _i ={y _i,t}_t=1^Ti for i=1, . . . , N

where {right arrow over (y)}_i is (T_i×1). This form is convenient for time series analysis but less desirable for data mining

In coordinate form, each observation (or row) represents a single reduced sequence, and each variable (or column) represents the reduced sequence value. Notice that the length of each reduced sequence, M, is fixed.

{right arrow over (r)} _i ={r _i,m}_m=1^M for i=1, . . . , N

where {right arrow over (r)}_i is (1×M). This form is convenient for data mining but less desirable for time series analysis.

To reduce a single series, a univariate reduction transformation maps the varying longitudinal dimension to the fixed coordinate dimension.

{right arrow over (r)} _i=Reduce_i({right arrow over (y)}_i) for i=1, . . . , N

where {right arrow over (r)}_i is (1×M), Y_i is (T_i×1), and Reduce_i( ) is the reduction transformation (e.g., seasonal decomposition).

For multiple series reduction, more than one series is reduced to a single reduction sequence. The bivariate case is illustrated.

{right arrow over (r)} _i=Reduce_i({right arrow over (y)}_i,{right arrow over (x)}_i,k) for i=1, . . . , N

where {right arrow over (r)}_i is (1×M), {right arrow over (y)}_i is (T_i×1), {right arrow over (x)}_i,k is (T_i×1), and Reduce_i( ) is the reduction transformation (e.g., cross-correlations).

In the above discussion, the reduction transformation, Reduce_i( ), is indexed by the series index, i=1, . . . , N, but typically it does not vary and further discussion assumes it to be the same, that is, Reduce( )=Reduce_i( ).

Univariate Time Series Data Reductions

Given a time series, y_i,t, or cycle series, y_i,l, univariately reduce the time series using a time series data reduction. Typically, univariate reductions are vector-to-vector operations and have the form: {right arrow over (r)}_i=UnivariateReduction({right arrow over (y)}_i)

Following are some examples:

• • Autocorrelation, {right arrow over (r)}_i=ACF ({right arrow over (y)}_i)
  • Seasonal Decomposition, {right arrow over (r)}_i=SeasonalDecomposition({right arrow over (y)}_i) Multiple Time Series Data ReductionsGiven a dependent time series, y_i,t, or cycle series, y_i,l, and an independent time series, x_i,t, multivariately reduce the time series using a time series data reduction. Typically, multiple time series reductions are matrix-to-vector operations and have the form:

{right arrow over (r)} _i=MultipleReduction({right arrow over (y)}_i,{right arrow over (x)}_i)

For example,

• • Cross-Correlation, {right arrow over (r)}_i=CCF({right arrow over (y)}_i,{right arrow over (x)}_i)

Vector Time Series Data Reductions

Given a panel time series, y_i,t, or panel cycle series, y_i,t=y_i,l, multivariately reduce the panel series using a vector series reduction. Typically, the vector reductions are matrix-to-matrix operations and have the form: R=Vector Reduction(Y)

Many vector reductions include univariate reductions applied to each series individually. For each series index

{right arrow over (r)} _i=UnivariateReduction({right arrow over (y)}_i)i=1, . . . , N

Some vector reductions are applied to a vector series jointly.

For example:

• • Singular Value Decomposition, R=SVD(Y)A time series exploration system may perform certain attribute derivation operations. For example, given a panel series data set, attributes can be derived from the time series data.

Univariate Time Series Attribute Derivation

Given a time series, y_i,t, or cycle series, y_i,l, derive an attribute using a univariate time series computation. Typically, univariate attribute derivations are vector-to-scalar operations and have the form: g_i,j=UnivariateDerivedAttribute({right arrow over (y)}_i)

For example:

• • Sum, g_i,j=Sum({right arrow over (y)}_i)
  • Mean, g_i,j=Mean({right arrow over (y)}_i)

Multiple Time Series Attribute Derivation

Given a dependent time series, y_i,t, or cycle series, y_i,l, and an independent time series, x_i,t, derive an attribute using a multiple time series computation. Typically, multiple attribute derivations are matrix-to-scalar operations and have the form:

g _i,j=MultipleDerivedAttribute({right arrow over (y)}_i,{right arrow over (x)}_i)

Following are some examples:

• • Elasticity, g_i,j=Elasticity({right arrow over (y)}_i,{right arrow over (x)}_i)
  • Cross-Correlation, g_i,j=CrossCorr({right arrow over (y)}_i,{right arrow over (x)}_i)

Vector Series Attribute Derivation

Given a panel time series, Y^(t), or panel cycle series, Y^(t)=Y^(l), compute a derived attribute values vector associated with the panel series. Essentially, the vector attribute derivation summarizes or groups the panel time series. Typically, the vector series attribute derivations are matrix-to-vector operations and have the form:

G _j=VectorDerivedAttribute(Y^(t))

Many vector series attribute derivations are just univariate or multiple attribute derivation applied to each series individually. For each series indices,

g _i,j=UnivariateDerivedAttribute({right arrow over (y)}_i)i=1, . . . ,N

OR

g _i,j=MultipleDerivedAttribute({right arrow over (y)}_i,{right arrow over (x)}_i)i=1, . . . ,N

Some vector series attribute derivations are applied to a vector series jointly.For example:

• • Cluster, G_j=Cluster(Y^(t)) cluster the time series

Data provided to, utilized by, and outputted by a time series exploration system may be structured in a variety of forms. The following describes concepts related to the storage and representation of the time series data.

Storage of Panel Series Data

Table 1 describes the storage of the Panel Series Data.

TABLE 1
Panel Series Data Storage Example

Table 1 represents a panel series. Each shaded area of the table represents a multiple time series. Each analysis variable column in each shaded area of the table represents a univariate time series. Each analysis variable column represents a vector time series.

Internal Representation of Panel Series Data

The amount of data associated with a Panel Series may be quite large. The Panel Series may be represented efficiently in memory (e.g., only one copy of the data in memory is stored). The Panel Series, (Y^(t),X^(t)), contains several multiple series data, (y_i,t,{right arrow over (x)}_i,t), which contains a fixed number univariate series data, y_i,t or {right arrow over (x)}_i,t. The independent variables are the same for all dependent series though some may contain only missing values.

FIG. 20 depicts an example internal representation of the Panel Series Data.

Reading and Writing the Panel Series Data

The data set associated with a Panel Series may be quite large. It may be desirable to read the data set only once. The user may be warned if the data set is to be reread.

Reading/writing the Panel Series Data into/out of memory may be performed as follows:

For each Multiple Time Series (or by group), read/write all of the Univariate Time Series associated with the by group. Read/write y_i,t or {right arrow over (x)}_i,t to form (y_i,t,{right arrow over (x)}_i,t) for each by group. Read/write each by group (y_i,t,{right arrow over (x)}_i,t) to form (Y^(t),X^(t)).

FIG. 21 depicts reading/writing of the Panel Series Data.

A time series exploration may store and manipulate attribute data

Storage of Attribute Data

Table 2 describes example storage of Attribute Data.

TABLE 2
Example Storage of Attribute Data

Table 2 represents an attribute data set, A={{right arrow over (a)}_i}_i=1^N. Each shaded row of the table represents an attribute vector for single time series, {right arrow over (a)}_i={{right arrow over (a)}_i,k}_k=1^K. Each attribute variable column represents an attribute value vector across all time series, A_k={{right arrow over (a)}_i,k}_l=1^N. Each table cell represents a single attribute value, a_i,k.

Table 2 describes the shaded areas of following Table 3 associated with the Panel Series.

TABLE 3
Panel Series Data Example

Notice that the Panel Series has a time dimension but the Attributes do not. Typically, the attribute data set is much smaller than the panel series data set. Table 3 show that the series index, i=1, . . . , N, are a one-to-one mapping between the tables. The mapping is unique but there may be time series data with no associated attributes (missing attributes) and attribute data with no time series data (missing time series data).

Internal Representation of Attribute Data

The amount of data associated with the attributes may be quite large. The Attribute Data may be represented efficiently in memory (e.g., only one copy of the data in memory is stored). The attribute data, A={{right arrow over (a)}_i}_i=1^N, contains several attribute value vectors, A_k={{right arrow over (a)}_i,k}_i=1^N, which contains a fixed number attribute values, a_i,k, for discrete data and a range of values for continuous data. The attribute variables are the same for all time series.

FIG. 22 depicts an example internal representation of the Attribute Data.

Reading and Writing the Attribute Data

The data set associated with Attributes may be quite large. It may be desirable to only read data once if possible. The user may be warned if the data set is to be reread.

Reading/writing attribute data into/out of memory can be performed as follows:

For each attribute vector (or by group), read/write all of the attribute values associated with the by group.

FIG. 23 depicts reading/writing of the attribute data.

In some implementations it may be desirable to limit or reduce an amount of data stored. The following discussion describes some practical concepts related to the storage and representation of the reduced data.

Storage of Reduced Data

TABLE 4
Reduced Data Storage
Table 4 depicts storage of the Reduced Data.

Table 4 represents a reduced data set, R={{right arrow over (r)}_i}_i=1^N. Each shaded row of the table represents a reduced data vector for single time series, {right arrow over (r)}_i={{right arrow over (r)}_i,m}_m=1^M. Each reduced variable column represents a reduced value vector across all time series, R_m={{right arrow over (r)}_i,m}_l=1^N. Each table cell represents a single reduced value, r_i,m.

Table 5 describes the shaded areas of the following table associated with the Panel Series.

TABLE 5
Panel Series Data Example

Notice that the Panel Series has a time dimension but the Reduced Data do not. Sometimes, the reduced data set is much smaller than the panel series data set. Tables 4 and 5 show that the series index, i=1, . . . , N, are a one-to-one mapping between the tables. The mapping is unique but there may be time series data with no associated reduced data (missing attributes) and reduced data with no time series data (missing time series data).

Dimension reduction may transform the series table (T×N) to the reduced table (N×M) where T=max{T₁, . . . , T_N} and where typically M<T. The number of series, N, can be quite large; therefore, even a simple reduction transform may manipulate a large amount of data. Hence, it is important to get the data in the proper format to avoid the post-processing of large data sets.

Time series analysts may often desire to analyze the reduced table set in longitudinal form, whereas data miners often may desire analyze the reduced data set in coordinate form.

Transposing a large table from longitudinal form to coordinate form and vice-versa form can be computationally expensive.

In some implementations a time series exploration system may make certain distance computations. The following discussion describes some practical concepts related to the storage and representation of the distance.

Storage of Distance Matrix

Table 6 describes the storage of the Distance.

TABLE 6
Distance Storage

Table 6 represents a distance matrix data set, D={{right arrow over (d)}_i}_i=1^N. Each shaded row of the table represents a distance vector for single time series, {right arrow over (d)}_i={d_i,j}_j=1^N. Each distance variable column represents a distance measure vector across all time series, D_j={d_i,j}_i=1^N. Each table cell represents a single distance measure value, d_i,j.

Table 7 describes the shaded areas of the following table associated with the Panel Series.

TABLE 7
Panel Series Example

Notice that the Panel Series has a time dimension but the Distance Matrix does not. Typically, the distance matrix data set is much smaller than the panel series data set.

Table 7 shows that the series index, i=1, . . . , N, are a one-to-one mapping between the tables. The mapping is unique but there may be time series data with no associated distance measures (missing measures) and distance measures without time series data (missing time series data).

In some implementations a time series exploration system may store derived data. The following discussion describes some practical concepts related to the storage and representation of the attribute data.

Storage of Derived Attribute Data

Table 8 describes storage of the derived attribute data.

TABLE 8
Derived Attribute Data Storage

Table 8 represents a derived attribute data set, G={{right arrow over (g)}_i}_i=1^N. Each shaded row of the table represents a derived attribute vector for single time series, G_j={g_i,j}_i=1^N. Each attribute variable column represents a derived attribute value vector across all time series, G_j={g_i,j}_i=1^N. Each table cell represents a single derived attribute value, g_i,j.

Internal Representation of Attribute Data

The amount of data associated with the derived attributes may be quite large. The derived attribute data may be represented efficiently in memory (e.g., only one copy of the data in memory is stored).

The derived attribute data, G={{right arrow over (g)}_i}_i=1^N, contains several derived attribute value vectors, G_j={g_i,j}_i=1^N, which contains a fixed number derived attribute values, g_i,j, for discrete data and a range of values for continuous data. The derived attribute variables are the same for all time series.

FIG. 24 depicts an internal representation of derived attribute data.

Reading and Writing the Derived Attribute Data

The data set associated with Derived Attributes may be quite large. It may be desirable to only read data once if possible. The user may be warned if the data set is to be reread.

Reading/writing derived attribute data into/out of memory may be performed as follows:

For each Derived Attribute Vector (or by group), read/write all of the derived attribute values associated with the by group.

FIG. 25 depicts reading/writing of derived attribute data.

FIGS. 26A, 26B, and 26C depict example systems for use in implementing a time series exploration system. For example, FIG. 26A depicts an exemplary system 2600 that includes a standalone computer architecture where a processing system 2602 (e.g., one or more computer processors located in a given computer or in multiple computers that may be separate and distinct from one another) includes a time series exploration system 2604 being executed on it. The processing system 2602 has access to a computer-readable memory 2606 in addition to one or more data stores 2608. The one or more data stores 2608 may include unstructured time stamped data 2610 as well as time series analysis functions 2612.

FIG. 26B depicts a system 2620 that includes a client server architecture. One or more user PCs 2622 access one or more servers 2624 running a time series exploration system 2626 on a processing system 2627 via one or more networks 2628. The one or more servers 2624 may access a computer readable memory 2630 as well as one or more data stores 2632. The one or more data stores 2632 may contain an unstructured time stamped data 2634 as well as time series analysis functions 2636.

FIG. 26C shows a block diagram of exemplary hardware for a standalone computer architecture 2650, such as the architecture depicted in FIG. 26A that may be used to contain and/or implement the program instructions of system embodiments of the present invention. A bus 2652 may serve as the information highway interconnecting the other illustrated components of the hardware. A processing system 2654 labeled CPU (central processing unit) (e.g., one or more computer processors at a given computer or at multiple computers), may perform calculations and logic operations required to execute a program. A processor-readable storage medium, such as read only memory (ROM) 2656 and random access memory (RAM) 2658, may be in communication with the processing system 2654 and may contain one or more programming instructions for performing the method of implementing a time series exploration system. Optionally, program instructions may be stored on a non-transitory computer readable storage medium such as a magnetic disk, optical disk, recordable memory device, flash memory, or other physical storage medium.

A disk controller 2660 interfaces one or more optional disk drives to the system bus 2652. These disk drives may be external or internal floppy disk drives such as 2662, external or internal CD-ROM, CD-R, CD-RW or DVD drives such as 2664, or external or internal hard drives 2666. As indicated previously, these various disk drives and disk controllers are optional devices.

Each of the element managers, real-time data buffer, conveyors, file input processor, database index shared access memory loader, reference data buffer and data managers may include a software application stored in one or more of the disk drives connected to the disk controller 2660, the ROM 2656 and/or the RAM 2658. Preferably, the processor 2654 may access each component as required.

A display interface 2668 may permit information from the bus 2652 to be displayed on a display 2670 in audio, graphic, or alphanumeric format. Communication with external devices may optionally occur using various communication ports 2672.

In addition to the standard computer-type components, the hardware may also include data input devices, such as a keyboard 2673, or other input device 2674, such as a microphone, remote control, pointer, mouse and/or joystick.

Additionally, the methods and systems described herein may be implemented on many different types of processing devices by program code comprising program instructions that are executable by the device processing subsystem. The software program instructions may include source code, object code, machine code, or any other stored data that is operable to cause a processing system to perform the methods and operations described herein and may be provided in any suitable language such as C, C++, JAVA, for example, or any other suitable programming language. Other implementations may also be used, however, such as firmware or even appropriately designed hardware configured to carry out the methods and systems described herein.

The systems' and methods' data (e.g., associations, mappings, data input, data output, intermediate data results, final data results, etc.) may be stored and implemented in one or more different types of computer-implemented data stores, such as different types of storage devices and programming constructs (e.g., RAM, ROM, Flash memory, flat files, databases, programming data structures, programming variables, IF-THEN (or similar type) statement constructs, etc.). It is noted that data structures describe formats for use in organizing and storing data in databases, programs, memory, or other computer-readable media for use by a computer program.

The computer components, software modules, functions, data stores and data structures described herein may be connected directly or indirectly to each other in order to allow the flow of data needed for their operations. It is also noted that a module or processor includes but is not limited to a unit of code that performs a software operation, and can be implemented for example as a subroutine unit of code, or as a software function unit of code, or as an object (as in an object-oriented paradigm), or as an applet, or in a computer script language, or as another type of computer code. The software components and/or functionality may be located on a single computer or distributed across multiple computers depending upon the situation at hand.

It should be understood that as used in the description herein and throughout the claims that follow, the meaning of “a,” “an,” and “the” includes plural reference unless the context clearly dictates otherwise. Also, as used in the description herein and throughout the claims that follow, the meaning of “in” includes “in” and “on” unless the context clearly dictates otherwise. Further, as used in the description herein and throughout the claims that follow, the meaning of “each” does not require “each and every” unless the context clearly dictates otherwise. Finally, as used in the description herein and throughout the claims that follow, the meanings of “and” and “or” include both the conjunctive and disjunctive and may be used interchangeably unless the context expressly dictates otherwise; the phrase “exclusive or” may be used to indicate situation where only the disjunctive meaning may apply.

Claims

1. A system comprising: one or more processors;one or more computer-readable storage mediums containing instructions configured to cause the one or more processors to perform operations including: analyzing a distribution of time-stamped data to identify a plurality of potential time series data hierarchies for structuring the data, wherein a potential time series data hierarchy is a framework for structuring the data using multiple time series;performing an analysis of the potential time series data hierarchies,wherein performing the analysis of the potential time series data hierarchies includes determining an optimal time series frequency and a data sufficiency metric for each of the potential time series data hierarchies; selecting one of the potential time series data hierarchies based on a comparison of the data sufficiency metrics;deriving, in a single-read pass through the time-stamped data, multiple time series from the time-stamped data according to the selected time series data hierarchy, wherein a derived time series includes observations at intervals of time spaced in a manner commensurate with the optimal time series frequency determined for the selected time series data hierarchy; andgenerating a time series forecast corresponding to at least one of the derived time series.

2. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: providing recommendations for the potential time series data hierarchies using forecasting functions that are selected from a functions repository, wherein the selected forecasting functions affect what recommendations are to be provided.

3. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: providing a series of user display screens that are displayed through one or more graphical user interfaces, wherein the series of user display screens are configured to be displayed in a step-wise manner so that a user can specify an approach through a series of predetermined steps on how the unstructured time-stamped data is to be structured; andstoring in a tracking data structure configuration information the user has specified in the series of user display screens and where in the series of user display screens the user is,wherein the series of user display screens utilize data obtained via the single-read pass of the time-stamped data.

4. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: analyzing the derived time series using one or more selected time series analysis functions.

5. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: comparing the derived time series to a time series of interest to identify a similar portion of the derived time series; andextracting the similar portion of the derived time series, wherein a selected time series analysis function operates on the extracted similar portion.

6. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: identifying a characteristic of the derived time series; andselecting a data model for a selected time series analysis function based on the identified characteristic, wherein the selected time series analysis function is performed using the selected data model.

7. The system of claim 6, wherein the characteristic is a seasonal pattern, a trend pattern, a growth pattern, or a delay pattern.

8. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: transforming the derived time series; andgenerating a visualization of the transformed or reduced time series.

9. The system of claim 1, wherein a time series analysis function is applied to the derived time series, wherein the time series analysis function comprises a forecasting function.

10. The system of claim 9, wherein the derived time series are provided to a forecasting application for forecasting using the time series analysis function.

11. The system of claim 1, wherein the time-stamped data is analyzed by applying a seasonality test, an intermittency test, or a trending data test.

12. The system of claim 1, wherein the time-stamped data is analyzed by applying a user-defined test or a business-objective test.

13. The system of claim 1, containing further instructions configured to cause the one or more processors to perform operations including: providing a first portion of the derived time series to a first processor for performing a statistical analysis; andproviding a second portion of the derived time series to a second processor for performing the statistical analysis, wherein the first portion and the second portion are based upon a portion of a hierarchy in which the first portion and the second portion reside.

14. A non-transitory computer program product, tangible embodied in a non-transitory machine readable storage medium, including instructions operable to cause a data processing apparatus to: analyze a distribution of time-stamped data to identify a plurality of potential time series data hierarchies for structuring the data, wherein a potential time series data hierarchy is a framework for structuring the data using multiple time series;perform an analysis of the potential time series data hierarchies, wherein performing the analysis of the potential time series data hierarchies includes determining an optimal time series frequency and a data sufficiency metric for each of the potential time series data hierarchies;select one of the potential time series data hierarchies based on a comparison of the data sufficiency metrics;derive, in a single-read pass through the time-stamped data, multiple time series according to the selected time series data hierarchy, wherein a derived time series includes observations at intervals of time spaced in a manner commensurate with the optimal time series frequency determined for the selected time series data hierarchy; andgenerate a time series forecast corresponding to at least one of the derived time series.

15. The non-transitory computer program product of claim 14, including further instructions operable to cause a data processing apparatus to: provide recommendations for the potential time series data hierarchies using forecasting functions that are selected from a functions repository, wherein the selected forecasting functions affect what recommendations are to be provided.

16. The non-transitory computer program product of claim 14, including further instructions operable to cause a data processing apparatus to: provide a series of user display screens that are displayed through one or more graphical user interfaces, wherein the series of user display screens are configured to be displayed in a step-wise manner so that a user can specify an approach through a series of predetermined steps on how the unstructured time-stamped data is to be structured; andstore in a tracking data structure configuration information the user has specified in the series of user display screens and where in the series of user display screens the user is, wherein the series of user display screens utilize data obtained via the single-read pass of the time-stamped data.

17. The non-transitory computer program product of claim 14, including further instructions operable to cause a data processing apparatus to: analyze the derived time series using one or more selected time series analysis functions.

18. The non-transitory computer program product of claim 14, containing further instructions configured to cause the one or more processors to perform operations including: compare the derived time series to a time series of interest to identify a similar portion of the derived time series; andextract the similar portion of the derived time series, wherein a selected time series analysis function operates on the extracted similar portion.

19. The non-transitory computer program product of claim 14, including further instructions operable to cause a data processing apparatus to: identify a characteristic of the derived time series; andselect a data model for a selected time series analysis function based on the identified characteristic, wherein the selected time series analysis function is performed using the selected data model.

20. A computer-implemented method comprising: analyzing, using one or more data processors, a distribution of time-stamped data to identify a plurality of potential time series data hierarchies for structuring the data, wherein a potential time series data hierarchy is a framework for structuring the data using multiple time series;performing, using the one or more data processors, an analysis of the potential time series data hierarchies, wherein performing the analysis of the potential time series data hierarchies includes determining an optimal time series frequency and a data sufficiency metric for each of the potential time series data hierarchies;selecting one of the potential time series data hierarchies based on a comparison of the data sufficiency metrics;deriving, in a single-read pass through the time-stamped data, multiple time series according to the selected time series data hierarchy, wherein a derived time series includes observations at intervals of time spaced in a manner commensurate with the optimal time series frequency determined for the selected time series data hierarchy; andgenerating a time series forecast corresponding to at least one of the derived time series.