Patent Application
20180121942
In general, a customer segmentation system identifies subsets of customers based on characteristics associated with customers. In this regard, customer segmentation separates customers into different groups based on characteristics associated therewith. Customers can be segmented based on any number of characteristics. For example, in some cases, subsets of customers are identified based on their demographic attributes, such as origin, gender, age, income, etc. In other cases, subsets of customers are identified based on their online interaction, such as a particular device or services being used, e.g., browser types, mobile device models, search engines, etc.; and such as where the customer navigated from, e.g., search engine, previous exit page, etc. In yet other cases, subsets of customers are identified based on other features, such as transaction histories or online profiles, e.g., social network profiles.
A customer segmentation system is useful for automatically dividing the customers into meaningful segments to perform targeted marketing, uncover unmet client needs, design new products, develop customized programs, establish proper service, allocate resources, and so on. A technique commonly used to perform customer segmentation is cluster analysis, which aims to separate data points into several groups so that the data points in the same group are more similar than those in different groups. However, the massive and high-dimensional customer data routinely brings various challenges for obtaining robust and high-quality customer segments.
Embodiments of the present invention relate to systems and methods for customer segmentation. In particular, embodiments of the present disclosure relate to a customer segmentation system based on consensus clustering technologies. As described in embodiments herein, technical solutions are provided to automatically obtain partitions for customer segmentation from high-dimensional customer data.
In various embodiments, this process for customer segmentation includes receiving a target cluster number and a group of customers in an original feature space (e.g., a customer space with various features about the customers). This process further includes generating basic partitions of the customers (i.e., clusters of customers) in the original feature space, e.g., via multiple sequential partitioning stages. The term “partition” refers to the collection of objects in a cluster. Subsequently, the original feature space is transformed into an augmented partition space, for example, based on membership information of the customers in respective basic partitions. In this augmented partition space, consensus-based partitions of the customers are determined based on the target cluster number and multiple stages of the greedy K-means based dynamic partition process.
Embodiments will be readily understood by the following detailed description in conjunction with the accompanying drawings.
The massive and high-dimensional customer data today presents various technical challenges for customer segmentation. Cluster analysis is one of many techniques used for customer segmentation. Traditional approaches apply some clustering methods, such as K-means, spectral clustering, and so on, for customer segmentation. Different clustering methods have been proposed based on different assumptions. As an example, K-means is the widely used clustering method, which generally finds K centroids to represent the whole dataset. As another example, agglomerate hierarchy clustering (AHC) iteratively merges the nearest two points or clusters until all the points are in the same cluster together. As yet another example, density-based spatial clustering of applications with noise (DBSCAN) separates the points by the high-density regions.
However, because different methods provide different clustering results, it is difficult to choose the most suitable method for a specific application. For example, it is difficult to speculate the right choice for diverse customer segmentation problems, e.g., diverse customer datasets that have numerous/different factors to consider. By way of example, customer segmentation for flight passengers likely requires different factors (e.g., departure and arrival airports) compared to customer segmentation for college students in choosing their elective classes. Further, some clustering methods have many parameters to tune up, and thus become volatile when applied to diverse customer datasets.
Traditionally, consensus clustering is usually formalized into a combinational optimization problem, which sets a global objective function and adopts some heuristics to find approximate solutions. Many methods are developed to solve different objective functions, including nonnegative matrix factorization, kernel-based methods, simulated annealing, etc. There are also some methods without an explicit objective function, including graph-based algorithms, co-association matrix based methods, relabeling and voting methods, locally adaptive cluster based methods, genetic algorithm based methods, etc.
Some conventional techniques include configuring the consensus clustering problem into a K-means clustering problem via, e.g., a utility function. However, the performance of this kind of K-means clustering method is unstable due to its dependency on its initialization conditions. Further, such a method generally does not specify how to generate the basic partitions and choose the proper cluster number.
Technical solutions are disclosed here to resolve various technical issues stemming from traditional cluster analysis for customer segmentation, such as issues related to the complex data structure, the effective feature engineering, and the proper cluster number. At a high level, technical solutions are provided in a greedy K-means based consensus clustering (GKCC) system for consumer segmentation. There are two phases in the GKCC system, basic partition generation phase and consensus clustering phase. As used herein, “basic partitions” refers to partitions generated in the original feature space (e.g., a customer space with various features about the customers). As used herein, “consensus clustering” refers to generating partitions in the augmented partition space, which summarizes high-level information of customers. In some embodiments, the first phase involves generating basic partitions with the cluster number varying from 2 to 2K, where K is a user-defined cluster number, in an iterative partitioning process. In some embodiments, the second phase for the consensus clustering involves deriving a binary matrix from the basic partitions. In particular, deriving the binary matrix involves using the membership information of customers in the basic partitions. Like the process of basic partition generation, the consensus clustering phrase may also use a K-means based clustering process operated in the augmented partition space. Stated differently, K-means can be used to operate on the binary matrix to generate partitions of the customers.
The GKCC system is based on the greedy center allocation in an augmented partition space. There are many benefits of utilizing GKCC. For example, GKCC resolves the sensitivity of initialization with theoretical guarantee, incorporates the basic partition generation into a unified framework, and returns a set of partitions with different cluster numbers for practical use. GKCC conducts the cluster analysis on the augmented partition space, rather than the original feature space, which uses high-level information to capture more meaningful cluster structures and results in more robust results. By using the dynamic partition process, GKCC incrementally adds new partition centers and overcomes the sensitivity issue of K-means initialization. Further, GKCC employs a sampling strategy to search for new partition centers, e.g., with only a predetermined number of stages to generate a predetermined number of partitions. Even further, the consequent intermediate basic partitions are used later for determining the final set of basic partitions as well as the augmented partition space.
Extensive experimental results on benchmark datasets demonstrate that GKCC outperforms other state-of-the-art clustering methods in terms of objective function value and external measurements. The GKCC system outperforms traditional systems and returns the partition with small objective function value and small deviation. Advantageously, GKCC also permits the usage of user-defined or application-oriented cluster numbers so that a suitable cluster number can be chosen based on the specific customer segmentation problem. As an example, customer segmentation for worldwide Photoshop® users may requires a cluster number much greater than customer segmentation for all guests attending a state dinner in the White House.
Referring now to
The customer analysis system 100 can include technologies that can be used to empower digital research and marketing. The customer segmentation system 110 generally identifies customer segments 170 from customers 160 based on customer information, e.g., one or more customer attributes. The one or more customer attributes selected for customer segmentation can be tailored for specific tasks, e.g., based on the needs of reports and analytics 120, marketing cloud 130, ad hoc analysis 140, and target 150. In some embodiments, real-time visitor information of customers is used for customer segmentation, which results in real-time customer segmentation information.
Reports and analytics 120 generates analytics and reports on various data, e.g., related to a specific customer segment. Standard reports may provide analytics of website and visitor activity, traffic patterns, referral data, advertising campaigns, visitor retention, product data, etc., based on customer segmentation. Reports and analytics 120 can also provide tools for users to configure segments, metrics, etc. In various embodiments, reports and analytics 120 retrieves or receives customer segmentations from the customer segmentation system 110 based on website attributes, visitor attributes, traffic attributes, referral attributes, product attributes, etc.
Reports and analytics 120 may provide summary reports for a general overview of the data. Reports and analytics 120 may also provide conversion reports related to detailed analysis of customer activity, e.g., customer conversion related to e-commerce transactions, sources of sales, advertising effectiveness, customer loyalty, and more. Even more, reports and analytics 120 may provide traffic reports related to in-depth insight into how visitors interact with a website. In various embodiments, these reports are generated based on customer segmentation information, e.g., segmented by selected customer attributes.
Marketing cloud 130 can include a set of marketing solutions to build personalized campaigns, e.g., for a targeted customer segment. A business can aim its marketing efforts and expect a reasonable return of investment from the targeted customer segment. Further, customer segmentation information can be used to design new products, determine the manufacturer's suggested retail price (MSRP) or the recommended retail price (RRP) for a new product, or estimate the success of a product or service in the marketplace.
Ad hoc analysis 140 facilitates identification of high-value customer segments with unlimited real-time visitor information, e.g., drill down into the data to get deep, precise, and comprehensive views of the customers. Ad hoc analysis 140 may also provide analysis or visualizations for customer segments over time (e.g., minutes, hours, days, weeks, etc.).
Target 150 tracks progress against target goals, e.g., based on customer segmentation. For example, target goals can be set based on customer segmentations from a geographic region or customer segmentations associated with specific transactions. Target 150 can also be used to measure performance of a website. When a target is created, one or more specific attribute metrics are measured, or an entire website is measured against some selected metrics. As an example, one can measure the number of visitors to a website (i.e., customer segment to the website) and use it as a target. Meanwhile, the customer segment from a specific source (e.g., geographical region, demographic characteristic) can also be used if the target is further drilled down to the number of visitors to the website from the specific source.
Although
Customer segmentation system 200 uses GKCC for customer segmentation. As described herein, GKCC is based on greedy center allocation in an augmented partition space built from basic partitions generated in the original feature space. To generate the basic partitions, a greedy dynamic search process can be used to incrementally choose new partition centers, which mitigates the sensitivity issue (e.g., not able to select reasonable initial partition centers) related to the initialization stage of K-means clustering methods. A predetermined number (e.g., 59) is used in the sampling strategy to accelerate the greedy dynamic search process. Advantageously, customer segmentation system 200 overcomes the sensitivity issue of traditional K-means clustering and returns partitions with small objective function value and small deviation.
Customer segmentation system 200 utilizes basic partition constructor 210 to receive customer data with various customer attributes and to determine basic partitions of the customers in the original feature space.
The term “feature space” refers to an n-dimensional space for hosting numerical features that represent objects. As an example, an object is represented by an n-dimensional vector in the feature space. The term “original feature space” refers to the feature space associated with the original features extracted from the raw data of the objects, e.g., raw customer information. By way of example, customers may have various attributes, such as attributes of demographics, attributes of computing devices, attributes of online activities, etc. One or more of such customer attributes may be used as the original features for representing the customers and constructing the original feature space.
In various embodiments, the basic partitions are incrementally generated in multiple stages of partitioning in a dynamic partition process. An object in the feature space may be also referred as a point in the feature space. In one embodiment, 59 points from the original feature space are randomly selected as the candidates. The point with the minimum objective function value is further selected and added to the set of existing cluster centers for computing the new cluster centers.
In some embodiments, the basic partitions are generated with the cluster number varying from 2 to 2K, where K is a user-defined cluster number, e.g., defined based on the specific needs of reports and analytics 120, marketing cloud 130, ad hoc analysis 140, and target 150 in the customer analysis system 100. For a certain stage k in the GKCC process, one random selected point for K-means clustering is added to the cluster centers from the previous stage to determine a set of partitions and the object function value associated with the set of partitions. For the certain stage k, such process is repeated for a predetermined number of times (e.g., 59 times), which results in different sets of partitions (e.g., 59 sets) for later fusion and respective objective function values associated with those randomly selected points. The point associated with the minimum objective function value is added to the existing centers to further compute the partitions and return a new set of centers for the next stage.
The partition space transformer 220 is to transform the original feature space to an augmented partition space based on membership information of the customers in the basic partitions, e.g., including the final set of basic partitions and the intermediate basic partitions. The phrase “augmented partition space” refers to the augmented feature space, which summarizes high-level information compared with the original feature space. By way of example, membership information of the customers in basic partitions in the original feature space may be used to construct the augmented partition space. “Membership information” refers to whether a customer belongs to a basic partition. In some embodiments, membership information is represented by a binary value to indicate the customer either in or not in a partition. In some embodiments, membership information is represented by a probability to indicate the likelihood of the customer in the partition.
In some embodiments, partition space transformer 220 is to construct a data structure to represent the augmented partition space with elements in the data structure corresponding to membership information of the customers in the basic partitions, e.g., as shown in augmented partition space 330 in
The consensus clustering builder 230 generally determines consensus-based partitions of the customers in the augmented partition space (e.g., represented by the concatenated binary matrix). In one implementation, consensus-based partitions can be determined based on multiple stages of partitioning process operated in the augmented partition space. In various embodiments, after a binary matrix being derived from the basic partitions, K-means based clustering is conducted on the binary matrix. This dynamic partition process starts with two centers (i.e., from k=2). One can be the center of the binary matrix, and the other can be a randomly selected point in the binary matrix. For k=2, the dynamic partition process generates two partitions as well as an objective function value for the randomly selected point.
This dynamic partition process is repeated several times, e.g., with another randomly selected point to replace the previously randomly selected point. Subsequently, the randomly selected point with the minimum objective function value is selected to determine the partitions and the corresponding centers for the next stage of processing. Eventually, consensus clustering builder 230 stops when the consensus clustering process yields the predetermined cluster number K.
The customer manager 240 is to manage customers and customer attributes. In various embodiments, customer manager 240 provides different interfaces to other components in the customer analysis system 100 of
In other embodiments, customer segmentation system 200 can be implemented differently than what is depicted in
In some embodiments, customer segmentation system 200 is embodied as a specialized computing device. In some embodiments, customer segmentation system 200 can be embodied, for example, as an application or a mobile app. In some embodiments, customer segmentation system 200 can be a distributed system; for example, basic partition constructor 210, partition space transformer 220, consensus clustering builder 230, and customer manager 240 can be distributed across any number of servers. Regardless of the computing platform on which customer segmentation system 200 is implemented, customer segmentation system 200 can be embodied as a hardware component, a software component, or any combination thereof for managing customer segmentation.
The augmented partition space 330, in this embodiment, is constructed by concatenating the membership information of all basic partitions in a binary matrix, where the positive membership is represented as 1 and the negative membership is represented as 0. Subsequently, consensus-based clusters 340 are determined in this augmented partition space. GKCC uses high-level information to capture more meaningful cluster structures as GKCC conducts the cluster analysis on the augmented partition space 330, rather than the original feature space 310. Meanwhile, GKCC overcomes the sensitivity of K-means initialization based on its greedy dynamic search and incrementally adding new centers. Further, a fixed numbers of sampling (e.g., 59) is adopted in this GKCC process for acceleration and efficiency.
In various embodiments, the process begins at block 410, where basic partitions are generated in the original feature space, e.g., by basic partition constructor 210 of
At block 420, the original feature space is transformed into an augmented partition space, e.g., by partition space transformer 220 of
To further illustrate the example process 400 for customer segmentation, a particular embodiment, GKCC-59, is listed herein, with training data X, the cluster number K, and the predetermined sampling number 59 as the input to the GKCC. Let X={x1, x2, . . . , xn} be a set of n data points belonging to K clusters, denoted as C={C1, . . . , Ck} where Ck∩Ck′=Ø, ∀k≠k′, and ∪k=1KCk=X. Given r basic partitions represented as π={π1, π2, . . . , πr}, each of which partitions X into Ki clusters, and maps each data point to a cluster label ranged from 1 to Ki.
Step 2. Generate basic partitions and update centers.
π=π∪π′
end
Run K-means on X with the initial center C∪dargmin
Step 3. k=k+1.
end
Build the binary matrix B by Eq. 3;
Let B=Ø be the set of centers;
B=B∪ the center of B;
Set k=2;
while k≤K do
Step 1. Sampling 59 points from B, bi with 1≤i≤59.
Step 2. Generate consensus clustering and update centers.
end
Run K-means on B with the initial center
and return the partition as πi and the new centers as B.
Step 3. k=k+1.
End
Output: π and πi with 1≤i≤K.
The goal of consensus clustering is to find an optimal consensus partition π, in other words, to find the consensus partition sharing the maximum utility function value with basic partitions as shown in Eq. 1, where U is a utility function that measures the similarity between two partitions of (π, πi). In this way, the utility function (U) is used to measure the relationship (e.g., similarity) between one set of partitions (e.g., π) and another set of partitions (e.g., πi). In some embodiments, the Categorical Utility Function (CUF) in Eq. 2 is used as the utility function (U). In Eq. 2, pkj(i) is the joint probability of one instance simultaneously belonging to Ck and Cji. Here, Ck is the k-th cluster in final partition π, and Cji is the j-th cluster in πi. pk+ and p+j are the cluster portion of π and πi, respectively.
The complex consensus clustering problem with CUF can be mapped into a K-means clustering problem with a binary matrix. Let B={b(x)} be a binary dataset derived from the set of r basic partitions II as shown in Eq. 3. In some embodiments, B is the concatenated matrix of all the basic partitions in 1-of-Ki coding, where Ki is the cluster number of πi. The final consensus clustering is obtained by running K-means on B with squared Euclidean distance in one embodiment.
This embodiment, GKCC-59, handles three challenges together, namely how to generate basic partitions, set the proper cluster number, and handle the initialization sensitivity of K-means. At block 410, the basic partitions are generated from the original feature space. At block 430, the consensus clusters are generated from the augmented partition space. In one embodiment, in the first loop for generating the basic partitions, the cluster number is conditioned to be 2K, while in the second loop, the cluster number for consensus clustering is conditioned to be K. Therefore, the number of stages for generating the basic partitions in the original feature space is greater than the number of stages for obtaining the consensus-based final partitions in the augmented partition space.
In the example of GKCC-59, the process starts with one center, i.e., the center of the data X, and incrementally adds new centers by randomly selecting 59 points as the candidates and picking up a candidate for the next stage of partitioning per the objective function value of K-means. During this process, 59 clustering results are obtained for a certain cluster number, which are used as basic partitions for further consensus clustering. After obtaining the set of basic partitions π, the greedy strategy is still applied for the consensus partition. Further, sampling is limited to 59 points to accelerate the search process. In other embodiments, the predetermined number for sampling can be in a range of 40 to 80, or other suitable sampling numbers based on specific applications.
During the basic partition generation, the cluster number is varied from 2 to 2K to increase the diversity of the basic partitions. Moreover, for a certain cluster number k, the (k−1) centers from the previous stage and one additional randomly selected point are used for K-means clustering. As a result, a large number of basic partitions can be obtained to construct the augmented partition space in the GKCC process.
Here, 59 points are sampled to choose the optimal point to be used to obtain the new cluster centers for the next stage. By dynamically adding new centers, GKCC mitigates the sensitivity issue of K-means because the partition centers do not need to be selected at the same time. Further, the number of samples is limited to the predetermined number of 59 here to avoid the brute-force global search for an optimal new center. Even further, all resulting 59 partitions in each stage can be used to construct the augmented partition space. In this embodiment, there are 59*(2K−1) basic partitions, which is likely large enough to construct a feature-rich augmented partition space.
GKCC is suitable for large-scale clustering as its time complexity is linear to the number of customers. The time complexity for generating basic partitions is O(InK2m), where I is the average stage number, n is the number of points (i.e., customers), K is the cluster number and m is the number of features. The time complexity for consensus clustering is O(InK3). Since K<<n and m<<n, the overall time complexity of GKCC is linear to n. Therefore, GKCC is suitable for large-scale clustering. Further, GKCC generally returns the stable partitions with a small variance.
Referring now to
At block 510, a predetermined number of customers are randomly selected from a group of customers as a set of candidates for the present stage, e.g., enabled by basic partition constructor 210 of
At block 520, for each candidate of the set of candidates, the candidate is added to the existing centers to generate a set of basic partitions of the customers in the original feature space, e.g., based on the K-means clustering. As a result, each candidate will have a corresponding set of basic partitions.
At block 530, respective objective function values for the set of candidates are determined based on, e.g., the standard K-means objective function value associated with a set of basic partitions. In some embodiments, the objective function value indicates a distance measure of the customers from their respective partition centers, e.g., based on a squared error function. Subsequently, the objective function value associated with the candidate can be determined based on the distance measure.
At block 540, the candidate with a minimum objective function value is added to the set of partition centers determined at a prior stage, and basic partitions for the present stage are generated based on the current set of partition centers, e.g., after running a K-means clustering process. At block 550, the new set of centers are returned. Subsequently, the process is moving to the next stage or iteration.
At block 610, a binary matrix is constructed to represent the augmented partition space based on the membership information in the basic partitions, e.g., enabled by consensus clustering builder 230 of
At block 620, the process samples a predetermined number of points in the augmented partition space, e.g., enabled by consensus clustering builder 230 of
At block 630, consensus clustering is performed, e.g., based on K-means on the binary matrix, e.g., enabled by consensus clustering builder 230 of
Having briefly described an overview of embodiments of the present invention, an exemplary operating environment in which embodiments of the present invention is to be implemented is described below to provide a general context for various aspects of the present invention. Referring initially to
The disclosure is described in the general context of computer code or machine-useable instructions, including computer-executable instructions such as program modules, being executed by a computer or other machines, such as a smartphone or other handheld devices. Generally, program modules, including routines, programs, objects, components, data structures, etc., refer to code that perform particular tasks or implement particular abstract data types. The embodiments of this disclosure are to be practiced in a variety of system configurations, including handheld devices, consumer electronics, general-purpose computers, more specialty computing devices, etc. The embodiments of this disclosure are also to be practiced in distributed computing environments where tasks are performed by remote-processing devices that are linked through a communications network.
Regarding
Computing device 700 typically includes a variety of computer-readable media. Computer-readable media includes any available media to be accessed by computing device 700, and includes both volatile and nonvolatile media, and removable and non-removable media. By way of example, and not limitation, computer-readable media comprises computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which is used to store the desired information and which is accessed by computing device 700. Computer storage media does not comprise signals per se. Communication media typically embodies computer-readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media. Combinations of any of the above should also be included within the scope of computer-readable media.
Memory 720 includes computer storage media in the form of volatile and/or nonvolatile memory. In various embodiments, the memory is removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical-disc drives, etc. Computing device 700 includes one or more processors 730 that read data from various entities such as memory 720 or I/O components 760. Presentation component(s) 740 present data indications to a user or a device. Exemplary presentation components include a display device, speaker, printing component, vibrating component, etc.
In various embodiments, memory 720 includes, in particular, temporal and persistent copies of segmentation logic 722. Segmentation logic 722 includes instructions that, when executed by one or more processors 730, result in computing device 700 managing customer segmentation, such as, but not limited to, process 300, process 400, process 500, or process 600. In various embodiments, segmentation logic 722 includes instructions that, when executed by processors 730, result in computing device 700 performing various functions associated with, but not limited to, basic partition constructor 210, partition space transformer 220, consensus clustering builder 230, or customer manager 240, in connection with
In some embodiments, one or more processors 730 are to be packaged together with segmentation logic 722. In some embodiments, one or more processors 730 are to be packaged together with segmentation logic 722 to form a System in Package (SiP). In some embodiments, one or more processors 730 are integrated on the same die with segmentation logic 722. In some embodiments, processors 730 are integrated on the same die with segmentation logic 722 to form a System on Chip (SoC).
I/O ports 750 allow computing device 700 to be logically coupled to other devices including I/O components 760, some of which are built-in components. Illustrative components include a microphone, joystick, game pad, satellite dish, scanner, printer, wireless device, etc. In some embodiments, the I/O components 760 also provide a natural user interface (NUI) that processes air gestures, voice, or other physiological inputs generated by a user. In some embodiments, inputs are to be transmitted to an appropriate network element for further processing. Additionally, the computing device 700 is equipped with sensors (e.g., accelerometers or gyroscopes) that enable detection of motion. The output of the sensors is to be provided to the display of the computing device 700 to render immersive augmented reality or virtual reality.
Although certain embodiments have been illustrated and described herein for purposes of description, a wide variety of alternate and/or equivalent embodiments or implementations calculated to achieve the same purposes could be substituted for the embodiments shown and described without departing from the scope of the present disclosure. This application is intended to cover any adaptations or variations of the embodiments discussed herein. Therefore, it is manifestly intended that embodiments described herein be limited only by the claims.
An abstract is provided herein to facilitate the reader to ascertain the nature and gist of the technical disclosure. The abstract is submitted with the understanding that it will not be used to limit the scope or meaning of the claims. The following claims are hereby incorporated into the detailed description, with each claim standing on its own as a separate embodiment.
