Patent Application
20200320548
The present disclosed technology relates to systems, methods, and computer readable storage media containing instructions for producing estimator models of unknown consumer behaviors. The present disclosed technology further relates to methods and systems for attributing revenue to individual consumers, and automatically segmenting consumers into related groups.
Organizations who market products and services indirectly (“indirect businesses”) frequently struggle to collect complete usage data at the consumer level. That is, while the indirect business may know in the aggregate that they have a particular number of consumers and the total consumption of their products or services, they are unable to determine what amount of each product or service is consumed by a particular consumer. Even as organizations possess ever increasing volumes of information about their consumers, indirect businesses like film production companies, sports leagues, and similar organizations do not have complete information about how each consumer is using their products and services. This is, in part, because there are numerous intermediating companies between indirect businesses and their ultimate consumer, such as TV networks, internet streaming services, merchandise shops, and others. Frequently, these intermediating companies can and do collect direct usage metrics for individual users, but they are either unable or unwilling to share that data. While the problem is particularly acute for media companies, the problem of indirect sales is present across many other industries, such as movie studios who sell through movie theaters, etc. Often the only reliable source of usage metrics comes from survey companies that collect data from a reasonably sized and/or representative sample of consumers that agree to have certain behaviors monitored or recorded. Even if these survey companies are willing to share aggregate statistics, and extrapolated estimates of total behavior, many such companies are still unwilling or unable to share per-consumer survey data, even for their panel members.
Even without per-consumer consumption data, many indirect businesses nevertheless have considerable databases with other consumer information. This information can come from manually submitted preference or profile data, such as a consumer's name, address, birthdate, favorite team or show, etc. In addition, indirect businesses often log information regarding interactions consumers have with the organization's products and services that are under the organization's control, such as webpage views, merchandise purchased, digital subscriptions, etc. These organizations may also be able to obtain consumer data from affiliates and business partners whose incentives are aligned with the indirect business's or who have contractual arrangements that permit sharing of certain consumer data.
Nevertheless, extremely valuable information—per-user consumption of their primary products or services—often remains out of reach. Because these indirect businesses cannot identify each individual consumer's consumption levels, they are often unable to identify consumers that are high-volume purchasers and users, understand how each consumer's interests affects his or her consumption, and are missing important information to shape future decisions about marketing and new products and services for its consumers. This long felt need to estimate per-product or per-service consumption, while preserving consumer privacy, is solved by the present disclosed technology, among other benefits.
The technology disclosed herein meets this need by using data within an indirect business's possession, combined with per-consumer data in the possession of a survey company to train a machine learning model to estimate, based on information in the indirect business's possession, per-consumer consumption values. The present disclosed technology further provides for combining the indirect business's data and the survey company's data on a system isolated from the indirect business. The indirect business can define machine learning or statistical models to be trained on the combined data, validate the trained model, and use it to estimate consumption by consumers in its own records without accessing the combined information. In this way, the per-consumer consumption information in possession of a survey company is not exposed to the direct business, the survey panel member's privacy is preserved, and the indirect business is able to obtain accurate estimates of each consumer's consumption of their products and services.
The present disclosed technology further includes a system and method for using consumer data to produce actionable consumer segmentation, including in conjunction with the modeling process described above. The present disclosed technology further provides methods to attribute revenue and/or profit derived from numerous sources to each consumer on a per-consumer level. Aspects of the disclosure relate to a computing system, computer-implemented method, and computer readable storage media for predicting, for a consumer in a group of consumers assigned to a subset of consumers in a plurality of subsets at the present time, which subset the consumer will be assigned to at a predetermined time in the future. Further aspects of the disclosure relate to a computer-implemented method for determining the likelihood that consumers in a group of consumers will migrate between subsets in a plurality of subsets of consumers. Further aspects of the disclosure related to a computer-implemented method for determining an expected revenue for a group of consumers at a predetermined time in the future.
While certain embodiments of the present disclosure are described with respect to indirect businesses determining media consumption, and the use of survey panel data, it is understood that the full scope of the disclosed technology encompasses the use of similar techniques for indirect businesses determining consumption of other products or services. Indeed, embodiments of the present disclosure are suitable for producing models to estimate product and/or service consumption, or any other consumer behavior, based on combining internal available data with actual measured values of that consumer behavior in the possession of another organization.
Aspects of the technology disclosed herein include a computer-implemented method for training a model to estimate an unknown consumer behavior while preserving consumer privacy, comprising the steps of: calculating, by a modeling computing system, a vector representing a consumer's behavior by applying one or more vectorization rules to a set of behavioral attribute values for the consumer in a data set of a first organization, transforming the vector, by the modeling computing system, into an estimated unknown consumer behavioral attribute by applying a trained model, wherein the model was trained by: providing, to a second computing system separate from the modeling computing system, a first training data set of the first organization for a plurality of consumers, wherein the training data set comprises, for each consumer, common identifying information, and a set of behavioral attribute values, and a second training data set of a second organization for a plurality of consumers, wherein the second training data set comprises, for each consumer, common identifying information, and actual behavioral attribute values for the unknown behavioral data attribute, combining, by the second computing system, the first training data set and the second training data set into a joined data set by joining the one or more behavioral data attribute values of the first data set and the actual behavior attribute values for the unknown behavioral attribute of the second data set, for each consumer, where the common identifying information of the first data set and the common identifying information of the second set correspond, calculating, by the second computing system, a vectorized training data set by applying the one or more vectorization rules to vectorize the set of behavioral attribute values for each consumer in the joined data set into a set of training vectors, accepting, by the second computing system, a trainable model definition from an external source, wherein a trainable model defined by the trainable model definition accepts a training vector as input, and produces an estimated unknown behavioral data attribute as an output, creating in one or more memories of the second computing system the defined trainable model, training, by the second computing system, the defined trainable model on the training vectors corresponding to a first subset of consumers in the joined data set to produce the trained model.
In some embodiments, training the model further comprises validating, by the second computing system, the trained model by transforming each training vector in a second subset of consumers in the joined data set into an estimated unknown behavioral data attribute, comparing the estimated unknown behavioral data attribute to the actual behavioral attribute value to produce a performance metric, and determining whether the performance metric falls within a predetermined range of acceptable performance values, wherein the first subset of consumers and the second subset of consumers is mutually exclusive. In some embodiments, the trained model is a decision tree, and wherein the step of training the defined trainable model further comprises pruning each terminal node of the decisions tree wherein the terminal node includes fewer than a predetermined number of consumers, or a predetermined percentage of the total number of consumers. In some embodiments, the set of behavioral attributes of the first training data set is a subset of available behavioral attributes, wherein the set of behavioral attributes are selected such that each behavioral attribute in the set of behavioral attributes is substantially statistically independent. In some embodiments, each behavioral attribute in the first training data set was selected by: evaluating the statistical independence of each pair of available behavioral attribute values for consumers in a consumer data set comprising available behavioral attribute values for a plurality of consumers, aggregating the available behavioral attributes into substantially statistically independent subsets of behavioral attributes, and selecting one behavioral attribute from each substantially statistically independent subset of behavioral attributes. In some embodiments, the first organization cannot access the second training data set of a second organization, and the joined data set. In some embodiments, training the model further comprises the step of validating the trained model by calculating the sum of the estimated unknown consumer behavior for all consumers in a consumer data set, and comparing the sum to an estimated total behavior. In some embodiments, training the model further comprises the step of destroying the joined data set.
Aspects of the disclosed technology include a computer-implemented method for predicting, for a consumer in a group of consumers assigned to a subset of consumers in a plurality of subsets at the present time, which subset the consumer will be assigned to at a predetermined time in the future, the method comprising: receiving consumer data for each consumer in the group of consumers, wherein the consumer data can be represented as a point in n-dimensional space, and wherein the dimensions include behavioral, demographic, and financial data attributes for the consumer, and wherein the demographic data attributes comprises one or more age-related data attributes, receiving a subset assignment for each consumer in the group of consumers, wherein the consumers assigned to each subset have, on average, a shorter distance between their respective representative points relative to the distance between consumers in different subsets, according to a predetermined distance metric, identifying a look-alike group of consumers from the group of consumers for the consumer, wherein the demographic data attributes of each consumer in the look-alike group is substantially similar to the consumer, except that at least one age-related data attribute of each consumer in the look-alike group of consumers is similar to the sum of the consumer's age data attribute and the difference between a current time and the predetermined future point in time, and determining the probability that the consumer will be assigned to each subset in the plurality of subsets at the predetermined time in the future as a function of the proportion of consumers in the look-alike group in each subset relative to all consumers in the look-alike group.
In some embodiments, the method can further comprise: measuring the distance between the point representing the consumer and a centroid of the points representing consumers in a subset for each subset in the plurality of subsets, and wherein the step of determining the probability that a consumer will be assigned to each subset further comprises determining the probability as a function of the distance between the point representing the consumer and a centroid of the second subset relative to the distance between the point and the other subsets in the plurality of subsets.
In some embodiments, the method can further comprise determining a standard deviation for each subset in the plurality of subsets for the distances between the point representing each consumer in the subset of consumers and a centroid of the subset, and determining that the consumer will be assigned to the subset it is assigned to at the present time at the predetermined time in the future if the distance between the point representing the consumer and the centroid of the subset to which it is assigned at the current time is below a predetermined multiple of the standard deviation for that subset.
Aspects of the disclosed technology can include a computer-implemented method for determining the likelihood that consumers in a group of consumers will migrate between subsets in a plurality of subsets of consumers, comprising: performing a method to predict the likelihood that a consumer will migrate from their assigned subset to another subset for each consumer in the group of consumers, for each pair-wise set of subsets in the plurality of subsets, calculating a weighted average of the probability that consumers in the first subset in the pair-wise set of subsets will be assigned to the second subset in the pair-wise set of subsets at the predetermined time in the future.
In some embodiments, where each subset in the plurality of subsets has an expected value associated with the subset, the method can further comprise: predicting an expected value associated with the consumer at the predetermined time in the future by multiplying the probability that a consumer will belong to each subset at the predetermined time in the future by the expected value associated with each subset.
Aspects of the disclosed technology can further include a computer-implemented method for determining an expected revenue for a group of consumers at a predetermined time in the future, comprising: predicting an expected value for each consumer in the group of consumers, and summing the expected value for each consumer in the group of consumers.
In some embodiments, the one or more age-related data attribute is selected from the group of: age, educational attainment, job position, and length of engagement with a product or service.
Embodiments of the disclosed technology can further include computing systems configured to implement the methods disclosed herein, and non-transitory computer-readable storage media containing instructions that, when executed, perform such methods. Further, embodiments of the present disclosed technology can comprise combinations of the features described above, as well as the various features described in the description, figures, and claims.
Included in the present specification are figures that illustrate various embodiments of the present disclosed technology. As will be recognized by a person of ordinary skill in the art, actual embodiments of the disclosed technology need not incorporate each and every component illustrated, but may omit components, add additional components, or change the general order and placement of components. Reference will now be made to the accompanying figures and flow diagrams, which are not necessarily drawn to scale, and wherein:
The following detailed description is directed to systems, methods, and computer-readable media for privacy-preserving generation of predictive models for estimating consumer behavior.
Although example embodiments of the present disclosure are explained in detail, it is to be understood that other embodiments are contemplated. Accordingly, it is not intended that the present disclosure be limited in its scope to the details of construction and arrangement of components set forth in the following description or illustrated in the drawings. The present disclosure is capable of other embodiments and of being practiced or carried out in various ways.
It must also be noted that, as used in the specification and the appended claims, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise. Moreover, titles or subtitles may be used in this specification for the convenience of a reader, which shall have no influence on the scope of the present disclosure.
The term “comprising” or “containing” or “including” is meant that at least the named element, material, or method step is present in the composition or article or method, but does not exclude the presence of other elements, materials, or method steps, even if the other such elements, material, or method steps have the same function as what is named.
In describing example embodiments, terminology will be resorted to for the sake of clarity. It is intended that each term contemplates its broadest meaning as understood by those skilled in the art and includes all technical equivalents that operate in a similar manner to accomplish a similar purpose.
It is to be understood that the mention of one or more steps of a method does not preclude the presence of additional method steps or intervening method steps between those steps expressly identified. Steps of a method may be performed in a different order than those described herein. Similarly, it is also to be understood that the mention of one or more components in a device or system does not preclude the presence of additional components or intervening components between those components expressly identified.
In the following detailed description, references are made to the accompanying drawings that form a part hereof and that show, by way of illustration, specific embodiments or examples. In referring to the drawings, like numerals represent like elements throughout the several figures.
Various products and services provided by third parties are mentioned as example components of embodiments in accordance with the disclosed technologies. The use of trademarked (registered or common-law) names are intended for descriptive purposes only—no claim of ownership over those terms is asserted by the applicants by this application. Further, the mention of a trademarked product or service is as an example only. Other products and services providing equivalent functions, whether commercial, open-source, or custom-developed to support embodiments are contemplated in accordance with the disclosed technology.
Certain examples of the disclosed technology are discussed and shown herein using names, addresses, behavioral attributes, financial data, and other forms of personal data. All such data is fictitious. No actual personal data is provided herein. Any correspondence between data provided in this application and actual persons, living or dead, is purely coincidental. In addition, the examples of business metrics are merely examples. Embodiments of the present disclosed technology are not limited to merely these metrics.
The term “attribute” as used herein refers to a category of data kept for each record in a plurality of records. In most instances, each record is a consumer, unless the context indicates otherwise. Non-limiting examples of attributes can comprise a name, address, value of merchandise purchased, season ticket spend, etc.
The term “value” as used herein refers to a specific value for a specific attribute. For example, an “attribute” can be names, and an attribute value can be a particular consumer's name.
The term “trainable model” as used herein refers to a computer model that (1) estimates an output value or values as (2) a function of one or more input values, and (3) is configured by processing one or more training examples, each training example comprising a target output value or values and one or more input values corresponding to the target output value or values.
The terms “known behavioral attribute” and “unknown behavioral attribute” as used herein refer to whether the behavioral attribute is known to a first organization. That is, whether something is “known” is determined relative to the first organization. An “unknown behavioral attribute” as used herein refers to a behavioral attribute that is known to another organization, but unknown to the first organization.
The term “partner behavioral attribute” as used herein refers to a behavioral attribute that is known to a third-party and unknown to a first organization, but accessible by the first organization.
Referring now to
Thus, as configured in
Embodiments of the present technology can also be implemented using cloud-based technologies, such as those depicted in
Embodiments of the disclosed technology can be built using one or more elements of cloud computing technology as shown in
Application definition and development components 201 (“ADD”) enable developers to define and develop applications prior to deployment, and to refine those designs in subsequent versions. ADD components 201 can include database and data warehouse components 201a that provide data sets and data storage for application development. These database and data warehouse components 201a include relational and non-relational data stores, graph databases, flat files, and other data storage technologies. ADD components 201 can further include streaming components 201b that facilitate rapid distribution of data to numerous system endpoints, such as message queues, stream processing software, and other data distribution systems. ADD components 201 can further include source code management components 201c, such as Git, Mercurial, Subversion, and other similar source management systems. Source code management components 201c can also include cloud-based servers for version control, such as GitHub or GitLab. ADD components 201 can further include application definition and image build components 201d that allow developers to define cloud-based infrastructure, including configurations of application servers, software defined networks, and containerized services. ADD components 201 can further include continuous integration and continuous delivery (CI/CD) components 201e that automate the process of application testing and deployment. CI/CD components 201e can be configured to automatically run automated tests on application software (e.g. such as when a change is committed to a version control platform), and if the tests are successful, to deploy the application software to a production environment.
Orchestration & management (“OM”) components 202 facilitate the containerization and subsequent coordinated execution of application software. OM components 202 include scheduling and orchestration components 202a that schedule and run containerized software. Non-limiting examples of scheduling and orchestration components 202a include Kubernetes and Docker Swarm. OM components 202 can further include coordination and service discovery components 202b that allow software to automatically discover cloud-based resources, such as data stores, data streaming sources, etc. OM components can further include service management components 202c that can include load balancers, reverse proxy systems, auto scalers, and other components that facilitate autonomous or manual application scaling.
Runtime components 203 can include basic environments for the support execution of cloud-based application software. Runtime components 203 can include cloud-native storage 203a, such as object stores, virtual file systems, block storage, and other forms of cloud-centric data storage. Runtime components 203 can include container runtimes 203b that provide the foundation for containerized application software, such as Docker or Rkt. Runtime components 203 can further include cloud-native network components 203c that provide software-defined networking and virtual private cloud technologies that enable components of cloud-based systems to communicate with each other, as well as with the wider Internet.
Provisioning components 204 can include components intended for configuring cloud components and triggering the creation of cloud resources on various cloud platforms. Provisioning components can include Host Management and Tooling components 204a that define and deploy configurations of cloud components when executed. Provisioning components 204 can further include infrastructure automation components 204b that automate basic cloud infrastructure tasks. Provisioning components 204 can further include container registries 204c that provide storage for containerized cloud applications that are deployable by other provisioning components. Provisioning components can further include secure image components 204d that provide security and verification for container images to ensure consistent and reliable deployment of trusted container images. Provisioning components can further include key management systems 204e that provide for secure storage of cryptographic keys.
Serverless components 205 can include components for deploying cloud applications that do not rely upon a continuously running (or scheduled) runtime execution, but instead run discrete components of functionality given a condition. Serverless components 205 can include components 205a to simplify the development of serverless applications, such as components that convert server-centric software into serverless code, event simulators, and simulations of cloud-based serverless platforms. Serverless components 205 can also include frameworks 205b that are predefined systems that take code in certain configurations and deploy them as serverless applications in cloud environments. Serverless components 205 can also include security components 205c that help to secure serverless applications.
Observability & analysis components (“O&A”) 206 can include systems for monitoring running cloud applications, detecting and observing defects and errors, and logging system performance. O&A components 206 can include monitoring components 206a that monitor running systems to display and/or record performance metrics, error rates, and other application data. O&A components 206 can also include logging components 206b that collect system logs from cloud-based components and aggregate them in a single place or to a single access point to review system performance. O&A components 206 can also include tracing components 206c that collect detailed trace logs when cloud components run into errors, system exceptions, and other problematic behaviors to assist in the identification and remediation of problems in cloud-based systems.
In some embodiments, one or more methods are embodied in a set of instructions for one or more processors having access to one or more types of memory. The instructions could be coded in hardware or in software. Many kinds of platforms may be used, including but not limited to: computers, mobile telephones, tablet devices, game consoles, network management devices, field-programmable gate arrays, and cloud-based computer systems. Aspects of the disclosure could be deployed on multiple devices for concurrent operation. Embodiments may be used as a component of a larger system.
As will be appreciated by one skilled in the art, aspects of the present disclosed technology may be embodied as a system, method or computer program product. Accordingly, aspects of the present disclosed technology may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects that may all generally be referred to herein as a “circuit,” “module” or “system.” Furthermore, aspects of the present disclosed technology may take the form of a computer program product embodied in one or more computer readable medium(s) having computer readable program code embodied thereon.
Any combination of one or more computer readable medium(s) may be utilized. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In some embodiments, the computer readable medium can be a non-transitory storage system on a cloud platform, such as, for example, in a database or data warehouse component 201a, a source code management tool 201c, cloud-native storage component 203a, embodied in a container image stored locally or in a container registry 204c, or deployed in a container runtime 203b. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.
A computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.
Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
Computer program code for carrying out operations for aspects of the present disclosed technology may be written in any combination of one or more programming languages, including languages such as Java, Scala, Python, Julia, R, SAS, Go, Ruby, Javascript, Smalltalk, C++ or the like. As defined herein, computer program code also includes the build artifact of the above languages, or similar languages and environments, such as object code, byte- or word-code, or other compiled, interpreted, transpiled, or otherwise processed code. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on one or more remote computers, servers, or serverless cloud platforms. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).
Aspects of embodiments of the present disclosed technology are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.
The disclosed technology is disclosed in terms of modules and submodules, each of which are to be understood as discrete units of functionality, which can be embodied as classes, modules, functions, compilation or build artifacts, or other components of one or more programming languages used to implement embodiments of the disclosed technology. While the present description illustrates one organization of the various modules and submodules for implementing embodiments of the disclosed technology, the disclosed technology is not so limited. Embodiments of the present disclosed technology can include other organizations for implementing equivalent or overlapping functionality for the various modules described herein, such as by sharing functionality between modules, combining modules, separating modules into multiple modules, implementing class hierarchies and the like. Additionally, the accompanying drawings illustrate examples of relationships between various modules and submodules (such as by flowchart connectors or inclusion of modules as sub-modules of other modules), but these relationships are not limiting. As would be recognized by a person of ordinary skill in the art, the output of any given module is available to be included as part of the input of any other component in accordance with various embodiments.
These computer program instructions may also be stored in a computer readable medium that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the computer readable medium produce an article of manufacture including instructions that implement the function/act specified in the flowchart and/or block diagram block or blocks.
The computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions that execute on the computer or other programmable apparatus provide processes for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.
Technical effects and benefits include producing trained models for estimating an unknown consumer behavior based on one or more known consumer behaviors, and in applying the trained model to estimate unknown consumer behaviors.
Blind data modeling 310 can comprise a first organization that provides an internal data set 311 to a third-party. The internal data set 311 comprises one or more behavioral attributes and common identifying information (“CII”) for a plurality of the first organization's consumers. CII can comprise information that allows behavioral attributes from the internal data set and external data set for a single consumer to be identified. Non-limiting examples of CII can include alphanumeric identifiers, personally identifiable information (e.g. name/address), or values derived therefrom, such as hash values. A second organization can provide an external data set 312 to the same third-party that comprises CII for a plurality of consumers, and an actual value for an unknown behavioral attribute, where the internal data set and external data set have at least some consumers in common. The third-party can join or merge the two data sets and model the relationships between the internal and external data sets 313 to produce a trained estimator model 314.
In some embodiments, the first organization can use this trained estimator model 314 to produce a “360-View” 320 of consumers that includes all known and estimated behavioral attributes, demographic, and CII data for each consumer. The first organization can apply the trained predictive model 321 to the internal data set 311 to estimate the unknown behavioral attributes for each consumer and create an expanded data set 322 that contains estimated consumption data for each consumer. In some embodiments, the first organization can further join or merge 324 the expanded data set 322 with one or more external data sets 323 (e.g., partner data). The one or more external data sets 323 can include behavioral attributes, CII, and other demographic information for each consumer. The first organization can perform financial attribution 325 to calculate the amount of revenue generated by each consumer based upon available data. The first organization can then construct a 360-View data set 326 that can comprise each consumer's CII, demographic information, behavioral attributes, estimated unknown behavioral attributes, and financial attribution values.
In some embodiments, the first organization can also perform automatic segmentation 330 on the 360-View data set 326. The first organization can also perform clustering 331 on consumer data sets 360-View. The clustering 331 will produce distinct groups of consumers that the first organization can join or merge with the 360-View data set 326. The first organization can use the joined 360-View data set to conduct a business analysis 340.
Benefits of the present disclosed technology can include that the first organization and the second organization can preserve the privacy and CII of their consumers by providing the internal data set 311 and the external data set 312 to a third-party processor. The first organization cannot access the second organization's external data set 312 and the second organization cannot access the first organization's internal data set 311. In addition, neither the first organization nor the second organization can access the joined data set that is constructed during the blind join/merge and model 313 step. In some embodiments, the third-party processor can destroy the joined data set to preserve the privacy and CII of consumers. Nevertheless, the first organization is able to define, train, validate, and/or use machine learning and statistical models to produce estimates of per-consumer consumption for each consumer in the first organization's database.
The disclosed technology can provide useful and accurate data while preserving consumer trust by preserving their privacy. Additionally, blind data modeling 310 can minimize or prevent potential violations of privacy laws, policies, and obligations.
Here, the modeling computing system 420 is described as under the control of a “third-party,” but the only requirement for preserving consumer privacy is that the first organization not have access to the second organization's data or the joined data set 700. In some embodiments, the modeling computing system 420 can be under the control of the second organization, or otherwise accessible to the second organization. Further, while the description uses the term “first” and “second” computing system, the actual designation as first or second is reversible, so long as the other features are present.
In order for the first organization to make use of this confidential and identifiable information in the possession of the second organization, a blind match and modeling process can be performed to provide estimates of an unknown behavioral attribute value for consumers known to the first organization. That is, on a first computer system 420 separate from, and generally inaccessible to, the first organization, the first training data set 500 in the possession of the first organization can be joined or merged 403 with the second training data set 600. Once merged, the separate computer system can train a statistical or machine learning model 405 to estimate, based on behavioral attribute values in possession of the first organization, the approximate value, or an estimate, of the unknown behavioral attribute value. The blind match can comprise exact matching on a CII attribute, fuzzy matching on one or more CII attributes, looking across multiple CII attribute values for the consumer, identifying a match if any one or a subset of attribute values match, inferring additional CII based on existing CII that is known from correlation databases (e.g. address-to-zip code), or machine learning/probabilistic models (e.g. name disambiguation).
Returning to the TV panel data example, a third-party, for example, can receive consumer data from the first and second organization, and merge the data, such that for a plurality of consumers, the matched data contains behavioral attribute values known to the first organization, and TV viewership data, such as minutes of sports watched. Based on the assumption that TV viewership habits are correlated with data in the possession of the first organization, a statistical or machine learning model can be trained on the merged data to estimate TV viewership based on known behavioral attribute values.
Behavioral attributes, both in the first training data set 500 and second training data set 600 can take a variety of forms. In some embodiments, behavioral attribute values can comprise dollar amounts, such as shop purchase 521, or season ticket spend 522. In some embodiments, behavioral attribute values can comprise Boolean data, such as whether the user has a digital subscription 523. In some embodiments, behavioral attribute values can comprise categorical data, such as digital subscription level 524, which can take multiple values. In some embodiments, such categorical data can be ordered (e.g., small, medium, large) or un-ordered (e.g., favorite team, state of residence, etc.). In some embodiments, behavioral attribute values can be simple numerical values, such as video game minutes played 525 or fantasy page views 526. In some embodiments, behavioral attribute values can be compound data objects, such as arrays, structures, maps, or similar that have multiple values accessible by key, index, or similar (not depicted). For example, a behavioral attribute value can comprise shop transaction data, which is an array of individual transactions. Each transaction can comprise a transaction date, amount, a list of items purchased, or other data. An example of a compound data object is shown below, illustrated in JSON format:
The first training data set 500 and the second training data set 600 are combined in a join/merge step 403. In some embodiments of the present disclosed technology, at least one consumer's data must exist in both the first training data set 500, and the second training data set 600. In other words, the CII for at least one consumer in the first training data set 500 must correspond to the CII for at least one consumer in the second training data set 600. For example, in first training data set excerpt 500, the bottom five users 530 are identical to the top five users 630 in the second training data set excerpt 600.
This correspondence need not be an exact match on one particular attribute, but the processor performing the join/merge step 403 must be capable of correlating the consumer's information in the first training data set 500 and the second training data set 600. This correlation can comprise exact matching on a CII attribute, fuzzy matching on one or more CII attributes, looking across multiple CII attribute values for the consumer, identifying a match if any one or a subset of attribute values match, inferring additional CII based on existing CII that is known from correlation databases (e.g. address-to-zip code), or machine learning/probabilistic models (e.g. name disambiguation). Numerous techniques are known to a person of ordinary skill in the art, any of which can be used to perform the correlations.
The result of the join/merge step 403 is a joined data set 700.
Once the joined data set 700 is produced, a model training 405 process is performed to produce an estimator model. The model training 405 process can comprise two steps. The first is a vectorization step which converts the values in the joined data set 700 into a simple numeric representation suitable for training an estimator model. The second step is to train the estimator model on the vectorized data. Vectorization of the data can comprise applying one or more rules to one or more attributes in the joined data set 700. In some embodiments, vectorization rules can comprise encoding rules, feature selection rules, standardization rules, and dimensionality reduction rules. However, as would be recognized by a person of ordinary skill in the art, each of these categories of steps may be omitted in certain circumstances. For example, if all behavioral values are already numeric, no encoding is necessary. Or if models produced without standardization or dimensionality reduction are computationally feasible, and provide adequate performance, those rules are unnecessary.
Encoding steps convert one or more attribute values into a format suitable for statistical analysis.
If the attribute is ordered categorical data (e.g. small/medium/large, standard/premium/ultimate), the categorical data can be encoded as an integer value indicating its position. For example, if the value is T-shirt size, and possible sizes are small, medium, and large, the vectorization rule could convert small to 0, medium to 1, and large to 2. If the attribute is un-ordered categorical data (e.g. favorite team, state of residence, etc.), the vectorization rule can be a one-hot encoding rule, such that each potential value of categorical data is represented as an additional attribute, and the attribute value is either a 0 if it does not correspond to the value, or 1 if it does. For example, as depicted in
If the attribute is a compound data object, such an array, structure, map, or similar indexed/keyed data structure, the vectorization rule can be a function to perform on the data to extract a numeric value, such as picking a particular value from the structure, or calculating a statistic based on the structure. For example, if the attribute was merchandise transactions, the attribute value might be an array of transactions, each one with a date and a transaction amount. The vectorization rule can be total merchandise sales, which would sum all the transactions amounts. The vectorization rule could also be merchandise sales in the last thirty days, in which case the vectorization rule could analyze the transactions, extract just those having a transaction date in the last thirty days, and then sum the transaction amounts.
In some embodiments, the vectorization rules can produce multiple output columns for one or more input columns. For example, a single numeric value can be vectorized into a first column by using the numeric value directly and into a second column as a bin number. Alternatively, a vectorized column can be vectorized as a ratio or comparison between multiple input columns (e.g. whether more minutes of video games are played than page views).
The vectorization rules can also comprise standardization rules to standardize, scale, center, or perform other statistical numeric transformations. Such transformations can include scaling the values to a range (e.g. normalization between 0 and 1), centering on average value, scaling to a known distribution (e.g. Gaussian distribution), filling in missing data (e.g. either a zero or a computed value such as an average value), or eliminating outliers (e.g. data beyond a certain distance, such as number of standard deviations, from the mean).
The vectorization rules can also comprise dimensionality reduction techniques. As is known in the art, machine learning and statistical estimation often suffers from the “curse of dimensionality,” having too many dimensions or degrees of freedom to feasibly calculate. Therefore, vectorization can, in some embodiments, comprise dimensionality reduction techniques, such as principal component analysis (PCA), non-negative matrix factorization (NNMF), and latent Dirichlet allocation (LDA), and other techniques known in the art.
The vectorization rules can also comprise feature selection criteria. In some embodiments, not all behavioral attribute values are used in model training 405. In general, it is desirable for the selected behavioral attribute values to be highly correlated with the target unknown, but not highly correlated with each other. From the original behavioral attribute values, certain attributes can be included or excluded based on industry expertise or hypotheses about correlations in the data. For example, if fantasy participation is uncorrelated with TV viewership (the target unknown behavioral attribute), then the fantasy participation attribute values can be dropped. In some embodiments, after behavioral attribute values are encoded, the values can be analyzed to determine whether input behavioral attribute values are highly correlated. Using multiple highly correlated input values can increase the computational complexity of training the model without meaningfully improving performance. One method of eliminating correlated data is to calculate a correlation coefficient between each pairwise set of behavioral attribute values, aggregating behavioral attributes into groups or categories of correlated behavioral attribute values, and then dropping all but one or a few of the correlated behavioral attribute values in each of the groups of correlated behavioral attribute values. This correlation coefficient-based technique can be applied at any stage of the vectorization process, either to the raw behavioral attribute values, encoded behavioral attribute values, or standardized behavioral attribute values.
Once the data is vectorized, it can be used to train an estimator model. The estimator model can comprise a trainable model, such as a machine learning, logistic, probabilistic, or statistical model. Examples of trainable models include artificial neural networks, automatically produced decision trees, support vector models, and other similar models. Trainable models can comprise regression models that estimate a continuous (or semi-continuous) variable as a function of inputs. In some embodiments, the estimator model can comprise a decision tree. In some embodiments, the trained model could also include one or more of a support-vector model (SVM), an artificial neural network (ANN), a least absolute shrinkage and selection operator (LASSO), and an elastic net regression, logistic regression, probabilistic regression (e.g. Bayesian regression), or other trainable regression model. The decision tree produces numerous decision points, and ultimately provides estimates of the unknown behavioral attribute value for each consumer in the first training data set 500. The decision tree can be trained using a training method such as an Iterative Dichotomizer (ID), a Classification and Regression Tree (CART), an Automatic Interaction Detector (AID), a Chi-Squared Automatic Interaction Detector (CHAID), a Conditional Inference Tree, Multivariate Adaptive Regression Splines technique (MARS), or other decision tree training methodology, as would be known to a person of ordinary skill in the art.
In order to preserve consumer privacy, the model training 405 step uses a model definition 410 provided from an external computing system 409. In this way, the first organization can control the external computing system 409, and use it to provide a model definition 410. At the same time, the first organization cannot directly access the joined data set 700, which contains sensitive CII data from the second training data set 600. In some embodiments, the model definition 410 can comprise information sufficient to produce a trainable estimator model (e.g. executable code to produce it in memory, an identification of the model), and any relevant parameters for the model. In some embodiments, the model definition 410 can further comprise vectorization rules, such as where no vectorization is performed prior to joining the first training data set 500 and second training data set 600. In some embodiments, the first training data set 500 and second training data set 600 can be vectorized before they are joined to produce the joined data set 700, and then the model definition 410 need only comprise information sufficient to create a trainable estimator model. The model definition 410 can take a variety of forms, including, for example, computer source code or object code, configuration files, or other build artifacts. In some embodiments, for example, the model definition 410 could be a SAS proc file, a python script, or other computer program. In some embodiments, the computer system that executes the model training 405 can provide a user interface, such as a REST API, website, or graphical user interface to allow a model to be defined.
The model training 405 then outputs a trained model 900 by training the defined model 410 on at least a portion of the joined data set 700. For example, to facilitate training and validation, the joined data set 700 can be split into a training and testing subset, and the model trained only on the training set. Subsets of the joined data set 700 can be provided in other ways, such as k-folds splits, shuffles, mini-batches, etc. The model is then trained in accordance with the training method appropriate for the defined model. For example, if the model is a decision tree, the model training 405 can implement a Classification and Regression Tree algorithm (CART) algorithm, Automatic Interaction Detector (AID), or a Chi-Squared Automatic Interaction Detector (CHAD) algorithm, as is known in the art. As another example, if the model is an artificial neural network, the training can comprise a form of Stochastic Gradient Descent with Backpropagation, or other training method as is known in the art.
As depicted in
To produce an estimated TV viewership using this tree, a computer system begins at the root node 910 for each consumer, and evaluates the rule for the consumer. In this example, the first rule 901 is whether the total fantasy page views is <=0.651 (e.g. the number of fantasy page views is 0.651 standard deviations above the mean). If the rule is met by the data associated with a consumer, the computer moves down the tree along the “true” branch for that rule 920. The process repeats until the computer arrives at a terminal node such as the bottom-right terminal node 941. For example, if a consumer has the following attributes:
The computer would evaluate the first node or the root node 910, and determine that the consumer's fantasy page views are not less than 0.651 standard deviations above the mean, and advance to the right, down the false branch to false node 921. The computer would then evaluate whether the value for season tickets is less than 0.5 (alternatively, whether the value is a 0 or a 1), and move down the false branch to parent node 931. The computer would then evaluate whether the fantasy page views were less than or equal to 1.424, which is false, moving to the bottom-right terminal node 941. As shown, in the training examples, there were 40 examples that would match this same route through the decision tree, and the average viewership numbers for those 40 samples would be 7414 minutes/year. The computer would therefore assign a TV viewership metric of 7414 for the consumer with the attributes above.
In some embodiments, a decision tree can be an effective choice for the trained model 900. Decision trees frequently work well for training on few training examples, produce explainable rule sets that can be understood by humans and evaluated for reasonableness (as compared to, for example, a very large matrix of weights for a neural network). Decision trees can also be adjusted prior to application against unseen examples. For example, in the decision tree depicted in 900, the bottom-right terminal node 941 consists of only 40 of the 1000 training examples used to train this decision tree. Such a small number of examples may indicate that these are unrepresentative outliers. Thus, the bottom-right terminal node 941 can be “pruned” and removed from the tree altogether, also eliminating the other bottom-right terminal node 940 sharing the common parent node 931. Thus, the pruned terminal nodes' parent node 931 becomes a new terminal node. As depicted, the parent node 931 consists of 118 training examples, or a little over 10% of the training examples. With the parent node 931 as a terminal node, consumers whose behavioral attributes arrive at that node are assigned a viewership value of 5360.2. This pruning process can be performed manually, or the model definition may specify a minimum terminal node size, either in absolute terms or as a percentage of the size of the training set.
To maintain privacy, and because model evaluation 407 requires access to the complete training examples, including sensitive data, the model evaluation 407 can be performed by the modeling computing system 420. In some embodiments, model evaluation 407 can comprise calculating performance metrics that compare the actual unknown behavioral attribute values to the estimated behavioral attribute values produced by the trained model 900. In some embodiments, model evaluation 407 can comprise comparing the total of the estimated behavioral attribute for a plurality of consumers to an estimate of the total of that behavioral attribute for the plurality of consumers. For example, if the total viewership of a television program is known, and the percentage of that viewership within the internal data set is known, then the total viewership for the internal data set can be estimated by multiplying the total viewership by the percentage of the viewership in the data set. That estimate of the total can be compared to the total of the estimated viewership for each individual consumer in the internal data set. In some embodiments, the performance metric can be a mean absolute error (MAE), or mean absolute percentage error (MAPE), as calculated between the estimated output value produced by the model and the actual target output value.
Other performance criteria can include accuracy measures, such as calculations of average error, variance, or other measurements related to the distribution of estimated output values. Performance criteria can also include compute performance criteria, such as the number of training epochs necessary to produce the model, time to evaluate the model on new input data, etc. Performance criteria can also relate to explainability or generalization criteria, such as, for example, with decision tree models, terminal node size. The evaluation criteria may be determined by the organization controlling the modeling computing system 420, or the first organization. In some embodiments, in the same way that model definitions 410 are provided from the external computing system 409, evaluation criteria can also be provided by the external computing system 409. Once the evaluation criteria is calculated by the modeling computing system 420, the modeling computing system 420 can deliver the evaluation data to the second computing system. The evaluation data provided to the modeling computing system 420 can comprise both the estimates and the per-consumer error. For example, the validation data 1000 represent the estimates as estimated minutes viewed 1041 and the per-consumer error as APE 1042. In some embodiments, to further preserve privacy and prevent reverse-engineering of the actual behavioral attributes 1030, the evaluation data 408 provided to the modeling computing system 420 can comprise only aggregate evaluation data 408, such as mean APE 1042 across all validation examples.
If the first organization is not satisfied with the evaluation data 408, the first organization can use the external computing system 409 to redefine the model 410 to attempt to improve performance, such as by adjusting parameters of the model, using a different model, or changing the vectorization rules. If the first organization is satisfied with the evaluation data 408, then the first organization can apply the trained model 900 to the first training data set 500 using its own computing system to estimate the unknown behavioral attribute value. During the blind data modeling process 400, the first organization cannot access the second training data set 600 or the joined data set 700. Similarly, the second organization cannot access the first training data set 500 or the joined data set 700. The “blind” creation of the trained model 900 preserves each consumer's CII, but allows the first organization to estimate an unknown behavioral attribute value for each of its consumers included in the first training data set 500. Alternatively, the first organization can provide the third-party modeling computing system 420 with a plurality of different model definitions 410 and receive multiple evaluation data 408 and performance statistics back at the same time. Then the first organization can pick the best model.
Because training examples can be reverse-engineered from trainable computer models in certain circumstances, some embodiments include countermeasures to prevent reverse-engineering training examples from the model. For example, where a decision tree is used, if the tree is fully grown, such that each terminal node comprises a single training example, it may be possible to reverse-engineer from the model the actual values of the target value for each consumer in the training set. To prevent this reverse-engineering, the modeling computing system 420 can enforce requirements on the models. For example, if the model is a decision tree, the first computer can enforce a minimum terminal node size.
With these estimated behavioral attribute values, the first organization can enhance each consumer's experience on the first organization's website, tailor advertisements and marketing campaigns that can reduce the time a consumer spends searching for a product or service, and improve loyalty programs for its consumers. For example, if the NFL can estimate whether a consumer participates in fantasy, then it could provide the consumer with advertisements tailored to fantasy participation. As an additional example, the NFL can use a trained model to estimate the number of minutes each consumer watched NFL games on TV or the Internet, or the number of minutes a consumer watched a specific NFL team during the NFL season on TV or the Internet. The NFL can use the number of minutes watched by each consumer to categorize the consumers into groups that can help the NFL provide the consumer with relevant products. Alternatively, if the NFL had a loyalty program that provided its consumers with perks for purchasing from products, tickets, subscriptions, or passes, then the NFL could provide the consumer with perks tailored to each consumer's unknown behavioral attribute. Prior to the conception of the present disclosed technology, sports leagues, like the NFL, could not acquire this information. Overall, the addition of this unknown behavioral attribute can increase a sports league's understanding of its fans, can improve a fan's experience by offering the fan products, programs, subscriptions, tickets, and ads that are tailored to that fan's behavioral attributes, and reduce the time it takes a fan to search for products, programs, subscriptions, and tickets that relate to the fan's team or a specific sports league.
Once the first organization receives a satisfactory trained model 900, it can apply the model to consumer data using its own computing system within its own first-party data 1101 to estimate the unknown behavioral attribute for each consumer in its first-party data 1101. Because the input behavioral attributes used to train the model 900 are also present in the first-party data 1101, the trained model 900 is compatible with the first-party data 1101.
The first-party data 1101 belongs to a first-party, and can comprise CII and behavioral attributes that are known and available to the first-party. For example, the first-party is the first organization that requested the blind data modeling 400 and received the trained model 900. Additionally, in order for the first-party to make use of the trained model 900, the first-party data 1101 should be the same data set as the first training data set 500. That is, the first-party data 1101 and the first training data set 500 include the same CII and behavioral attributes that were used during the blind data modeling 400 to train the trained model 900. At a minimum, the first-party data 1101 should include the CII and behavioral attributes used to train the trained model 900. Applying the trained model 900 to an alternative data set that was not used to train the trained model 900 could produce inaccurate and misleading estimations of the unknown behavioral attribute values of each consumer included in the first-party data 1101. For the foregoing reasons, an example of the first-party data 1101 is illustrated in the data set excerpt 500 in
After the first-party receives the trained model 900, it can apply the model 1102 to the first-party data 1101 to output annotated data 1103. The first-party receives the model via an input from a graphic user interface or on a source code or object code file and uploads the input to a computer that also stores the first-party data 1101. The model estimates an unknown behavioral attribute value for each consumer based on each consumer's behavioral attribute values contained in the first-party data 1101 and outputs annotated data 1103.
As a practical example, the NFL can apply a trained model developed during the blind data modeling 400 to estimate how many minutes each consumer watches NFL games each year. The model may estimate that a consumer that purchased an NFL Sunday Ticket subscription and Los Angeles Rams season tickets watches more minutes of the NFL season than a consumer that did not purchase a Sunday Ticket subscription or season tickets. As yet another example, a model may estimate that a consumer that plays a substantial number of minutes of the Madden NFL video game watches more minutes of the NFL season than a consumer that does not own or play the Madden NFL video game. Alternatively, the estimated unknown behavioral attribute can comprise minutes viewed of an NBA, MLB, NHL, or MLS game. Applying the model to the first-party data 1101 results in the output of the annotated data 1103.
The annotated data 1103 can comprise the first-party data 1101 (i.e. CII and behavioral attributes for each consumer) and an estimated unknown behavioral attribute value for each consumer. The annotated consumer data 1103 is similar to the joined data set 700 as illustrated in
The first party can further expand the annotated data 1103 by joining or merging 1105 the annotated data 1103 with partner data 1104. The partner data 1104 belongs to third parties and can comprise, for one or more consumers, CII sufficient to identify the consumer, demographic information, and one or more partner behavioral attributes values for one or more partner behavioral attributes. These third parties can include partners, subsidiaries, affiliates, franchisees, or any other party that can legally share its partner data 1104 without violating a consumer's privacy rights.
Unlike the second training data set 600 included in the blind data modeling 400, the first-party can legally and contractually access the partner data without violating a consumer's CII. Typically, the first-party will enter into a contract with a partner that permits the sharing of the partner data 1104, including the CII, the demographic information, and partner behavioral attribute values of each consumer. Additionally, the partner has likely made an agreement with its consumers that allows the partner to share the CII, demographic information, and behavioral attribute values for each consumer. Preferably, some of the consumers included in the partner data 1104 are also included in the annotated data 1103.
Additionally, the partner data 1104 or first-party data 1101 can comprise demographic information about each consumer. An example of demographic information from partner data 1104 is illustrated in the 360-View data set excerpt 1200. As illustrated in
The partner data 1104 can also comprise one or more partner behavioral attribute values for one or more partner behavioral attributes. In some embodiments, the partner behavioral attribute values can comprise dollar amounts, such as shop purchase 1231. Shop purchase 1231 can comprise the amount of money a consumer spent on one or more categories of products (e.g., pizza, clothing, or shoes). In some embodiments, the partner behavioral attribute values can comprise simple numerical values, such as video game minutes played 1234, or fantasy page views 1235. In some embodiments, partner behavioral attribute values can comprise Boolean data, such as whether the user has a digital subscription 1233, a music subscription, or accounts on specific websites like ESPN or the NFL's website. In some embodiments, the partner behavioral attribute values can comprise categorical data, such as digital or music program subscription level, which can take multiple values.
In some embodiments, the first-party data 1101 can include the partner data 1104. In some embodiments, including the partner data 1104 in the first-party data 1101 may increase the performance of a trained model created during the blind data modeling process 400. For example, increasing the number of consumers considered during the training of a trained model may improve the reliability and accuracy of the trained model because it has more information to consider about each consumer. Alternatively, the partner data 1104 may comprise, for one or more consumers, one or more of the same behavioral attribute values for one or more behavioral attributes included in the first-party data 1101. In this embodiment, the partner data 1104 can increase the accuracy of the first-party data 1101. Accordingly, the blind data modeling process 400 would likely generate a more accurate and reliable trained model 900. If the partner data 1104 is included in the first-party data 1101, then the first-party does not have to perform the join or merge 1105 step during the construction of the 360-View because the partner data 1104 is already included. In a further embodiment, the first-party could decide not to join or merge 1105 the partner data 1104 with the first-party data 1101 or the first-party may not have access to partner data 1104 to join or merge 1105 with the first-party data. However, as previously mentioned, the inclusion of the partner data 1104 may allow the first-party to create more definitive clusters during segmentation 330.
Joining or merging the first-party data 1101 and the partner data 1104 outputs first-party-partner data 1106. At this point, the first-party could perform segmentation 330 to cluster its consumers based on their behavioral attribute values, unknown behavioral attribute values, demographic information, and partner behavioral attribute values.
In some embodiments, a first-party can use its financial data 1107 to calculate financial attribution 1108 values for each consumer included in the first-party-partner data 1106. This financial attribution 1108 is a mechanism for attributing a particular dollar amount of revenue or profit to each consumer and aggregating all sources of revenue or profit for that consumer. Such sources of revenue can come in a variety of forms. As depicted in
Direct revenue 1251 is revenue collected directly by the first-party from a consumer. Direct revenue 1251 can be, for example, membership fees in fan clubs, digital access subscriptions, etc. Calculating direct revenue 1251 can be as simple as summing the value of all direct revenue transactions. For example, in
Direct licensed revenue 1252 is revenue that results from a license with a third-party or affiliate. An example of direct licensed revenue 1252 would be the sale of licensed merchandise, from which the first-party receives a royalty. Direct licensed revenue can be calculated by multiplying each licensed transaction or event by the relevant royalty rate. For example, in
Media revenue 1254 is a share of a media contract revenue attributed to each consumer. In general, organizations like sports leagues or other media companies enter into agreements to sell licensed content for a lump sum to distributors, such as TV networks, movie studios, or digital streaming services. In order to get a 360-View that includes media revenue 1254 attributable to each consumer, there must be an estimate of (1) total consumption of the media, and (2) consumption of that media by each consumer. As depicted in
Sponsorship revenue, or indirect licensed revenue 1253, is a share of sponsorship or advertising revenue attributed to each consumer. Organizations often receive revenue from third parties for advertising or sponsorship. However, organizations would like to be able to attribute the value of sponsorship deals to each individual consumer. There are a variety of ways of performing such attribution. For example, the simplest attribution would be to take the sponsorship value and divide it by the total number of consumers in the organization's consumer database. In some embodiments, sponsorship revenue attribution can be calculated by estimating the total number of consumers (including those outside the consumer database) using survey data, dividing the sponsorship agreement revenue by the total number of consumers, and attributing an equal share to all consumers in the database. In some embodiments, sponsorship revenue attribution can be calculated by dividing consumers into a “target” group and a “non-target” group, based on behavioral or demographic data. Many sponsors have target demographics for their advertising and sponsorship, and are primarily interested in advertising to certain segments of the organization's consumers. Thus, the revenue can be first distributed between “target” and “non-target” groups (e.g. a 70%/30% split), subdivided into the total number of consumers in each group (in-database, or total), and then attributed to each consumer. In some embodiments, the total sponsorship revenue can be allocated only to target consumers (e.g. a 100%/0% split).
Techniques for attributing sponsorship revenue can also be combined with techniques for estimating media viewership. For example, if the sponsorship agreement is tied to a specific distribution channel, then the sponsorship revenue can be attributed only to those consumers who participate in the distribution channel, either equally or proportional to their consumption of the media.
While the above calculations discuss attributing revenue from various channels to specific consumers, a similar process can be performed for attributing margin, or profit, to individual consumers. In embodiments that attribute margin to each consumer, each revenue number above is adjusted downward in proportion to the overhead associated with each revenue source (e.g. cost of goods, administrative costs, third-party contractor expenditures, etc.).
Once a 360-View is produced, the organization can use the 360-View in various business analytics such as clustering the consumers in the 360-View into groups, calculating and evaluating return on investment, developing targeted acquisition or retention campaigns, or determining the lifecycle and life-stage of behavioral attribute values for each consumer.
One application of the 360-View is to automatically group consumers into relevant market segments. Therefore, some embodiments of the disclosed technology further comprise applying a clustering algorithm to consumer data, such as the 360-View, to group consumers according to behavior. By grouping consumers this way, organizations can focus outreach and marketing efforts on groups of similar users that may respond in similar ways. Within groups, the grouping can be used for targeted advertising to consumers whose interactions with the organization's products and services vary from their group norms, with the assumption that the consumer is likely to be interested in that activity. Segmentation 330 allows an organization to create strategic marketing campaigns, target retention campaigns, enhanced experiences for each consumer, tailored advertising for each consumer, and financial planning based on the clustering results.
In order for such an automatic segmentation mechanism to function, consumers must be divided into groups, where each member of the group is relatively similar to others in the group, and is relatively dissimilar from members of other groups. A variety of techniques exist for clustering data in this manner. Automatic clustering algorithms that are known in the art include k-Means, Affinity Propagation, Mean Shift, Spectral Clustering, Ward Clustering, Agglomerative Clustering, DB SCAN, Birch, and Gaussian Mixture. Each of these techniques first require as input a plurality of points in n-dimensional space, similar to the trained models explained above. Two candidate algorithms that tend to be particularly well suited for consumer segmentation include mini batch K-means and Agglomerative Clustering, both of which have complexities between O(n log n) and O(n2).
Next, clustering 1304 is performed. This comprises applying a clustering algorithm (as explained above) to the vectorized data 1303. In some embodiments, the clustering algorithm can be provided a predetermined number of groups into which to subdivide the consumers. In some embodiments, the number of groups can be programmatically selected by repeating the clustering algorithm with various numbers of groups, and evaluating the quality of the clusters (as described herein), and selecting a number of groups that provides the best performance. Many clustering algorithms, likewise, can use various distance measures to evaluate the quality of individual clusters. While the default in most circumstances is Euclidean distance (e.g. straight line distance), many other distance metrics can be used, such as squared Euclidean distance, standardized Euclidean distance, cosine distance, Manhattan distance, Bray-Curtis distance, Canberra distance, Chebyshev distance, Jensen-Shannon distance, Mahalanobis distance, Minkowski distance, and other distance metrics as are known in the art.
Once a set of vectorization rules, algorithm, and distance metric the algorithm can be executed to cluster consumers 1305 into the desired number of groups. The result is a set of cluster IDs 1305 for consumers in the 360-View 1301. The clusters can then be evaluated 1306 to ensure adequate performance. Evaluation criteria can include, for example, determining the average density of each cluster, distortion (mean sum of squared distances to centers), intercluster distances, calculating the Variance Ratio Criterion (Calinski Harabaz score), calculation of a silhouette score (mean ratio of intra-cluster and nearest-cluster distance), or other similar metrics. These performance criteria can be compared to a minimum or maximum acceptable value. Alternatively, the clustering can be repeated using different vectorization rules, algorithms, desired number of groups, and/or distance measures, and the scores recalculated to see if they improve or worsen. Techniques for such evaluation can also comprise generating an elbow chart (e.g. mapping one or more fit scores/computational performance metrics against choices for number of groups), silhouette visualizations, or intercluster distance maps.
If the performance is acceptable 1307, the clustered data can be merged/joined 1309 with the 360-View data 1301 to produce a clustered 360-View 1310. Alternatively, if the clustering method did not produce adequate clusters, the clustering process can be revised 1308 with different sets of parameters (e.g. different number of desired groups, different vectorization rules, inclusion/exclusion of behavioral attributes or derived/calculated behavioral attributes, different clustering algorithm, etc.).
The result of executing the algorithm is a group ID associated with each consumer in the consumer database, which can be merged or joined 1309 to the 360-View 1301 to produce a clustered 360-View 1310.
An organization can use the clustered 360-View 1310 to improve marketing strategies, enhance consumer experiences, provide consumers with tailored advertisements to reduce consumer search time for products or services, and decrease advertising or marketing costs. Segmentation 1560 can also help organizations develop a long-term plan regarding its consumers and how best to serve them over a period of years. Finally, segmentation 1560 can help an organization focus its business development, its product development, the services it provides, and the relationships it has with its consumers.
In some embodiments, the 360-View 1310 can be used to predict the likelihood that a consumer will migrate from one cluster to another. This information can be useful for many applications, such as, for example, identifying specific consumers that have a high likelihood of moving to a cluster with a higher expected revenue, profit, or margin for additional ad spend, promotional giveaways, or other targeted marketing. In some embodiments, a migration can be predicted based on cluster affinity. In other words, a likely migration can be predicted based on the proximity of the vector corresponding to a specific consumer to a representative point of each of the clusters of consumers. In some embodiments, a migration can be predicted based on historical data. In some embodiments, the likelihood that a consumer will move to another cluster can be calculated, at least in part, based on a “stickiness” of the consumer to their assigned cluster. In some embodiments, a move can be predicted based on “look-alike” consumers that differ in one or more an age related attribute. In some embodiments, the likelihood that a consumer will move between clusters can be used to provide additional insights into both individual consumers and populations of consumers. In some embodiments, a group of consumers can be analyzed to determine average migration statistics over time. Aspects of the present invention for predicting the likelihood that a consumer will migrate to a different cluster can be combined in various ways to provide additional estimates.
The space of visualization 1800 is divided into cells indicating the boundaries between clusters (e.g. Voroni Cells, depicted by the dashed lines), indicating the boundaries between clusters 1820, 1840, and 1860. As shown by the cluster boundaries, each of consumers 1810-1815 are assigned to cluster 1820, each of consumers 1830-1832 are assigned to cluster 1840, and each of consumers 1850-1854 are assigned to cluster 1860. Each of clusters 1820, 1840, and 1860 have a point indicating an average or typical point within the cluster, such as a centroid, shown at points 1821, 1841, and 1861, respectively. From those distances, a standard deviation can be calculated, indicating a radius around the representative point within which a majority of the other consumers belonging to that cluster will be placed (˜68% if a normal distribution is assumed). The standard deviations of each of clusters 1820, 1840, and 1860 are depicted graphically in visualization 1800, with rings 1822, 1842, and 1862 depicting one standard deviation, and rings 1823, 1843, and 1863 depicting two standard deviations from the average points 1821, 1841, and 1861 of clusters 1820, 1842, and 1844, respectively.
Returning to
In some embodiments, the likelihood that a consumer will move between future clusters 2200, as explained above, can be used to provide additional insights into both individual consumers 2210 and populations of consumers in future clusters 2200. For example, estimates of likely cluster migration can be used to calculate an expected value 2213 of an individual consumer 2210 over a predetermined time interval. In some embodiments, an expected value for each cluster can be calculated by, for example, calculating an average of the attributed revenue, margin, or profit of each consumer belonging to the cluster. Other statistics, such as the min/max, percentiles, standard deviation, fitting a normal, gaussian, or other probabilistic distribution to the actual distribution of attributed revenue, margin, or profit of each consumer can likewise be calculated to characterize the distribution for each cluster. These metrics can then be combined with the likelihood that a consumer will migrate into each future cluster 2220 to produce an expected value 2230 of the individual consumer 2210 at a predetermined point in time. The expected value 2230 can then be calculated by, for example, multiplying the expected value 2230 of each future cluster 2220 by the likelihood that the individual consumer 2210 will migrate to that future cluster 2220. Other calculations are possible to predict ranges of expected values 2230, such as performing the same calculation on the minimums and maximums, percentiles, or a predetermined number of standard deviations above and below the expected value of each cluster. This calculation can produce a metric by which an individual consumer 2210 can be analyzed to predict whether the individual consumer's 2210 attributable revenue is likely to increase or decrease, and to estimate the magnitude of such a change, shown as the percentage increase 2214 in example table 2200. This information can be used to inform marketing strategies to identify either individual consumers 2210 that are likely to increase in value for specific advertising or promotional targeting, or identify individual consumers 2210 at risk of decreasing expected value to minimize those possible losses of revenue.
In some embodiments the method can comprise determining a probability 2505 that the consumer will be assigned to each subset in the plurality of subsets at the predetermined time in the future. As would be understood by persons of ordinary skill in the art, embodiments also include determining a probability that the consumer will be assigned to a single, or a subset of the subsets in the plurality of subsets at a predetermined time in the future. In some embodiments, the step of determining a probability 2505 can comprise determining a probability as a function of the proportion of consumers in the look-alike group in each subset relative to all consumers in the look-alike group. In some embodiments, the step of determining a probability 2505 can comprise determining the probability as a function of the distance between the point representing the consumer and a centroid of the second subset relative to the distance between the point and the other subsets in the plurality of subsets. In some embodiments, the step of determining a probability 2504 can comprise determining a standard deviation for each subset in the plurality of subsets for the distances between the point representing each consumer in the subset of consumers and a centroid of the subset, and determining that the consumer will be assigned to the subset it is assigned to at the present time at the predetermined time in the future if the distance between the point representing the consumer and the centroid of the subset to which it is assigned at the current time is below a predetermined multiple of the standard deviation for that subset.
The flowchart and/or block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosed technology. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims below are intended to include any structure, material, or act for performing the function in combination with other claimed elements as specifically claimed. The description of the present disclosed technology have been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the invention. The embodiment was chosen and described in order to best explain the principles of the disclosed technology and the practical application, and to enable others of ordinary skill in the art to understand the disclosed technology for various embodiments with various modifications as are suited to the particular use contemplated.
Any flow diagrams depicted herein show just one example. There may be many variations to this diagram or the steps (or operations) described therein without departing from the spirit of the disclosed technology. For instance, the steps may be performed in a differing order or steps may be added, deleted or modified. All of these variations are considered a part of the claimed disclosed technology.
While the preferred embodiment to the disclosed technology had been described, it will be understood that those skilled in the art, both now and in the future, may make various improvements and enhancements which fall within the scope of the claims which follow. These claims should be construed to maintain the proper protection for the invention first described.
This application is a continuation-in-part of co-pending U.S. application Ser. No. 16/374,594, entitled “Systems and Methods for Privacy-Preserving Generation of Models for Estimating Consumer Behavior”, filed Apr. 3, 2019, issued as U.S. Pat. No. 10,430,727 on Oct. 1, 2019, which is hereby incorporated by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16374594 | Apr 2019 | US |
| Child | 16583975 | US |
