Information processing system, computer program, and information processing method
The information processing system addresses biases in human behavior data by using propensity scores and environmental adjustments to correct and supplement data sets, improving the accuracy and completeness of purchasing behavior analysis.
Patent Information
- Application Number
- JP2023011858
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2023-01-30
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2043-01-30
AI Technical Summary
Existing data analysis of human behavior is prone to information bias due to limited sampling and environmental factors, leading to inaccurate representation of population behavior.
An information processing system that includes a first acquisition unit, a second acquisition unit, and a correction unit to acquire and correct data sets using propensity scores and environmental adjustments to suppress biases, supplementing missing data based on estimation models.
The system effectively corrects and supplements data sets to reduce biases, enhancing the accuracy and completeness of human behavior analysis, particularly in purchasing behavior data.
Smart Images

Figure 0007590472000001 
Figure 0007590472000002 
Figure 0007590472000003
Abstract
Description
[Technical field]
[0001] The present disclosure relates to an information processing system and an information processing method. [Background technology]
[0002] It has been customary to analyze human behavior. For example, consumer purchasing behavior is analyzed to estimate potential buyers of a product, and advertisements for the product are then delivered to the estimated buyers.
[0003] A technology is also already known for estimating a first behavior of a second consumer group based on first data regarding a first behavior of a first consumer group and second data regarding second behaviors of the first consumer group and the second consumer group (see Patent Document 1). [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2020-161038 A Summary of the Invention [Problem to be solved by the invention]
[0005] However, data on human behavior collected for analysis may contain undesired information bias. For example, the data collected for analysis is data on the behavior of a limited group of people as a sample.
[0006] The behavior of a limited group of people may be biased with respect to the behavior of the population. In addition, there may be information bias due to the behavioral environment, including the behavior observation environment. Therefore, in order to grasp the characteristics of human behavior with high accuracy through sample-based analysis, it is desirable to be able to correct the information bias contained in the collected data.
[0007] Therefore, according to one aspect of the present disclosure, it is desirable to provide an information processing technique suitable for correcting bias in information regarding human behavior contained in data. [Means for solving the problem]
[0008] According to one aspect of the present disclosure, there is provided an information processing system including a first acquisition unit, a second acquisition unit, and a correction unit. The first acquisition unit acquires a first dataset related to a plurality of first subjects as a first sample. The second acquisition unit acquires a second dataset related to a plurality of second subjects as a second sample. The correction unit corrects the first dataset based on the second dataset.
[0009] The first data set comprises, for each of a plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as the first behavioral data, and the second data set comprises, for each of a plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as the second behavioral data.
[0010] The first behavior data, for a plurality of first behaviors, numerically describes, for each first behavior, a corresponding behavior of a corresponding first subject. The second behavior data, for a plurality of second behaviors, numerically describes, for each second behavior, a corresponding behavior of a corresponding second subject. The plurality of second behaviors includes a plurality of first behaviors.
[0011] The correction unit corrects the first dataset based on the second dataset by correcting the numerical values for each first behavior for multiple first subjects in the first dataset in a direction to suppress a first bias, which is a numerical bias due to bias in the multiple first subjects in the first dataset, and a second bias, which is a numerical bias due to the behavioral environment of each of the multiple first subjects in the first dataset.
[0012] Data sets on human behavior can contain information bias due to multiple factors. For example, human behavior is thought to depend on the environment in which a person is placed, in addition to the person's attributes.
[0013] Therefore, by correcting the first dataset in a direction to suppress the first bias and the second bias, it is possible to appropriately suppress multiple biases contained in the first dataset related to the first sample. Therefore, according to one aspect of the present disclosure, it is possible to appropriately correct bias in information related to human behavior included in collected data due to multiple factors.
[0014] According to one aspect of the present disclosure, the first sample may be a set of people limited to people who belong to a particular group. The second sample may be a set of people who do not belong to a particular group or a set of people who are not limited to people who belong to a particular group. Common variables may be used between the first behavioral data and the second behavioral data to explain the behavior.
[0015] The correction unit may correct the numerical values for each first behavior for the multiple first subjects in the first dataset in a direction that suppresses a first bias resulting from the first sample being limited to a set of people belonging to a specific group, based on a propensity score corresponding to the probability that a person conditioned on a common variable belongs to a specific group.
[0016] According to an aspect of the present disclosure, the first behavioral data and the second behavioral data may have attribute information that represents an attribute of a subject who is the subject of the behavior by at least one variable. A common variable that is a common variable for explaining the attribute may be used between the first behavioral data and the second behavioral data.
[0017] The correction unit may use the common variable related to this attribute to correct the numerical values for each first behavior for the multiple first subjects in the first dataset in a direction that suppresses a first bias resulting from the first sample being limited to a set of people belonging to a specific group, based on a propensity score corresponding to the probability that a person conditioned on the common variable belongs to a specific group.
[0018] According to one aspect of the present disclosure, each of the plurality of first behaviors and the plurality of second behaviors may be a purchasing behavior. The first sample may be a set of consumers who participate in a particular reward program for purchases or consumers who use a particular electronic payment service. The first behavior data may be data describing the purchasing behavior of the corresponding first subject observed through the provision of identification information from the corresponding first subject.
[0019] When collecting data on behavior, a specific reward program or electronic payment service may be used, and in this case, the collected data may include bias due to the limited consumer set and the limited observation method. According to one aspect of the present disclosure, it is possible to appropriately correct collected data including such bias.
[0020] According to an aspect of the present disclosure, each of the plurality of first actions and the plurality of second actions may be a purchasing action distinguished by at least one of a purchase object and a purchase store. According to an aspect of the present disclosure, each of the plurality of first actions and the plurality of second actions may be an action through an information terminal.
[0021] According to one aspect of the present disclosure, the correction unit may determine, for each of the multiple regions, the magnitude of the opportunity to perform each of the multiple first behaviors, and correct the numerical values for each of the first behaviors for the multiple first subjects in the first dataset in a direction to suppress a second bias resulting from the magnitude of the opportunity to perform each of the multiple first subjects in their respective active regions.
[0022] There may be regional bias in the behavioral environment. There may also be regional bias in the magnitude of the opportunity to perform a behavior. Therefore, by correcting the behavioral dataset in a direction that suppresses the bias caused by the magnitude of the opportunity to perform, it is possible to appropriately correct the bias in the information on human behavior contained in the dataset.
[0023] According to one aspect of the present disclosure, each of the plurality of first actions and the plurality of second actions is a purchasing action that is differentiated at least by a purchasing store, and may be a purchasing action in one of the plurality of stores. The one store may include a group of stores of the same type.
[0024] The correction unit may correct the numerical values for each first behavior for a plurality of first subjects in the first dataset based on the geographic distribution of a group of similar stores in a direction that suppresses a second bias caused by the geographic distribution.
[0025] The geographic distribution of similar stores may affect the likelihood of purchases at the corresponding stores, so it is useful to correct the first data set to suppress bias due to the geographic distribution.
[0026] According to one aspect of the present disclosure, the correction unit may correct the numerical values for each first behavior for multiple first subjects in the first dataset based on the characteristics of the payment method used in each of the multiple stores, in a direction to suppress a second bias resulting from differences in the characteristics of the payment method among the multiple stores.
[0027] Depending on the payment method, bias may occur in the observation of purchasing behavior. Therefore, it is meaningful to correct the first data set based on the characteristics of the payment method.
[0028] According to an aspect of the present disclosure, the plurality of second behaviors may include a plurality of third behaviors different from the plurality of first behaviors. The second behavior data may be data that numerically describes, for each of the plurality of third behaviors not explained in the first behavior data, a corresponding behavior of the corresponding second subject.
[0029] In this case, the correction unit may construct an estimation model for calculating estimated values of numerical values related to the plurality of third behaviors from numerical values related to the plurality of first behaviors based on the second dataset. The correction unit may correct the first dataset so as to complement information on the plurality of third behaviors missing in the first dataset by adding, for each of the plurality of first subjects, the estimated values related to the plurality of third behaviors of the corresponding first subject calculated based on the estimation model to the first behavior data.
[0030] According to one aspect of the present disclosure, the calculation of the estimated value based on the estimation model may be performed based on numerical values related to the multiple first behaviors in the first behavior data that have been corrected in a direction to suppress the first bias and the second bias. By calculating the estimated value based on the corrected numerical values, information related to a highly reliable third behavior can be added to the first data set, thereby increasing the information value of the first data set.
[0031] According to one aspect of the present disclosure, a computer program may be provided for causing a computer to realize at least a portion of the functions of the information processing system described above. According to one aspect of the present disclosure, a computer program may be provided for causing a computer to realize the functions of the first acquisition unit, the second acquisition unit, and the correction unit in the information processing system described above.
[0032] According to an aspect of the present disclosure, there may be provided an information processing method. The information processing method may be executed by a computer. According to an aspect of the present disclosure, the information processing method may be a method for executing at least a part of the processing executed by the information processing system described above.
[0033] According to one aspect of the present disclosure, an information processing method may include obtaining a first dataset for a plurality of first subjects as a first sample, obtaining a second dataset for a plurality of second subjects as a second sample, and correcting the first dataset based on the second dataset.
[0034] The first data set comprises, for each of a plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as the first behavioral data, and the second data set comprises, for each of a plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as the second behavioral data.
[0035] The first behavior data, for a plurality of first behaviors, numerically describes, for each first behavior, a corresponding behavior of a corresponding first subject. The second behavior data, for a plurality of second behaviors, numerically describes, for each second behavior, a corresponding behavior of a corresponding second subject. The plurality of second behaviors includes a plurality of first behaviors.
[0036] The correction may include correcting the first dataset by correcting the numerical values for each first behavior of the multiple first subjects in the first dataset based on the second dataset in a direction to suppress a first bias, which is a numerical bias caused by a bias of the multiple first subjects in the first dataset, and a second bias, which is a numerical bias caused by the behavioral environment of each of the multiple first subjects in the first dataset. This information processing method can produce the same effects as the above-mentioned information processing system. [Brief description of the drawings]
[0037] [Figure 1] FIG. 2 is a block diagram illustrating functions realized in the information processing system. [Diagram 2] FIG. 2 is a block diagram illustrating an electrical configuration of the information processing system. [Diagram 3] FIG. 2 is a diagram illustrating the configuration of a first data set. [Figure 4] FIG. 13 is a diagram illustrating the configuration of a second data set. [Diagram 5] FIG. 13 is a diagram illustrating the configuration of a complemented first data set. [Figure 6]1 is a flowchart (part 1) showing a correction complement process executed by a processor of the information processing system. [Figure 7] 13 is a flowchart (part 2) illustrating the correction complement process executed by the processor of the information processing system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0038] Exemplary embodiments of the present disclosure will now be described with reference to the drawings. [First embodiment] 1 is configured to correct a first data set DS1 describing the behavior of a first group of people using a second data set DS2 describing the behavior of a second group of people. The information processing system 1 is further configured to complement missing data in the corrected first data set DS1 and output a complemented data set DC.
[0039] As shown in FIG. 2, the information processing system 1 includes a processor 11, a memory 12, a storage 13, a user interface 15, an input / output interface 17, and a communication device 19.
[0040] For the above correction and complementation, the processor 11 is configured to execute processing in accordance with a computer program stored in the storage 13. The memory 12 includes a RAM, and is used as a working area when the processor 11 executes processing. The memory 12 temporarily stores the computer program and data read from the storage 13.
[0041] The storage 13 is configured to store computer programs and data used in the above processes. Examples of the storage 13 include a hard disk drive and a solid state drive.
[0042] The user interface 15 is configured to display various information to the user. The user interface 15 is further configured to receive an operation from the user and input a corresponding operation signal to the processor 11.
[0043] The user interface 15 may include, for example, a liquid crystal display to display various information to the user, and may include, for example, a keyboard and a pointing device to receive operations from the user.
[0044] The input / output interface 17 is configured to be connectable to an external device. When an external storage device is connected, the input / output interface 17 is configured to be able to read data from the external storage device and to be able to write data to the external storage device based on a command from the processor 11. The input / output interface 17 may be, for example, a universal serial bus (USB) interface.
[0045] The communication device 19 is connected to a local area network and / or a wide area network as a communication network, and is configured to be controlled by the processor 11 to communicate with communication equipment in the communication network.
[0046] The processor 11 of the information processing system 1 executes processing in accordance with the computer program stored in the storage 13, and thereby functions as a data acquisition unit 111, a bias correction unit 113, a model construction unit 115, a loss completion unit 117, and an output unit 119.
[0047] The data acquisition unit 111 acquires a data set to be corrected and complemented as a first data set DS1. The data acquisition unit 111 further acquires a reference data set to be referred to during the correction and complementation as a second data set DS2.
[0048] The first data set DS1 and the second data set DS2 may be acquired from an external storage device through the input / output interface 17. Alternatively, the first data set DS1 and the second data set DS2 may be acquired from an external communication device through the communication device 19.
[0049] According to this embodiment, the first data set DS1 is a data set describing the purchasing behavior of a first group of consumers, and the second data set DS2 is a data set describing the purchasing behavior of a second group of consumers.
[0050] The first group corresponds to a particular consumer group, specifically, the first group is a group of consumers who participate in a rewards program in which rewards are provided to consumers according to their purchasing behavior.
[0051] A reward program is, for example, a program that provides consumers with valuable points as a reward in accordance with purchasing behavior. When a consumer makes a purchase, the consumer can receive valuable points in accordance with the purchase by presenting the member ID of the reward program on a physical card (a so-called point card) or on the screen of an information terminal.
[0052] The second group is a group of consumers who participate in a panel study on purchasing behavior. The second group is a group of consumers that is different from the first group, but may include some overlap with the first group.
[0053] 3 includes a record for each consumer. Each record in the first dataset DS1 includes attribute data of the corresponding consumer and first purchase data of the corresponding consumer.
[0054] Hereinafter, the consumers corresponding to each record in the first data set DS1 are also referred to as “first consumers.” A group of first consumers is a group of consumers who participate in the reward program.
[0055] The attribute data describes attributes of the corresponding first consumer. The attributes include at least one of demographic attributes and psychographic attributes. The attribute data can include, for example, information describing gender, age, and geographic region as demographic attributes of the corresponding first consumer.
[0056] The first purchasing data describes the purchasing behavior of a corresponding first consumer. The first purchasing data describes a feature A for each combination of a store and a product for multiple stores H1, H2, ... and multiple products G1, G2, ... that belong to a specific store group. This feature A is a feature A of purchasing behavior that is distinguished by a combination of the product as the purchase target and the purchasing store. For example, the first purchasing data describes a feature A that indicates whether or not the corresponding first consumer purchased the corresponding product at the corresponding store for each combination of store and product. The feature A is expressed as a binary value, for example, the numerical values 0 and 1.
[0057] A specific store group is a group of affiliated stores of a reward program. Stores H1, H2, ... shown in Fig. 3 correspond to affiliated stores H1, H2, ... of the reward program. In other words, the first purchase data describes feature quantity A for each combination of affiliated store and product for the group of affiliated stores H1, H2, ....
[0058] A first consumer who participates in the reward program can receive rewards according to the purchase by presenting his / her membership ID at an affiliated store. Affiliated stores and non-affiliated stores are classified according to whether or not rewards can be received for purchases at the store.
[0059] The first data set DS1 describes only the consumer purchasing behavior observed with the presentation of the member ID, which is the consumer's identification information, among the consumer purchasing behaviors. The first data set DS1 can be obtained, for example, from the operating company of the reward program.
[0060] 4, the second data set DS2 also includes a record for each consumer. Each record in the second data set DS2 includes attribute data of the corresponding consumer and second purchase data of the corresponding consumer.
[0061] Hereinafter, the consumers corresponding to each record in the second data set DS2 are also referred to as “second consumers.” A group of second consumers is a group of consumers who participate in a panel survey on purchasing behavior.
[0062] The second consumer population is a consumer population that is different from the first consumer population, which may include consumers that belong to the first consumer population, but includes consumers that do not belong to the first consumer population. The second consumer population may also be a consumer population that does not include any first consumers.
[0063] The attribute data of the second dataset DS2, like the first dataset DS1, describes attributes of the corresponding second consumer. The attribute data includes, for example, information describing the gender, age, and residential area as demographic attributes of the corresponding second consumer.
[0064] The variable X shown in Figures 3 and 4 indicates a consumer attribute and is used in common in the first dataset DS1 and the second dataset DS2. That is, the second dataset DS2 explains the attributes of each of the second consumers at least in part by using the variable X common to the first dataset DS1. The common variable X may be the gender, age, and residential area of the consumer. Hereinafter, the consumer attribute indicated by the variable X is also referred to as the consumer attribute X.
[0065] The second purchase data is data explaining the purchase behavior of the corresponding second consumer. The second purchase data describes a feature A related to the purchase behavior for each combination of store and product for multiple stores H1, H2, ..., L1, L2, ... and multiple products G1, G2, ... that are not limited to the specific store group.
[0066] The multiple stores H1, H2, ..., L1, L2, ... that are not limited to a specific store group are multiple stores H1, H2, ..., L1, L2, ... including affiliated stores H1, H2, ... and non-affiliated stores L1, L2, .... In other words, the second purchasing data describes feature A for each combination of store and product for a larger set of combinations of stores and products that includes the set of combinations of stores and products for which the first purchasing data describes feature A.
[0067] The second purchase data, like the first purchase data, can be configured to describe, for each combination of store and product, a feature A that indicates whether or not a corresponding second consumer purchased the corresponding product at the corresponding store. The feature A is expressed as a binary value, for example, 0 or 1.
[0068] The bias correction unit 113 is configured to correct the first data set DS1 and the second data set DS2 acquired by the data acquisition unit 111. The first and second groups of consumers whose purchasing behaviors are explained in the first and second data sets DS1 and DS2 are first and second samples extracted from a set of consumers in the entire market who may purchase a product that is a population.
[0069] In the first data set DS1, a first set of consumers whose purchasing behavior is to be explained is limited to the set of rewards program participants from the universe, while in the second data set DS2, a second set of consumers whose purchasing behavior is to be explained is limited to the set of panel survey participants from the universe.
[0070] In this way, the first dataset DS1 and the second dataset DS2 explain the purchasing behavior of different samples extracted from the population. For this reason, it is not appropriate to simply compare the feature quantity A of the purchasing behavior between the first dataset DS1 and the second dataset DS2.
[0071] That is, the first dataset DS1 and the second dataset DS2 have a bias in the feature A according to the bias between the first consumer set and the second consumer set. Hereinafter, the bias in the feature A according to the bias of the consumer set is referred to as a first bias.
[0072] More specifically, in this embodiment, one feature A collectively represents the purchasing behavior at multiple affiliated stores of the same chain. Convenience stores, drug stores, and the like have multiple affiliated stores that belong to the same chain. One feature A represents the purchasing behavior of consumers at the corresponding store, regarding multiple affiliated stores of the same chain as one store.
[0073] In this case, a single store corresponding to one feature A has a geographical distribution of multiple corresponding affiliated stores, and this geographical distribution differs depending on the region. In other words, the environment regarding the product purchasing behavior of each consumer who uses each store is different. Therefore, the first dataset DS1 and the second dataset DS2 have a second bias of the feature A according to the above-mentioned environment of each consumer.
[0074] For example, while a first region may have 50% first-line drug stores, 30% second-line drug stores, and 20% third-line drug stores, a second region may have 20% first-line drug stores, 30% second-line drug stores, and 50% third-line drug stores. In this case, the purchasing environment and purchasing opportunities at each line of drug stores will differ between the first region and the second region.
[0075] From this, it can be said that the feature A(i,j) for each combination of store and product for the corresponding consumer, which is described in each record in the first dataset DS1 and the second dataset DS2, is the feature A(i,j)=p(i,j) corresponding to the observed value p(i,j)=q(i,j)·p(i|x)·p(i,j|z) obtained by adding the first bias p(i|x) and the second bias p(i,j|z) to the true value q(i,j).
[0076] where i is the index of a consumer and j is the index of a combination of store and product. p(i|x) represents a first bias contained in the observed value p(i,j) of the purchasing behavior of the consumer corresponding to index i, and value p(i,j|z) represents a second bias contained in the observed value p(i,j) of the purchasing behavior of the combination of store and product corresponding to index j of the consumer corresponding to index i.
[0077] Feature A(i,j) represents feature A related to purchasing behavior of a consumer corresponding to index i purchasing a product corresponding to index j at a store corresponding to index j, for example, feature A indicating whether the corresponding product was purchased. Similarly, true value q(i,j) and observed value p(i,j) respectively represent the true value and observed value of feature A related to purchasing behavior of a consumer corresponding to index i purchasing a product corresponding to index j at a store corresponding to index j.
[0078] The bias correction unit 113 estimates the first bias p(i|x) and the second bias p(i,j|z) included in each feature A(i,j) of the first data set DS1 and the second data set DS2 based on the first data set DS1 and the second data set DS2, and corrects the feature A(i,j) to a value C(i,j)=A(i,j) / (p(i|x)·p(i,j|z)) using the reciprocals of the first bias p(i|x) and the second bias p(i,j|z). In this way, the bias correction unit 113 corrects the feature A(i,j)=p(i,j) to a value C(i,j) corresponding to the true value q(i,j).
[0079] The model construction unit 115 constructs an estimation model for estimating feature quantity C(i=i1, j=j2) related to the purchasing behavior of a first consumer at a non-affiliated store from feature quantity C(i=i1, j=j1) at the affiliated store of the first consumer, based on corrected feature quantity C(i=i2, j) obtained by correcting feature quantity A(i=i2, j) related to the purchasing behavior of a second consumer at multiple stores H1, H2, ..., L1, L2, ... including affiliated stores H1, H2, ... and non-affiliated stores L1, L2, ... described in the second dataset DS2.
[0080] Here, index i1 is the index of the first consumer, index i2 is the index of the second consumer, index j1 is the index of a combination of a member store and a product, and index j2 is the index of a combination of a non-member store and a product.
[0081] The missing part completion unit 117 inputs, for each first consumer, the feature C(i1, j1) of the purchasing behavior of the first consumer at an affiliated store identified from the first dataset DS1 of the corresponding first consumer into an estimation model, and estimates the feature C(i1, j2) of the purchasing behavior of the corresponding first consumer at a non-affiliated store.
[0082] The missing data completion unit 117 completes the missing purchasing data of a non-affiliated store in each record included in the corrected first dataset DS1 by adding, as missing data, estimated purchasing data that explains the purchasing behavior of the corresponding first consumer at the non-affiliated store.
[0083] The estimated purchase data added to each record describes, for each combination of a non-affiliated store and a product, a feature C(i1, j2) of the purchasing behavior of the corresponding first consumer regarding the corresponding combination of a non-affiliated store and a product.
[0084] Fig. 5 illustrates an example of a supplemented data set DC that is constructed by supplementing the corrected first data set DS1 with estimated purchase data of non-affiliated stores. According to Fig. 5, the supplemented data set DC has not only purchase data that describes the feature amount C(i1,j1) for affiliated stores H1, H2, ..., but also purchase data that describes the feature amount C(i1,j2) for non-affiliated stores L1, L2, ....
[0085] The output unit 119 outputs the complemented dataset DC generated by the loss complement unit 117. For example, the output unit 119 records the complemented dataset DC in the storage 13. Alternatively, the output unit 119 displays the complemented dataset DC to a user of the information processing system 1 through the user interface 15.
[0086] In this manner, the information processing system 1 of this embodiment performs bias correction and loss completion on the first data set DS1, and outputs the completed data set DC.
[0087] Next, details of the correction and complement process executed by the processor 11 based on an instruction from a user will be described with reference to Fig. 6 and Fig. 7. The processor 11 executes the correction and complement process, thereby functioning as the data acquisition unit 111, bias correction unit 113, model construction unit 115, loss complement unit 117, and output unit 119 described above.
[0088] When the correction and complement process shown in FIG. 6 is started, the processor 11 acquires a first data set DS1 to be corrected and complemented (S110), and further acquires a second data set DS2 for reference (S120).
[0089] Then, the processor 11 estimates a first bias p(i|x) contained in each feature A(i, j) of the first dataset DS1 and the second dataset DS2 by calculating a propensity score θ corresponding to the probability that a consumer conditioned on the consumer attribute X is a participant of the reward program based on the first dataset DS1 and the second dataset DS2 (S130).
[0090] That is, the processor 11 calculates the consumer attribute X=X i For consumer i with p(i|x i )≒p(i|θ) is the propensity score θ=e i (X i )=p(x i =1|X i ) where x i corresponds to the above x and indicates whether or not the consumer i is a participant of the reward program. i = 1 indicates that consumer i is a participant of the rewards program, and x i = 0 indicates that consumer i is not a participant of the rewards program. The first consumer i = 1, and the second consumer basically has x i =0.
[0091] p(x i =1|X i ) is consumer attribute X=X i A consumer i who has a rewards program participant (x i =1), p(i|θ)=p(i|e i (X i ) is the propensity score θ=e i (X i ) is the first bias p(i|θ) contained in the feature A(i,j) of consumer i in the model, which corresponds to the probability that a purchase is observed when the propensity score has the value θ.
[0092] When a reward program participant has made a purchase at an affiliated store, and the probability of observing that purchase is α, the first bias p(i|θ) corresponds to α·θ. If we assume that α=1, then the first bias p(i|θ)=θ. α=1 corresponds to the case where the consumer remembers to present his / her membership ID.
[0093] In the next step S140, the processor 11 estimates a second bias p(i,j|z) included in each feature A(i,j) in the first data set DS1 and the second data set DS2. Here, the method of estimating the second bias p(i,j|z) will be described in detail.
[0094] As described above, the first dataset DS1 and the second dataset DS2 in this embodiment describe feature A(i, j) for each combination of store and product, but the stores are distinguished by store chain, and a group of stores in the same chain are treated as a single store.
[0095] Therefore, the feature value A(i,j) of the purchasing behavior of the store h[j] and product combination corresponding to the index j of the consumer i is the share z r[i]g[j] The bias p(i,j|z r[i]g[j] ) can be said to be occurring.
[0096] The store category here corresponds to a business type such as a drug store or a convenience store. When the store category is a drug store, the share of a drug store of a particular chain among drug stores in a residential area corresponds to the share within the store category here. In the present embodiment, the share within a store category will hereinafter be referred to simply as the share.
[0097] For example, when the share of drug stores in a consumer's area is smaller than that of drug stores in a first chain, the consumer will have fewer opportunities to engage in purchasing behavior at drug stores in the first chain, and the observed amount of purchasing behavior at drug stores in the first chain explained in the first dataset DS1 will be smaller, compared to when the share of drug stores in the first chain is larger than that of drug stores in the second chain.
[0098] For this reason, the processor 11 calculates, for each feature A(i,j), the share z r[i]g[j] Calculate the share z r[i]g[j] Based on the second bias p(i,j|z r[i]g[j] ) is estimated.
[0099] Specifically, Sharez r[i]g[j] The second bias p(i,j|z r[i]g[j] )=z r[i]g[j] It is estimated that the share z r[i]g[j] can be said to represent the magnitude of the opportunity for a consumer to engage in purchasing behavior through store h[j] in area r[i].
[0100] Sharez r[i]g[j] To calculate this, the storage 13 can store a table that describes, for each region and store category, the total number of stores M in the corresponding store category in the corresponding region, and the number N of affiliated stores for each chain in the corresponding store category.
[0101] Processor 11 calculates the share z of store h[j]. r[i]g[j] Using the number of affiliated stores N corresponding to store h[j] in store category g[j] corresponding to store h[j] in the area corresponding to residential area r[i] and the total number of stores M in store category g[j], the formula z r[i]g[j] =N / M.
[0102] In this way, the processor 11 estimates the first bias p(i|x) and the second bias p(i,j|z) contained in each feature A(i,j) in the first data set DS1 and the second data set DS2 in S130 and S140, and then executes processing in S150.
[0103] In S150, the processor 11 corrects each feature A(i, j) explained by the first data set DS1 and the second data set DS2 in a direction to suppress the first bias p(i|x) and the second bias p(i, j|z).
[0104] That is, the processor 11 calculates a correction value C(i,j)=A(i,j) / (p(i|x)·p(i,j|z)) corresponding to the feature A(i,j) by applying to each feature A(i,j) the inverse of a first bias p(i|x) due to the attribute of the corresponding consumer i and also the inverse of a second bias p(i,j|z) due to the share of the store h[j] in the residential area r[i] of the corresponding consumer i. The residential area corresponds to the area where the corresponding consumer is active.
[0105] In the next step S160, the processor 11 constructs an estimation model for estimating a feature C(i1, j2) related to the purchasing behavior of the first consumer at a non-affiliated store from the feature C(i1, j1) of the first consumer at an affiliated store, based on a corrected second dataset DS2 in which each feature A(i2, j) described by each record of the second dataset DS2 is corrected to a value C(i2, j).
[0106] The corrected second dataset DS2 describes, for each consumer i2, the consumer attributes X of the consumer i2, a feature C(i2, j1) related to the purchasing behavior of each product at each affiliated store of the corresponding consumer i2, and a feature C(i2, j2) related to the purchasing behavior of each product at each non-affiliated store of the corresponding consumer i2.
[0107] Therefore, if the corrected second data set DS2 is used as training data, for example, by machine learning, consumer attribute X of consumer i = X i and the feature quantity C(i,j1) related to the purchasing behavior of the consumer i at the affiliated store, an estimation model can be constructed for estimating the feature quantity C(i,j2) related to the purchasing behavior of the consumer i at the non-affiliated store. The estimation model may be constructed, for example, by regression, or may be constructed using a neural network, or may be constructed using other machine learning models.
[0108] In the next step S170, the processor 11 uses the corrected first dataset DS1 and the above estimation model to calculate the feature quantity C(i1, j2) for each combination of store and product at each non-affiliated store for each first consumer i1.
[0109] Specifically, for each record of the corrected first data set DS1, the processor 11 calculates the consumer attribute X of the consumer indicated by the corresponding record, i and feature quantity C(i1,j1) relating to the purchasing behavior of each combination of store and product at affiliated stores are input into an estimation model, and feature quantity C(i1,j2) relating to the purchasing behavior of the corresponding consumer for each combination of store and product at non-affiliated stores is obtained as output from the estimation model.
[0110] Then, the processor 11 adds estimated purchase data describing feature C(i1, j2) related to the purchasing behavior of the corresponding consumer for each combination of store and product at non-affiliated stores to each record of the corrected first data set DS1, thereby completing the missing parts of the first data set DS1 (S180).
[0111] In S180, the processor 11 generates a supplemented data set DC corresponding to the first data set DS1 by adding estimated purchase data to each record of the corrected first data set DS1. Then, the processor 11 outputs the generated supplemented data set DC (S190) and ends the correction and supplementation process.
[0112] According to the information processing system 1 of the present embodiment described above, the first data set DS1 relating to consumer purchasing behavior can be corrected so as to appropriately remove multiple biases contained in the feature A that explains the consumer purchasing behavior. Furthermore, based on the feature C from which the biases have been removed, data missing in the first data set DS1 is complemented by using the second data set DS2.
[0113] Therefore, according to this embodiment, bias correction and missingness complementation can be performed well on a dataset related to purchasing behavior that includes biased information, and the accuracy of the information in the dataset can be improved. Specifically, it is possible to suppress bias in information included in a dataset related to purchasing behavior collected using a reward program, which is dependent on the collection format, and further, it is possible to complement purchasing data of non-member stores that cannot be obtained by collecting data through a reward program, thereby generating a dataset that is useful for analyzing consumer purchasing behavior.
[0114] [Second embodiment] Next, the information processing system 1 of the second embodiment will be described. However, the information processing system 1 of the second embodiment differs from the first embodiment in the contents of the first data set DS1 and the second data set DS2, and therefore only some of the functions realized by the bias correction unit 113, the model construction unit 115, and the missing part completion unit 117 are different from the first embodiment. Therefore, hereinafter, as the configuration of the information processing system 1 of the second embodiment, the configuration different from the first embodiment will be selectively described.
[0115] The first dataset DS1 acquired by the data acquisition unit 111 (see Figure 1) in this embodiment is a dataset that describes the purchasing behavior of consumers who use a specific electronic payment service, rather than the purchasing behavior of reward program participants, as the behavior of a first group of people.
[0116] Specifically, the first data set DS1 is a data set that describes the purchasing behavior of consumers corresponding to credit card users at stores that accept credit card payments. The first data set DS1 is created, for example, from credit card payment records. A credit card user may be a consumer corresponding to the account holder of a credit card issued by a specific company.
[0117] Specifically, the first dataset DS1 (see FIG. 3) has a record for each consumer corresponding to a credit card user that describes the purchasing behavior of the corresponding consumer by credit card payment. Each record describes, as feature A related to the purchasing behavior of the corresponding consumer using a credit card, feature A for each combination of store and product for multiple stores H1, H2, ... that accept credit card payments, together with consumer attributes X.
[0118] On the other hand, the second data set DS2 (see FIG. 4) acquired by the data acquisition unit 111 has a record describing the purchasing behavior of the corresponding consumer, regardless of the payment method, for each participant of the panel survey. Each record describes, as a feature A related to the purchasing behavior of the corresponding consumer, a feature A related to the purchasing behavior for each combination of store and product for a plurality of stores H1, H2, ..., L1, L2, ..., including stores that accept credit card payments and stores that do not accept credit card payments, together with consumer attributes X.
[0119] According to this embodiment, similar to the first embodiment, each feature A(i, j) explained by the first dataset DS1 and the second dataset DS2 has a first bias p(i|x) corresponding to the difference in consumer sets between the first dataset DS1 and the second dataset DS2.
[0120] In addition, the choice of payment method by the consumer, i.e., whether or not to choose credit card payment, varies depending on the store. Therefore, the feature A(i,j) has a second bias p(i,j|z) depending on the consumer and the store.
[0121] For this reason, the bias correction unit 113 calculates the first bias p(i|x) by using the propensity score θ, similarly to the first embodiment. i (X i )=p(x i =1|X i ), consumer attribute X=X i Consumer i with a credit card (x i=1) is the probability p(x i =1|X i ) can be calculated, which allows a first bias p(i|θ)=θ to be estimated.
[0122] The bias correction unit 113 further calculates the credit card usage rate V(i,j) of the corresponding consumer i at the store corresponding to the index j as a second bias p(i,j|z) (p(i,j|z)=V(i,j)).
[0123] The credit card usage rate at a certain store means the percentage of payments in which a credit card is used as a payment method among all payments in the store, regardless of the payment method. The credit card usage rate represents the characteristics of the use of payment methods at the corresponding store. In other words, the second bias p(i,j|z) in this embodiment is a bias caused by differences in the characteristics of the determination method at each store.
[0124] Credit card payment data from each store and POS (Point of Sales) data, which is a sales record from each store, may be used to calculate the credit card usage rate. ID-POS data from panel survey participants may also be used to calculate the credit card usage rate.
[0125] Based on the ID-POS data, an estimation model can be constructed that estimates the credit card usage rate V of consumers having corresponding consumer attributes in each store from the consumer attributes. Based on the constructed estimation model, the credit card usage rate V(i,j) in each store of each consumer corresponding to the first data set DS1 can be estimated.
[0126] The bias correction unit 113 corrects each feature amount A(i,j) in the first data set DS1 by a value C(i,j)=A(i,j) / (p(i|θ)*p(i,j|z)) using the first bias p(i|θ)=θ and the second bias p(i,j|z)=V(i,j). The bias correction unit 113 corrects each feature amount A(i,j) in the second data set DS2 by a value C(i,j)=A(i,j) / p(i|θ).
[0127] Based on the second dataset DS2 corrected in this manner, the model construction unit 115 constructs an estimation model that estimates the feature C(i, j2) of a store where credit cards cannot be used from the consumer attribute X and the feature C(i, j1) of a store where credit cards can be used.
[0128] The missing part complementing unit 117 uses this estimation model to complement the missing parts of the first data set DS1 by adding, as missing data, estimated purchase data describing the feature value C(i1, j2) of the corresponding consumer i1 at the store where credit cards cannot be used, to each record of the first data set DS1. The output unit 119 outputs the thus generated complemented data set DC (see FIG. 5).
[0129] According to this embodiment, it is possible to effectively perform bias correction and missing complementation for a dataset on purchasing behavior that includes biased information and is prepared based on payment records using limited payment methods, and to increase the accuracy of the information in the dataset. Specifically, it is possible to suppress bias in information according to the collection form contained in a dataset on purchasing behavior collected using credit card payment records, and further, it is possible to complement purchase data of stores that do not accept credit card payments, which cannot be obtained only by using credit card payment records, and thus it is possible to generate a dataset that is useful for analyzing consumer purchasing behavior.
[0130] [Third embodiment] Next, the information processing system 1 of the third embodiment will be described. However, the information processing system 1 of the third embodiment differs from the first embodiment in the contents of the first data set DS1 and the second data set DS2, and therefore only some of the functions realized by the bias correction unit 113, the model construction unit 115, and the missing part completion unit 117 are different from the first embodiment. Therefore, hereinafter, as the configuration of the information processing system 1 of the third embodiment, the configuration different from the first embodiment will be selectively described.
[0131] In this embodiment, the first data set DS1 acquired by the data acquisition unit 111 has a record for each user of an information terminal who uses a specific online service. Each record describes a feature amount A for each type of use of the information terminal of the corresponding user.
[0132] A user who uses a particular online service may be, for example, a user who uses a particular online service (such as an online shopping mall) by using particular application software.
[0133] The information terminal is, for example, a smartphone or a tablet terminal. The feature amount A for each type is specifically a feature amount related to the use of each application software installed in the information terminal.
[0134] That is, each record of the first data set DS1 describes a feature A related to the use of each application software in the information terminal of the corresponding user. The feature A can be measured through the specific application software used when using the specific online service.
[0135] The second dataset DS2 has a record for each user of an information terminal corresponding to a participant of the panel survey. Each record describes a feature A for each type of information terminal usage for the corresponding user. Specifically, each record describes a feature A for the usage of each application software on the information terminal of the corresponding user. With regard to the second dataset DS2, the feature A for the usage of each application software for each user can be measured through application software for measurement distributed to participants of the panel survey.
[0136] The feature A described in the first dataset DS1 and the second dataset DS2 may be a feature A related to the use of an online service used through the corresponding application software. In this case, the feature A may be said to be a feature related to the online behavior of a user.
[0137] As in the first embodiment, the bias correction unit 113 calculates the propensity score θ=e i (X i )=p(x i =1|X i ), consumer attribute X=X i A user i who has i =1) is the probability p(x i =1|X i ) and use this to estimate the first bias p(i|θ)=θ.
[0138] The bias correction unit 113 further calculates the category share W(i,j) of the application software corresponding to index j in the operating system (OS) of the information terminal used by the corresponding user i as a second bias p(i,j|z) (p(i,j|z)=W(i,j)).
[0139] The category share of a given application software on an OS is a value calculated by dividing the total usage (e.g., the number of installations or the total usage time) of the given application software on that OS by the total usage of all application software in the same category as the given application software on the same OS. The total usage here refers to the total usage of multiple users who use the same OS. The category share of each application software on each OS is determined, for example, through panel survey data.
[0140] The share of each application software within a category varies depending on the OS. Therefore, a second bias p(i,j|z) may occur in the feature A related to the usage of each application software by each user depending on the OS of the information terminal used by the user.
[0141] For this reason, the bias correction unit 113 removes the second bias p(i,j|z) caused by the share in the feature A(i,j) by dividing, for each user i, the feature A(i,j) for each application software j of the corresponding user i by the share in the category of the corresponding application software j in the OS of the information terminal of the corresponding user i.
[0142] Specifically, the bias correction unit 113 corrects the feature A(i, j) for each user's application software in the first dataset DS1 and the second dataset DS2 by the value C(i, j) = A(i, j) / (p(i|θ) * p(i, j|z)).
[0143] In addition, the application software handled in this embodiment includes not only first application software in which the first dataset DS1 describes feature A, but also second application software in which the first dataset DS1 does not describe feature A.
[0144] That is, the first data set DS1 describes the feature quantity A(i,j) for each user and each application software for the first application software. The second data set DS2 describes the feature quantity A(i,j) for each user and each application software for the first and second application software.
[0145] The model construction unit 115 calculates the feature quantity C(i,j) for each application related to the first application software and the corresponding consumer attribute X=X i An estimation model is constructed to estimate the feature quantity C(i,j) for each application related to the second application software of the corresponding user i from the above.
[0146] The missing part complementing unit 117 adds the feature quantity C(i,j) for each user and for each application software related to the second application software estimated using the estimation model to the corrected first data set DS1, thereby realizing missing part complementing for the first data set DS1. The output unit 119 outputs the thus generated complemented data set DC.
[0147] According to this embodiment, bias correction and missingness completion can be performed well on a dataset that describes the feature amount related to the use of application software, including information bias, prepared based on the behavior records of users who use limited online services, and the information accuracy of the dataset can be improved. Specifically, it is possible to suppress the bias of information according to the collection form contained in the dataset that describes the feature amount related to the use of application software collected using the behavior records through the information terminals of limited users, and further, it is possible to complement the behavior records through the information terminals of limited users that cannot be obtained from the information terminals of the limited users, and therefore it is possible to generate a dataset that is useful for analyzing the behavior through the information terminals of consumers.
[0148] [Other embodiments] The present disclosure is not limited to the above-described first to third embodiments, and various aspects can be adopted. The configurations of the first data set DS1 and the second data set DS2 handled by the information processing system 1 are not limited to the first to third embodiments. The information processing system 1 may be configured not to perform missing data complementation, and the output unit 119 may be configured to output the first data set DS1 corrected by the bias correction unit 113.
[0149] The function of one component in the above embodiments may be distributed among multiple components. The functions of multiple components may be integrated into one component. Part of the configuration of the above embodiments may be omitted. At least part of the configuration of the above embodiments may be added to or substituted for the configuration of another of the above embodiments. All aspects included in the technical idea identified from the wording of the claims are embodiments of the present disclosure.
[0150] [Technical idea disclosed in this specification] It can be understood that the present specification discloses the following technical ideas. [Item 1] A first acquisition unit that acquires a first dataset related to a first plurality of subjects as a first sample; A second acquisition unit that acquires a second dataset related to a second plurality of subjects as a second sample; a correction unit that corrects the first data set based on the second data set; Equipped with The first data set comprises, for each of the plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as first behavioral data; the second data set comprises, for each of the plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as second behavioral data; The first behavior data describes, for each of a plurality of first behaviors, a corresponding behavior of the corresponding first subject by using a numerical value; The second behavior data describes, for each of a plurality of second behaviors, a corresponding behavior of the corresponding second subject by using a numerical value; the plurality of second actions includes the plurality of first actions; The correction unit corrects the first dataset by correcting the numerical values for each of the first behaviors of the multiple first subjects in the first dataset in a direction to suppress a first bias, which is a bias in the numerical values due to a bias in the multiple first subjects in the first dataset, and a second bias, which is a bias in the numerical values due to the behavioral environment of each of the multiple first subjects in the first dataset. [Item 2] the first sample is a set of individuals limited to individuals belonging to a particular group; The second sample is a group of people not belonging to the specific group or a group of people not limited to people belonging to the specific group, a common variable is used between the first behavioral data and the second behavioral data to explain the behaviors; 2. The information processing system according to item 1, wherein the correction unit corrects the numerical values for each of the first behaviors for the plurality of first subjects in the first dataset, based on a propensity score corresponding to a probability that a person conditioned by the common variable based on the first dataset and the second dataset belongs to the specific group, in a direction to suppress the first bias resulting from the first sample being limited to a set of people belonging to the specific group. [Item 3] the first sample is a set of individuals limited to individuals belonging to a particular group; The second sample is a group of people not belonging to the specific group or a group of people not limited to people belonging to the specific group, the first behavioral data and the second behavioral data have attribute information expressing an attribute of a subject who is a subject of the behavior by at least one variable; a common variable is used between the first behavioral data and the second behavioral data to explain the attribute; 2. The information processing system according to item 1, wherein the correction unit corrects the numerical values for each of the first behaviors for the plurality of first subjects in the first dataset, based on a propensity score corresponding to a probability that a person conditioned by the common variable based on the first dataset and the second dataset belongs to the specific group, in a direction to suppress the first bias resulting from the first sample being limited to a set of people belonging to the specific group. [Item 4] each of the plurality of first behaviors and the plurality of second behaviors is a purchasing behavior; The first sample is a set of consumers who participate in a particular reward program for purchases or who use a particular electronic payment service; An information processing system described in any one of items 1 to 3, wherein the first behavioral data is data describing the purchasing behavior of the corresponding first subject observed through the provision of identification information from the corresponding first subject. [Item 5] 5. The information processing system according to any one of items 1 to 4, wherein each of the plurality of first actions and the plurality of second actions is a purchasing action that is distinguished by at least one of a purchasing object and a purchasing store. [Item 6] 4. The information processing system according to any one of items 1 to 3, wherein each of the plurality of first actions and the plurality of second actions is an action taken through an information terminal. [Item 7] 6. The information processing system according to any one of items 1 to 5, wherein the correction unit determines, for each of a plurality of regions, a magnitude of an opportunity to perform each of the plurality of first behaviors, and corrects the numerical values for each of the first behaviors for the plurality of first subjects in the first dataset in a direction to suppress the second bias resulting from the magnitude of the opportunity to perform each of the plurality of first subjects in an active region. [Item 8] each of the plurality of first actions and the plurality of second actions is a purchasing action distinguished at least by a purchasing store, and is a purchasing action in one store among a plurality of stores; The one store includes a group of stores of the same type, The information processing system according to any one of items 1 to 5, wherein the correction unit corrects the numerical values for each of the first behaviors for the plurality of first subjects in the first dataset based on the geographical distribution of the group of stores of the same business type in a direction to suppress the second bias caused by the geographical distribution. [Item 9] each of the plurality of first actions and the plurality of second actions is a purchasing action distinguished at least by a purchasing store, and is a purchasing action in one store among a plurality of stores; The information processing system according to any one of items 1 to 5, wherein the correction unit corrects the numerical values for each of the first actions for the plurality of first subjects in the first dataset in a direction that suppresses the second bias caused by differences in characteristics of the payment means used at each of the plurality of stores, based on the characteristics of the payment means used at each of the plurality of stores. [Item 10] the plurality of second actions includes a plurality of third actions different from the plurality of first actions; The second behavior data describes, for each of the plurality of third behaviors not explained in the first behavior data, a corresponding behavior of the corresponding second subject by using a numerical value; The correction unit is constructing an estimation model based on the second data set for calculating estimated values of the numerical values related to the plurality of third behaviors from the numerical values related to the plurality of first behaviors; correcting the first dataset to supplement information regarding the third behaviors missing in the first dataset by adding, for each of the first subjects, the estimated values regarding the third behaviors of the corresponding first subject calculated based on the estimation model to the first behavioral data; The information processing system according to any one of items 1 to 9, wherein the calculation of the estimated value based on the estimation model is performed based on the numerical values related to the plurality of first behaviors in the first behavior data that have been corrected in a direction to suppress the first bias and the second bias. [Item 11] 11. A computer program for causing a computer to realize functions as the first acquisition section, the second acquisition section, and the correction section in the information processing system according to any one of items 1 to 10. [Item 12] 1. A computer-implemented information processing method, comprising: Obtaining a first dataset relating to a first plurality of subjects as a first sample; Obtaining a second dataset relating to a second plurality of subjects as a second sample; correcting the first data set based on the second data set; Including, The first data set comprises, for each of the plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as first behavioral data; the second data set comprises, for each of the plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as second behavioral data; The first behavior data describes, for each of a plurality of first behaviors, a corresponding behavior of the corresponding first subject by using a numerical value; The second behavior data describes, for each of a plurality of second behaviors, a corresponding behavior of the corresponding second subject by using a numerical value; the plurality of second actions includes the plurality of first actions; The information processing method includes correcting the first dataset by correcting the numerical values for each of the first behaviors of the multiple first subjects in the first dataset in a direction to suppress a first bias, which is a bias in the numerical values due to bias in the multiple first subjects in the first dataset, and a second bias, which is a bias in the numerical values due to the behavioral environment of each of the multiple first subjects in the first dataset. [Explanation of symbols]
[0151] 1...information processing system, 11...processor, 12...memory, 13...storage, 15...user interface, 17...input / output interface, 19...communication device, 111...data acquisition unit, 113...bias correction unit, 115...model construction unit, 117...missing part completion unit, 119...output unit.
Claims
1. A first acquisition unit that acquires a first dataset related to a first plurality of subjects as a first sample; A second acquisition unit that acquires a second dataset related to a second plurality of subjects as a second sample; a correction unit that corrects the first data set based on the second data set; Equipped with The first data set comprises, for each of the plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as first behavioral data; the second data set comprises, for each of the plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as second behavioral data; The first behavior data describes, for each of a plurality of first behaviors, a corresponding behavior of the corresponding first subject by using a numerical value; The second behavior data describes, for each of a plurality of second behaviors, a corresponding behavior of the corresponding second subject by using a numerical value; the plurality of second actions includes the plurality of first actions; the first sample is a set of individuals limited to individuals belonging to a particular group; The second sample is a group of people not belonging to the specific group or a group of people not limited to people belonging to the specific group, a common variable is used between the first behavioral data and the second behavioral data to explain the behaviors; The correction unit is Identifying, for each of the first subjects, a first bias, which is a bias in the numerical values in the first dataset due to the first sample being limited to a set of people belonging to the specific group, based on a propensity score corresponding to a probability that a person conditioned on the common variables based on the first dataset and the second dataset belongs to the specific group; Identifying a second bias, which is a bias in the numerical values caused by differences in behavioral environments among the plurality of first subjects in the first dataset, for each combination of the first subject and the first behavior; correcting the first dataset in a direction to suppress the first bias and the second bias by correcting the numerical values for each of the first behaviors for the first subjects in the first dataset with the reciprocal of the first bias and the reciprocal of the second bias; the first bias corresponds to a likelihood of observing the first behavior of the first subject based on the propensity score in the common variable of the first subject; An information processing system in which the second bias corresponds to the magnitude of a factor by which a first behavior corresponding to the combination is observed in an environment in which a first subject corresponding to the combination is placed.
2. A first acquisition unit that acquires a first dataset related to a first plurality of subjects as a first sample; A second acquisition unit that acquires a second dataset related to a second plurality of subjects as a second sample; a correction unit that corrects the first data set based on the second data set; Equipped with The first data set comprises, for each of the plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as first behavioral data; the second data set comprises, for each of the plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as second behavioral data; The first behavior data describes, for each of a plurality of first behaviors, a corresponding behavior of the corresponding first subject by using a numerical value; The second behavior data describes, for each of a plurality of second behaviors, a corresponding behavior of the corresponding second subject by using a numerical value; the plurality of second actions includes the plurality of first actions; the first sample is a set of individuals limited to individuals belonging to a particular group; The second sample is a group of people not belonging to the specific group or a group of people not limited to people belonging to the specific group, the first behavioral data and the second behavioral data have attribute information expressing an attribute of a subject who is a subject of the behavior by at least one variable; a common variable is used between the first behavioral data and the second behavioral data to explain the attribute; The correction unit is Identifying, for each of the first subjects, a first bias, which is a bias in the numerical values in the first dataset due to the first sample being limited to a set of people belonging to the specific group, based on a propensity score corresponding to a probability that a person conditioned on the common variables based on the first dataset and the second dataset belongs to the specific group; Identifying a second bias, which is a bias in the numerical values caused by differences in behavioral environments among the plurality of first subjects in the first dataset, for each combination of the first subject and the first behavior; correcting the first dataset in a direction to suppress the first bias and the second bias by correcting the numerical values for each of the first behaviors for the first subjects in the first dataset with the reciprocal of the first bias and the reciprocal of the second bias; the first bias corresponds to a likelihood of observing the first behavior of the first subject based on the propensity score in the common variable of the first subject; An information processing system in which the second bias corresponds to the magnitude of a factor by which a first behavior corresponding to the combination is observed in an environment in which a first subject corresponding to the combination is placed.
3. each of the plurality of first behaviors and the plurality of second behaviors is a purchasing behavior; The first sample is a set of consumers who participate in a particular reward program for purchases or who use a particular electronic payment service; An information processing system as described in claim 1 or claim 2, wherein the first behavioral data is data describing the purchasing behavior of the corresponding first subject observed through the provision of identification information from the corresponding first subject.
4. 3. The information processing system according to claim 1, wherein each of the plurality of first actions and the plurality of second actions is a purchasing action that is distinguished by at least one of a purchasing object and a purchasing store.
5. 3. The information processing system according to claim 1, wherein each of the plurality of first actions and the plurality of second actions is an action taken through an information terminal.
6. the correction unit, for each of a plurality of regions, determines the magnitude of an opportunity to perform each of the plurality of first behaviors, and, for each combination of the first subject and the first behavior, identifies the second bias resulting from the magnitude of the opportunity to perform the first behavior corresponding to the combination in an activity region of the first subject corresponding to the combination; 3. The information processing system according to claim 1, wherein the second bias corresponds to a magnitude of the execution opportunity.
7. each of the plurality of first actions and the plurality of second actions is a purchasing action distinguished at least by a purchasing store, and is a purchasing action in one store among a plurality of stores; The one store includes a group of stores of the same type, the correction unit determines, for each of a plurality of regions, a share of each of the plurality of stores, and for each combination of the first subject and the first behavior, identifies the second bias resulting from a share of a store in which the first behavior corresponding to the combination was performed in an activity area of the first subject corresponding to the combination; 3. The information processing system according to claim 1, wherein the second bias corresponds to a size of the share.
8. each of the plurality of first actions and the plurality of second actions is a purchasing action distinguished at least by a purchasing store, and is a purchasing action in one store among a plurality of stores; the correction unit, for each of the plurality of stores, determines the amount of use of a specific payment method for each consumer attribute in the corresponding store, and for each combination of the first subject and the first behavior, identifies the second bias resulting from the amount of use of a specific payment method in the store where the first behavior corresponding to the combination was performed, related to the consumer attribute of the first subject corresponding to the combination; 3. The information processing system according to claim 1, wherein the second bias corresponds to the magnitude of the amount of use.
9. Each of the plurality of first actions and the plurality of second actions is an action that is distinguished by a type of application software used in an information terminal, and is an action that uses one type of application software among the plurality of types of application software, the correction unit determines, for each type of information terminal, a share of each type of application software in a corresponding type of information terminal, and, for each combination of the first subject and the first behavior, identifies the second bias caused by a share of application software used in the first behavior corresponding to the combination in a type of information terminal used by the first subject corresponding to the combination; 3. The information processing system according to claim 1, wherein the second bias corresponds to a size of the share.
10. the plurality of second actions includes a plurality of third actions different from the plurality of first actions; The second behavior data describes, for each of the plurality of third behaviors not explained in the first behavior data, a corresponding behavior of the corresponding second subject by using a numerical value; The correction unit is constructing an estimation model based on the second data set for calculating estimated values of the numerical values related to the plurality of third behaviors from the numerical values related to the plurality of first behaviors; correcting the first dataset to supplement information regarding the third behaviors missing in the first dataset by adding, for each of the first subjects, the estimated values regarding the third behaviors of the corresponding first subject calculated based on the estimation model to the first behavioral data; An information processing system as described in claim 1 or claim 2, wherein the calculation of the estimated value based on the estimation model is performed based on the numerical values regarding the plurality of first behaviors in the first behavior data that have been corrected in a direction to suppress the first bias and the second bias.
11. 3. A computer program for causing a computer to realize functions as the first acquisition unit, the second acquisition unit, and the correction unit in the information processing system according to claim 1.
12. 1. A computer-implemented information processing method, comprising: Obtaining a first dataset relating to a first plurality of subjects as a first sample; Obtaining a second dataset relating to a second plurality of subjects as a second sample; correcting the first data set based on the second data set; Including, The first data set comprises, for each of the plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as first behavioral data; the second data set comprises, for each of the plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as second behavioral data; The first behavior data describes, for each of a plurality of first behaviors, a corresponding behavior of the corresponding first subject by using a numerical value; The second behavior data describes, for each of a plurality of second behaviors, a corresponding behavior of the corresponding second subject by using a numerical value; the plurality of second actions includes the plurality of first actions; the first sample is a set of individuals limited to individuals belonging to a particular group; The second sample is a group of people not belonging to the specific group or a group of people not limited to people belonging to the specific group, a common variable is used between the first behavioral data and the second behavioral data to explain the behaviors; The correction is Identifying, for each of the first subjects, a first bias, which is a bias in the numerical values in the first dataset due to the first sample being limited to a set of people belonging to the specific group, based on a propensity score corresponding to a probability that a person conditioned on the common variables based on the first dataset and the second dataset belongs to the specific group; Identifying a second bias, which is a bias in the numerical values caused by differences in behavioral environments among the plurality of first subjects in the first dataset, for each combination of the first subject and the first behavior; correcting the first dataset in a direction to suppress the first bias and the second bias by correcting the numerical values for each of the first behaviors for the first subjects in the first dataset with the reciprocal of the first bias and the reciprocal of the second bias; the first bias corresponds to a likelihood of observing the first behavior of the first subject based on the propensity score in the common variable of the first subject; An information processing method in which the second bias corresponds to the magnitude of a factor by which a first behavior corresponding to the combination is observed in the environment in which a first subject corresponding to the combination is placed.
13. 1. A computer-implemented information processing method, comprising: Obtaining a first dataset relating to a first plurality of subjects as a first sample; Obtaining a second dataset relating to a second plurality of subjects as a second sample; correcting the first data set based on the second data set; Including, The first data set comprises, for each of the plurality of first subjects, behavioral data describing a behavior of the corresponding first subject as first behavioral data; the second data set comprises, for each of the plurality of second subjects, behavioral data describing a behavior of the corresponding second subject as second behavioral data; The first behavior data describes, for each of a plurality of first behaviors, a corresponding behavior of the corresponding first subject by using a numerical value; The second behavior data describes, for each of a plurality of second behaviors, a corresponding behavior of the corresponding second subject by using a numerical value; the plurality of second actions includes the plurality of first actions; the first sample is a set of individuals limited to individuals belonging to a particular group; The second sample is a group of people not belonging to the specific group or a group of people not limited to people belonging to the specific group, the first behavioral data and the second behavioral data have attribute information expressing an attribute of a subject who is a subject of the behavior by at least one variable; a common variable is used between the first behavioral data and the second behavioral data to explain the attribute; The correction is Identifying, for each of the first subjects, a first bias, which is a bias in the numerical values in the first dataset due to the first sample being limited to a set of people belonging to the specific group, based on a propensity score corresponding to a probability that a person conditioned on the common variables based on the first dataset and the second dataset belongs to the specific group; Identifying a second bias, which is a bias in the numerical values caused by differences in behavioral environments among the plurality of first subjects in the first dataset, for each combination of the first subject and the first behavior; correcting the first dataset in a direction to suppress the first bias and the second bias by correcting the numerical values for each of the first behaviors for the first subjects in the first dataset with the reciprocal of the first bias and the reciprocal of the second bias; the first bias corresponds to a likelihood of observing the first behavior of the first subject based on the propensity score in the common variable of the first subject; An information processing method in which the second bias corresponds to the magnitude of a factor by which a first behavior corresponding to the combination is observed in the environment in which a first subject corresponding to the combination is placed.
Citation Information
Patent Citations
Information processor, information processing system, information processing method and program
JP2016062511A
Estimation system, estimation method and computer program
JP2020161038A
Estimation system and estimation method
JP2022086981A
Reducing processing requirements to correct for bias in ratings data having interdependencies among demographic statistics
US20170017975A1
