JP2024002569A - Information processing device, information processing system, information processing method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve prediction accuracy of an intervention effect in direct marketing.

SOLUTION: An information processing device includes: an acquisition part for acquiring customer information and purchase information; a learning part which calculates a predicted value of a conditional intervention effect for each user and a predicted value of a future intervention effect by inputting the customer information and purchase information acquired by the acquisition part into a learning model in which data showing conditional intervention effects for each user due to campaign intervention and data showing future intervention effects due to the campaign intervention is learned as a learning data set; and a prediction unit for correcting the prediction value of the future intervention effect by using a trend.

SELECTED DRAWING: Figure 2

COPYRIGHT: (C)2024,JPO&INPIT

Description

The present invention relates to an information processing device, an information processing system, an information processing method, and a program.

Direct marketing is a method in which companies communicate directly with customers to encourage them to make purchases. For example, a department store issues a catalog to customers and accepts orders via the catalog. In this case, the department store can quantify the return on costs invested in marketing measures by comparing the costs required to create and distribute the catalog with the sales and profits generated through the catalog. It is possible to measure the

The structure of break-even in direct marketing activities is related to fixed production costs, variable production costs, and promotional costs. When making decisions in direct marketing, given the structure of fixed production costs and variable production costs, promotion costs are determined to maximize one's profits. On the other hand, in order to make accurate decisions regarding promotion costs, it is necessary to understand the impact of promotion costs on demand. The impact of promotional costs on demand varies depending on the campaign's targets, timing, creative, offer content, etc. Therefore, direct marketing as a whole consists of two decision-making elements: planning a campaign that maximizes demand acquisition per promotional cost, and determining the total promotional cost that maximizes final profit.

Here, it is assumed that iROAS (incremental return on ad spend: revenue increased by advertising expenditure) can be accurately calculated for each customer when targeted by a campaign. In this case, among the decision-making factors, the total amount of optimal targeting and promotion costs can be determined at once. For example, all customers whose iROAS is positive are targeted for a campaign, and the cost required in that case becomes the optimal total promotional cost.
However, in actual direct marketing, such ideal decision-making is not always carried out. The main factors include:

(1) Prediction of Uplift Due to fundamental problems in causal inference, it is difficult to estimate and predict the amplification effect (Uplift) of a campaign.

(2) Consideration of consumer heterogeneity The Uplift effect of a campaign differs depending on each customer, but in order to predict this with high accuracy, a method such as machine learning is required. If there is no organizational know-how to utilize machine learning, a method is often adopted in which consumer heterogeneity is taken into account using segmentation based on RFM (Recency Frequency Monetary). However, in most cases, heterogeneity is only predicted based on the response. There are several previous studies that utilize Uplift Modeling in direct marketing to make decisions that take into account consumer heterogeneity in the Uplift effect.

(3) Consideration of temporal heterogeneity Not only does the Uplift effect differ for each customer, but it also differs depending on the timing of campaign implementation. There are many techniques, mainly in the field of economics, for estimating intervention effects based on numerical changes in outcomes between the intervention and non-intervention periods. However, there is no technology for making direct marketing decisions based on predictions of future Uplift effects, assuming that Uplift effects themselves vary over time.
In actual direct marketing, trends in seasonal campaign results are quantified based on Response rather than Uplift, and different thresholds for RFM segments are used for each season, and targets are selected based on campaign results from the same period the previous year. There is also a technique for devising ways to reflect seasonality and trends as much as possible.

(4) Quantifying the impact of creative offer content on Uplift To predict marketing effectiveness in advance, consider not only the target and its timing, but also the impact of what kind of offer and creative will be presented. There is a need. On the other hand, when modeling customer reactions, it is difficult to quantify elements such as offers and creatives. The difficulty of predicting campaigns that have characteristics such as the target product in an offer being different each time or the creative being different each time increases as the variation factors due to offers and creatives that cannot be included in the model are large.

For example, in the technique described in Non-Patent Document 1, simulation data exhibits a certain degree of accuracy.

Shu Li and Peter Buhlmann, Estimating heterogeneous treatment effects in non-stationary time series with state-space models, December, 2018.

However, in the technique described in Non-Patent Document 1, when verified using actual data, seasonal components may disappear when estimating future trends. For this reason, trends in direct marketing could not be calculated. In addition, the effect of the intervention could not be predicted due to the disappearance of the seasonal component. As described above, there has been a problem in that the accuracy of predicting intervention effects in direct marketing is not sufficient.

The present invention has been made to solve the above problems, and its purpose is to provide an information processing device, an information processing system, an information processing method, and a program that can improve the prediction accuracy of intervention effects in direct marketing. It's about doing.

One aspect of the present invention has been made in view of the above problems, and includes an acquisition unit that acquires customer information and purchase information, data indicating conditional intervention effects for each user due to campaign intervention, and future predictions due to campaign intervention. By inputting the customer information and purchase information acquired by the acquisition unit into a learning model that has been trained using the data indicating the intervention effect as a learning data set, the predicted value of the conditional intervention effect for each user and the future intervention can be calculated. The information processing apparatus includes a learning unit that calculates a predicted value of an effect, and a prediction unit that corrects the predicted value of the future intervention effect using a trend.

Further, one aspect of the present invention has been made in view of the above problems, and includes data indicating a conditional intervention effect for each user in a predetermined period including at least a period in which the campaign intervenes and a period in which the campaign does not intervene; a learning unit that calculates a predicted value of a future intervention effect using a learning model that learns data indicating a future intervention effect as a learning data set, and data indicating a conditional intervention effect for each user during a period in which the campaign does not intervene. and a prediction unit that calculates a trend for correcting the predicted value of the future intervention effect by suppressing the influence of the intervention effect.

Further, one aspect of the present invention has been made in view of the above problems, and includes an acquisition unit that acquires customer information and purchase information, data indicating a conditional intervention effect for each user due to campaign intervention, and data indicating the conditional intervention effect for each user due to campaign intervention. By inputting the customer information and purchase information acquired by the acquisition unit into a learning model trained using data indicating future intervention effects as a learning data set, the predicted value of the conditional intervention effect for each user and the future The information processing system includes a learning unit that calculates a predicted value of the intervention effect in the future, and a prediction unit that corrects the predicted value of the future intervention effect using a trend.

Further, one aspect of the present invention has been made in view of the above problems, and includes data indicating a conditional intervention effect for each user in a predetermined period including at least a period in which the campaign intervenes and a period in which the campaign does not intervene; a learning unit that calculates a predicted value of a future intervention effect using a learning model that learns data indicating a future intervention effect as a learning data set, and data indicating a conditional intervention effect for each user during a period in which the campaign does not intervene. and a prediction unit that calculates a trend for correcting the predicted value of the future intervention effect by suppressing the influence of the intervention effect.

Further, one aspect of the present invention has been made in view of the above problems, and includes an acquisition process in which a computer of an information processing device acquires customer information and purchase information, and a conditional intervention effect for each user due to campaign intervention. By inputting the customer information and purchase information acquired by the acquisition unit into a learning model that has been trained using data indicating the future intervention effect of campaign intervention as a learning data set, the data indicating the future intervention effect of campaign intervention can be input into the learning model. This information processing method includes a learning process of calculating a predicted value of an effect and a predicted value of a future intervention effect, and a prediction process of correcting the predicted value of the future intervention effect using a trend.

Further, one aspect of the present invention has been made in view of the above-mentioned problems, and the computer of the information processing device sets conditions for each user during a predetermined period that includes at least a period in which a campaign intervenes and a period in which the campaign does not intervene. A learning process that calculates a predicted value of the future intervention effect using a learning model that learns data showing the intervention effect and data showing the future intervention effect as a learning data set, and An information processing method comprising: a prediction step of calculating a trend for suppressing the influence of data indicating a conditional intervention effect and correcting a predicted value of the future intervention effect.

Further, one aspect of the present invention has been made in view of the above problems, and includes an acquisition step of acquiring customer information and purchase information, and a conditional intervention effect for each user due to campaign intervention, in a computer of an information processing device. By inputting the customer information and purchase information acquired by the acquisition unit into a learning model that has been trained using data indicating the future intervention effect of campaign intervention as a learning data set, the data indicating the future intervention effect of campaign intervention can be input into the learning model. This is a program for executing a learning step of calculating a predicted value of an effect and a predicted value of a future intervention effect, and a prediction step of correcting the predicted value of the future intervention effect using a trend.

Further, one aspect of the present invention has been made in view of the above-mentioned problems, and is configured to set a condition for each user in a predetermined period including at least a period in which a campaign intervenes and a period in which the campaign does not intervene in the computer of the information processing device. a learning step of calculating a predicted value of the future intervention effect by a learning model that learns data showing the intervention effect and data showing the future intervention effect as a learning data set; This is a program for executing a prediction step of calculating a trend for suppressing the influence of data indicating a conditional intervention effect and correcting the predicted value of the future intervention effect.

According to at least one aspect described above, it is possible to improve the prediction accuracy of intervention effects in direct marketing.

1 is a schematic diagram showing an example of the configuration of an information processing system according to a first embodiment. FIG. 1 is a schematic block diagram showing the functional configuration of an information processing device according to the present embodiment. It is a schematic diagram showing an example of CCEUT concerning this embodiment. It is a schematic diagram showing an example of CCEPTA concerning this embodiment. It is a table showing an example of the structure of customer information according to the present embodiment. It is a table showing an example of the structure of purchasing information according to the present embodiment. It is a table showing an example of the structure of DM information according to the present embodiment. It is a table showing an example of the structure of destination information according to the present embodiment. 1 is a schematic block diagram showing a hardware configuration of an information processing device according to an embodiment. FIG. 7 is a flowchart illustrating an example of learning processing according to the present embodiment. It is a flow chart which shows an example of prediction processing concerning this embodiment. It is a flow chart which shows an example of prediction processing concerning this embodiment. It is a figure showing an example of verification data concerning this embodiment. It is a figure showing an example of distribution of each covariate concerning this embodiment. It is a figure showing another example of distribution of each covariate concerning this embodiment. It is a figure showing another example of distribution of each covariate concerning this embodiment. FIG. 2 is a diagram showing an example of the relationship between each covariate and a potential outcome according to the present embodiment. FIG. 7 is a diagram showing an example of a time-series plot of the true intervention effect for each individual according to the present embodiment. It is a figure showing an example of data specifications of a learning/prediction period concerning this embodiment. It is a figure which shows an example of the data sampled by each policy based on this embodiment. It is a figure showing an example of a trend in each policy concerning this embodiment. It is a figure showing an example of the verification result of sampling concerning this embodiment. FIG. 3 is a diagram illustrating an example of a verification result of the Calibration method according to the present embodiment.

(First embodiment)
Hereinafter, the first embodiment will be described in detail with reference to the drawings.

<Information processing system>
FIG. 1 is a schematic diagram showing an example of an information processing system Sys according to the first embodiment.
The information processing system Sys includes an information processing device 1, a company B, a user U1, and a user U2. The information processing device 1 receives various information IDs from company B, such as customer information, purchase information, direct message (DM) information, and other information. The information processing device 1 performs machine learning based on information corresponding to the received various information IDs, and outputs predicted information OD, such as the destination of a direct message, etc., as a learning result.

Company B sends a direct message to the user based on the prediction information OD. For example, company B sends a direct message DM1 such as product introduction to user U1. Further, company B sends a direct message DM2 such as product introduction to user U2. In contrast, it is assumed that the user U1 does not purchase a product as a response R1. On the other hand, it is assumed that the user U2 receives the direct message DM2 introducing the product as a response R2, and purchases the product.
In direct marketing, if company B can send more direct messages to users, such as user U2, who take actions such as purchasing in response to direct messages, high advertising effectiveness can be achieved.

The information processing device 1 plans a campaign that maximizes the acquisition of demand for advertising costs, and determines the total amount of advertising costs that will maximize the final profit. Further, the information processing device 1 performs prediction of conditional intervention effects for each user and prediction of future intervention effects in consideration of trends. Here, the trend includes trend components and seasonal components. Moreover, considering the trend means correcting the predicted value using the trend. The information processing device 1 predicts the destination of the direct message and the timing of sending the direct message based on the intervention effect. In other words, the information processing device 1 optimizes targeting, which is an element of decision-making in direct marketing, and extracts users who are expected to have high advertising effectiveness, that is, high affinity with the advertising content. Next, we predict the conditional intervention effect for each user. Furthermore, the information processing device 1 determines the destination of the direct message so that only users with whom it is expected to have a high affinity can receive the direct message. Therefore, in estimating the intervention effect (Uplift Modeling), which is the core process of this optimization, the information processing device 1 takes into account differences between individuals (consumer heterogeneity) and differences due to timing (temporal heterogeneity). By considering both of them together, more accurate Uplift Modeling can be performed.

Here, the accuracy of the Uplift Model when predicting the amount of Uplift due to advertisements for each individual can ideally be measured by the square error between the true Uplift value and the predicted value. However, when trying to evaluate and measure the accuracy of the Uplift Model in an actual campaign, it is impossible to directly measure the squared error etc. due to the fundamental problem of causal inference described above. AUUC (Area Under the Uplift Curve) or the like may be used to evaluate the accuracy of the Uplift Model in such an actual campaign. However, since the verification in this application uses artificial data that allows one to directly know the true Uplift value, indicators such as MSE and MAE that are used in normal evaluation of machine learning are used. In this case, the closer the predicted Uplift value is to the actual Uplift value of the artificial data, the higher the accuracy is evaluated.

Here, the customer information is data including user name, name, address, telephone number, email address, gender, age, household composition, etc. The purchase information is data on the purchase history of each user, and includes the date of purchase, genre, product name, store of purchase, and the like. The DM information is data including DM identification information, genre, request source, target, type, etc. for each direct message. Note that the customer information may be classified in advance into groups and subgroups based on customer attributes, purchase history of each customer, and the like.

Furthermore, direct messages include e-mails sent by specifying an addressee, mailings such as catalogs, letters, and postcards, and sales calls. Direct messages do not include newspaper insert flyers, anonymous direct mail, area postings, outdoor advertisements, magazine advertisements, newspaper advertisements, radio commercials, television commercials, Internet commercials, etc. that are targeted at an unspecified number of people. Moreover, a trend is a time-limited trend element (component) such as a seasonal element, a time element, or a fashionable element.

<Information processing device>
FIG. 2 is a functional block diagram showing an example of the information processing device 1 according to the present embodiment.
The information processing device 1 includes a control section 110, a storage section 130, an input section 140, and an output section 150.

The storage unit 130 is, for example, a storage device such as a hard disk drive, SSD, or memory. The storage unit 130 stores various programs to be executed by the control unit 110, such as firmware and application programs, and results of processing executed by the control unit 110.

The control unit 110 is a processor such as a central processing unit (CPU). The control unit 110 generates output information for the input information, for example, based on the input information input from the input unit 140 and the information stored in the storage unit 130.
Specifically, the control unit 110 reads purchase information, DM information, and destination information stored in the storage unit 130, and generates a learning data set based on the read information. The control unit 110 learns individual conditional intervention effects based on the generated learning data set. The learning model uses Uplift Modeling, for example. Details will be described later.

The control unit 110 estimates the conditional intervention effect (CATE) for each user, that is, for each individual, based on statistics such as customer information, purchase information, DM information, and destination information. Further, the control unit 110 models the seasonal component/trend component of an index that is considered to show a similar trend to the Uplift value, and corrects the predicted value of the Uplift value. That is, the control unit 110 estimates the conditional intervention effect for each individual and predicts the future intervention effect as a Calibration Score, thereby predicting the Adjusted CATE for each individual. The control unit 110 determines the destination of the direct message based on the predicted Adjusted CATE.

The input unit 140 is, for example, an input device such as a keyboard or a touch panel. The input unit 140 receives input information based on user operations. The input unit 140 outputs the received input information to the control unit 110.

The output unit 150 is, for example, a display unit, and displays output information input from the control unit 110.

Note that instead of or in addition to the input section 140 and the output section 150, a communication section (not shown) may accept input information from an external device and output information to the external device.

The control unit 110 will be explained in more detail.

The control unit 110 includes an acquisition unit 111, a learning unit 112, a prediction unit 113, and an output processing unit 114.
The acquisition unit 111 acquires input information via the input unit 140. Further, the acquisition unit 111 reads and acquires various information stored in the storage unit 130. The acquisition unit 111 outputs the acquired information to the learning unit 112 and the prediction unit 113.

The learning unit 112 performs machine learning based on the information acquired by the acquisition unit 111 and the information stored in the storage unit 130. Specifically, the learning unit 112 reads purchase information, DM information, and destination information stored in the storage unit 130, and generates a learning data set based on the read information. The learning unit 112 learns individual conditional intervention effects based on the generated learning data set. Here, Uplift Modeling is used as a learning model for individual conditional intervention effects.

Uplift Modeling is a model that models and predicts intervention effects for each individual in order to maximize iROAS. In this embodiment, consumer heterogeneity and temporal heterogeneity are taken into consideration as factors, and the individual Uplift value, which is the future intervention effect for each individual, is predicted without setting up a complicated model such as a state space model. .

In this embodiment, a CALIBRATED CATE Estimator is used as a model that satisfies these properties. The CALIBRATED CATE Estimator includes two models. The first model is CATE Estimator, which learns consumer heterogeneity from past data. The second model is Timeseries Forecaster, which captures changes in Uplift behavior in past data over time and indicates the level of Uplift at a desired point in the future. The control unit 110 corrects the individual Uplift value of the CATE Estimator using a Calibration Score, which is the difference in the Uplift level between the CATE Estimator's learning time and the prediction time obtained by the Timeseries Forecaster. Thereby, the control unit 110 performs prediction that takes into account temporal heterogeneity from the CATE Estimator learning time point to the prediction time point. Here, CATE refers to Conditional Average Treatment Effect, ie, conditional intervention effect, and includes temporal heterogeneity.

There are many methods for Uplift Modeling that takes into account consumer heterogeneity, although their accuracy and constraints on data vary. In this embodiment, consumer heterogeneity is estimated using X-Learner, which is a type of Meta-Learner, as a CATE Estimator. Note that the scope of application of the present invention is not limited to X-Learner, and any algorithm for estimating conditional intervention effect (CATE) can be used. Due to this property, analysts can use the amount of data set they use depending on its complexity, or select and use an algorithm with which they are familiar.

Meta Learner does not estimate the intervention effect within a single algorithm, but rather estimates the intervention effect by adding meta operations to the results of a regular machine learning model. Meta learners include T-learner, S-learner, and X-learner. T-learner estimates the intervention effect by comparing the outputs of a model learned only by a group that received the intervention and a model learned only by a group that did not receive the intervention. S-learner simultaneously learns both the target group and the treatment group, and compares the output when the intervention condition w is set to w=0 and the output when w=1. X-learner models the T-learner for each intervention group, then calculates the treatment effect for each intervention group, and calculates the CATE of the individual you want to predict by combining the CATE of both models for each group. Weight by propensity score.

A feature of Meta-Learner is that good estimation accuracy can be expected even when the amount of data between the intervention group and the non-intervention group is imbalanced. Another feature of Meta-Learner is that since the base learning device can be freely specified, an algorithm that matches the characteristics of the data can be used. The latter can be used in different ways, such as using a linear model when there is a small amount of data where linearity can be assumed, and conversely, using a highly expressive algorithm such as gradient boosting when a nonlinear relationship is expected and the amount of data is large. Can be done.

Temporal heterogeneity can be estimated using general time series data analysis, SARIMA, ETS, Prophet, etc. In this embodiment, a case will be described in which Prophet is used as the Timeseries Forecaster because it is easy to capture and interpret seasonal trends and exogenous variables, and has good accuracy.

In order to predict trends and seasonality in Timeseries Forecaster, data for a certain period of time in the past to be predicted is required. At this time, past data will include periods in which there was an intervention and periods in which there was no intervention. In CCEPTA, if an indicator on which parallel trends are assumed is affected by an intervention, the trend estimation may include changes in the indicator due to the intervention that are unrelated to the original trend. As a countermeasure, it is also possible to add the presence or absence of intervention and its scale as exogenous variables when estimating trends. However, if there are only one or two interventions, or if they are concentrated on a specific day of the week, it is difficult to accurately exclude the influence of the intervention from the original trend fluctuations. Furthermore, when estimating a trend using only data that did not receive an intervention, the average of all data during the period without the intervention is used, but during the period with the intervention, the average of the data without the intervention is used. I end up. For this reason, bias occurs due to changes in the distribution of the population, which is the basis for calculating the average, before and after the intervention.

In order to avoid this bias, Propensity Based Sampling (Equation (1)) is used for the data used for learning the Timeseries Forecaster to predict trends and seasonality. In a time window in which no intervention occurs, sampling is performed with a probability that is the probability of receiving an intervention, propensity score, e(X), which is subtracted from 1. During the intervention period, all individuals who did not receive the intervention will be sampled. Thereby, the information processing device 1 can estimate trends without bias even when the learning period includes periods with intervention and periods without intervention.

Converting the predicted value learned at one point by the CATE Estimator in a manner that takes into account temporal heterogeneity can be considered as a problem of Concept Shift, which is a type of Dataset Shift. Concept Shift is defined as Equation (2), where the distribution of y (target) when conditioned by x (covariate) and the distribution of x when conditioned by y change between learning and prediction. This is the situation.

When a Dataset Shift occurs, the model is generally retrained frequently or the most recent training samples are given stronger weight for learning. In this embodiment, by predicting the level of Ptest(y|x) at a future time, Ptraining(y|x) is converted to calculate a score z that matches Ptest(y|x) Addressing Concept Shift issues when considering temporal heterogeneity.

Note that in equation (3), Calibration Score is used to correct the prediction result, but it is also possible to learn a value corrected by Calibration Score when learning past data.

In this case, x _ts used to calculate the score z and x _cate used to predict Y are different. It is desirable that the x _cate for predicting detailed uplift values (effects) for each individual basically include all feature quantities considered to be effective, as in a normal machine learning process. If Unconfoundedness is a premise for CATE calculation, it is desirable that the range of feature values be wide in order to ensure that there are no unobserved confounding factors.

On the other hand, x _ts when calculating the Calibration Score (score z) is only a grouping unit that indicates "in which unit the time series trends differ." Here, even if x _ts is made smaller than necessary, it will not contribute to the estimation accuracy of the score z and may cause overfitting. As an example of x _ts , in the case of travel-related marketing, it is also possible to include a variable such as "presence or absence of children in elementary and junior high school" as a grouping with different temporal heterogeneity with respect to summer vacation and winter vacation.

To estimate the temporal heterogeneity used in this embodiment, due to the constraints of available data, we use CALIBRATED CATE Estimator with Uplift Trend (CCEUT) and CALIBRATED CATE Estimator with Parallel Trend Assump. tion (CCEPTA).

CCEUT is a method of calculating a Calibration Score by treating the transition of the Uplift value itself as time series data and modeling the trend. CCEUT is applicable when data sufficient for estimating the Uplift value is available over multiple periods to some extent.

On the other hand, CCEPTA assumes that an observable indicator other than the Uplift value itself (such as the sales level of the Control group) has a parallel trend with the Uplift value trend, and models the trend using the changes in the indicator as time series data. This is a method of calculating the Calibration Score by

CCEUT will be explained in more detail.
FIG. 3 is a schematic diagram showing an example of a CCEUT used in this embodiment.
CCEUT is a method for calculating a Calibration Score, which is a correction value for the individual conditional intervention effect (CATE) estimated by Meta Learner used as a CATE Estimator. CCEUT predicts individual Calibration Scores, that is, future intervention effects, by modeling Uplift seasonal components and trend components from Uplift results of past campaigns. Adjusted CATE is obtained by multiplying the CATE estimated by Meta Learner by the Clibration Score calculated by CCEUT.

Next, CCEPTA will be explained in more detail.
FIG. 4 is a schematic diagram showing an example of CCEPTA used in this embodiment.
CCEPTA is a method for calculating a Calibration Score, which is a correction value for the individual conditional intervention effect (CATE) estimated by Meta Learner used as a CATE Estimator. CCEPTA predicts an individual's Calibration Score, that is, the future intervention effect, by modeling the seasonal component and trend component of an index that is considered to show a similar trend to the Uplift value. Adjusted CATE is obtained by multiplying the CATE estimated by Meta Learner by the Citation Score calculated by CCEPTA.

CCEPTA can be applied even when there is not enough past data to treat the transition of the Uplift value itself as time series data. The assumption of parallel trends can be a seemingly strong constraint. However, when considering the promotional effects of a product, for example, the effect will be limited even if the promotion is carried out at a time when there is no demand at all. On the other hand, if the promotion is carried out at a time when there is a certain level of demand, a larger uplift effect can be expected.

Considering such a case, for example, it is expected that the raw demand trend excluding promotions will have a positive correlation to some extent with the trend of promotion effects. For practitioners with domain knowledge, finding observable indicators with favorable characteristics for such CCEPTA is easier than iteratively testing the effectiveness of campaigns and accumulating data for CCEUT. . However, as an exception, in a situation where the raw demand trend has increased so much that there is no longer any room for uplift through promotions, there is a possibility that the raw demand trend and the promotion effect will have a negative correlation. In this way, there are situations in which it is difficult to assume parallel trends under certain conditions, so it is necessary to carefully judge whether the assumption of parallel trends really holds true.
In this embodiment, an example in which CCEPTA is used will be described.

Returning to FIG. 2, the learning unit 112 performs machine learning based on customer information, purchase information, DM information, and destination information to learn conditional intervention effects for each individual. The learning unit 112 causes the storage unit 130 to store the learning results. Further, the learning unit 112 models Uplift by learning seasonal components and trend components from Uplift results of past campaigns, and learns individual Calibration Scores.

The prediction unit 113 calculates Adjusted CATE based on the learning result by the learning unit 112 using the individual CATE and the Calibration Score. The prediction unit 113 determines the destination of the direct message based on the Adjusted CATE, and stores it in the storage unit 130.

The output processing unit 114 reads out destination information stored in the storage unit 130 and causes the output unit 150 to output the destination information.

Next, the storage unit 130 will be explained in more detail.

The storage unit 130 includes a customer information storage unit 131, a purchase information storage unit 132, a DM information storage unit 133, a destination information storage unit 134, and a learning result storage unit 135.
The customer information storage unit 131 stores customer information. The customer information is information in which identification information, user name, name, address, telephone number, email address, gender, age, household structure, etc. are associated. This will be explained with reference to FIG.

FIG. 5 is a table showing an example of the structure of customer information stored in the customer information storage unit 131.
In the illustrated example, the identification information is associated with a user name, name, address, telephone number, email address, gender, age, and household structure.
Identification information is information for identifying a user. The user name is the user's registered name. The name is the user's name. The address is the user's address. The telephone number is the user's telephone number. The email address is the user's email address. Gender and age are the user's gender and age. The household composition is the household composition of the household to which the user belongs.
Information on each of these items is, for example, information registered in advance by each user.

Returning to FIG. 2, the purchase information storage unit 132 stores purchase information. Purchase information is information indicating the purchase history of each customer, and is information in which information such as purchase date, genre, product name, and store of purchase are associated with each other. This will be explained with reference to FIG.

FIG. 6 is a table showing an example of the configuration of purchase information stored in the purchase information storage unit 132.
In the illustrated example, each piece of identification information that identifies a user is associated with a purchase date, genre, product name, and store of purchase.
The purchase date is the date on which the product was purchased. The genre is the genre of the product. The product name is the name of the product. The purchase store is the store where the product was purchased.
Information on each of these items may be registered in advance by each user, or may be created based on payment history information or the like.

Returning to FIG. 2, the DM information storage unit 133 stores DM information. DM information is direct message information, and is information in which DM identification information, genre, request source, target, type, and other information are associated with each other. This will be explained with reference to FIG.

FIG. 7 is a table showing an example of the structure of DM information stored in the DM information storage section 133.
In the illustrated example, genre, request source, target, and type are associated with DM identification information.
DM identification information is identification information that identifies a direct message. The genre is the genre to which the direct message is related. The request source is a request source that requests sending of a direct message. The target is the target to whom the direct message is sent. The type is the direct message sending method.
Information on each of these items is created based on a request from the requester.

Returning to FIG. 2, destination information storage section 134 stores destination information. The destination information is information in which identification information of a user who is a destination is associated with each DM identification information. This will be explained with reference to FIG.

FIG. 8 is a table showing an example of the structure of the destination information stored in the destination information storage unit 134.
In the illustrated example, each piece of DM identification information is associated with the identification information of the user who is the destination.
The destination information is generated by being determined by the prediction unit 113 as a destination candidate for the direct message.

Returning to FIG. 2, the learning result storage unit 135 stores the learning results by the learning unit 112.

<Hardware configuration>
Next, the hardware configuration of the information processing device 1 will be explained. FIG. 9 is a block diagram showing an example of the hardware configuration of the information processing device 1 according to the present embodiment.
The information processing device 1 includes a CPU 11, a drive section 12, a storage medium 13, an input section 14, an output section 15, a ROM 16 (Read Only Memory), a RAM 17 (Random Access Memory), and an auxiliary storage section 18. , and an interface section 19. The CPU 11, the drive section 12, the input section 14, the output section 15, the ROM 16, the RAM 17, the auxiliary storage section 18, and the interface section 19 are interconnected via a bus.

Note that the CPU 11 here refers to a processor in general, and includes not only a device called a CPU in a narrow sense, but also includes, for example, a GPU, a DSP, and the like. Furthermore, the CPU 11 referred to herein is not limited to being implemented by one processor, but may be implemented by combining multiple processors of the same or different types.

The CPU 11 reads and executes programs stored in the auxiliary storage unit 18, ROM 16, and RAM 17, reads out various data stored in the auxiliary storage unit 18, ROM 16, and RAM 17, and writes various data to the auxiliary storage unit 18 and RAM 17. By writing , the information processing device 1 is controlled. Further, the CPU 11 reads various data stored in the storage medium 13 via the drive section 12, and writes various data to the storage medium 13. The storage medium 13 is a portable storage medium such as a magneto-optical disk, a flexible disk, or a flash memory, and stores various data.

The drive section 12 is a reading device for a storage medium 13 such as an optical disk drive or a flexible disk drive.

The input unit 14 is an input device such as a mouse, a keyboard, a touch panel, a power button, and a setting button.

The output unit 15 is an output device such as a display unit or a speaker.

The ROM 16 and the RAM 17 store programs and various data for operating each functional section of the display device 10.

The auxiliary storage unit 18 is a hard disk drive, a flash memory, or the like, and stores programs and various data for operating each functional unit of the information processing device 1.

The interface unit 19 has a communication interface and is connected to the network NW by wire or wirelessly.

For example, the control unit 110 in the functional configuration of the information processing device 1 in FIG. 2, which will be described later, corresponds to the CPU 11 in FIG. Further, the input section 140 in FIG. 2 corresponds to the input section 14 in FIG. 9. Further, the output section 150 in FIG. 2 corresponds to the output section 15 in FIG. 9. Furthermore, the storage unit 130 in FIG. 2 corresponds to the storage medium 13 in FIG. 9.

<About operation>
FIG. 10 is a flowchart illustrating an example of the operation of the information processing device 1 according to this embodiment. This figure shows the operation of the information processing device 1 in the learning stage.

(Step S100) The acquisition unit 111 reads customer information, DM information, purchase information, destination information, and learning results from the storage unit 130. Upon completion of the process, the acquisition unit 111 advances the process to step S102.
(Step S102) The acquisition unit 111 creates a learning data set based on each piece of read information, and stores it in the storage unit 130. When the acquisition unit 111 completes the process, it advances the process to step S104.
(Step S104) The learning unit 112 reads the learning data set from the storage unit 130 and performs learning based on the read information. When the learning is completed, the learning unit 112 causes the storage unit 130 to store the learning result. After that, the operation shown in this figure ends.

FIG. 11 is a flowchart showing another example of the operation of the information processing device 1 according to this embodiment. This figure shows the operation of the information processing device 1 in the CATE and Calibration Score estimation (calculation) stages.

(Step S200) The learning unit 112 acquires customer information from the customer information storage unit 131. DM information is acquired from the DM information storage unit 133. Upon completion of the process, the learning unit 112 advances the process to step S202. Steps S202 to S204 are performed for each individual (for each user).
(Step S202) The learning unit 112 acquires purchase information corresponding to the customer identification information included in the customer information from the purchase information storage unit 132. Upon completion of the process, the learning unit 112 advances the process to step S204.

(Step S204) The learning unit 112 predicts CATE and Calibration Score based on the learning result and purchase history. The learning unit 112 causes the storage unit 130 to store the learning results and prediction results. When the learning unit 112 completes the processing for all users, the learning unit 112 ends the operation shown in this figure.

FIG. 12 is a flowchart showing another example of the operation of the information processing device 1 according to this embodiment.
(Step S300) The prediction unit 113 acquires customer identification information included in the customer information from the customer information storage unit 131. The prediction unit 113 acquires purchase information corresponding to the customer's identification information from the purchase information storage unit 132. The prediction unit 113 acquires DM information from the DM information storage unit 133. Upon completion of the process, the prediction unit 113 advances the process to step S302.
(Step S302) The prediction unit 113 calculates CATE for each customer based on the DM information and the purchase information. Upon completion of the process, the prediction unit 113 advances the process to step S304.

(Step S304) The prediction unit 113 calculates a Calibration Score for each customer based on the DM information and the purchase information. Upon completion of the process, the prediction unit 113 advances the process to step S306.
(Step S306) The prediction unit 113 calculates Adjusted CATE based on CATE and Calibration Score. When the processing is completed, the information processing device 1 advances the processing to step S308.

(Step S308) The prediction unit 113 predicts the destination based on the identification information included in the customer information and the calculated Adjusted CATE. When the processing is completed, the information processing device 1 advances the processing to step S310.

(Step S310) The output processing unit 114 acquires destination information from the destination information storage unit 134 and generates an output image. The output processing unit 114 causes the output unit 150 to output the generated image. When the processing is completed, the information processing device 1 ends the processing shown in this figure.

In this way, the information processing device 1 according to the present embodiment includes the acquisition unit 111 that acquires customer information and purchase information, data indicating conditional intervention effects for each user due to campaign intervention, and future intervention effects due to campaign intervention. By inputting the customer information and purchase information acquired by the acquisition unit 111 into a learning model that has been trained using the data indicating the learning data as a learning data set, the predicted value of the conditional intervention effect for each user and the predicted value of the future intervention effect can be calculated. The learning unit 112 includes a learning unit 112 that calculates a predicted value, and a prediction unit 113 that corrects a predicted value of a future intervention effect using a trend.

Thereby, the information processing device 1 can optimize targeting, which is one element of decision-making in direct marketing, by correcting the predicted value of future intervention effects using trends. Therefore, based on the predicted value of the conditional intervention effect for each user and the predicted value of the future intervention effect, it is possible to extract people who are expected to have a high advertising effect, that is, a high affinity with the content of the advertisement. can. Furthermore, the prediction accuracy of intervention effects in direct marketing can be improved.

Furthermore, the information processing device 1 according to the present embodiment stores data indicating a conditional intervention effect for each user in a predetermined period including at least a period in which the campaign intervenes and a period in which the campaign does not intervene, and data indicating the future intervention effect. A learning unit 112 calculates a predicted value of the future intervention effect using a learning model that learns as a learning data set, and a learning unit 112 that calculates the predicted value of the future intervention effect by suppressing the influence of data indicating the conditional intervention effect for each user during the period in which the campaign does not intervene. and a prediction unit 113 that calculates a trend for correcting the predicted value of the intervention effect.

As a result, even if the learning period includes a period in which there is a campaign intervention and a period in which there is no campaign intervention, the information processing device 1 can calculate the distribution of the population, which is the premise for calculating the average before and after the intervention, even if there is a mixture of periods with campaign intervention and periods without campaign intervention in the learning period. can suppress change. Therefore, it is possible to estimate trends without bias due to changes in population distribution. Therefore, it is possible to improve the prediction accuracy of intervention effects in direct marketing.

Next, the verification results for the contents explained above will be explained.

The prerequisites for using the Calibrated CATE Estimator are that the number of individuals is sufficient to estimate heterogeneity, the panel data format is available over a period long enough to perform time-series prediction, and the accuracy of the estimated Uplift value is For evaluation, it is necessary to use RCTs or artificial data with measurable counterfactual outcomes. However, there is no data that has been conducted under such ideal conditions and that has been made available for verification. Therefore, in this verification, we will use artificial data.

On the other hand, even if the accuracy of time-series prediction using artificial data generated under ideal conditions is high, it is not necessarily possible to capture changes in complex trends in actual marketing. For this reason, in this verification, we used response data to actual marketing campaigns as the base for trend components, and added artificially generated individual-level heterogeneity to it, creating a hybrid panel of real data and artificial data. Use the data as an analysis target. This makes it possible to verify the Uplift value at the individual level, which is an advantage of artificial data in that counterfactual outcomes can be observed, and to verify the prediction of time series components under conditions that are to some extent equivalent to real data.

Geoexperiments Research version 1.0.3 is used as the base data for the time series components. This will be explained with reference to FIG. FIG. 13 is a diagram showing an example of verification data according to this embodiment. In this verification, 100 regions were randomly assigned to a treatment group and a control group so that the total sales were approximately equal, and daily sales in each region were observed in three periods: pretest, intervention, and cooldown. It is assumed that online advertisements are distributed only during the intervention period of the treatment group, and there is no intervention for other targets or periods. The data is panel data from an RCT over a relatively long period from January 1, 2015 to April 7, 2015. This data is close to ideal data in this verification, but since there are only 100 regions, it is not a sufficient number of samples to model heterogeneity. During the Pretest period, the Sales movements of the Treatment group and Control group matched with extremely high accuracy, but during the Intervention period starting from February 15th, the Sales of the Treatment group increased, and during the Cooldown period, the Sales movements of the Control group and Return to parity.

Furthermore, while the level of Sales has a periodicity that appears to be weekly, it shows a gradual upward trend towards the latter half of the data collection period. In time-series data, data in which the mean and variance of observed data change over time is called non-stationary data, and independent and simultaneous distribution (i.i.d.) is assumed as a premise for analysis. It is difficult to analyze using normal machine learning. For this reason, before analysis, it is often transformed to provide stationarity using difference transformation, logarithmic difference transformation, or the like.

With only the data of Geoexperiments Research version 1.0.3, there are only 100 regions (100 types) of individuals that can be used for estimation of heterogeneity. Here, data in which 10,000 customers randomly belong to one of 100 regions and each customer has a randomly generated covariate is generated using equations (4) and (5). The distribution of each generated covariate is shown in FIGS. 14 to 16. FIGS. 14 to 16 are diagrams showing examples of distributions of each covariate according to this embodiment.

Each customer has potential outcomes Y ₍₀₎ and Y ₍₁₎ , respectively, and what is actually observed is set to Y _(w) = Y, which matches the intervention condition. The latent outcomes are generated as shown in Equation (6), Equation (7), and Equation (8), respectively. Note that base _it is the value of Sales at time t of the group to which individual i belongs in the data in Figure 13, and w, which expresses temporal heterogeneity, indicates the presence or absence of the intervention that x _it actually received. represent.

The relationship between each covariate and the latent outcome is shown in a DAG (Directed Acyclic Graph) as shown in FIG. 17. FIG. 17 is a diagram showing an example of the relationship between each covariate and a potential outcome according to this embodiment.

FIG. 18 shows a time series plot of the true intervention effect τ _it for each individual generated from this DAG. FIG. 18 is a diagram showing an example of a time-series plot of the true intervention effect for each individual according to the present embodiment.
Each individual has a different level of intervention effect (consumer heterogeneity), and there is a variety of temporal heterogeneity, ranging from individuals whose intervention effects vary greatly depending on the day of the week to individuals who have almost no intervention effect.

The data specifications of the learning/prediction period used in this verification are as shown in FIG. FIG. 19 is a diagram showing an example of data specifications of the learning/prediction period used in the verification according to this embodiment.
Regarding the intervention, we verified two cases: one in which each individual was assigned to each region in an RCT, and the other in which x _iu was used as the probability of receiving the intervention (propensity score).

Calibrated CATE Estimator is a meta-learning algorithm, and any algorithm can be used for each of CATE Estimator and Timeseries Forecaster.
In this verification, as in the embodiment, X-Learner, which is a type of Meta Learner, is used as a CATE Estimator. Lightgbm version 3.2.1 developed by Microsoft (registered trademark) was used for the X-Learner outcome model and CATE model.

Further, in this verification, version 0.7.1 of Prophet developed by Facebook Inc. (registered trademark) was used as Timeseries Forecaster. Note that although the data used for learning Prophet was obtained after propensity-based sampling, the effectiveness of propensity-based sampling itself will also be verified.

Evaluation in the verification was performed by learning and predicting in the following three patterns, and using the following three indicators to compare the predicted intervention effect with the true value of the intervention effect.

(1) Mean Absolute Error (MAE)
MAE is expressed by equation (9). Since MAE is the average of the absolute values of errors, it is an index that can be easily interpreted as accuracy.

(2) Mean Squared Error (MSE)
MSE is expressed by equation (10). Since MSE is the average of squared errors, it is an index that is more susceptible to outliers than MAE. In this verification, we use it as an auxiliary indicator to check whether the number of cases in which we missed the mark is extremely large compared to other trials.

(3) Mean Error (ME)
ME is expressed by equation (11). ME is not usually used much, but in this case, we are evaluating whether the bias caused by temporal heterogeneity can be reduced by Timeseries Forecaster, and we are determining how much bias remains in the predicted values. Used as an indicator for If the error follows a normal distribution, it can be interpreted that the closer it is to 0, the more unbiased the prediction is.

First, the effectiveness of Propensity Based Sampling will be verified. Data sampling and Calibration Score calculation were performed using the following four patterns.

(1) Ground truth
The Ground Truth is not the actual observed value Y _it but is aggregation of data Y _{(0) it} when no intervention occurs for all individuals, which is available only as simulation data.

(2) Propensity Based Sampling
Propensity Based Sampling is data sampled in the same manner as in the embodiment.

(3) All Samples
All Samples is an extraction of all data regardless of whether the intervention was received or not.

(4) Untreated Samples
Untreated Samples is data extracted from all Y _it in which there was no intervention, regardless of the period with or without intervention.

Data sampled under each of Policies (1) to (4) was visualized. FIG. 20 is a diagram showing an example of data sampled by each policy according to the present embodiment.
In the illustrated example, All Samples completely matches the Ground Truth during the non-intervention period, but deviates upward compared to the Ground Truth when entering the intervention period. This is simply due to the effect of the intervention, which causes the observed numbers to be stronger than the actual trend.
On the other hand, although Untreated Samples completely agrees with the Ground Truth during the non-intervention period, it deviates downward from the Ground Truth when entering the intervention period. This is because in the current verification data, individuals with higher sales expectations tend to be more likely to receive intervention, so the average sales of individuals who did not receive intervention during the intervention period was the average of all samples before the intervention. This is a bias that arises from falling below the sales average.

For the same reason, the data of Propensity Based Sampling deviates downward from the Ground Truth in both the non-intervention period and the intervention period. However, unlike the above, there is no change in the trend even when the intervention period begins, and a stable trend can be extracted from the visualization results.

FIG. 21 is a diagram showing an example of trends in each policy according to the present embodiment.
As shown in FIG. 21, the trends extracted using the data in (3) and (4) deviate upward and downward compared to trend 1 extracted from the true data. The trend extracted from the data in (2) also constitutes a trend parallel to the true trend.

FIG. 22 is a diagram illustrating an example of sampling verification results according to this embodiment.
As mentioned above, when calculating the Calibration Score in the learning period and prediction period based on the extracted trends, Propensity Based Sampling is closest to the Ground Truth, with an error of 0.0017, which is 10 minutes faster than other methods. An error of less than 1 was achieved.

From this, Propensity Based Sampling is more effective in estimating trends in situations where intervention periods and non-intervention periods intersect, compared to other samplings (1), (3), and (4) that were compared this time. This is a great method.

Next, the intervention effect for each individual in the prediction period was estimated using six patterns of combinations of the following two verification assumptions and three calibrations, and the results were compared with positive values.

Verification assumptions (1) RTC data with completely random intervention allocation (2) Quasi-experimental data where only the allocation probability for each individual is known Calibration method (1) Calibrate the predicted results (Post-Train)
(2) Calibrate the learning data (Pre-Train)
(3) No Calibration

FIG. 23 is a diagram showing an example of a verification result of Calibration according to this embodiment.
As shown in the figure, Calibrated CATE Estimator has improved accuracy compared to the normal model No Calibration that does not perform calibration and Pre-train that calibrates learning data in both RCT data and quasi-experimental data under the verification premise. It is shown that

In this way, as a result of verifying the effectiveness of the content explained in the above embodiment using artificial data, it was found that consumer heterogeneity can be estimated with high accuracy just by specifying feature values and training/prediction data. While retaining the advantage of machine learning-based methods that it is possible, we have demonstrated that bias in predicted values caused by non-stationarity of the objective variable can be suppressed compared to normal machine learning-based methods, and that highly accurate predictions can be achieved. confirmed.

Although one embodiment of the present invention has been described above in detail with reference to the drawings, the specific configuration is not limited to that described above, and various design changes etc. may be made without departing from the gist of the present invention. It is possible to

For example, in each of the embodiments described above, an example in which the information processing device 1 learns and predicts is explained, but the present invention is not limited to this. For example, in the information processing device 1, a learning device that performs learning and a prediction device that performs prediction may be separate devices.

Furthermore, a part of the information processing device 1 in each of the embodiments described above, for example, the control unit 110, may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Note that the "computer system" herein refers to a computer system built into the information processing device 1, and includes an OS (Operating System) and hardware such as peripheral devices.

Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into computer systems. Furthermore, a "computer-readable recording medium" refers to a medium that dynamically stores a program for a short period of time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, it may also include a device that retains a program for a certain period of time, such as a volatile memory inside a computer system that is a server device or a client. Further, the above-mentioned program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

Further, part or all of the information processing device 1 in the embodiment described above may be realized as an integrated circuit such as an LSI (Large Scale Integration). Each functional unit of the server device 20 (20a) may be made into a processor individually, or some or all of them may be integrated into a processor. Further, the method of circuit integration is not limited to LSI, but may be implemented using a dedicated circuit or a general-purpose processor. Further, if an integrated circuit technology that replaces LSI emerges due to advances in semiconductor technology, an integrated circuit based on this technology may be used.

Sys Information processing system 1 Information processing device B Company U1, U2 User 110 Control unit 111 Acquisition unit 112 Learning unit 113 Prediction unit 114 Output processing unit 130 Storage unit 131 Customer information storage unit 132 Purchase information storage unit 133 DM information storage unit 134 Sending Ahead information storage unit 135 Learning result storage unit 140 Input unit 150 Output unit 11 CPU
12 Drive section 13 Storage medium 14 Input section 15 Output section 16 ROM
17 RAM
18 Auxiliary storage section 19 Interface section

Claims (11)

an acquisition unit that acquires customer information and purchasing information;
The customer information and purchase information acquired by the acquisition unit are applied to a learning model that has been trained using data indicating a conditional intervention effect for each user due to campaign intervention and data indicating future intervention effects due to campaign intervention as a learning data set. a learning unit that calculates a predicted value of a conditional intervention effect and a predicted value of a future intervention effect for each user by inputting the information;
a prediction unit that corrects the predicted value of the future intervention effect using a trend;
An information processing device comprising: A learning model that learns data indicating the conditional intervention effect for each user during a predetermined period including at least the period in which the campaign intervened and the period in which the campaign did not intervene, and data indicating the future intervention effect as a learning dataset, a learning department that calculates the predicted value of the intervention effect;
a prediction unit that calculates a trend for correcting the predicted value of the future intervention effect by suppressing the influence of data indicating the conditional intervention effect for each user during the period in which the campaign is not intervened;
An information processing device comprising: The prediction unit changes the number of samples of data indicating the conditional intervention effect for each user in the period in which the campaign is not intervened, and the number of samples of data indicating the conditional intervention effect for each user in the period in which the campaign intervenes. to suppress the said influence,
The information processing device according to claim 2. The trend is a predicted value obtained by modeling similar results,
The information processing device according to claim 1 or claim 2. The trend is a predicted value obtained by modeling indicators showing similar trends;
The information processing device according to claim 1 or claim 2. an acquisition unit that acquires customer information and purchasing information;
The customer information and purchase information acquired by the acquisition unit are applied to a learning model that has been trained using data indicating a conditional intervention effect for each user due to campaign intervention and data indicating future intervention effects due to campaign intervention as a learning data set. a learning unit that calculates a predicted value of a conditional intervention effect and a predicted value of a future intervention effect for each user by inputting the information;
a prediction unit that corrects the predicted value of the future intervention effect using a trend;
An information processing system equipped with. A learning model that learns data indicating the conditional intervention effect for each user during a predetermined period including at least the period in which the campaign intervened and the period in which the campaign did not intervene, and data indicating the future intervention effect as a learning dataset, a learning department that calculates the predicted value of the intervention effect;
a prediction unit that calculates a trend for correcting the predicted value of the future intervention effect by suppressing the influence of data indicating the conditional intervention effect for each user during the period in which the campaign is not intervened;
An information processing system equipped with. The computer of the information processing device is
an acquisition process for acquiring customer information and purchasing information;
The customer information and purchase information acquired through the acquisition process are added to a learning model that has been trained using data indicating the conditional intervention effect for each user due to the campaign intervention and data indicating the future intervention effect due to the campaign intervention as a learning data set. A learning process that calculates a predicted value of a conditional intervention effect and a predicted value of a future intervention effect for each user by inputting
a prediction process of correcting the predicted value of the future intervention effect using a trend;
An information processing method having The computer of the information processing device is
A learning model that learns data indicating the conditional intervention effect for each user during a predetermined period including at least the period in which the campaign intervened and the period in which the campaign did not intervene, and data indicating the future intervention effect as a learning dataset, a learning process that calculates a predicted value of the intervention effect;
a prediction step of calculating a trend for correcting the predicted value of the future intervention effect by suppressing the influence of data indicating the conditional intervention effect for each user during the period in which the campaign does not intervene;
An information processing method having In the computer of the information processing device,
an acquisition step of acquiring customer information and purchasing information;
The customer information and purchase information obtained in the acquisition step are added to the learning model that has been trained using data indicating the conditional intervention effect for each user due to the campaign intervention and data indicating the future intervention effect due to the campaign intervention as a learning data set. a learning step of calculating a predicted value of a conditional intervention effect and a predicted value of a future intervention effect for each user by inputting the information;
a prediction step of correcting the predicted value of the future intervention effect using a trend;
A program to run. In the computer of the information processing device,
A learning model that learns data indicating the conditional intervention effect for each user during a predetermined period including at least the period in which the campaign intervened and the period in which the campaign did not intervene, and data indicating the future intervention effect as a learning dataset, a learning step of calculating a predicted value of the intervention effect;
a prediction step of calculating a trend for correcting the predicted value of the future intervention effect by suppressing the influence of data indicating the conditional intervention effect for each user during the period in which the campaign is not intervened;
A program to run.

Images