WO2025041235A1 - Optimal intervention presentation device, optimal intervention presentation method, and program - Google Patents

Abstract

An optimal intervention presentation device according to one aspect of the present disclosure comprises: a static characteristic acquisition unit that acquires static characteristic information representing a static characteristic that is a personal characteristic having a low frequency of occurrence of a change among personal characteristics of a user; a dynamic characteristic estimation unit that, by using information on action records of the user, estimates dynamic characteristic information representing a dynamic characteristic that is a personal characteristic having a high frequency of occurrence of a change compared to the static characteristic among the personal characteristics; an intervention effect prediction unit that, by using the static characteristic information and the dynamic characteristic information, predicts an intervention effect representing an effect expected when and whenever a predetermined intervention is presented to the user; a selection unit that selects an intervention with the highest intervention effect as an optimal intervention; and an intervention presentation unit that presents the optimal intervention to the user.

Description

OPTIMAL INTERVENTION PRESENTATION DEVICE, OPTIMAL INTERVENTION PRESENTATION METHOD, AND PROGRAM

This disclosure relates to an optimal intervention presentation device, an optimal intervention presentation method, and a program.

Technologies that support behavioral change in users by presenting some kind of intervention to the user have been known for some time (e.g., Non-Patent Document 1, Non-Patent Document 2, Patent Document 1, etc.). With these technologies, it is important to select and present an appropriate intervention for each target user (hereinafter also referred to as the target) in order to realize behavioral change in the target user. On the other hand, to select an appropriate intervention for each target, not only information directly related to the target's goals, such as the target's behavioral performance, but also information on the target's personal characteristics (e.g., social characteristics such as home environment, living environment, etc., psychological characteristics such as motivation for behavioral change, etc.) is required.

JP 2021-86282 A

Asuken | Asuken Diet - A nutritionist will support you with your diet for free, on the Internet <URL: https://www.asken.jp/> Chernozhukov, Victor, et al. "Double/debiased machine learning for treatment and structural parameters." (2018): C1-C68.

However, some subjects' personal characteristics (e.g., psychological characteristics such as motivation to change behavior) may change frequently. For this reason, if the user is required to input changes in personal characteristics each time in order to select an appropriate intervention, a problem arises in that it places a heavy burden on the user.

The present disclosure has been made in consideration of the above points, and provides technology that can reduce the burden on users required to present optimal interventions to them.

The optimal intervention presentation device according to one aspect of the present disclosure includes a static characteristic acquisition unit that acquires static characteristic information representing static characteristics, which are personal characteristics that change less frequently among the personal characteristics of a user; a dynamic characteristic estimation unit that uses information on the behavioral track record of the user to estimate dynamic characteristic information representing dynamic characteristics, which are personal characteristics that change more frequently than the static characteristics among the personal characteristics; an intervention effect prediction unit that uses the static characteristic information and the dynamic characteristic information to predict, for each predetermined intervention, an intervention effect that represents an expected effect when the intervention is presented to the user; a selection unit that selects the intervention that maximizes the intervention effect as an optimal intervention; and an intervention presentation unit that presents the optimal intervention to the user.

Technology is provided that can reduce the burden on users required to present optimal interventions to them.

FIG. 1 is a diagram illustrating an example of a causal relationship in behavioral change. FIG. 13 is a diagram showing an example of intervention presentation using static and dynamic characteristics. 1 is a diagram illustrating an example of the overall configuration of a behavior change support system according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a hardware configuration of the optimal intervention presentation device according to the present embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of the optimal intervention presentation device according to the present embodiment. 13 is a flowchart illustrating an example of an optimal intervention presentation process according to the present embodiment. FIG. 13 is a diagram illustrating an example of dynamic characteristic estimation. FIG. 13 is a diagram illustrating an example of an effect of a change in dynamic characteristics on a behavioral performance.

Below, one embodiment of the present invention will be described.

<Causal relationships in behavior change>
First, a causal relationship in behavioral change that is the premise of the following embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of a causal relationship in behavioral change.

When selecting an intervention to present to a subject, it is necessary to select the most suitable intervention by considering the subject's personal characteristics (Figure 1 (1)). Here, the most suitable intervention is the intervention that is expected to produce the most effective behavioral results to achieve a certain goal in line with the purpose of behavioral change, depending on the purpose of the behavioral change. Specific examples of the purpose of behavioral change include "health improvement," "dieting," and "passing the entrance exam for the school of one's choice." Specific examples of goals include "weight 70 kg" when the purpose of behavioral change is "health improvement" or "dieting," and "passing the entrance exam for University A" when the purpose of behavioral change is "passing the entrance exam for the school of one's choice." Specific examples of behavioral change include "walking" or "eating a healthy diet" when the purpose is "health improvement" or "dieting," and "studying" when the purpose is "passing the entrance exam for the school of one's choice." On the other hand, examples of personal characteristics include social characteristics such as family environment and living environment, and psychological characteristics such as motivation for behavioral change. Other examples of personal characteristics include occupation, personality, health check results, and mock exam results. However, the individual characteristics that are taken into account when selecting the most appropriate intervention will vary depending on the goal of behavior change.

When an intervention is presented to a subject, the intervention may result in a change in behavioral performance (Figure 1 (2)). For example, if the purpose of behavioral change is "improving health," an intervention that sends a message such as "Let's walk today to stay healthy!" may result in a change in the user's behavioral performance, such as the number of steps taken, and the number of steps may increase.

On the other hand, personal characteristics change due to various factors such as changes in the subject's environment and the passage of time, and these changes may also result in changes in behavioral performance (Figure 1 (3)). For example, if the purpose of behavioral change is "health improvement," when the subject's work style changes from "working at the office" to "working from home," the user's behavioral performance, such as the number of steps taken, may also change, and the number of steps may decrease. Similarly, for example, if the subject's motivation decreases over time, the number of steps may decrease, and conversely, if motivation increases, the number of steps may increase.

In this way, the subject's behavioral performance can change depending on the intervention and changes in personal characteristics. Furthermore, when selecting the optimal intervention, the subject's personal characteristics must be taken into consideration. On the other hand, because personal characteristics can change due to various factors such as changes in the subject's environment and the passage of time, in order to select the optimal intervention, it is necessary to have the user input the changes in personal characteristics. However, there are also personal characteristics that can change frequently, such as changes in motivation. For this reason, if the user is required to input changes in personal characteristics each time, this places a heavy burden on the user.

<Proposed method>
Hereinafter, a method (hereinafter, also referred to as the proposed method) capable of reducing the user load required to present the optimal intervention to the user is proposed. In this proposed method, in order to reduce the user load caused by having the user (subject) input the changes in the personal characteristics, the personal characteristics are classified into two types, namely, "static characteristics" and "dynamic characteristics". Here, the static characteristics are personal characteristics that change infrequently. On the other hand, the dynamic characteristics are personal characteristics other than the static characteristics, and are personal characteristics that change relatively frequently. Specific examples of static characteristics include social characteristics such as the family environment and living environment, occupation, personality, results of health checkups, and mock exam results. On the other hand, specific examples of dynamic characteristics include psychological characteristics such as motivation for behavioral change, and the schedule of the subject.

Intervention presentation using static and dynamic characteristics will be described with reference to Figure 2. Figure 2 shows an example of intervention presentation using static and dynamic characteristics.

First, in the proposed method, the static characteristics of the subject are acquired (S1). Next, in the proposed method, at each predetermined time t, the dynamic characteristics are estimated from the subject's behavioral record (S2), and the optimal intervention is selected taking into consideration the static and dynamic characteristics and presented to the subject (S3). At this time, in the proposed method, the optimal intervention is selected using the following formula (1).

　ここで、Ｉは最適な介入、Ｌ＝｛Ｉ_１，・・，Ｉ_Ｎ｝は介入集合、Ｎは介入の総数、Ｉ_ｎ（１≦ｎ≦Ｎ）はｎ番目の介入、ｉは対象者を一意に識別するインデックス、Ｒ_ｉｔは対象者ｉの時刻ｔにおける行動実績である。ｏ_ｉ ^ｓは対象者ｉの静的特性に関する情報（以下、静的特性情報ともいう。）を表すベクトル形式のデータ、ｏ_ｉ ^ｄは対象者ｉの動的特性に関する情報（以下、動的特性情報ともいう。）を表すベクトル形式のデータである。また、ｄｏ（Ｉ＝Ｉ_ｎ）は介入Ｉ_ｎを提示すること、ｄｏ（Ｉ≠Ｉ_ｎ）は介入Ｉ_ｎを提示しないことをそれぞれ表し、Ｅは条件付き期待値である。上記の式（１）は、静的特性情報ｏ_ｉ ^ｓ及び動的特性情報ｏ_ｉ ^ｄが与えられたときに、介入Ｉ_ｎを対象者ｉに提示したときの行動実績Ｒ_ｉｔの期待値と、介入Ｉ_ｎを対象者ｉに提示しないとき（つまり、介入Ｉ_ｎ以外の介入を対象者ｉに提示したとき）の行動実績Ｒ_ｉｔの期待値との差が最大となる介入Ｉ_ｎを最適な介入Ｉとして選択することを意味している。

Here, I is the optimal intervention, L={I ₁ ,...,I _N } is the intervention set, N is the total number of interventions, I _n (1≦n≦N) is the nth intervention, i is an index that uniquely identifies a subject, and R _it is the behavioral performance of subject i at time t. _{o i} ^s is vector data representing information on the static characteristics of subject i (hereinafter also referred to as static characteristic information), and o _i ^d is vector data representing information on the dynamic characteristics of subject i (hereinafter also referred to as dynamic characteristic information). In addition, do(I=I _n ) represents the presentation of intervention I _n , and do(I≠I _n ) represents the non-presentation of intervention I _n , and E is the conditional expectation. The above formula (1) means that, when static characteristic information ^{o is} _and dynamic characteristic information o _i ^d are given, the intervention I _n that maximizes the difference between the expected value of the behavioral performance R _it when intervention I _n is presented to subject i and the expected value of the behavioral performance R _it when intervention I _n is not presented to subject i (i.e., when an intervention other than intervention I _n is presented to subject i) is selected as the optimal intervention I.

In addition, the proposed method predicts the intervention effect (i.e., the behavioral performance expected from the intervention) when the optimal intervention is presented to the subject, and detects a change in static characteristics from the discrepancy between the predicted intervention effect and the actual behavioral performance (S4). Then, in the proposed method, if a change in static characteristics is detected, the static characteristics of the subject are acquired again.

As a result, the proposed method eliminates the need for the user (subject) to input changes in dynamic characteristics, reducing the burden on the user. Furthermore, the proposed method requires the user (subject) to input changes in static characteristics when they are detected, making it possible to select more appropriate interventions. In other words, the proposed method estimates dynamic characteristics that may change frequently from past behavior, eliminating the need to obtain them from the user, while static characteristics that may change infrequently are obtained from the user only when changes are detected. This not only reduces the burden on the user, but also prevents situations in which an inappropriate intervention is selected due to changes in static characteristics.

When selecting the optimal intervention, the optimal intervention may be selected using the following formula (2) instead of the above formula (1).

　ここで、ｚ_ｉは対象者ｉの個人特性に関する情報（以下、個人特性情報ともいう。）を表すベクトル形式のデータ、ｐは条件付き確率分布である。上記の式（２）は、静的特性情報ｏ_ｉ ^ｓ及び動的特性情報ｏ_ｉ ^ｄが与えられたときに、条件付き確率分布ｐ（ｚ_ｉ｜ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）に従って個人特性情報ｚ_ｉをサンプリングすることを意味している。また、上記の式（２）は、サンプリングされた個人特性情報ｚ_ｉの下で、介入Ｉ_ｎを対象者ｉに提示したときの行動実績Ｒ_ｉｔの期待値と、介入Ｉ_ｎを対象者ｉに提示しないときの行動実績Ｒ_ｉｔの期待値との差が最大となる介入Ｉ_ｎを最適な介入Ｉとして選択することを意味している。なお、条件付き確率分布ｐ（ｚ_ｉ｜ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）は、例えば、ｐ（ｚ_ｉ，ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）及びｐ（ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）の分布として正規分布等といった既知の分布を仮定した上で、ｐ（ｚ_ｉ｜ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）＝ｐ（ｚ_ｉ，ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）／ｐ（ｏ_ｉ ^ｓ，ｏ_ｉ ^ｄ）により求めることができる。

Here, z _i is vector data representing information on the personal characteristics of the subject i (hereinafter also referred to as personal characteristic information), and p is a conditional probability distribution. The above formula (2) means that when static characteristic information o _i ^s and dynamic characteristic information o _i ^d are given, the personal characteristic information z _i is sampled according to the conditional probability distribution p (z _i |o _i ^s , o _i ^d ). In addition, the above formula (2) means that, under the sampled personal characteristic information z _i , an intervention I _n that maximizes the difference between the expected value of the behavioral performance R _it when the intervention I _n is presented to the subject i and the expected value of the behavioral performance R _it when the intervention I _n is not presented to the subject i is selected as the optimal intervention I. In addition, the conditional probability distribution p(z _i | _o ^is , _o ^id ) can be obtained, for example, by assuming that the distributions of p(z _i , _o ^is , _o ^id ) and p( _o ^is , _o ^id ) are known distributions such as normal distributions, and then obtaining p(z _i | _o ^is , _o ^id ) = p(z _i , _o ^is , _o ^id )/p( _o ^is , _o ^id ).

Below, we will explain the behavior change support system 1 that supports the behavior change of a user (subject) by presenting the optimal intervention to the user (subject) using the above proposed method.

<Example of overall configuration of behavior change support system 1>
First, an example of the overall configuration of the behavior change support system 1 according to this embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the overall configuration of the behavior change support system 1 according to this embodiment.

As shown in FIG. 3, the behavior change support system 1 according to this embodiment includes an optimal intervention presentation device 10 and one or more subject terminals 20. Furthermore, the optimal intervention presentation device 10 and each subject terminal 20 are communicatively connected via a communication network 30 including, for example, the Internet.

The optimal intervention presentation device 10 is a computer or computer system, such as a general-purpose server, that selects the optimal intervention using the proposed method described above and presents it on the subject terminal 20. Note that the optimal intervention presentation device 10 may be, for example, configured with multiple general-purpose servers, may be realized by a virtual machine, or may be a PC (personal computer), etc.

The subject terminal 20 is various information terminals such as a smartphone, a tablet terminal, a wearable device, etc., used by the subject. The subject terminal 20 may be, for example, a PC (personal computer), etc., or various devices such as a weight scale. Hereinafter, the subject terminal 20 used by subject i will be represented as "subject terminal 20 _i ."

Note that the overall configuration of the behavior change support system 1 shown in FIG. 3 is an example and is not limited to this. For example, the optimal intervention presentation device 10 and the subject terminal 20 may be configured as an integrated unit.

<Example of Hardware Configuration of Optimal Intervention Presentation Device 10>
Next, an example of the hardware configuration of the optimal intervention presentation device 10 according to the present embodiment will be described with reference to Fig. 4. Fig. 4 is a diagram showing an example of the hardware configuration of the optimal intervention presentation device 10 according to the present embodiment.

As shown in FIG. 4, the optimal intervention presentation device 10 according to this embodiment has an external I/F 101, a communication I/F 102, a RAM (Random Access Memory) 103, a ROM (Read Only Memory) 104, an auxiliary storage device 105, and a processor 106. Each of these pieces of hardware is connected to each other so as to be able to communicate with each other via a bus 107.

The external I/F 101 is an interface with external devices such as a recording medium 101a. Examples of recording media 101a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

The communication I/F 102 is an interface for connecting to the communication network 30. The RAM 103 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 104 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 105 is a non-volatile storage device (storage device) such as a HDD (Hard Disk Drive), SSD (Solid State Drive), flash memory, etc. The processor 106 is an arithmetic device such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit).

Note that the hardware configuration shown in FIG. 4 is an example, and the hardware configuration of the optimal intervention presentation device 10 is not limited to this. For example, the optimal intervention presentation device 10 may have an input device such as a keyboard mouse or a touch panel, or a display device such as a display or a display panel. Also, for example, the optimal intervention presentation device 10 may have multiple auxiliary storage devices 105 or multiple processors 106. Furthermore, for example, the optimal intervention presentation device 10 may not have some of the hardware shown in FIG. 4, or may have various hardware other than the hardware shown in FIG. 4.

<Example of functional configuration of optimal intervention presentation device 10>
Next, an example of the functional configuration of the optimal intervention presentation device 10 according to the present embodiment will be described with reference to Fig. 5. Fig. 5 is a diagram showing an example of the functional configuration of the optimal intervention presentation device 10 according to the present embodiment.

As shown in FIG. 5, the optimal intervention presentation device 10 according to this embodiment has a static characteristic acquisition unit 111, a behavioral performance acquisition unit 112, a dynamic characteristic estimation unit 113, an intervention effect prediction unit 114, an intervention selection unit 115, an intervention presentation unit 116, an evaluation unit 117, and a static characteristic change detection unit 118. Each of these units is realized, for example, by a process in which one or more programs installed in the optimal intervention presentation device 10 are executed by the processor 106 or the like. The optimal intervention presentation device 10 according to this embodiment also has a static characteristic storage unit 119, a behavioral performance storage unit 120, an intervention effect storage unit 121, and an intervention storage unit 122. Each of these storage units is realized, for example, by a storage area of the auxiliary storage device 105 or the like. However, at least one of these storage units may be realized by a storage area of a storage device or the like connected to the optimal intervention presentation device 10 so as to be able to communicate with it.

The static characteristic acquisition unit 111 acquires the static characteristic information o _i ^s from the subject terminal 20 _i . In addition, the static characteristic acquisition unit 111 stores the static characteristic information o _i ^s acquired from the subject terminal 20 _i in the static characteristic storage unit 119. Note that what kind of personal characteristic is to be the static characteristic is set in advance.

The behavioral record acquiring unit 112 acquires the behavioral record R _it from the subject terminal 20 _i at each predetermined time t. The behavioral record acquiring unit 112 also stores the behavioral record R _it acquired from the subject terminal 20 _i in the behavioral record storage unit 120. Typical examples of the time t include "a predetermined time every day,""a predetermined time every week,""the time when a certain predetermined website is logged in,""the time when a certain predetermined application is started,""the time when a certain predetermined event occurs," and the like.

The dynamic characteristic estimation unit 113 estimates dynamic characteristic information o _id from the history of the behavioral record R _it stored in the behavioral record storage unit 120 (i.e., the sequence data of the behavioral record R _it ) for each time ^t . Here, the dynamic characteristic is, for example, a psychological characteristic such as motivation for behavioral change, or an individual characteristic that may change frequently, such as a schedule of a subject, but the dynamic characteristic is not explicitly set, and an individual characteristic that is not set as a static characteristic becomes the dynamic characteristic. A specific example of a method for estimating the dynamic characteristic information o _id ^will be described later.

The intervention effect prediction unit 114 predicts the intervention effect when an intervention I _n (n=1, ..., N) is presented to a subject i using the static characteristic information o _i ^s and the dynamic characteristic information o _i ^d estimated by the dynamic characteristic estimation unit 113. The intervention effect prediction unit 114 predicts, for each intervention I _n , a conditional average treatment effect (CATE), which is a conditional causal effect of the intervention I _n , as the intervention effect. Here, the conditional average treatment effect CATE _in when the intervention _{I n} is presented to the subject i is given by the following.

CATE _in = E(R _it | o _i ^s , o _i ^d , do (I=I _n )) - E(R _it | o _i ^s , o _i ^d , do (I≠I _n ))
That is, the conditional average treatment effect CATE _in is given by the equation that is the subject of argmax in the above equation (1). This allows each of the N conditional average treatment effects CATE _i1 , ..., CATE _iN to be obtained as the intervention effect.

Note that the conditional average treatment effect CATE _in when intervention I _n is presented to subject i may be given by:

CATE _in =E (R _it | z _i , do (I=I _n )) - E (R _it | z _i , do (I≠I _n )), z _i ~ p (z _i | o _i ^s , o _i ^d )
That is, the conditional average treatment effect CATE _in may be given by the equation subject to argmax in equation (2) above.

The intervention selection unit 115 uses the intervention effect predicted by the intervention effect prediction unit 114 to select an optimal intervention I from the intervention set L={I ₁ ,...,I _N } stored in the intervention storage unit 122. That is, the intervention selection unit 115 selects an intervention I n' corresponding to the highest value CATE _in' ( _n'∈ {1,...,N}) among CATE _i1 ,...,CATE _iN as the optimal intervention I for the subject i. In other words, the intervention selection unit 115 uses n'=argmax _{n∈{1,...,N}} CATE _in to set the optimal intervention I for the subject i as I=I _n' . However, if all of CATE _i1 ,...,CATE _iN are negative, the intervention selection unit 115 does not select the optimal intervention I for the subject i. In this case, the optimal intervention I is not presented to the subject i. Hereinafter, the optimal intervention for subject i at time t will also be represented as I ^(i,t) .

Furthermore, the intervention selection unit 115 assigns time t to the intervention effect CATE _in' corresponding to the optimal intervention I ^{(i, t)} = I _n' for the subject i at time t, and stores the assigned time t in the intervention effect storage unit 121. Hereinafter, the intervention effect CATE _in' assigned time t will be represented as T _it . This intervention effect T _it represents the intervention effect (conditional average treatment effect) predicted when the optimal intervention I ^{(i, t)} at time t is presented to the subject i.

The intervention presentation unit 116 presents the optimal intervention I ^(i,t) selected by the intervention selection unit 115 to the subject terminal _20i . That is, the intervention presentation unit 116 transmits the optimal intervention I ^(i,t) to the subject terminal _20i . As a result, a message or the like representing the optimal intervention I ^(i,t) is displayed on a display or the like provided on the subject terminal _20i .

The evaluation unit 117 evaluates whether or not a discrepancy occurs between the behavioral performance of the subject i and the intervention effect, using the behavioral performance R _it in the most recent predetermined period and the intervention effect T _it in the same period. That is, the evaluation unit 117 evaluates whether or not a discrepancy occurs between the behavioral performance of the subject i and the intervention effect, using the series of behavioral performance {R _it |tεΔ} stored in the behavioral performance storage unit 120 and the series of intervention effects {T _it |tεΔ} stored in the intervention effect storage unit 121. Here, Δ is the most recent predetermined period (that is, a period expressed as Δ=[t _p , t _{c ] when the current time is t=t c and} a certain time in the past is t=t _p ). The length of the period Δ is set in advance.

The static characteristic change detection unit 118 detects that a change has occurred in the static characteristics of the subject i when the evaluation unit 117 evaluates that a discrepancy has occurred between the behavioral performance of the subject i and the intervention effect. On the other hand, the static characteristic change detection unit 118 determines that no change has occurred in the static characteristics of the subject i when the evaluation unit 117 does not evaluate that a discrepancy has occurred between the behavioral performance of the subject i and the intervention effect.

The static characteristic storage unit 119 stores, for each subject i, the static characteristic information o _i ^s of the subject i.

The behavior record storage unit 120 stores, for each subject i, the behavior record R _it of the subject i at each time t.

The intervention effect storage unit 121 stores, for each subject i, the intervention effect information T _it at each time t for that subject i.

The intervention storage unit 122 stores an intervention set L = {I ₁ , ..., I _N }. Typical examples of each intervention I _n ∈ L include data for displaying on the subject terminal 20 text (message), images, videos, etc., that encourage the subject to change their behavior.

<Optimal intervention presentation process>
Next, the optimal intervention suggestion process according to this embodiment will be described with reference to Fig. 6. Fig. 6 is a flowchart showing an example of the optimal intervention suggestion process according to this embodiment. In the following, a case where an optimal intervention is suggested for a certain subject i will be described.

The static characteristic acquisition unit 111 acquires static characteristic information o _i ^s from the subject terminal 20 _i and stores it in the static characteristic storage unit 119 (step S101). This step is executed, for example, when the subject i starts using the subject terminal 20 _i to use the behavior change support service provided by the behavior change support system 1. Note that the static characteristics are input by the subject i on the subject terminal 20 _i , for example.

The behavior record acquiring unit 112 acquires the behavior record R _it from the subject terminal 20 _i and stores it in the behavior record storage unit 120 (Step S102).

The following describes the process of steps S103 to S110, assuming that the current time t= _tc . Note that steps S103 to S110 are repeatedly executed for each time t. For example, typically, when the unit of time t is "days," steps S103 to S110 are executed every day.

The dynamic characteristic estimation unit 113 estimates dynamic characteristic information o _i ^d from the history of the behavior record R _it stored in the behavior record storage unit 120 (sequence data of the behavior record R _it ) (step S103). For example, the dynamic characteristic estimation unit 113 can realize a variational autoencoder using a recurrent neural network (RNN) or the like, and use an intermediate representation vector h _t of a model in which the input and output of the variational autoencoder are extended to sequence data as the dynamic characteristic information o _i ^d . This is because the variational autoencoder can obtain a vector representation that takes into account differences between elements of a sequence as an intermediate representation vector due to the nature of the regularization term of the loss function during learning.

For example, when the unit of time t is "days" and dynamic characteristic information o _{i d is estimated from sequence data {R it |t = t c -3, t c -2, t c -1} of behavior record R it for the past three days, the dynamic characteristic information o i} ^d _{can be estimated by an encoder} ⁽ RNN) of a variational autoencoder as shown in Fig. 7. That is, the sequence data {R _it |t = t _c -3, t _c -2, t _c -1} of behavior record R _it is input to an encoder of a variational autoencoder realized by an RNN, and the mean μ and covariance Σ are obtained as the output. After that, an intermediate representation vector h _t is sampled from a multivariate normal distribution N (μ, Σ) of mean μ and covariance Σ, and this intermediate representation vector h _t is set as dynamic characteristic information o _i ^d . However, using the series data of the behavior result R _it for the past three days is just an example, and it is possible to appropriately set how much of the past period of the series data of the behavior result R _it is to be used.

Here, when the variational autoencoder is trained, for example, the parameters of the encoder and the decoder of the variational autoencoder may be updated so as to minimize a known predetermined loss function using sequence data of the behavioral performance R _it in a certain predetermined period in the past. The variational autoencoder may be trained for each subject i, may be trained commonly for a plurality of subjects i, or may be trained commonly for a plurality of subjects i and then retrained for each subject i. The loss function may be, for example, a function consisting of a term representing the error (prediction error) between the sequence data input to the encoder and the sequence data output from the decoder, and a regularization term representing the Kullback-Leibler divergence between the multivariate normal distribution N(μ,Σ) of mean μ and covariance Σ output from the encoder and the standard normal distribution.

In the above step S103, the dynamic characteristic information o _id is estimated by a model in which the input and output of the variational autoencoder is extended to ^sequence data, but this is only one example, and the dynamic characteristic information o _id may be estimated by other methods. For example, the sequence data of the behavioral record R _it may be clustered by time-series clustering, and a vector representing the cluster to ^which the sequence data belongs may be used as the dynamic characteristic information o ^id _.

The intervention effect prediction unit 114 predicts the intervention effect when an intervention I _n (n=1, ..., N) is presented to a subject i (step S104). That is, for each intervention I _n , the intervention effect prediction unit 114 calculates the conditional average treatment effect CATE _in of that intervention I _n as the intervention effect.

The intervention selection unit 115 selects the intervention I _n' ∈L corresponding to the highest value of CATE _in' (n' ∈ {1, ..., N}) among the conditional average treatment effects CATE _i1 , ..., CATE _iN as the optimal intervention I ^{(i, t)} = I _n' for subject i (step S105).

The intervention selection unit 115 stores the intervention effect T _it in the intervention effect storage unit 121, where the intervention effect CATE _in′ corresponding to the optimal intervention I ^{(i, t)} = I _n′ is assigned the current time t = _tc (step S106).

The intervention presenting unit 116 presents the optimal intervention I ^(i,t) to the subject terminal _20i (step S107).

The behavioral record acquiring unit 112 acquires the behavioral record R _it representing the behavioral record after the optimal intervention I ^{(i, t)} was presented to the subject i from the subject terminal 20 _i and stores it in the behavioral record storage unit 120 (step S108).

The evaluation unit 117 evaluates whether or not a discrepancy occurs between the behavioral performance of the subject i and the intervention effect by using the behavioral performance R _it in the most recent predetermined period Δ=[t _p , t _c ] and the intervention effect _T it in the same period Δ (step S109). The evaluation unit 117 may evaluate whether or not a discrepancy occurs by, for example, the following steps 1 to 3. Hereinafter, it is assumed that t _p =t _c -M+1, and the number of times included in the period Δ is M.

Step 1: The evaluation unit 117 calculates, for each tεΔ, the number m that satisfies |R _it -T _it |≧th ₁ , where th ₁ is a threshold value, and its value is set in advance.

Step 2: The evaluation unit 117 uses the number m calculated in the above step 1 to determine whether m/M is equal to or greater than _th2 . _th2 is a threshold value that satisfies 0< _th2 <1, and its value is set in advance.

Step 3: If the evaluation unit 117 determines that m/M is greater than or equal to th ₂ , it evaluates that a discrepancy has occurred between the behavioral performance of the subject i and the intervention effect, and if not, it evaluates that no discrepancy has occurred between the behavioral performance of the subject i and the intervention effect.

Here, the length of the period Δ is set in advance, but it is preferable that it be at least several days (i.e., M≧2) rather than a single day (i.e., M=1). The reason for this is as follows. For example, when considering dynamic characteristics such as motivation, schedule, habits, etc., as shown in Figure 8, behavioral performance can change due to an increase or decrease in motivation, habit formation, schedule conflicts, etc. In such cases, for example, temporary hindrance due to schedule conflicts, etc., often also results in a temporary decline in behavioral performance. For this reason, if the period Δ is set to a single day, etc., the prediction error of the intervention effect may be overestimated.

The static characteristic change detection unit 118 determines whether or not a change in the static characteristics of the subject i has been detected (step S110). If it is assessed in step S109 above that a discrepancy has occurred between the behavioral performance of the subject i and the intervention effect, the static characteristic change detection unit 118 detects that a change has occurred in the static characteristics of the subject i. On the other hand, if it is not assessed in step S109 above that a discrepancy has occurred between the behavioral performance of the subject i and the intervention effect, the static characteristic change detection unit 118 determines that no change has occurred in the static characteristics of the subject i.

If a change in the static characteristic is detected in step S110 above, the static characteristic acquisition unit 111 acquires the static characteristic information o _i ^s from the subject terminal 20 _i , and updates the static characteristic information o _i ^s stored in the static characteristic memory unit 119 with the acquired static characteristic information o _i ^s (step S111).

If no change in the static characteristics is detected in step S110 above, or if step S111 above is executed, step S103 above is executed at the next time t+1.

<Example>
An example of the behavior change support system 1 according to this embodiment will be described below.

Example 1
In the first embodiment, a health care service is assumed as the behavior change support service provided by the behavior change support system 1, and a case will be described in which a message for encouraging a subject i to walk is personalized.

In Example 1, each intervention is a message, and the intervention set is assumed to be L = {"Why not try walking far today?", "Let's walk today for your health!", "Even just cleaning your room is exercise."}. In addition, the static characteristics are set as health check results at the time of using the healthcare service, past exercise experience, personality diagnosis results, and working style. On the other hand, dynamic characteristics are not explicitly set, but are assumed to be schedules, frequency of laundry, physical strength, etc.

The unit of time t is "days", and the behavioral record R _it of the subject i is the number of steps of the previous day obtained at midnight every day. Furthermore, the behavioral record of the most recent three days is used to estimate the dynamic characteristics and predict the intervention effect.

As an example, suppose that the static characteristics of a certain subject i are "health check result: abnormal", "exercise experience: yes", "personality test result: high conscientiousness", and "work style: telecommuting", and that the recent number of steps is on the rise. In this case, it can be estimated that the subject has started walking due to a sense of crisis about the health check result and conscientiousness, and that the dynamic characteristic of physical strength has returned based on the past exercise experience. Therefore, for example, the intervention effect of a message such as "Why not try walking far today?" is maximized, and the message is presented on the subject terminal 20 _i .

The above example will be explained below in relation to each step in Figure 6.

First, when the use of the health care service is started, the static characteristic information o _i ^s is acquired from the subject terminal 20 _i by the static characteristic acquisition unit 111 and stored in the static characteristic storage unit 119 (step S101). The static characteristic information o _i ^s is stored in a format such as o _i ^s = (1, 1, 1, 0). Here, the first element of o _i ^s represents the "health check result", and takes "0" if there is no abnormality, and "1" if there is an abnormality. Similarly, the second element represents the "exercise experience", and takes "1" if there is an exercise experience, and takes "0" if there is no exercise experience. The third element represents the "personality diagnosis result", and takes a value determined in advance according to the diagnosis result. The fourth element represents the "work style", and takes "0" if there is telecommuting, and takes "1" if there is not. In addition, the behavioral record acquisition unit 112 acquires the most recent number of steps as the behavioral record R _it , and stores it in the behavioral record storage unit 120 (step S102).

Next, steps S103 to S111 are executed every day while the health care service is being used. That is, first, the dynamic characteristic estimation unit 113 estimates dynamic characteristic information o _i ^d from the series data of the behavioral performance R _it for the past three days (step S103). Since the most recent number of steps of the subject i is on the rise, a vector expression representing the change in dynamic characteristic "return of physical strength" that influences the rising trend is estimated as dynamic characteristic information o _i ^d . Next, the intervention effect prediction unit 114 predicts the intervention effect from the static characteristic information o _i ^s and the dynamic characteristic information o _i ^d (step S104), and then the intervention selection unit 115 selects the optimal intervention I ^{(i, t)} = "Why don't you walk far today?" (step S105). In addition, the intervention effect T _it corresponding to the optimal intervention I ^{(i, t)} is stored in the intervention effect storage unit 121 by the intervention selection unit 115 (step S106), and the optimal intervention I ^{(i, t)} is presented to the subject terminal 20 _i by the intervention presentation unit 116 (step S107). As a result, for example, a message such as "Why don't you walk farther today?" is displayed on the display of the subject terminal 20 _i .

Next, at midnight on the day after the optimal intervention I ^(i,t) was presented to the subject terminal _20i , the behavioral record acquisition unit 112 acquires the number of steps for the previous day as the behavioral record _Rit , and stores it in the behavioral record storage unit 120 (step S108).

The evaluation unit 117 also determines whether or not there is a discrepancy between the behavioral performance R _it for the most recent three days and the intervention effect T _it (step S109). For example, the behavioral performance R _it for the most recent three days is [8210, 8643, 8811], the intervention effect T _it for the most recent three days is [8200, 8510, 8500], and th ₁ = 300, th ₂ = 0.5. In this case, the number m for which |R _it -T _it | ≧ th ₁ is 1, so m/M = 1/3. Therefore, it is evaluated that there is no discrepancy between the behavioral performance of the subject i and the intervention effect, and the static characteristic change detection unit 118 determines that there is no change in the static characteristic of the subject i (step S110).

For example, if the number of steps of a subject i is decreasing every 2 or 3 days, a vector expression representing "the frequency of washing is every 2 or 3 days" is estimated as the dynamic characteristic information o _i ^d in step S103. Therefore, for example, "Even just cleaning the room is a good exercise" is selected as the optimal intervention I ^(i,t) in step S104.

Example 2
In Example 2, a video learning service for students taking exams is assumed as the behavior change support service provided by the behavior change support system 1, and a case will be described in which daily suggestions of study materials are personalized.

In Example 2, when subject i logs in to a system for using a behavior change support service and when he or she finishes watching a video of a learning material, the next learning material is presented as an intervention. The learning materials are assumed to be diverse in terms of field, subject, difficulty level, instructor, etc.

In addition, as static characteristics, mock test scores, whether or not the student participates in club activities, and preferred schools at the time of using the video learning service are set. On the other hand, while dynamic characteristics are not explicitly set, schedules, viewing environments, etc. are assumed.

The unit of time t is the "number of times a video has been viewed," and each time a subject i finishes viewing a video of a learning material, the time t of the subject i advances. Furthermore, the behavioral record R _it of the subject i is set to indicate whether or not the learning material presented as the intervention has been viewed after the intervention is presented. The most recent 10 behavioral records are used to estimate the dynamic characteristics and predict the intervention effect.

As an example, suppose that the static characteristics of a certain subject i are that their grades are insufficient to pass the entrance exam for their preferred school, and that they are currently participating in club activities. In addition, suppose that their most recent behavioral records (whether or not they have watched the learning materials presented as an intervention) are sparse, such as [1, 0, 0, 1, 0, 0, 0, 1, 0, 0]. Here, "1" indicates that they have watched, and "0" indicates that they have not watched. In this case, a change in the dynamic characteristics could be that they are unable to secure sufficient time in their schedule, and so learning materials that summarize the main points in a short amount of time are more likely to be presented as an intervention.

The above example will be explained below in relation to each step in Figure 6.

First, when the use of the video learning service is started, the static characteristic information o _i ^s is acquired from the subject terminal 20 _i by the static characteristic acquisition unit 111 and stored in the static characteristic storage unit 119 (step S101). The static characteristic information o _i ^s is stored in a format such as o _i ^s = (50, 1, 30). Here, the first element of o _i ^s represents "score of mock exam" and takes a numerical value representing a deviation value. Similarly, the second element represents "whether or not club activities are held", taking "1" if club activities are being held and "0" if not. The third element represents "desired school" and takes a numerical value representing the desired school. In addition, the behavioral record acquisition unit 112 acquires whether or not the learning material presented at the start of use has been viewed as a behavioral record R _it , and stores it in the behavioral record storage unit 120 (step S102).

Next, each time the subject i finishes watching the video of the learning material, steps S103 to S111 are executed. That is, first, the dynamic characteristic estimation unit 113 estimates dynamic characteristic information o _i ^d from the sequence data of the most recent 10 past behavioral results R _it (step S103). Since the sequence data of the most recent 10 past behavioral results R _it of the subject i is [1, 0, 0, 1, 0, 0, 0, 1, 0, 0], a vector expression representing the change in dynamic characteristic that influences it, "cannot secure enough schedule", is estimated as dynamic characteristic information o _i ^d . Next, the intervention effect prediction unit 114 predicts the intervention effect from the static characteristic information o _i ^s and the dynamic characteristic information o _i ^d (step S104), and then the intervention selection unit 115 selects "learning material in which the main points are summarized in a short time" as the optimal intervention I ^{(i, t)} (step S105). In addition, the intervention effect T _it corresponding to the optimal intervention I ^{(i, t)} is stored in the intervention effect storage unit 121 by the intervention selection unit 115 (step S106), and the optimal intervention I ^{(i, t)} is presented to the subject terminal 20 _i by the intervention presentation unit 116 (step S107). As a result, for example, learning materials in which the main points are summarized in a short time are presented as the next learning materials on the display of the subject terminal 20 _i .

Next, after the optimal intervention I ^(i,t) is presented to the subject terminal _20i , the behavioral record acquisition unit 112 acquires the presence or absence of viewing of the learning material presented in the intervention I ^(i,t) as the behavioral record _Rit , and stores it in the behavioral record storage unit 120 (step S108).

The evaluation unit 117 also determines whether or not there is a discrepancy between the most recent past 10 behavioral results R _it and the intervention effect T _it (step S109). For example, the most recent past 10 behavioral results R _it are [0,0,1,0,0,0,1,0,0,1], the most recent past 10 intervention effects T _it are [0.7,0.4,0.7,0.3,0.6,0.7,0.6,0.4,0.6,0.6], and th ₁ =0.5 and th ₂ =0.5. In this case, the number m for which |R _it -T _it |≧th ₁ is 4, so m/M=4/10. Therefore, it is evaluated that there is no discrepancy between the behavioral results and the intervention effect of the subject i, and the static characteristic change detection unit 118 determines that there is no change in the static characteristic of the subject i (step S110).

For example, the most recent past 10 behavioral results R _it are [1, 0, 0, 0, 0, 1, 1, 0, 0, 0], and the most recent past 10 intervention effects T _it are [0.7, 0.4, 0.5, 0.5, 0.4, 0.6, 0.6, 0.5, 0.5, 0.6]. In this case, m/M=5/10. This is, for example, when the static characteristics of the subject i, such as the school of choice, have changed. For this reason, the static characteristic change detection unit 118 detects that a discrepancy has occurred between the behavioral results of the subject i and the intervention effect (step S110), and the static characteristic acquisition unit 111 acquires and updates the static characteristic information o _i ^s (step S111).

<Summary>
As described above, in the method proposed in this embodiment, the problem setting for selecting an intervention in behavior change is formulated by the causal relationships among "personal characteristics,""intervention," and "behavioral performance," and the causal relationships among various factors that change the personal characteristics. In addition, the personal characteristics are separated into static characteristics and dynamic characteristics, and the dynamic characteristics enable intervention selection that explicitly takes into account changes in the subject's state.

Therefore, the optimal intervention presentation device 10 that realizes the above proposed method can estimate dynamic characteristic information _o ^id from the sequence data of the behavioral performance R _it for each time t, and then select and present the optimal intervention I ^{(i, t) from} the dynamic characteristic information o _id and the static characteristic information o _id ^s acquired in advance. As a result, even if the dynamic characteristic information o _id ^, which can change frequently, changes, it is not necessary for the subject i to input the dynamic characteristic information o _id ^, and therefore the load on the user can be reduced.

Furthermore, the optimal intervention presentation device 10 according to the present embodiment can predict the intervention effect when the optimal intervention I ^(i,t) is presented to the subject i, and detect a change in ^the static characteristic information _ois from the deviation between the intervention effect _Tit and the actual behavioral performance _Rit . Therefore, when a change in the static characteristic information _ois is detected, it is possible to have the subject i input the change, and it is possible to prevent a decrease in the intervention effect due to a change in ^the static characteristic information _ois ^.

The present invention is not limited to the specifically disclosed embodiments above, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims.

REFERENCE SIGNS LIST 1 Behavioral change support system 10 Optimal intervention presentation device 20 Subject terminal 30 Communication network 101 External I/F
101a Recording medium 102 Communication I/F
103 RAM
104 ROM
105 Auxiliary storage device 106 Processor 107 Bus 111 Static characteristic acquisition unit 112 Behavioral performance acquisition unit 113 Dynamic characteristic estimation unit 114 Intervention effect prediction unit 115 Intervention selection unit 116 Intervention presentation unit 117 Evaluation unit 118 Static characteristic change detection unit 119 Static characteristic storage unit 120 Behavioral performance storage unit 121 Intervention effect storage unit 122 Intervention storage unit

Claims (6)

a static characteristic acquisition unit that acquires static characteristic information representing a static characteristic, which is a personal characteristic that changes infrequently among personal characteristics of a user;
a dynamic characteristic estimation unit that estimates dynamic characteristic information representing a dynamic characteristic, which is an individual characteristic that changes more frequently among the individual characteristics compared to the static characteristic, using information regarding the behavioral record of the user;
an intervention effect prediction unit that predicts, for each predetermined intervention, an intervention effect that represents an expected effect when the intervention is presented to the user, using the static characteristic information and the dynamic characteristic information;
A selection unit that selects an intervention that has the highest intervention effect as an optimal intervention;
an intervention presentation unit that presents the optimal intervention to the user;
An optimal intervention presentation device having the above-mentioned. The intervention effect prediction unit is
The optimal intervention presentation device according to claim 1 , wherein an expected value of a behavioral performance when the intervention is presented to the user is predicted as the intervention effect for each of the interventions. The dynamic characteristic estimation unit is
Calculating the mean and covariance obtained by inputting the sequence data of the user's behavioral performance to an encoder of a trained variational autoencoder;
The optimal intervention presentation device according to claim 1 , wherein the static characteristic information is estimated as a vector sampled from a multivariate normal distribution determined by the mean and the covariance. an evaluation unit that evaluates a deviation between the behavioral performance and the intervention effect in a predetermined past period by using the behavioral performance in the period and the intervention effect corresponding to the optimal intervention in the period;
a static characteristic change detection unit that detects a change in the static characteristic when the evaluation unit evaluates that the deviation has occurred,
The static characteristic acquisition unit is
The optimal intervention presentation device according to claim 1 , wherein when the static characteristic change detection unit detects that a change has occurred in the static characteristic, the static characteristic is reacquired. A static characteristic acquisition step of acquiring static characteristic information representing a static characteristic, which is a personal characteristic that changes infrequently among the personal characteristics of the user;
a dynamic characteristic estimation step of estimating dynamic characteristic information representing a dynamic characteristic, which is an individual characteristic that changes more frequently among the individual characteristics compared to the static characteristic, using information regarding the behavioral record of the user;
an intervention effect prediction step of predicting, for each predetermined intervention, an intervention effect that represents an expected effect when the intervention is presented to the user, using the static characteristic information and the dynamic characteristic information;
A selection procedure for selecting an intervention with the highest intervention effect as an optimal intervention;
an intervention suggestion step of suggesting the optimal intervention to the user;
The computer then executes the optimal intervention presentation method. A static characteristic acquisition step of acquiring static characteristic information representing a static characteristic, which is a personal characteristic that changes infrequently among the personal characteristics of the user;
a dynamic characteristic estimation step of estimating dynamic characteristic information representing a dynamic characteristic, which is an individual characteristic that changes more frequently among the individual characteristics compared to the static characteristic, using information regarding the behavioral record of the user;
an intervention effect prediction step of predicting, for each predetermined intervention, an intervention effect that represents an expected effect when the intervention is presented to the user, using the static characteristic information and the dynamic characteristic information;
A selection procedure for selecting an intervention with the highest intervention effect as an optimal intervention;
an intervention suggestion step of suggesting the optimal intervention to the user;
A program that causes a computer to execute the following.

Images