Selection device, training method, and program

Abstract

The purpose of the present disclosure is to allow a designer of an application, a system, or a service for prompting behavior change to provide an intended target person with an optimal intervention method for behavior change.　The present disclosure is a selection device for selecting an intervention method for behavior change to be provided to a target person, the selection device having a learning model construction unit that constructs a machine learning model by causing, in a learning phase, a machine learning model to undertake machine learning by using a rank learning procedure on the basis of a score indicating the degree of leading to the behavior change and at least one of the personal characteristics of the target person and the intervention method.

Description

Selection device, learning method and program

[0001] The present disclosure relates to the selection of intervention methods for promoting behavioral change taking into account individual characteristics.

[0002] There are many different types of intervention methods that can be used to influence people from the outside world to change their behavior. In the fields of healthcare and medicine, applications, systems, and services have been developed to encourage people to adopt healthy behaviors, and these applications incorporate some type of intervention method to encourage behavioral change as part of their functionality. For example, a feature that awards points upon achieving a goal is an application function that embodies the intervention method known as "reward." The effectiveness of an implemented behavioral change intervention method is heavily influenced by the user's personal characteristics. Therefore, designers must make careful selections, as their intended encouragement of behavior may actually have the opposite effect (dismotivate users). However, it is difficult for designers to identify and select the most appropriate intervention method for the intended target audience from the vast number of intervention methods available, and there is a need for assistance in selecting an intervention method.

[0003] The Transtheoretical Model (TTM) has been developed to change an individual's behavior to quit smoking. It defines five stages of behavior (indifference, interest, preparation, action, and maintenance) based on the individual's readiness and level of behavior. The TTM points out that effective intervention methods vary depending on the stage the individual is currently in, and suggests what intervention methods are best for moving the individual to the next stage or maintaining the "maintenance stage" (see Non-Patent Document 1). For example, intervention methods for moving from the "indifference stage" to the "interest stage" include "raising awareness" through information provision, "emotional experiences" regarding the health threats of not changing behavior, and "environmental reassessment" by reassessing the impact of the behavior on those around them. It suggests that intervention methods that primarily influence the individual's way of thinking may be effective in moving the individual to the next stage.

[0004] Additionally, in the field of behavioral science, there is a method known as a mega-study, which actually investigates what kind of intervention method will bring about behavioral change in subjects (Non-Patent Document 2). This is a method of verifying which of various intervention methods is superior in effectiveness over a certain period of time in a group of tens of thousands of people.

[0005] Reed, GR, Velicer, WF, Prochaska, JO, Rossi, JS, & Marcus, BH (1997). What makes a good staging algorithm: examples from regular exercise. American Journal of Health Promotion, 12(1), 57-66.Milkman, KL, Gromet, D., Ho, H., Kay, JS, Lee, TW, Pandiloski, P., ... & Duckworth, AL (2021). Megastudies improve the impact of applied behavioral science. Nature, 600(7889), 478-483.

[0006] However, although Non-Patent Document 1 takes into consideration individual characteristics such as the stage of behavioral change, it is difficult to accurately grasp the individual characteristics of the subject based on that alone, and there are cases where it is not possible to determine the optimal intervention method. The stage of behavioral change merely indicates the subject's current level of interest in a desirable behavior and how long they have been working on it, and just because the subjects are at the same stage does not necessarily mean that a specific intervention method will be equally effective for all subjects.

[0007] Furthermore, if the method described in Non-Patent Document 2 were to be applied to a mobile application for verification, it would be necessary to implement all of the hundreds of possible intervention methods as application functions and then conduct large-scale user testing over a certain period of time. This would require enormous development and research costs, making it unrealistic. In reality, it is common for an application to initially implement only a few main functions, with the functionality gradually expanded. Designers must identify, prioritize, and implement the most appropriate intervention methods for the intended service target users from the vast number of intervention methods. This decision is extremely difficult, so assistance is required.

[0008] The present disclosure has been made in consideration of the above circumstances, and aims to provide designers of applications, systems, or services for promoting behavioral change with optimal intervention methods for behavioral change for the intended target population.

[0009] In order to solve the above problem, the disclosure of claim 1 is a selection device that selects an intervention method for behavioral change to be provided to a subject, and the selection device has a learning model construction unit that constructs a machine learning model by machine learning the machine learning model using a rank learning method in a learning phase based on at least one of the subject's personal characteristics and the intervention method, and a score indicating the degree to which it leads to the behavioral change.

[0010] As described above, the present disclosure has the effect of enabling designers of applications, systems, or services for encouraging behavioral change to provide optimal intervention methods for behavioral change to intended subjects.

[0011] 1 is an overall configuration diagram of a selection device according to an embodiment; FIG. 2 is an electrical hardware configuration diagram of a selection device according to an embodiment; FIG. 3 is a conceptual diagram showing source data; FIG. 4 is a conceptual diagram of a learning dataset; FIG. 5 is a diagram showing an overview of the processing of each part of the selection device; FIG. 6 is a flowchart showing overall processing or operation according to an embodiment; FIG. 7 is a detailed flowchart of a learning data generation process; FIG. 8 is a detailed flowchart of a learning model construction process; FIG. 9 is a detailed flowchart of an inference process; and FIG. 10 is a detailed flowchart of a ranking output process.

[0012] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0013] [Overview of the Selection Device] The selection device 10 of this embodiment inputs personal characteristics of an application, system, or service (hereinafter sometimes referred to as "service, etc.") envisioned by a designer, and introduces a ranking learning framework into the selection of intervention methods, thereby enabling output of a prioritized list of intervention methods. Rank learning is a supervised machine learning technique that trains a ranking model for extracting documents highly relevant to a search query from multiple documents. Rank learning is often used in the field of web search, but in this embodiment, it is formulated to distinguish the relative magnitude of the effectiveness of intervention methods according to personal characteristics, and is used to clarify which intervention methods should be prioritized for consideration or implementation in the services, etc. envisioned by the designer in order to effectively promote behavioral change.

[0014] The selection device 10 is a selection device that selects a behavioral change intervention method to be provided to a subject (a specified individual), and is configured with one or more computers. When the selection device 10 is configured with multiple computers, it may be referred to as a "selection device" or a "selection system."

[0015] [Functional Configuration of Selection Device] Next, the functional configuration of the selection device 10 will be described using Figure 1. Figure 1 is a diagram showing the overall configuration of the selection device according to an embodiment. Here, we will explain how to output a prioritized list of intervention methods when the purpose is to promote walking in a behavior change application. The problem is formulated as ranking a group of candidate intervention methods using rank learning.

[0016] 1, the selection device 10 includes an original data acquisition unit 11, a learning data generation unit 12, a learning model construction unit 13, an inference unit 14, and a ranking output unit 15. Each of these units has a function realized by an instruction from a CPU 101 (described later in FIG. 2) based on a program. In addition, a original data storage unit 21, a learning data storage unit 22, and a learning model storage unit 23 are configured in the RAM 103 or the SSD 104.

[0017] (Original Data Acquisition Unit) The original data acquisition unit 11 first acquires original data for generating training data as preparation and stores it in the original data storage unit 21. The original data is assumed to be N pieces of one-to-many data, each of which is an arbitrary measure of personal characteristic q and a list r ∈ R_q of M_q candidate intervention methods for that characteristic. That is, the set of candidate intervention method lists r (not a single intervention method, but a group consisting of multiple intervention methods) for each personal characteristic is represented as R_q, and the number of elements in the set R_q is represented as M_q. This also indicates that there are q_1, ..., q_N individual characteristics q, and corresponding sets R_q_1, ..., R_q_N of candidate intervention methods.

[0018] Furthermore, a score y_(q,r) is assigned to indicate whether the intervention will lead to behavioral change in a subject with individual characteristic q. The original data can be collected using methods such as questionnaire survey results. For example, in the case of promoting walking, the original data can be obtained by presenting an intervention method to respondents and asking them to evaluate whether it will lead to motivation to walk.

[0019] Fig. 3 is a conceptual diagram showing the original data. As shown in Fig. 3, the original data is associated with information such as a user ID (identification) for identifying the target user, personal identification of the target user, and the degree to which each intervention method leads to behavioral change (score).

[0020] "Personal characteristics" include, for example, basic attribute information such as age and gender, stage of behavioral change, usual number of steps walked, target number of steps walked, or personality traits such as the Big-Five factor structure: Big 5 (extroversion, agreeableness, conscientiousness, neuroticism, openness to experience).

[0021] "Intervention methods" include, for example, rewards, cooperation, penalties, bonuses, competition in persuasion strategies, simulation, customization, personalization, points in gamification, badges, leaderboards, events, or levels.

[0022] "Simulation" refers to intervention methods that analyze factors that affect user behavior and present them. "Customization" refers to allowing users to configure the system's content and functions to suit their preferences. "Personalization" refers to presenting content that is personalized to the user's characteristics. "Badge" refers to presenting milestones that indicate the achievement of certain achievements. "Leaderboard" refers to a board or ranking that displays the user's score or points. "Level" refers to encouraging the achievement of difficult goals by dividing them into stages.

[0023] As the "score regarding whether it leads to behavioral change," for example, a value that evaluates whether it leads to motivation to walk (such as a 7-point Likert scale) can be used.

[0024] (Learning data generation unit) The learning data generation unit 12 uses the source data stored in the source data storage unit 21 to generate Σ_N M_q pairs of features X_(q, r) obtained from the input personal characteristics q and the list of candidate intervention methods r, and scores y_(q, r) indicating whether they will lead to behavioral change.

[0025] There are three main patterns for creating features, as shown below, and each feature is as follows: That is, the training data generation unit 12 may generate, as training data, pairs of features obtained from at least one of the input personal characteristics and the intervention method candidate group list, and a score regarding whether or not the feature leads to behavioral change. (1) Features obtained from personal characteristics (2) Features obtained from the intervention method candidate group list (3) Features obtained from pairs of personal characteristics and intervention candidate group list Here, as an example of (1), when "age" is used as the personal characteristic, a specific age value such as "25 years old" or a generation such as "20s" is used as the feature.

[0026] As an example of (2), if the strategy is "bonus," it falls under the category of "gamification" in behavioral change theory, so that classification can be used as a feature. Alternatively, the text sentence with the definition "Extra award that requires exceptional effort" can be used as a feature. Alternatively, a one-hot expression can be used to indicate whether a specific word is included in the document.

[0027] As an example of (3), for example, "TF-IDF" or "BM25" is used, which is a matching score between personal characteristics and a list of intervention candidate groups.

[0028] The score indicates a real number from 1 to 7 when using a seven-point Likert scale.

[0029] Then, the learning data generating unit 12 stores the learning data, which is the data set shown in (Equation 1), in the learning data storage unit 22.

[0030] Dataset = {(X_(q,r), y_(q,r)):q = 1, ... , N, r = 1, ... , M_q} ... (Equation 1) (Learning model construction unit) The learning model construction unit 13 constructs a machine learning model of a neural network by performing learning using the learning data stored in the learning data storage unit 22, and stores it in the learning model storage unit 23.

[0031] In this case, the learning model construction unit 13 performs machine learning using a rank learning method (see Reference 1). <Reference 1> T.-Y. Liu, Learning to Rank for Information Retrieval, Springer, 2011. Here, the rank learning method used in this embodiment will be described in more detail. Rank learning methods are classified into three types: pointwise, pairwise, and listwise, depending on the definition of the loss function, and any of the methods may be used in this embodiment.

[0032] (a) The pointwise method is based on the assumption that if the machine learning model's output value f(X_(q,r)) can be correctly inferred as a score y_(q,r) indicating whether it will lead to behavioral change, then sorting the output results will yield a correct ranking. The loss is defined as the error between the predicted value and the correct value, as shown in the following (Equation 2): L_pointwise = Σ_(q,r) L(f(X_(q,r)), y_(q,r)) ... (Equation 2) (b) The pairwise method is based on the hypothesis that if the ordering relationship between any two pairs can be correctly inferred (predicted), then that ordering relationship will result in a correct ranking. The loss is defined as the probability value or other expression of the plausibility of the ordering relationship between the two pairs, as shown in (Equation 3). L_pairwise = Σ_q Σ_(y_(q,i)>y_(q,j):(i,j)<=M_q) L(f(X_(q,i)), X_(q,j)) ... (Equation 3) (c) The listwise method is a method in which the output ranking itself is directly optimized using a ranking evaluation index such as MAP. The loss is defined as an expression of the error between all predicted values ​​for the query and the correct value, as shown in (Equation 4). L_listwise = Σ_q L(f(X_(q,1)), y_(q,1), ... ,f(X_(q,M_q)), y_(q,M_q) ) ... (Equation 4) From the above, a neural network machine learning model is constructed by updating the rank learning model so that the loss function is minimized. For example, using the squared error in the pointwise method, the problem can be solved as minimizing L(f(x_(q,r)),y_(q,r)) = (y_(q,r) - f(x_(q,r))^2).

[0033] (Inference unit) The inference unit 14 uses a trained machine learning model to infer and output a score y_(q,r) indicating whether the input personal characteristics and X_(q,r) in (Equation 1) will lead to behavioral change.

[0034] (Ranking Output Unit) The ranking output unit 15 outputs the ranking results by sorting the inference results based on the scores (for example, scores 1 to 7).

[0035] [Hardware Configuration of Selection Device] Next, the electrical hardware configuration of the selection device 10 will be described with reference to Fig. 2. Fig. 2 is a diagram showing the electrical hardware configuration of the selection device.

[0036] The selection device 10 is a computer that includes a CPU 101 , a ROM 102 , a RAM 103 , an SSD 104 , an external device connection I / F (Interface) 105 , a network I / F 106 , a display 107 , an input device 108 , a media I / F 109 , and a bus line 110 .

[0037] Of these, the CPU 101 controls the overall operation of the selection device 10. The ROM 102 stores programs such as IPL used to drive the CPU 101. The RAM 103 is used as a work area for the CPU 101.

[0038] The SSD 104 reads or writes various data under the control of the CPU 101. Note that instead of the SSD 104, a hard disk drive (HDD) may be used.

[0039] The external device connection I / F 105 is an interface for connecting various external devices, such as a display, a speaker, a keyboard, a mouse, a USB memory, and a printer.

[0040] The network I / F 106 is an interface for performing data communication via the communication network 100 .

[0041] The display 107 is a type of display means such as a liquid crystal display or organic electroluminescence (EL) display that displays various images.

[0042] The input device 108 is a keyboard, a pointing device, or the like, and is an example of an input means for accepting input operations such as characters, numbers, and various instructions.

[0043] The media I / F 109 controls reading and writing (storing) of data from and to a recording medium 109m such as a flash memory, etc. The recording medium 109m includes DVDs and Blu-ray Discs (registered trademarks).

[0044] The bus line 110 is an address bus, a data bus, etc. for electrically connecting the components such as the CPU 101 shown in FIG.

[0045] [Processing or Operation of the Embodiment] Next, processing or operation of the present embodiment will be described with reference to Fig. 5 to Fig. 10. Fig. 5 is an image diagram of the overall processing or operation according to the present embodiment. Fig. 6 is a flowchart showing the overall processing or operation according to the embodiment.

[0046] S11: In the learning phase, the source data acquisition unit 11 first acquires source data for generating learning data as preparation, and stores the source data in the source data storage unit 21.

[0047] S12: In the learning phase, the learning data generation unit 12 generates, as learning data, pairs of feature values ​​obtained from at least one of the input personal characteristics and the list of intervention method candidates, and scores indicating whether the feature values ​​will lead to behavioral change, as shown in Fig. 5. In Fig. 5, personal characteristics such as "indifference stage" and "male in his 40s" are shown. Also, scores (1, 2, 2, 7, 4, 5, etc.) are shown for each intervention method candidate.

[0048] S13: In the learning phase, the learning model construction unit 13 constructs a neural network machine learning model f capable of ranking candidate intervention methods by learning using the learning data stored in the learning data storage unit 22, and stores the model in the learning model storage unit 23.

[0049] S14: In the inference phase, the inference unit 14 uses the trained machine learning model f for scores to infer and output a score regarding whether the input predetermined personal characteristic will lead to behavioral change. In Figure 5, "maintenance phase female" is shown as the predetermined personal characteristic.

[0050] S15: In the inference phase, the ranking output unit 15 outputs the ranking results by sorting the inference results in descending order of score. In Fig. 5, the intervention method "bonus" has a score of "0.9", the intervention method "self-monitoring" has a score of "0.6", and the intervention method "competition" has a score of "0.3", and the ranking list is shown in descending order of score from top to bottom.

[0051] <Detailed Description of Each Process or Operation> Next, detailed description will be given of the processes or operations of steps S12 to S15 with reference to FIGS.

[0052] (Learning Data Generation Processing) FIG. 7 is a detailed flowchart of the learning data generation processing.

[0053] S21: The learning data generation unit 12 acquires the individual characteristics, intervention method, and score as source data from the source data storage unit 21.

[0054] S22: The learning data generating unit 12 generates learning data based on the individual characteristics, the intervention method, and the score.

[0055] S23: The learning data generating unit 12 stores the learning data in the learning data storage unit 22.

[0056] (Learning Model Construction Process) FIG. 8 is a detailed flowchart of the learning model construction process.

[0057] S31: The learning model construction unit 13 acquires learning data from the learning data storage unit 22.

[0058] S32: The learning model construction unit 13 constructs a machine learning model using the learning data.

[0059] S33: The learning model construction unit 13 stores the trained machine learning model in the learning model storage unit 23.

[0060] (Inference Processing) FIG. 9 is a detailed flowchart of the inference processing.

[0061] S41: The inference unit 14 acquires the personal characteristics of a predetermined subject.

[0062] S42: The inference unit 14 acquires the trained machine learning model from the training model storage unit 23.

[0063] S43: The inference unit 14 infers an inference result (score for each intervention method) corresponding to a predetermined individual characteristic.

[0064] S44: The inference unit 14 outputs the inference result to the ranking output unit 15.

[0065] (Ranking Output Processing) FIG. 10 is a detailed flowchart of the ranking output processing.

[0066] S51: The ranking output unit 15 acquires the inference result from the inference unit 14.

[0067] S52: The ranking output unit 15 sorts the intervention method candidates in a predetermined order (highest rank, lowest rank, etc.) based on the inference result.

[0068] S53: The ranking output unit 15 outputs the final result (a ranking list of behavioral modifications to be provided to the predetermined subject).

[0069] [Major Effects of the Embodiments] As described above, the embodiments have the effect of enabling designers of applications, systems, or services for promoting behavioral change to provide the optimal intervention method for behavioral change to the intended target (specified individual).

[0070] [Supplementary Note] The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations).

[0071] (1) The selection device 10 can be realized by a computer and a program, but this program can also be recorded on a (non-transitory) recording medium or provided via the communication network 100.

[0072] (2) The CPU 101 may be a single CPU or multiple CPUs.

[0073] REFERENCE SIGNS LIST 10 Selection device 11 Source data acquisition unit 12 Learning data generation unit 13 Learning model construction unit 14 Inference unit 15 Ranking output unit 21 Source data storage unit 22 Learning data storage unit 23 Learning model storage unit

Claims

1. A selection device that selects an intervention method for behavioral change to be provided to a subject, the selection device having a learning model construction unit that constructs a machine learning model in a learning phase by machine learning the machine learning model using a rank learning technique based on at least one of the subject's personal characteristics and the intervention method, and a score indicating the degree to which the intervention method leads to the behavioral change.

2. The selection device according to claim 1, wherein the rank learning technique is a pointwise technique, a pairwise technique, or a listwise technique.

3. The selection device according to claim 1, wherein the personal characteristics include basic attribute information such as age and gender, a stage of behavioral change, usual number of steps walked, target number of steps, or information on personality traits.

4. The selection device of claim 1, wherein the intervention methods include rewards, cooperation, penalties, bonuses, competition in persuasion strategies, simulation, customization, personalization, points, badges, leaderboards, events, or levels in gamification.

5. A selection device as described in any one of claims 1 to 4, comprising: an inference unit that, in an inference phase, uses a trained machine learning model constructed by the learning model construction unit to infer the score for each of the candidate intervention methods from the personal characteristics of a specified subject; and a ranking output unit that outputs a ranking list of intervention methods for behavior change to be provided to the specified subject by rearranging the inferred intervention candidates in a specified order.

6. A learning method executed by a selection device that selects an intervention method for behavioral change to be provided to a subject, the learning method comprising: a learning model construction process that executes, in a learning phase, a learning model construction process that constructs a machine learning model by machine learning using a rank learning technique based on the subject's personal characteristics and at least one of the intervention method, and a score indicating the degree to which it leads to the behavioral change.

7. A program for causing a computer to execute the method according to claim 6.

Images