Defection behavior prediction device, defection behavior prediction method, and program

Abstract

The purpose of the present disclosure is to predict the occurrence probability of an estrangement act with high accuracy.　Provided is an estrangement act prediction device for predicting the occurrence probability of an estrangement act, said device evaluating conversation data, including conversation partner information for identifying a partner of conversation, that indicates conversation content and a conversion state, so as to obtain a conversation content evaluation value and a conversation state evaluation value; evaluating a relationship with the partner of conversation on the basis of relationship data that indicates a human relationship with a user of interest and the conversation partner information, so as to obtain a relationship evaluation value; calculating a social interaction effect on the basis of the conversation content evaluation value, the conversation state evaluation value, and the relationship evaluation value; calculating a temporal attenuation value of the social interaction effect on the basis of a temporal attenuation function; and calculating the occurrence probability of an estrangement act.

Description

[Technical field]

[0001] The present disclosure relates to a churn behavior prediction device, a churn behavior prediction method, and a program. [Background technology]

[0002] In recent years, there has been a demand for predicting with high accuracy the behavior of users (defective behavior) that goes against the achievement of goals set by the users themselves, taking into account the social interaction of users.

[0003] For example, a person may take in more than 3500 kcal per day despite trying to limit their calorie intake to 2000 kcal or less, or may gain more than 5 kg despite trying to limit their monthly weight gain to less than 1 kg. Such defection behavior is known to be closely related to social interactions. For example, Non-Patent Document 1 showed that the frequency of actions such as following and replying to users in a weight management app and information about the other person (e.g., whether they are of the opposite sex or not, whether they are thin or not, etc.) are important for predicting the achievement rate of the target weight. In addition, Non-Patent Document 2 showed that people who successfully maintain their weight after losing weight often have experiences of being praised and working together with family and friends to lose weight, rather than simply being encouraged or urged, in their interactions with family and friends. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Wang, Zhiwei, et al. "Understanding and predicting weight loss with mobile social networking data." Proceedings of the 2017 ACM on Conference on Information and Knowledge Management. 2017. DOI: https: / / doi.org / 10.1145 / 3132847.3133019 [Non-Patent Document 2] Karfopoulou, Eleni, et al. "The role of social support in weight loss maintenance: results from the MedWeight study." Journal of behavioral medicine 39.3 (2016): 511-518. Summary of the Invention [Problem to be solved by the invention]

[0005] However, although these conventional techniques and knowledge are certainly useful for predicting defection behavior, they still have the following problems.

[0006] First, the prediction performance of churn behavior is insufficient (Non-Patent Document 1). Information on a user's follow relationship and reply frequency certainly improves the prediction performance of churn behavior, but Non-Patent Document 1 reports that the prediction ability is lower than basic attribute information such as the user's age, gender, and BMI (Body Mass Index).

[0007] Next, the evaluation of social interactions is limited. In other words, the evaluation of the relationship with the other person and the quality of the conversation is not comprehensive. Specifically, while basic information about the other person is taken into account in the evaluation of the relationship (Non-Patent Documents 1 and 2), the psychological or physical distance between the other person in social interactions, such as the sense of closeness to the other person, the sense of trust, the strength of the connection, the length of time the other person has known the other person, and the distance in communication, is not evaluated. In addition, the content of the conversation is not taken into account in the first place (Non-Patent Document 1), or only the presence or absence of interactions directly related to weight loss is taken into account (Non-Patent Document 2). In other words, the means of the conversation, the time required for the conversation, the topic of the conversation, and the impression of the conversation are not evaluated.

[0008] Finally, the fact that the influence of social interactions attenuates (changes) over time is not taken into consideration (Non-Patent Documents 1, 2). If the same social interaction occurred once in the near past and once in the distant past from the target defection behavior, the conventional technology considers these social interactions to have the same influence on the defection behavior. In reality, it is assumed that the influence of the one that occurred in the distant past will attenuate over time, so the influence of the one that occurred near and in the past will be relatively large, but the conventional technology does not take this into account.

[0009] In other words, the problems that the present invention aims to solve are: (1) the evaluation of the people involved in social interactions and the nature of the conversation is not comprehensive; (2) the decay of the influence of social interactions over time is not taken into account; and (3) the predictive performance of defection behavior is low.

[0010] The object of the present invention is to solve the problems of the above-mentioned conventional technology by evaluating the nature of the other parties and conversations in a user's social interactions and taking into account the time-dependent decay of the influence of social interactions, thereby accurately predicting the probability of occurrence of defection behavior. [Means for solving the problem]

[0011] In order to achieve the above object, the invention of claim 1 provides a defection behavior prediction device for predicting the probability of defection behavior, comprising: a conversation data evaluation unit that evaluates conversation data including conversation partner information for identifying a conversation partner, and indicating a conversation content and a conversation situation, to obtain a conversation content evaluation value and a conversation situation evaluation value; a relationship data evaluation unit that evaluates a relevance to a conversation partner based on relationship data indicating a human relationship with a target user and the conversation partner information, to obtain a relationship evaluation value; a social interaction effect calculation unit that calculates a social interaction effect based on the conversation content evaluation value, the conversation situation evaluation value, and the relationship evaluation value; a social interaction effect time decay processing unit that calculates a time decay value of the social interaction effect based on a time decay function; and a defection behavior prediction unit that calculates the probability of defection behavior based on the social interaction effect that has been time decay processed. The defection behavior prediction device has the above-mentioned features. Effect of the Invention

[0012] As described above, according to the present invention, by evaluating the nature of the other party and conversation in a user's social interactions and taking into account the time-dependent decay of the influence of social interactions, it is possible to predict with high accuracy the probability of occurrence of defection behavior. [Brief description of the drawings]

[0013] [Figure 1] 1 is a diagram showing a mechanical configuration of a defection behavior prediction device according to an embodiment of the present invention. [Diagram 2] 1 is a diagram showing a mechanical configuration of a defection behavior prediction device according to an embodiment of the present invention. [Diagram 3] FIG. 2 is a hardware configuration diagram of the defection behavior prediction device. [Figure 4] 13 is a flowchart showing the processing of the defection behavior prediction device (learning mechanism). [Diagram 5] 13 is a flowchart showing the processing of the defection behavior prediction device (learning mechanism). [Figure 6] 11 is a flowchart showing the processing of a defection behavior prediction device (prediction mechanism). [Figure 7]3 is a diagram showing an example of a storage format of a conversation data storage unit 102. FIG. [Figure 8] FIG. 13 is a diagram showing an example of an output format of a conversation data evaluation value. [Figure 9] FIG. 13 is a diagram showing an example of an output format of conversation partner information. [Figure 10] FIG. 13 is a diagram illustrating an example of an output format of a relationship evaluation value. [Figure 11] FIG. 13 is a diagram showing an example of an output format of a social interaction effect. [Figure 12] 13 is a diagram showing an example of a storage format of a target behavior data storage unit when a daily calorie intake is set as a target behavior. FIG. [Figure 13] 1A and 1B are diagrams showing examples of output formats of churn levels, in which (a) shows a case where the churn level is expressed by a discrete value, and (b) shows a case where the churn level is expressed by a continuous value. [Figure 14] FIG. 13 illustrates an example of an output format for time-decayed social interaction effects. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0015] [Functional configuration] 1 and 2 are diagrams showing the configuration of a defection behavior prediction device according to an embodiment of the present invention. The functional configuration comprises a learning mechanism (FIG. 1) and a prediction mechanism (FIG. 2).

[0016] The learning mechanism of the defection behavior prediction device 10 (Figure 1) is composed of a target behavior data storage unit 101, a conversation data storage unit 102, a relationship data storage unit 103, a defection behavior standard constant input unit 104, a defection level calculation unit 105, a conversation data evaluation unit 106, a relationship data evaluation unit 107, a social interaction effect calculation unit 108, a time decay function input unit 109, a social interaction effect time decay processing unit 110, a time decay function storage unit 111, a defection behavior prediction model learning unit 112, and a defection behavior prediction model storage unit 113 shown in Figure 1, and outputs a model that has learned the relationship between the magnitude of the effect of social interaction taking into account time decay and the defection level of the target behavior and its parameters.

[0017] Of these, the target behavior data storage unit 101 stores the target value for the target behavior, the actual measured value of the target behavior, and the time when the target behavior was recorded, linked to a user ID (identification) and a behavior ID.

[0018] The conversation data storage unit 102 stores information about the content of the conversation, information about the conversation partner, and the time when the conversation was recorded, linked to a user ID and a conversation ID.

[0019] The relationship data storage unit 103 stores identification information of people who are connected to the target user (have a human relationship) in association with the user ID.

[0020] The prediction mechanism of the defection behavior prediction device 10 (Figure 2) is composed of a conversation data evaluation unit 201, a relationship data evaluation unit 202, a social interaction effect calculation unit 203, a time decay function storage unit 204, a social interaction effect time decay processing unit 205, a defection behavior prediction model storage unit 206, and a defection behavior prediction unit 207, and predicts defection behavior (probability of occurrence) based on the input conversation data and relationship data.

[0021] [Hardware configuration] Next, a hardware configuration of the churn behavior prediction device 10 will be described with reference to Fig. 3. Fig. 3 is a hardware configuration diagram of the churn behavior prediction device.

[0022] 3, the churn behavior prediction device 10 includes a processor 401, a memory 402, an auxiliary storage device 403, a connection device 404, a communication device 405, and a drive device 406. The hardware components constituting the churn behavior prediction device 10 are connected to each other via a bus 407.

[0023] The processor 401 serves as a control unit that controls the entire defection behavior prediction device 10, and includes various arithmetic devices such as a CPU (Central Processing Unit). The processor 401 reads various programs onto the memory 402 and executes them. The processor 401 may include a GPGPU (General-purpose computing on graphics processing units).

[0024] The memory 402 has a main storage device such as a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The processor 401 and the memory 402 form a so-called computer, and the processor 401 executes various programs read onto the memory 402, causing the computer to realize various functions.

[0025] The auxiliary storage device 403 stores various programs and various information used when the various programs are executed by the processor 401 .

[0026] The connection device 404 is a connection device that connects an external device (for example, a display device 410, an operation device 411) and the defection behavior prediction device 10.

[0027] The communication unit 405 is a communication device for transmitting and receiving various types of information to and from other devices.

[0028] The drive unit 406 is a device for setting the recording medium 430. The recording medium 430 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM (Compact Disc Read-Only Memory), a flexible disk, a magneto-optical disk, etc. The recording medium 430 may also include semiconductor memories that record information electrically, such as a ROM (Read Only Memory), a flash memory, etc.

[0029] The various programs to be installed in the auxiliary storage device 403 are installed, for example, by setting the distributed recording medium 430 in the drive device 406 and reading out the various programs recorded in the recording medium 430 by the drive device 406. Alternatively, the various programs to be installed in the auxiliary storage device 403 may be installed by being downloaded from a network via the communication device 405.

[0030] [Processing or Operation of the Present Embodiment] Next, the processing or operation of this embodiment will be described with reference to FIGS.

[0031] 4 to 6 are flowcharts showing the processing of the defection behavior prediction device. The processing of the defection behavior prediction device differs between the learning mechanism (FIGS. 4 and 5) and the prediction mechanism (FIG. 6).

[0032] <Processing procedure in the learning mechanism> S100: The churn behavior criterion constant input unit 104 accepts an input of a churn behavior criterion constant by the operating subject of the churn behavior prediction device 10, and outputs the input churn behavior criterion constant to the churn level calculation unit 105. The definition of the churn behavior criterion constant and the specific processing procedure of S100 will be described later.

[0033] S101: The conversation data evaluation unit 106 receives conversation data from the conversation data storage unit 102, evaluates the conversation data, and outputs information about the conversation partner to the relationship data evaluation unit 107 and an evaluation value about the content and situation of the conversation to the social interaction effect calculation unit 108. The conversation data is data in which the content of the conversation, the situation, and the information of the other party are recorded in chronological order. The data format about the content of the conversation does not matter (text, voice, image, etc.). The data about the conversation partner is identification information of the conversation partner, and the format does not matter as long as it is information that can uniquely identify the conversation partner (name, My Number number used in Japan, etc.). The conversation data evaluation unit 106 evaluates the conversation content, the conversation partner, and the conversation situation from the conversation data. Examples of evaluation items of the conversation content include the time required for the conversation, the number of characters, the topic, and the impression, but are not limited to these and may be arbitrarily determined by the operator of the defection behavior prediction device 10. Examples of evaluation items of a conversation partner include familiarity, reliability, the time since acquaintance, and frequency of contact, but are not limited to these and may be arbitrarily determined by the operator of the defection behavior prediction device 10. Examples of evaluation items of a conversation situation include physical distance from the conversation partner, the means of conversation, and the place of conversation, but are not limited to these and may be arbitrarily determined by the operator of the defection behavior prediction device 10. FIG. 7 shows an example of a storage format of the conversation data storage unit 102. FIG. 8 shows an example of an output format of the conversation data evaluation value from the conversation data evaluation unit 106 to the social interaction effect calculation unit 108 in S101. FIG. 9 shows an example of an output format of conversation partner information from the conversation data evaluation unit 106 to the relationship data evaluation unit 107 in S101. Specific processing procedures will be described later.

[0034] S102: The relationship data evaluation unit 107 receives relationship data from the relationship data storage unit 103 and conversation partner information from the conversation data evaluation unit 106, evaluates the relationship with the conversation partner, and outputs the evaluation result to the social interaction effect calculation unit 108. The relationship data is identification information of a person with whom the target user has a connection, and the format does not matter as long as the information can uniquely identify the person. Examples of evaluation items of the relationship include closeness, reliability, acquaintance period, and contact frequency, but are not limited to these and may be arbitrarily determined by the operator of the defection behavior prediction device 10. FIG. 10 shows an example of the output format of the evaluation value of the relationship by S102. The specific processing procedure will be described later.

[0035] S103: The social interaction effect calculation unit 108 receives the conversation content evaluation value and the conversation situation evaluation value from the conversation data evaluation unit 106, and the relationship evaluation value from the relationship data evaluation unit 107, calculates a social interaction effect, and outputs it to the social interaction effect time decay processing unit 110. The social interaction effect corresponding to a certain target behavior of a certain user is calculated from the conversation content evaluation value, conversation situation evaluation value, and relationship evaluation value, or a part of them, in a conversation that occurred before the target behavior.

[0036] 11 shows an example of the output format of the social interaction effect by S103 (an example in which the magnitude of the social interaction effect is determined by the product of the evaluation value of the impression of the conversation content and the evaluation value of the closeness of the conversation partner). The specific processing procedure will be described later.

[0037] S104: The time decay function input unit 109 accepts an input of a time decay function by the subject operating the churn behavior prediction device 10, and outputs the input time decay function to the social interaction effect time decay processing unit 110 and the time decay function storage unit 111. The time decay function may be arbitrarily determined by the subject operating the churn behavior prediction device 10. Specific processing will be described later.

[0038] S105: The defection level calculation unit 105 receives the target behavior data from the target behavior data storage unit 101 and the defection behavior reference constant from the defection behavior reference constant input unit 104, calculates the defection level of the target behavior, and outputs it to the defection behavior prediction model learning unit 112. The defection level may be calculated as a discrete value or a continuous value. FIG. 12 shows an example of the storage format of the target behavior data storage unit 101. FIG. 13 shows an example of the output format of the defection level by S105. Specific processing procedures will be described later.

[0039] S106: The social interaction effect time decay processing unit 110 receives the social interaction effect from the social interaction effect calculation unit 108 and the time decay function from the time decay function input unit 109, calculates a time decay value of the social interaction effect, and outputs it to the defection behavior prediction model learning unit 112. Fig. 14 shows an example of the output format of the social interaction effect after time decay processing. The specific processing procedure will be described later.

[0040] S107: The defection behavior prediction model learning unit 112 receives the defection level of the target behavior from the defection level calculation unit 105 and the time-decayed social interaction effect from the social interaction effect time-decay processing unit 110, learns a model predicting the defection level from the time-decayed social interaction effect, and outputs the learned model to the defection behavior prediction model storage unit 113. The model used here may be any model based on supervised learning. Specific processing procedures will be described later.

[0041] <Processing procedure in prediction mechanism> S200: The conversation data evaluation unit 201 receives conversation data as input, evaluates the conversation data, and outputs evaluation values ​​of the conversation content and conversation situation to the social interaction effect calculation unit 203 and conversation partner information to the relationship data evaluation unit 202.

[0042] S201: The relationship data evaluation unit 202 receives relationship data as input, and also receives conversation partner information from the conversation data evaluation unit 201, evaluates the relationship with the conversation partner, and outputs the evaluation to the social interaction effect calculation unit 203.

[0043] S202: The social interaction effect calculation unit 203 receives the conversation content evaluation value and the conversation situation evaluation value from the conversation data evaluation unit 201 and the relationship evaluation value from the relationship data evaluation unit 202, calculates a social interaction effect based on these, and outputs it to the social interaction effect time decay processing unit 205.

[0044] S203: The social interaction effect time decay processing unit 205 receives the social interaction effect from the social interaction effect calculation unit 203 and the time decay function from the time decay function storage unit 204, calculates the time decay value of the social interaction effect, and outputs it to the defection behavior prediction unit 207.

[0045] S204: The defection behavior prediction unit 207 receives the time-decay processed social interaction effect from the social interaction effect time decay processing unit 205 and the defection behavior prediction model from the defection behavior prediction model storage unit 206, calculates and outputs the probability of defection behavior occurring.

[0046] <Specific processing steps> The specific processing procedure is described below.

[0047]

number

[0048]

number

[0049]

number

[0050]

number

[0051]

number

[0052]

number

[0053]

number

[0054]

number

[0055]

number

[0056]

number

[0057]

number

[0058] The set of conversation data for all users belonging to the user set U is

[0059]

number

[0060]

number

[0061]

number

[0062] Here, the process in the defection behavior criterion constant input unit 104 (FIG. 4: S100) will be described. The defection behavior criterion constant is a real value equal to or greater than 0 that determines how much a measured value of a certain target behavior must deviate from its target value to be considered a defection behavior. Hereinafter, the defection behavior criterion constant is denoted as k. The defection behavior criterion constant is a value for determining the percentage by which the measured value of the target behavior deviates from the target value. For example, when the operator of the defection behavior prediction device 10 determines a rule that "if the measured value deviates from the target value by +20% or more, the target behavior that recorded the measured value is considered to be a defection behavior," the operator of the defection behavior prediction device 10 sets k=0.2 and inputs it to the defection behavior criterion constant input unit 104. The defection behavior criterion constant input unit 104 outputs the input defection behavior criterion constant k to the defection level calculation unit 105.

[0063] The process (S101 in FIG. 4) in the conversation data evaluation unit 106 will be described. The conversation data evaluation unit 106 receives conversation data C from the conversation data storage unit and evaluates the contents and situation of the conversation. A function G_C is used to derive a conversation content evaluation value using conversation data c_(u, m) as an input, and a function G_S is used to derive a conversation situation evaluation value. Examples of evaluation items for the conversation content evaluation value include conversation time, number of characters, topic, and impression, but are not limited to these, and are arbitrarily determined by the operator of the defection behavior prediction device 10. Evaluation methods include a questionnaire method in which the user u is actually asked to evaluate, an automatic collection method using a Bluetooth sensor, a GPS sensor, dialogue topic extraction, dialogue summarization technology, or a combination thereof, but are not limited to these, and are arbitrarily determined by the operator of the defection behavior prediction device 10. Examples of evaluation items for the conversation situation evaluation value include distance, means, and location from the conversation partner, but are not limited to these, and are arbitrarily determined by the operator of the defection behavior prediction device 10. The evaluation method is mainly a questionnaire method in which the user is asked to evaluate, but is not limited to this and can be arbitrarily determined by the operator of the churn behavior prediction device 10.

[0064] For a conversation m of a certain user u, let j_C∈J_C be an evaluation item with the conversation content, and let j_S∈J_S be an evaluation item with the conversation situation. c The evaluation value of e u,m,js Then, it can be written as follows (Equation 1) and (Equation 2).

[0065]

number

[0066]

number

[0067] The process in the relationship data evaluation unit 107 (FIG. 4: S102) will be described. The relationship data evaluation unit 107 receives relationship data R from the relationship data storage unit 103 and conversation partner information V from the conversation data evaluation unit 106, and evaluates the relationship with the person who became the conversation partner. The evaluation items of the relationship include, but are not limited to, closeness, reliability, duration, and frequency of contact, and are arbitrarily determined by the operator of the defection behavior prediction device 10. The partner of a certain conversation m of a certain user u is v u,m In this case, the set of all conversation partners of user u is

[0068]

number

[0069] A conversation partner of a user

[0070]

number

[0071]

number

[0072]

number

[0073] The process in the social interaction effect calculation unit 108 (FIG. 4: S103) will be described. The social interaction effect refers to the strength of the influence of a conversation on a user. The social interaction effect calculation unit 108 receives the conversation content evaluation value G C (c u,m ) and conversation situation evaluation value G S (c u,m ) is output from the relationship data evaluation unit 107 as a relationship evaluation value G R (v u ) and calculate the social interaction effect based on these values. Here, as an example, the evaluation value of the impression of the conversation content of a certain user u in a certain conversation m is

[0074]

number

[0075]

number

[0076] Let z be the magnitude of the social interaction effect that a conversation m of a certain user u has on a certain user u. u,m Then, it can be written as follows (Equation 4).

[0077]

number

[0078] The process in the time decay processing function input unit 119 (FIG. 4: S104) will be described. The time decay function is a function that expresses the manner in which the magnitude of the social interaction effect decays over time. The time decay function is determined by a function such as a linear function, a step function, a hyperbolic function, or a sigmoid function, but is not limited to these, and can be arbitrarily determined by the operating subject of the churn behavior prediction device 10 as long as it follows the above definition. Hereinafter, the time decay function will be represented as d. Here, as an example, a case in which the sigmoid function shown in (Equation 5) is adopted will be shown. The operating subject of the churn behavior prediction device 10 inputs the following time decay function d to the time decay function input unit 109, and the time decay function input unit 109 outputs the input time decay function d to the social interaction effect time decay processing unit 110.

[0079]

number

[0080]

number

[0081]

number

[0082]

number

[0083] In addition, if the defection levels are defined as three levels according to the rule that "if the measured value of the target behavior is +20% or more of the target value, the defection level is 1; if it is +50% or more, the defection level is 2; otherwise, the defection level is 0," then k = 0.2, n = {1, 2.5, ∞} are set.

[0084] The defection level calculation unit 105 calculates defection levels y u,i and outputs it to the defection behavior prediction model learning unit 112.

[0085] The process in the social interaction effect time decay processing unit 110 (FIG. 4: S106) will be described. The social interaction effect time decay processing unit 110 receives the time interaction effect z u,m The time decay function d is received from the time decay function input unit 109, and the time decay value x of the social interaction effect is calculated. u,m Here, we assume that the social interaction effect decays as determined by the time decay function depending on the time between the recording time of the conversation and the recording time of the target behavior, and calculate x u,m Calculate.

[0086]

number

[0087] The process in the defection behavior prediction model learning unit 112 (FIG. 4: S107) will be described. The defection behavior prediction model learning unit 112 receives the defection level y u,i The time decay processed social interaction effect x u,m Receive x u,m From u,i Here, as an example, a case where a binary churn level calculated as a discrete value is received from the churn level calculation unit 105 will be described.

[0088] The structure of the prediction model can be, for example, a gistic regression, a support vector machine, or a long short term memory (LSTM), but can be arbitrarily determined by the operator of the defection behavior prediction device 10 as long as it is applicable to supervised learning. However, when learning the prediction model, y u,i x to predict u,m Yes u,i In other words, the defection behavior prediction model learning unit 112 must use the one recorded earlier (in the past).

[0089]

number

[0090] Here is an example:

[0091]

number

[0092] where x u,i A social interaction effect vector with elements

[0093]

number

[0094]

number

[0095]

number

[0096]

number

[0097]

number

[0098]

number

[0099]

number

[0100]

number

[0101]

number

[0102]

number

[0103]

number

[0104]

number

[0105]

number

[0106] The defection behavior prediction model learning unit 112 obtains the parameter θ * Calculate the defection prediction model F and the optimal parameters θ * to the defection behavior prediction model storage unit 113.

[0107] Next, the process in the conversation data evaluation unit 201 (FIG. 6: S200) will be described. The conversation data evaluation unit 201 receives conversation data as input, and evaluates the content, situation, and other person of the conversation. Here, the input conversation data is assumed to be c. The conversation data c is calculated based on the conversation content evaluation function G C and the conversation situation evaluation function G S and outputs the evaluation result to the social interaction effect calculation unit 203. In addition, the conversation partner information v of the conversation partner is output to the relationship data evaluation unit 202.

[0108] The process in the relationship data evaluation unit 202 (FIG. 6: S201) will be described. The relationship data evaluation unit 202 receives relationship data as an input, and also receives conversation partner information v from the conversation data evaluation unit 201, and evaluates the relationship data of conversation partner v using a relationship evaluation function G R and outputs the evaluation result to the social interaction effect calculation unit 203.

[0109] The process in the social interaction effect calculation unit 203 (FIG. 6: S202) will be described. The social interaction effect calculation unit 203 receives the conversation content evaluation value G C (c) and conversation situation evaluation value G S (c) is output from the relationship data evaluation unit 202 as a relationship data evaluation value G R (v) is received and the social interaction effect is calculated.

[0110] As an example, we use the social interaction effect z as the evaluation value of the impression of the conversation content.

[0111]

number

[0112]

number

[0113]

number

[0114] The processing in the social interaction effect time decay processing unit 205 (FIG. 6: S203) will be described. The social interaction effect time decay processing unit 205 receives the social interaction effect z from the social interaction effect calculation unit 203 and the time decay function d from the time decay function storage unit 204, and calculates the time decay value of the social interaction effect. When the time when the defection behavior to be predicted occurred is t, the time when the conversation used to calculate the social interaction effect took place is s, and the time decay function d is a sigmoid function, the time decay processed social interaction effect x is calculated as follows.

[0115]

number

[0116] The process in the defection behavior prediction unit 207 (FIG. 6: S204) will be described. The defection behavior prediction unit 207 receives the time-decayed social interaction effect x from the social interaction effect time-decay processing unit 205, the prediction model F and the learned parameters θ * and predicts the value (probability) y of the defection behavior as follows:

[0117]

number

[0118] This concludes the description of the specific processing procedure.

[0119] [Main Effects of the Embodiments] As described above, according to this embodiment, the defection behavior prediction device 10 comprehensively evaluates the relationship with the other party in the social interaction and the content of the conversation, and captures the time decay of the influence of the social interaction. This has the effect of improving the prediction performance of defection behavior.

[0120] 〔supplement〕 The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations).

[0121] The defection behavior prediction device 10 can be realized by a computer and a program, and this program can be recorded on a (non-transitory) recording medium or provided via a network such as the Internet. [Explanation of symbols]

[0122] 10. Defection behavior prediction device 101 Target behavior data storage unit 102 Conversation data storage unit 103 Relationship Data Storage 104 Defection behavior standard constant input section 105 Defection Level Calculation Department 106 Conversation Data Evaluation Department 107 Relational Data Evaluation Department 108 Social Interaction Effect Calculation Department 109 Time decay function input section 111 Time decay function storage section 112 Defection behavior prediction model learning part 113 Defection behavior prediction model storage section 201 Conversation Data Evaluation Department 202 Relational Data Evaluation Department 203 Social Interaction Effect Calculation Department 204 Time decay function storage section 205 Social interaction effect time decay processing unit 206 Defection behavior prediction model storage section 207 Defection Behavior Prediction Department

Claims

1. A defection behavior prediction device for predicting a defection behavior occurrence probability, a conversation data evaluation unit that evaluates conversation data including conversation partner information for identifying a conversation partner and indicating a conversation content and a conversation situation to obtain a conversation content evaluation value and a conversation situation evaluation value; a relationship data evaluation unit that evaluates a relevance with the conversation partner based on relationship data indicating a human relationship with the target user and the conversation partner information to obtain a relationship evaluation value; a social interaction effect calculation unit that calculates a social interaction effect based on the conversation content evaluation value, the conversation situation evaluation value, and the relationship evaluation value; a social interaction effect time decay processing unit that calculates a time decay value of the social interaction effect based on a time decay function; a defection behavior prediction unit that calculates a defection behavior occurrence probability based on the time-decayed social interaction effect; The defection behavior prediction device has the following features.

2. The defection behavior prediction device according to claim 1, A defection behavior criterion constant input unit that receives an input of a defection behavior criterion constant; a defection level calculation unit that receives target behavior data from the target behavior data storage unit, receives a defection behavior criterion constant from the defection behavior criterion constant input unit, and calculates a defection level of the target behavior; a time decay function input unit that accepts an input of a time decay function; having the conversation data evaluation unit receives the conversation data from the conversation data storage unit, evaluates the conversation data, and obtains the conversation content evaluation value and the conversation situation evaluation value; the relationship data evaluation unit receives the relationship data from the relationship data storage unit, receives the conversation partner information from the conversation data evaluation unit, and evaluates a relevance with the conversation partner based on the relationship data and the conversation partner information to obtain a relationship evaluation value; the social interaction effect calculation unit calculates the social interaction effect based on the conversation content evaluation value, the conversation situation evaluation value, and the relationship evaluation value; the social interaction effect time decay processing unit receives the social interaction effect from the social interaction effect calculation unit, receives the time decay function from the time decay function input unit, and calculates a time decay value of the social interaction effect; Furthermore, the defection behavior prediction device A defection behavior prediction device having a defection behavior prediction model learning unit that receives the defection level of the target behavior from the defection level calculation unit, receives the social interaction effect after time decay processing from the social interaction effect time decay processing unit, learns a model that predicts the defection level from the social interaction effect after time decay processing, and outputs the learned model to a defection behavior prediction model storage unit.

3. A method executed by a defection behavior prediction device for predicting a defection behavior occurrence probability, comprising: a conversation data evaluation step of evaluating conversation data including conversation partner information for identifying a conversation partner and indicating a conversation content and a conversation situation to obtain a conversation content evaluation value and a conversation situation evaluation value; a relationship data evaluation step of evaluating a relationship with the conversation partner based on relationship data indicating a human relationship with the target user and the conversation partner information to obtain a relationship evaluation value; a social interaction effect calculation step of calculating a social interaction effect based on the conversation content evaluation value, the conversation situation evaluation value, and the relationship evaluation value; a social interaction effect time decay processing step of calculating a time decay value of the social interaction effect based on a time decay function; A defection behavior prediction step of calculating a probability of defection behavior occurrence based on the time-decayed social interaction effect; A method for predicting defection behavior.

4. A program causing a computer to execute the method according to claim 3.

Images