Action prediction program, action prediction device and action prediction method

Abstract

To provide a specific action during an actable time.SOLUTION: An action prediction device 100 acquires information on a user's internal state, predicts the user's action pattern in at least one of a plurality of time sections obtained by dividing a preset period into predetermined time units by analyzing the acquired information on the user's action pattern, and based on the predicted action pattern, specifies, from among the plurality of time sections, a time section in which the predetermined action is recommended to the user. Processing of the action prediction device 100 can be applied to a service that recommends exercise or the like during free time.SELECTED DRAWING: Figure 1

Description

[Technical Field]

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

[0002] Conventionally, there is a technology that presents a user's future status in diary format based on predicted leisure time and total task time (see, for example, Non-Patent Document 1). This technology generates a linear regression model that takes the total time of a set schedule for each user as input and outputs the result of adding the leisure time and total task time. This technology then inputs the set schedule for the next day into the generated model, and predicts leisure time and task time. By using this technology, it is possible to suggest exercise as a predetermined activity during the predicted leisure time. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2021 / 161469 [Patent Document 2] Japanese Patent Application Laid-Open No. 2016-189059 [Patent Document 3] Japanese Patent Application Publication No. 2019-004924 [Patent Document 4] Japanese Patent Application Laid-Open No. 2012-143345 [Non-patent literature]

[0004] [Non-Patent Document 1] Task management method using future predictions based on life logs and schedules https: / / cir.nii.ac.jp / crid / 1050282812882358912 Summary of the Invention [Problem to be solved by the invention]

[0005] However, with conventional technology, even if exercise is suggested as a predetermined activity during predicted leisure time, the user may not actually take the activity. In other words, conventional technology is unable to suggest a specific activity during a time when the user can actually take the activity.

[0006] In one aspect, the goal is to suggest specific actions when action is available. [Means for solving the problem]

[0007] In one aspect, the behavior prediction program causes a computer to execute a process of acquiring information regarding a user's internal state, analyzing the acquired information regarding the user's internal state, predicting the user's behavior pattern in at least one of a plurality of time segments obtained by dividing a preset period into predetermined time units, and identifying a time segment among the plurality of time segments in which a predetermined behavior will be recommended to the user based on the predicted behavior pattern. [Effects of the Invention]

[0008] On the one hand, it can suggest specific actions at times when action is possible. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram showing an example of the flow of behavior prediction according to the first embodiment. [Figure 2] FIG. 2 is a diagram illustrating an example of the configuration of the behavior prediction system according to the first embodiment. [Figure 3] FIG. 3 is a diagram illustrating an example of the configuration of the behavior prediction device according to the first embodiment. [Figure 4] FIG. 4 is a diagram illustrating an example of the configuration of the learning unit according to the first embodiment. [Figure 5] FIG. 5 is a diagram illustrating an example of the action log. [Figure 6] FIG. 6 is a diagram illustrating an example of cumulative calculation according to the first embodiment. [Figure 7]FIG. 7 is a diagram showing an example of the expected correlation feature extraction process according to the first embodiment. [Figure 8] FIG. 8 is a diagram showing an example of the internal behavior factor feature quantity extraction process according to the first embodiment. [Figure 9] FIG. 9 is a diagram illustrating an example of the configuration of a prediction unit according to the first embodiment. [Figure 10] FIG. 10 is a diagram showing an example of the data flow of the learning process according to the first embodiment. [Figure 11] FIG. 11 is a diagram showing an example of the data flow of the prediction process according to the first embodiment. [Figure 12] FIG. 12 is a diagram showing an example of a flowchart of the learning process according to the first embodiment. [Figure 13] FIG. 13 is a diagram illustrating an example of a flowchart of the prediction process according to the first embodiment. [Figure 14] FIG. 14 is a diagram illustrating an example of the configuration of a learning unit according to the second embodiment. [Figure 15] FIG. 15 is a diagram illustrating an example of the configuration of a prediction unit according to the second embodiment. [Figure 16] FIG. 16 is a diagram showing an example of the behavioral factor feature amount extraction process according to the second embodiment. [Figure 17] FIG. 17 is a diagram showing an example of an output screen according to the first and second embodiments. [Figure 18] FIG. 18 is a diagram showing a reference example of an output screen according to the prior art. [Figure 19] FIG. 19 is a diagram illustrating an example of the hardware configuration of a behavior prediction device according to the first and second embodiments. DETAILED DESCRIPTION OF THE INVENTION

[0010] Examples of the behavior prediction program, behavior prediction device, and behavior prediction method according to the present embodiment will be described in detail below with reference to the accompanying drawings. Note that the present embodiment is not limited to these examples. Furthermore, the examples can be combined as appropriate within a consistent range.

[0011] [Embodiment 1] (Flow of behavior prediction) FIG. 1 is a diagram showing an example of a flow of behavior prediction according to the first embodiment. The behavior prediction device 100 that performs behavior prediction analyzes information about the user's inner state to predict a behavior pattern of the user in at least one of a plurality of time segments obtained by dividing a preset period into predetermined time units. For example, the behavior prediction device 100 that performs behavior prediction analyzes information about the user's inner state to predict a behavior pattern of the user in each of a plurality of time segments obtained by dividing a preset period into predetermined time units. Then, based on the predicted behavior pattern, the behavior prediction device 100 identifies a time segment among the plurality of time segments for which a predetermined behavior of the user is to be recommended. The time segment here refers to, for example, a period (segment) obtained by dividing one day into 15-minute units.

[0012] FIG. 1 illustrates a case where the behavior prediction device 100 predicts the user's behavior for the next day. The behavior prediction device 100 acquires the user's past to current behavior log from the user's user terminal 300. The behavior log here is data recording the user's behavior patterns for each time segment. The behavior pattern for each time segment indicates whether or not the behavior occurred for each behavior type. The presence or absence of the behavior is represented, for example, by "1" indicating that the behavior was performed and "0" indicating that the behavior was not performed. Examples of behavior types include sleeping, eating, and work. The user's past to current behavior log is an example of information regarding the user's internal state.

[0013] Then, the behavior prediction device 100 calculates the cumulative behavior log from the behavior logs from the past to the current day (s1). For example, the behavior prediction device 100 uses the behavior patterns of each time segment from the past to the current day to accumulate the number of times that a behavior has been performed for each behavior type from the past to the current day.

[0014] The cumulative behavior log is calculated to obtain the cumulative past behaviors in order to extract the user's internal state. Examples of internal states include fatigue, stress, and motivation. For example, accumulated lack of sleep leads to increased fatigue, which can lead to oversleeping and lazy behavior. Accumulated work hours can lead to increased fatigue and oversleeping. On the other hand, accumulated sleep hours can lead to increased motivation and more diligent behavior. It is believed that internal states such as fatigue, stress, and motivation are generated by the accumulation of past behaviors. In other words, behaviors caused by the accumulation of past behaviors are thought to contain internal states such as fatigue, stress, and motivation that are difficult for the user to notice. Therefore, the behavior prediction device 100 accumulates behavior logs in order to extract the user's internal state.

[0015] Then, the behavior prediction device 100 inputs the accumulated behavior log data and the user's schedule for the next day into the prediction model M0 (s2, s3). The user's schedule for the next day is data obtained by converting the currently known schedule for the next day into the format of a behavior log. The prediction model M0 here is a model trained to estimate the degree of implementation of each behavior type for the time segment of the prediction target day from the accumulated behavior logs from the past and the schedule. In other words, the prediction model M0 inputs the accumulated behavior logs from the past to the current day and the schedule for the next day to predict the behavior of the next day. The prediction model M0 can be said to be a model that can take into account accumulated internal behavioral factors by using the accumulated past behavior logs.

[0016] Then, the prediction model M0 predicts the degree of performance of each activity type for each time slot of the next day (S4). Here, the prediction model M0 predicts the performance probability of each activity type as a prediction result for the time slot period "0:15 to 0:30." As an example, if the activity types are sleeping, eating, working, and free time, the predicted performance probability for sleeping is "0.6," the predicted performance probability for eating is "0.04," the predicted performance probability for working is "0.06," and the predicted performance probability for free time is "0.3." In other words, the prediction model M0 probabilistically predicts which activity will be performed in each time slot.

[0017] Then, the behavior prediction device 100 outputs the behavior prediction result for the next day to the user terminal 300 of the user (s5). For example, the behavior prediction device 100 predicts the behavior for the next day based on the degree of implementation of each behavior type for each time segment on the next day. In addition, the behavior prediction device 100 identifies, from the multiple time segments, time segments for which exercise is recommended during the user's free time, based on the degree of implementation of each behavior type for each time segment on the next day. Then, the behavior prediction device 100 outputs, for example, the behavior prediction result for the next day, including recommendations for the identified time segments, to the user terminal 300 of the user. In this case, the period from 21:00 to 23:00 is predicted to be free time, and therefore the time segments included in this period are identified as time segments for which exercise is recommended.

[0018] In this way, the behavior prediction device 100 predicts behavior in consideration of the internal behavioral factors accumulated by the user, and can recommend a specific behavior to the user at a time when the user is able to act.

[0019] (Example of behavior prediction system configuration) FIG. 2 is a diagram showing an example of the configuration of a behavior prediction system according to this embodiment. As shown in FIG. 2, the behavior prediction system 1 is a system in which a behavior prediction device 100 and user terminals 300-1 to n (n is any natural number; hereinafter, collectively referred to as "user terminals 300") are connected to each other via a network N so that they can communicate with each other. Note that the network N can be any of various communication networks, such as an intranet, whether wired or wireless. Furthermore, the network N may not be a single network, but may be configured, for example, by connecting an intranet and the Internet via a network device such as a gateway or other device (not shown).

[0020] The behavior prediction device 100 may be, for example, a server computer managed by a service provider that provides a behavior prediction support service. The behavior prediction device 100 receives, for example, past and current behavior logs and next day's schedule data from the user terminal 300 in order to predict the user's behavior for the next day. The behavior prediction device 100 also transmits, for example, the next day's behavior prediction result to the user terminal 300. Note that communication between the behavior prediction device 100 and the user terminal 300 may be performed, for example, via a user interface (UI) displayed in an application pre-installed on the user terminal 300. Alternatively, the communication may be performed via a UI displayed on a website or web application provided by the behavior prediction device 100 or the like.

[0021] 2, the behavior prediction device 100 is shown as a single computer, but may be, for example, a distributed computing system configured with multiple computers. Furthermore, the behavior prediction device 100 may be, for example, a cloud computing device managed by a service provider that provides cloud computing services.

[0022] The user terminal 300 is, for example, an information processing terminal used by each user. The user terminal 300 may be a mobile terminal such as a smartphone or a tablet PC (Personal Computer), or may be a desktop PC or a notebook PC. An application for performing behavior prediction and the like, provided by the behavior prediction device 100 or the like, is pre-installed in the user terminal 300. Alternatively, if the user terminal 300 uses a website or web application provided by the behavior prediction device 100 or the like that has functions equivalent to those of the application, the application does not need to be installed. The user transmits their own behavior logs from the past to the current day and schedule data for the next day to the behavior prediction device 100 via the user terminal 300.

[0023] (Configuration example of a behavior prediction device) FIG. 3 is a diagram illustrating an example of the configuration of a behavior prediction device according to the first embodiment. As illustrated in FIG. 3, the behavior prediction device 100 includes a learning unit 100A and a prediction unit 100B. The learning unit 100A learns a prediction model M0 that receives an accumulated past to current day behavior logs, a schedule for the next day, and correct answer data for the next day, and outputs a result of predicting the next day's behavior. The prediction unit 100B inputs the accumulated past to current day behavior logs and the schedule for the next day into the prediction model M0, predicts the user's behavior for the next day, and outputs the result of predicting the next day's behavior. Details of the learning unit 100A will be described later. Details of the prediction unit 100B will be described later.

[0024] (Example of learning unit configuration) FIG. 4 is a diagram illustrating an example of the configuration of a learning unit according to the first embodiment. As illustrated in FIG. 4, the learning unit 100A includes a cumulative calculation unit 110, a prediction model M0, and an update unit 180. The prediction model M0 includes a first feature extraction unit 120, a second feature extraction unit 130, a third feature extraction unit 140, a scheduled correlation feature extraction unit 150, an internal behavior factor feature extraction unit 160, and a probability prediction unit 170. The cumulative calculation unit 110 is an example of an acquisition unit. The internal behavior factor feature extraction unit 160, the probability prediction unit 170, and the update unit 180 are examples of a prediction unit. The update unit 180 is an example of an identification unit.

[0025] The accumulation calculation unit 110 calculates the accumulation of past to current day action logs. For example, the accumulation calculation unit 110 acquires past to current day action logs from the user's user terminal 300. Then, the accumulation calculation unit 110 accumulates the number of times an action was performed for each action type based on the action pattern for each time segment included in the acquired action log.

[0026] An example of an action log will now be described with reference to FIG. 5. FIG. 5 is a diagram illustrating an example of an action log. As illustrated in FIG. 5, the action log stores action patterns for each time segment. The time segment indicates, for example, a period obtained by dividing a day into 15-minute increments. The action pattern indicates whether or not an action was performed for each action type. Here, "1" is set to indicate that the action was performed, and "0" is set to indicate that the action was not performed. The action types are set to, for example, sleep, meal, and work. Note that the division unit of the time segment is not limited to 15-minute increments, but may also be 30-minute increments or one-hour increments. The action types are not limited to "sleeping," "meal," and "work," but may include various actions such as "travel." The action types may also include "free time." One time segment may also be referred to as a period.

[0027] As an example, when the time division is "0:00~0:15" (0 period), "1" (yes) is stored for whether or not "sleep" has occurred, "0" (no) for whether or not "meal" has occurred, and "0" (no) for whether or not "work" has occurred.

[0028] An example of the cumulative calculation will be described with reference to FIG. 6. FIG. 6 is a diagram showing an example of the cumulative calculation according to the first embodiment. As shown in FIG. 6, the cumulative calculation unit 110 accumulates the number of times an action was actually performed for each action type using action logs from the past to the current day. Alternatively, the cumulative calculation unit 110 calculates, for each action type, a cumulative ratio indicating how many times an action was actually performed among the total number of time segments included in the action logs from the past to the current day using action logs from the past to the current day. Here, a case where a cumulative ratio is calculated is shown. As a result, for each action type, a ratio indicating the cumulative number of actions to the entire time length N (total number of time segments) from the past to the current day used as the action log is set in the last time segment. Here, the cumulative ratio of the action type "sleeping" in the last time segment is set to 28 / N. The cumulative ratio of the action type "eating" in the last time segment is set to 12 / N. The cumulative ratio of the action type "work" in the last time segment is set to 32 / N. The vector (28 / N, 12 / N, 32 / N) represents data corresponding to the accumulation of behavior.

[0029] Returning to Fig. 4, the first feature extraction unit 120 extracts features from the accumulation of actions based on the encoder of the prediction model M0. For example, the first feature extraction unit 120 extracts features from the ratio of the accumulation of past to current action logs acquired by the accumulation calculation unit 110. Note that the extraction of features can be performed by applying existing deep learning, and therefore a detailed description thereof will be omitted.

[0030] The second feature extraction unit 130 extracts features from the next day's schedule based on the encoder of the prediction model M0. For example, the second feature extraction unit 130 extracts features from data obtained by converting the next day's schedule into an action log format. Note that feature extraction can be performed by applying existing deep learning, so a detailed description will be omitted.

[0031] When learning behavior for the t-th period (t=an integer equal to or greater than 1) of the next day, the third feature extraction unit 140 extracts features from the behavior log from 0 to t-1 periods of the next day based on the encoder of the prediction model M0. Note that when learning behavior for the 0th period of the next day, which is the first time, the third feature extraction unit 140 does not have an input behavior log, so it only needs to extract features from the behavior log for the last period of the day. Furthermore, since existing deep learning can be applied to extract features, a detailed description will be omitted.

[0032] The expected correlation feature extraction unit 150 extracts a feature that is correlated with a schedule and a past action log. For example, the expected correlation feature extraction unit 150 extracts the expected correlation feature by using the feature extracted by the third feature extraction unit 140 and the feature extracted by the second feature extraction unit 130. The expected correlation feature extraction process performed by the expected correlation feature extraction unit 150 will be described later.

[0033] The internal behavior factor feature extraction unit 160 extracts internal behavior factor feature amounts. For example, the internal behavior factor feature extraction unit 160 extracts internal behavior factor feature amounts using the feature amounts extracted by the third feature extraction unit 140 and the feature amounts extracted by the first feature extraction unit 120. The internal behavior factor feature extraction process performed by the internal behavior factor feature extraction unit 160 will be described later.

[0034] The probability prediction unit 170 predicts the behavior pattern for each behavior type in the t-th period of the next day. For example, the probability prediction unit 170 predicts the probability that the behavior indicated by each behavior type will be performed (behavior probability) in the t-th period of the next day, using the internal behavior factor feature amounts extracted by the internal behavior factor feature amount extraction unit 160 and the expected correlation feature amounts extracted by the expected correlation feature amount extraction unit 150.

[0035] The update unit 180 updates the parameters (weights) of the prediction model M0. For example, the update unit 180 compares the behavior probability prediction result for the t-th period of the following day predicted by the probability prediction unit 170 with the correct data for the t-th period of the following day. As an example, the update unit 180 compares the probability for each behavior type for the t-th period of the following day predicted by the probability prediction unit 170 with data obtained by converting the correct data for the t-th period of the following day into a behavior log format. The update unit 180 then calculates an error from the comparison result. The update unit 180 then updates the weights (parameters) of the attention mechanism of the prediction model M0 based on the calculated error. In addition, the update unit 180 combines the data obtained by converting the correct data for the t-th period of the following day into a behavior log format with the behavior logs for the 0 to t-1 periods of the following day and increments t by 1. The combined behavior log then becomes the behavior log for the 0 to t-1 periods of the following day to be used in the learning process for the next period.

[0036] (An example of estimated correlation feature extraction) Here, an example of the expected correlation feature extraction process executed by the expected correlation feature extraction unit 150 will be described with reference to Fig. 7. Fig. 7 is a diagram showing an example of the expected correlation feature extraction process according to the first embodiment. Note that Fig. 7 illustrates a case where behavior in the t-th period of the next day is to be learned. An attention layer L1 shown in Fig. 7 corresponds to the expected correlation feature extraction unit 150.

[0037] As shown in Fig. 7, the third feature extraction unit 140 extracts features from the action log from 0 to t-1 period of the next day. Then, the extracted features are converted into feature vectors corresponding to the action log from 0 to t-1 period of the next day and output to the Attention layer L1. The second feature extraction unit 130 extracts features from the schedule data for the next day. Then, the extracted features are converted into feature vectors corresponding to the schedule data for the next day and output to the Attention layer L1.

[0038] Then, the attention layer L1 calculates the inner product of the feature vector corresponding to the action log from 0 to t-1 period of the next day and the feature vector corresponding to the schedule data of the next day, and generates a weight matrix M1. The generated weight matrix M1 is generated by calculating the inner product for each time segment, for example, with each time segment corresponding to the schedule data on the horizontal axis and each time segment corresponding to the action log from 0 to t-1 period on the vertical axis.

[0039] Here, for example, if there is a "meeting" scheduled for "8:00" the next morning, there is a causal relationship in which the wake-up behavior (e.g., "breakfast") occurs earlier than usual. In such a case, the weight matrix is trained so that the values of the stronger relationships become larger, as shown in M1. For example, the matrix is trained so that the values of the causal schedules "meeting (8:00-8:15)" and "meeting (8:15-8:30)" and the resulting daily behaviors "breakfast (6:45-7:00)" and "breakfast (7:00-7:15)" become larger.

[0040] The attention layer L1 then adds the weighted sum of the feature vectors corresponding to the action log from 0 to t-1 of the next day to the weight matrix M1 to generate the expected correlation feature (characteristic <2> ) is extracted (output). In other words, the attention layer L1 extracts behavioral factor features that are influenced by already scheduled behaviors.

[0041] Thereafter, if there are no other characteristics, the probability prediction unit 170 uses the prediction layer L10 to output a prediction result from the expected correlation feature. That is, the probability prediction unit 170 predicts the behavior probability for each behavior type in the t-th period of the next day from the expected correlation feature.

[0042] (An example of internal behavioral factor feature extraction) Here, an example of the internal behavior factor feature extraction process executed by the internal behavior factor feature extraction unit 160 will be described with reference to Fig. 8. Fig. 8 is a diagram showing an example of the internal behavior factor feature extraction process according to the first embodiment. Note that Fig. 8 illustrates a case where behavior for the t-th period of the next day is to be learned. The attention layer L2 shown in Fig. 8 corresponds to the internal behavior factor feature extraction unit 160.

[0043] As shown in FIG. 8, the accumulation calculation unit 110 calculates the accumulation of the past to current day action logs (f1). For example, using the past to current day action logs, the accumulation calculation unit 110 calculates, for each action type, the cumulative ratio indicating how many times an action was actually performed among the total number of time segments included in the past to current day action logs (see FIG. 6). As an example, if the past to current day action logs are for one day, the total number of time segments N is 96, which is 24 hours x 4 (0-15 minutes, 15-30 minutes, 30-45 minutes, 45-60 minutes). If there are 28 time segments in which "sleep" is represented by "1," the cumulative ratio of "sleep" is calculated as "28 / 96." Similarly, if there are 12 time segments in which "meal" is represented by "1," the cumulative ratio of "meal" is calculated as "12 / 96." If there are 32 time segments in which "work" is represented by "1," the cumulative ratio of "work" is calculated as "12 / 96." The cumulative percentages of these (sleep, meals, work) are data corresponding to the accumulation of actions. Then, the accumulation calculation unit 110 outputs the calculated cumulative percentages.

[0044] Then, a feature is extracted for the cumulative ratio by the first feature extraction unit 120. Then, the extracted feature is converted into a feature vector corresponding to the cumulative ratio and output to the attention layer L2.

[0045] Furthermore, for the action log from 0 to t-1 period of the next day, features are extracted by the third feature extraction unit 140. Then, the extracted features are converted into feature vectors corresponding to the action log from 0 to t-1 period of the next day, and output to the attention layer L2.

[0046] Then, the attention layer L2 calculates the inner product of a feature vector corresponding to the action log from 0 to t-1 of the next day and a feature vector corresponding to the cumulative ratio, for example, to generate a weight matrix.

[0047] The attention layer L2 then adds the weighted sum of the feature vectors corresponding to the behavior log from 0 to t-1 of the next day to the weight matrix to generate the internal behavior factor features (characteristics). <1> ) is extracted (output). In other words, under the assumption that there is a causal relationship in which future behavior is determined by the accumulation of past behaviors, the Attention layer L2 extracts the causal characteristics between the behavior from period 0 to period t-1 of the following day and the result of calculating the accumulation from the time series data of past to current behaviors. For example, under the assumption that there are internal behavioral factors that accumulate due to the accumulation of behaviors, such as continuous overtime work from the past to the current day and long sleep time the next day, the Attention layer L2 extracts (outputs) the user's internal state. In other words, the Attention layer L2 estimates the degree of psychological fatigue accumulated in the user.

[0048] Then, the probability prediction unit 170 calculates the characteristic <2> If the (expected correlation feature) is extracted, the internal behavior factor feature (characteristic <1> ) and characteristics <2> The combined result is input to the prediction layer L10 and the prediction result is output. That is, the probability prediction unit 170 predicts the behavior probability for each behavior type in the t-th period of the next day from the internal behavior factor feature amount and the scheduled correlation feature amount.

[0049] (Example of prediction unit configuration) Next, an example of the configuration of the prediction unit 100B will be described with reference to FIG. 9. FIG. 9 is a diagram showing an example of the configuration of a learning unit according to the first embodiment. As shown in FIG. 9, the prediction unit 100B has an accumulation calculation unit 110, a prediction model M0, and an output unit 190. The accumulation calculation unit 110 performs the same processing as the accumulation calculation unit 110 of the learning unit 100A shown in FIG. 4, and therefore its description will be omitted. The prediction model M0 has the same configuration as the prediction model M0 of the learning unit 100A shown in FIG. 4, and therefore the same reference numerals will be used, and its description will be omitted. The difference between the learning unit 100A and the prediction unit 100B is that an output unit 190 is added.

[0050] The output unit 190 outputs the behavior prediction result for the next day to the user terminal 300 of the user. For example, the output unit 190 receives the behavior probability prediction result for the t-th period of the next day predicted by the probability prediction unit 170. If the t-th period is before the final period, the output unit 190 combines data obtained by converting the behavior probability prediction result for the t-th period of the next day into the format of a behavior log with the behavior log from 0 to t-1 hours of the next day and increments t by 1. The combined behavior log then becomes the behavior log from 0 to t-1 hours of the next day. Furthermore, if the t-th period is the final period, the output unit 190 predicts the behavior of the next day based on each behavior probability prediction result from 0 to t periods of the next day. In addition, the output unit 190 identifies, from among multiple time segments, a time segment in which, for example, exercise is recommended during the user's free time, based on each behavior probability prediction result from 0 to t periods of the next day. Then, the output unit 190 outputs the next day's behavior prediction result including the recommendation for the identified time segment to the user terminal 300 of the user.

[0051] This allows the output unit 190 to suggest to the user a specific activity, such as exercise, during the available time.

[0052] (Data flow for learning process) Here, the data flow of the learning process executed by the learning unit 100A will be described with reference to Fig. 10. Fig. 10 is a diagram showing an example of the data flow of the learning process according to the first embodiment. Note that Fig. 10 does not show the next day's schedule and past to present action logs input to the learning unit 100A, but it is assumed that the learning unit 100A uses these data.

[0053] The prediction model M0 outputs a prediction result for the tth period using behavior logs from 0 to t-1 (t is an integer equal to or greater than 1). The prediction result for the tth period here refers to a behavior probability prediction result for each behavior type in the tth period of the following day. The learning unit 100A then compares the behavior probability prediction result for the tth period of the following day with the correct data for the tth period of the following day. The learning unit 100A then calculates an error from the comparison result and updates the weights of the attention mechanism of the prediction model M0 based on the calculated error. In addition, the learning unit 100A converts the correct data for the tth period of the following day into a behavior log format, combines this data with the behavior logs for the 0th to t-1th period of the following day, and increments t by 1. The combined behavior log then becomes the behavior log for the 0th to t-1th period of the following day. Note that although the weights have been described as being updated for each prediction result, this is not limited thereto and may be updated for prediction results for multiple predetermined periods.

[0054] Here, first, let us assume that t is 1. Let us assume that an action log for each time segment ("sleep", "meal", "work") is generated. The prediction model M0 uses the action log (1, 0, 0) for "0:00 to 0:15" (0 (= t-1) period) to output (a1) the action probability prediction result for "0:15 to 0:30" (1 period). The action probability prediction result for the first period is shown as (0.8, 0.1, 0.1).

[0055] The learning unit 100A then compares the behavior probability prediction result for "0:15 to 0:30" (one period) on the following day with the correct data for "0:15 to 0:30" (one period) on the following day (a2). Here, the behavior probability prediction result is (0.8, 0.1, 0.1), and the correct data is (1, 0, 0). The learning unit 100A then calculates the error from the comparison result and updates the weights of the attention mechanism of the prediction model M0 based on the calculated error (a3).

[0056] Additionally, the learning unit 100A combines the correct answer data (1,0,0) for "0:15 to 0:30" (one period) of the next day with the action log for the period 0 to t-1 of the next day (a4). Then, the learning unit 100A adds 1 to t to make it 2.

[0057] Next, when t is 2, the prediction model M0 uses the behavior log for period 0 and the behavior log (1,0,0) for "0:15 to 0:30" (period 1) to output a behavior probability prediction result for "0:30 to 0:45" (period 2) (a1). The behavior probability prediction result for period 2 is shown as (0.5,0.3,0.2).

[0058] The learning unit 100A then compares the behavior probability prediction results for "0:30 to 0:45" (two periods) on the next day with the correct data for "0:30 to 0:45" (two periods) on the next day (a2).The learning unit 100A then calculates an error from the comparison results and updates the weights of the attention mechanism of the prediction model M0 based on the calculated error (a3).

[0059] Additionally, the learning unit 100A combines the correct answer data (1,0,0) for "0:30 to 0:45" (period 2) on the next day with the action log for period 0 to t-1 on the next day (a4). Then, the learning unit 100A adds 1 to t to make it 3.

[0060] In this way, the learning unit 100A executes the learning process by repeating the processes a1 to a4 until t reaches the final period.

[0061] (Data flow for prediction process) Next, the prediction process executed by the prediction unit 100B will be described with reference to Fig. 11. Fig. 11 is a diagram showing an example of the data flow of the prediction process according to the first embodiment. Note that Fig. 11 does not show the next day's schedule and past to current day action logs input to the prediction unit 100B, but it is assumed that the prediction unit 100B uses these data.

[0062] The prediction model M0 outputs a prediction result for the tth period using behavior logs from 0 to t-1 (t is an integer equal to or greater than 1). The prediction result for the tth period here refers to the behavior probability prediction result for each behavior type in the tth period of the next day. Then, the prediction unit 100B converts the prediction result for the tth period of the next day into a behavior log format, combines this data with the behavior log for the 0 to t-1 period of the next day, and increments t by 1. The combined behavior log then becomes the behavior log for the 0 to t-1 period of the next day.

[0063] Here, first, let us assume that t is 1. Let us assume that the behavior log for each time segment ("sleep", "meal", "work") is generated. The prediction model M0 uses the behavior log (1, 0, 0) for "0:00 to 0:15" (0 (= t-1) period) to output the behavior probability prediction result for "0:15 to 0:30" (1 period) (b1). The behavior probability prediction result for the first period is shown as (0.8, 0.1, 0.1).

[0064] Then, the prediction unit 100B combines the behavior probability prediction result (0.8, 0.1, 0.1) for "0:15 to 0:30" (one period) on the next day with the behavior log from 0 to t-1 on the next day (b2). Then, the prediction unit 100B adds 1 to t to make it 2.

[0065] Next, when t is 2, the prediction model M0 uses the behavior log for period 0 and the behavior log (1,0,0) for "0:15 to 0:30" (period 1) to output the behavior probability prediction result for "0:30 to 0:45" (period 2) (b1). The behavior probability prediction result for period 2 is shown as (0.7,0.1,0.2).

[0066] Then, the prediction unit 100B combines the behavior probability prediction result (0.7, 0.1, 0.2) for "0:30 to 0:45" (two periods) on the next day with the behavior log from 0 to t-1 on the next day (b2). Then, the prediction unit 100B adds 1 to t to make it 3.

[0067] In this way, the prediction unit 100B executes the prediction process by repeating the processes of b1 and b2 until t becomes the final period. Then, the prediction unit 100B predicts the behavior of the next day based on the behavior probability prediction results for each period from 0 to t of the next day.

[0068] (Flowchart of learning process) Next, a flowchart of the learning process executed by the learning unit 100A will be described with reference to Fig. 12. Fig. 12 is a diagram showing an example of a flowchart of the learning process according to the first embodiment.

[0069] 12, the learning unit 100A acquires a past action log from the user terminal 300 (step S11). The learning unit 100A determines whether the learning process for the period is the first time (step S12). If it is determined that the learning process for the period is the first time (step S12; Yes), the learning unit 100A extracts the last period (segment) from the past action log as the 0th period (step S13).

[0070] Then, the learning unit 100A calculates the accumulation of past action logs (step S14). For example, the learning unit 100A uses the past action logs to accumulate the number of times that an action was actually performed for each action type. Alternatively, the learning unit 100A uses the past action logs to calculate, for each action type, an accumulation ratio indicating how many times an action was actually performed out of the total number of time segments included in the past action logs.

[0071] Then, the learning unit 100A calculates attention between the action log of the 0th period and the cumulative total of past action logs (step S15). For example, the learning unit 100A calculates the inner product of the feature for the action log of the 0th period and the feature for the cumulative ratio to generate a weight matrix. Then, the learning unit 100A calculates attention feature by weighting the feature for the action log of the 0th period to the weight matrix. Such attention feature is an internal behavior factor feature. Then, the learning unit 100A proceeds to step S18.

[0072] On the other hand, if it is determined that the learning process for the period is not the first time (step S12; No), the learning unit 100A calculates the cumulative amount of past action logs (step S16). The cumulative calculation is the same as in step S14.

[0073] Then, the learning unit 100A calculates attention between the action logs for the 0th to tth periods (t is an integer equal to or greater than 1) and the accumulation of past action logs (step S17). For example, the learning unit 100A calculates the inner product of the feature amounts for the action logs for the 0th to tth periods and the feature amount for the accumulation ratio to generate a weight matrix. Then, the learning unit 100A calculates attention features by weighting the feature amounts for the action logs for the 0th to tth periods in the weight matrix. These attention features are internal behavior factor features. Then, the learning unit 100A proceeds to step S18.

[0074] In step S18, the learning unit 100A predicts the behavior probability for the t+1th period from the attention feature (step S18). For example, the learning unit 100A predicts the probability (behavior probability) that the behavior indicated by each behavior type will be performed in the t+1th period using the internal behavior factor feature.

[0075] The learning unit 100A then determines whether t+1 is smaller than the final period (step S19). If it is determined that t+1 is smaller than the final period (step S19; Yes), the learning unit 100A calculates the error between the predicted behavior probability and the correct label (step S20). The learning unit 100A then updates the parameters (weights) of the prediction model M0 based on the calculated error (step S21).

[0076] Then, the learning unit 100A combines the action log up to the t-th period with the correct label for the t+1-th period (step S22). Then, the learning unit 100A increments t by 1 (step S23). Then, the learning unit 100A proceeds to step S17 to perform learning processing for the next period.

[0077] On the other hand, if it is determined that t+1 is equal to or greater than the final period (step S19; No), the learning unit 100A combines the action log up to the t-th period with the action log for the t+1-th period (step S24). Then, the learning unit 100A ends the learning process.

[0078] (Prediction process flowchart) Next, a flowchart of the prediction process executed by the prediction unit 100B will be described with reference to Fig. 13. Fig. 13 is a diagram showing an example of a flowchart of the prediction process according to the first embodiment.

[0079] 13, the prediction unit 100B acquires a past action log from the user terminal 300 (step S31). The prediction unit 100B determines whether the period prediction process is the first time (step S32). If it is determined that the period prediction process is the first time (step S32; Yes), the prediction unit 100B extracts the last period (segment) from the past action log as the 0th period of the next day (step S33).

[0080] Then, the prediction unit 100B calculates the accumulation of past action logs (step S34). Note that the calculation of the accumulation may be the same as that of the learning unit 100A.

[0081] Then, the prediction unit 100B calculates attention between the behavior log of the 0th period of the next day and the cumulative past behavior logs (step S35). The calculation of attention may be the same as that of the learning unit 100A. The attention feature, which is the calculation result, is an internal behavior factor feature. Then, the prediction unit 100B proceeds to step S38.

[0082] On the other hand, if it is determined that the period prediction process is not the first time (step S32; No), the prediction unit 100B calculates the cumulative value of past action logs (step S36). The cumulative calculation is the same as in step S34.

[0083] Then, the prediction unit 100B calculates attention between the behavior log for the next day's 0th to tth period (t is an integer equal to or greater than 1) and the cumulative past behavior log (step S37). The calculation of attention is the same as in step S35. The attention feature amount, which is the calculation result, is an internal behavior factor feature amount. Then, the prediction unit 100B proceeds to step S38.

[0084] In step S38, the prediction unit 100B predicts the behavior probability for the t+1th period of the next day from the attention feature amount (step S38). Note that the calculation of the behavior probability may be the same as that of the learning unit 100A.

[0085] Then, the prediction unit 100B determines whether t+1 is smaller than the final period (step S39). If it is determined that t+1 is smaller than the final period (step S39; Yes), the prediction unit 100B combines the behavior probability up to the t-th period with the behavior probability for the t+1-th period (step S40). Then, the prediction unit 100B increments t by 1 (step S41). Then, the prediction unit 100B proceeds to step S37 to perform prediction processing for the next period.

[0086] On the other hand, if it is determined that t+1 is equal to or longer than the final period (step S39; No), the prediction unit 100B combines the behavior probability up to the t-th period and the behavior probability for the t+1-th period (step S42).

[0087] Then, the prediction unit 100B outputs the behavior prediction result for the next day to the user terminal 300 based on the behavior probability prediction result (step S43). For example, the prediction unit 100B identifies a period among the multiple periods for which, for example, exercise will be recommended to the user in his / her free time based on the behavior probability for each of the periods 0 to t+1 for the next day. Then, the prediction unit 100B outputs the behavior prediction result for the next day including the recommendation for the identified period to the user terminal 300 of the user. Then, the prediction unit 100B ends the prediction process.

[0088] In the first embodiment, it has been described that the learning unit 100A of the behavior prediction device 100 performs learning using inner behavior factor feature quantities and expected correlation feature quantities. However, the learning unit 100A may perform learning using feature quantities other than inner behavior factor feature quantities and expected correlation feature quantities. Therefore, in the following second embodiment, a case will be described in which the learning unit 100A performs learning using, for example, behavior pattern feature quantities in addition to inner behavior factor feature quantities and expected correlation feature quantities.

[0089] [Embodiment 2] Fig. 14 is a diagram showing an example of the configuration of a learning unit according to the second embodiment. Note that, with regard to the functional configuration of the learning unit 100A according to the second embodiment, the same components as those of the learning unit 100A shown in Fig. 4 are denoted by the same reference numerals, and a description of the overlapping configurations and operations will be omitted. The first embodiment differs from the second embodiment in that a fourth feature extraction unit 210 and a behavioral factor feature extraction unit 220 are added.

[0090] The fourth feature extraction unit 210 extracts features from the past to the current day's action logs based on the encoder of the prediction model M0. Note that the feature extraction can be performed by applying existing deep learning, and therefore detailed explanations will be omitted.

[0091] The behavioral factor feature extraction unit 220 extracts features caused by the previous day's behavior or the periodicity of behavior. For example, the behavioral factor feature extraction unit 220 extracts behavioral factor features using the features extracted by the fourth feature extraction unit 210 and the features extracted by the third feature extraction unit 140. The behavioral factor feature extraction process performed by the behavioral factor feature extraction unit 220 will be described later.

[0092] Fig. 15 is a diagram showing an example of the configuration of a prediction unit according to the second embodiment. Note that, with regard to the functional configuration of the prediction unit 100B according to the second embodiment, the same components as those of the prediction unit 100B shown in Fig. 9 are denoted by the same reference numerals, and a description of the overlapping configurations and operations will be omitted. The first embodiment differs from the second embodiment in that a fourth feature extraction unit 210 and a behavioral factor feature extraction unit 220 are added.

[0093] The fourth feature extraction unit 210 performs the same processing as the fourth feature extraction unit 210 of the learning unit 100A shown in FIG. 14, and therefore its description will be omitted.

[0094] The behavior factor feature quantity extraction unit 220 performs the same processing as the behavior factor feature quantity extraction unit 220 of the learning unit 100A shown in FIG. 14, and therefore a description thereof will be omitted.

[0095] (An example of behavioral factor feature extraction) Here, an example of the behavioral factor feature extraction processing executed by the behavioral factor feature extraction unit 220 will be described with reference to Fig. 16. Fig. 16 is a diagram showing an example of the behavioral factor feature extraction processing according to the second embodiment. Note that Fig. 16 illustrates a case where behavior in the t-th period of the next day is to be learned. The attention layer L3 shown in Fig. 16 corresponds to the behavioral factor feature extraction unit 220.

[0096] As shown in Fig. 16, a third feature extraction unit 140 extracts features from the action log from 0 to t-1 period of the next day. Then, the extracted features are converted into feature vectors corresponding to the action log from 0 to t-1 period of the next day, and output to the Attention layer L3. A fourth feature extraction unit 210 extracts features from the action logs from the past to the current day. Then, the extracted features are converted into feature vectors corresponding to the action logs from the past to the current day, and output to the Attention layer L3.

[0097] Then, the Attention layer L3 calculates the inner product of, for example, a feature vector corresponding to the action log from 0 to t-1 period of the next day and a feature vector corresponding to the action log from the past to the current day, to generate a weight matrix M2. For example, if the next day is day N, the generated weight matrix M2 is generated by calculating the inner product for each time segment, with the horizontal axis representing each time segment corresponding to the previous day (day N-1) and the vertical axis representing each time segment corresponding to the action log from 0 to t-1 period of the next day (day N).

[0098] Here, for example, suppose there is a cyclical behavioral factor where working overtime on the previous day (day N-1) will result in working overtime the following day (day N). In such a case, the weight matrix is trained so that values where there is a strong relationship are larger, as shown in M2. For example, the matrix is trained so that the value of the matrix between "Work (8:00-8:15)" and "Work (8:15-8:30)" on day N-1 and "Work (8:00-8:15)" and "Work (8:15-8:30)" on day N is larger.

[0099] The Attention layer L3 then adds the weighted sum of the features of the behavior log from 0 to t-1 of the next day to the Weight matrix M2 to generate the behavior factor features (characteristics <3> ) is extracted (output). In other words, the attention layer L3 extracts behavioral factor features that are caused by the behavior on the previous day or by the periodicity of the behavior.

[0100] Then, the probability prediction unit 170 calculates the internal behavior factor feature amount (characteristic <1> ), expected correlation features (characteristics <2> ) and behavioral factor features (characteristics <3> ) is input to the prediction layer L10 and the prediction result is output. That is, the probability prediction unit 170 predicts the behavior probability for each behavior type in the t-th period of the next day from the internal behavior factor feature, the scheduled correlation feature, and the behavior factor feature.

[0101] (Example of output screen) Next, an output screen displayed on the user terminal 300 will be described with reference to FIG. 17. FIG. 17 is a diagram showing an example of an output screen according to the first and second embodiments. As shown in FIG. 17, the output unit 190 predicts the behavior of the next day based on the behavior probability prediction results for each of the periods 0 to t of the next day. In addition, the output unit 190 identifies, from among a plurality of time segments, a time segment for recommending exercise during the user's free time, based on the behavior probability prediction results for each of the periods 0 to t of the next day. Then, the output unit 190 outputs the behavior prediction results for the next day, including recommendations for the identified time segments, to the user terminal 300 of the user.

[0102] Here, the previous day's activity log, planned activities, tomorrow's activity prediction results, and recommended exercise plan are displayed. The previous day's activity log represents the actual activities obtained from the past to today's activity logs entered by the user. The planned activities represent the activities obtained from the next day's schedule entered by the user. The tomorrow's activity prediction results represent the next day's activities predicted based on the respective activity probability prediction results for the next day's period from 0 to t. Note that while only activities are displayed here, the activity probabilities indicated in the activity probability prediction results may also be displayed in addition to the activities. The recommended exercise plan also shows recommended exercises for the free time periods in tomorrow's activity prediction results. Here, "stretching" is displayed as an exercise between 8:00 PM and 10:00 PM in the tomorrow's activity prediction results. In other words, the period from 8:00 PM to 10:00 PM is identified as the time period for which exercise is recommended to the user.

[0103] The final line displays a comment about the predicted behavior for tomorrow. For example, the comment might read, "Since you plan to go to work tomorrow, you'll likely start your day by getting up earlier than usual. It's estimated that you'll have some free time between 8:00 PM and 11:00 PM. Since your free time is relatively short and you may be feeling tired from walking a lot on the way to work, I recommend you do some extended stretching to prepare for tomorrow."

[0104] This allows the behavior prediction device 100 to suggest a specific exercise to the user at a time when the user is able to do the activity.

[0105] (Output screen of prior art) Here, a reference example of an output screen of a conventional technique (a task management method that presents future predictions based on a life log and a schedule) is shown. FIG. 18 is a diagram showing a reference example of an output screen of the conventional technique. As shown in FIG. 18, a list of daily future diaries and the daily future diaries are displayed. The daily future diary shows "2 hours 0 minutes" for plans, "7 hours 22 minutes" for tasks, and "8 hours 13 minutes" for sleep, along with comments. The remaining time other than the time for plans, tasks, and sleep is leisure time, which is the total amount of free time for the day. However, the total amount of free time for the day does not indicate how long or what time period each block of free time falls within. Therefore, it is difficult to suggest a specific exercise based on the total amount of free time for the day. In other words, even if a specific exercise is suggested for the predicted leisure time, the user may not be able to find the block of free time and may not actually perform the suggested specific activity. In other words, with conventional techniques, it is difficult to suggest a specific activity for a time when the user can take action.

[0106] In contrast to this, the behavior prediction device 100 according to the first and second embodiments predicts, for example, the length of free time and the time period in which it occurs, and can therefore suggest specific exercises for the time when the user is able to perform the activity.

[0107] (effect) As described above, the behavior prediction device 100 acquires information about the user's internal state. By analyzing the acquired information about the user's internal state, the behavior prediction device 100 predicts the user's behavior pattern in at least one of a plurality of time segments obtained by dividing a preset period into predetermined time units. Then, based on the predicted behavior pattern, the behavior prediction device 100 identifies a time segment among the plurality of time segments in which a predetermined behavior is recommended to the user. This allows the behavior prediction device 100 to suggest a specific behavior to the user during a time when the behavior is possible.

[0108] Furthermore, in the behavior prediction device 100, the acquisition process acquires information about the user's behavioral record. The prediction process estimates the user's accumulated level of psychological fatigue by analyzing the acquired information about the user's behavioral record, and predicts the degree to which the user will perform each behavior type used in the behavior pattern in each of a plurality of time segments divided into predetermined time units based on the estimated level of the user's fatigue. The identification process identifies, from the plurality of time segments, a time segment for which exercise is to be recommended to the user based on the predicted degree to which each behavior type will be performed in each of the plurality of time segments. This allows the behavior prediction device 100 to suggest exercise during time periods when the user is able to exercise.

[0109] Furthermore, in the behavior prediction device 100, the analyzing process includes accumulating the number of times a behavior was performed for each behavior type used in the behavior pattern based on the behavior pattern for each of a plurality of time segments included in the information on the user's behavioral record, and generating cumulative information on the user's behavior. In this way, the behavior prediction device 100 can generate information related to the internal state created by the accumulation of past behaviors by generating the cumulative information on the user's behavior.

[0110] Furthermore, in the behavior prediction device 100, the accumulated information on the user's behavior and the correct answer data of the user's behavior pattern for one day are input into a machine learning model that predicts the user's behavior pattern to learn the user's inner state. In this way, the behavior prediction device 100 can learn the user's inner state by using information on the user's inner state.

[0111] Furthermore, in the behavior prediction device 100, the machine learning model converts input data into feature quantities indicating the inner state based on deep learning. This allows the behavior prediction device 100 to learn the inner state of the user by using the feature quantities indicating the inner state.

[0112] Furthermore, in the behavior prediction device 100, the learning process learns the user's internal state by updating the parameters of the machine learning model based on the error between the behavior pattern that indicates the output result when cumulative information on the user's behavior is input to the machine learning model and the correct answer data of the user's behavior pattern. This allows the behavior prediction device 100 to efficiently learn the user's internal state.

[0113] Furthermore, in the behavior prediction device 100, the learning process inputs at least one of information on the user's planned behavior on the same day as the correct answer data and information on the user's behavioral record, in addition to the accumulated information on the user's behavior, into the machine learning model to learn the user's internal state. This allows the behavior prediction device 100 to simultaneously learn correlations with schedules and past behavior.

[0114] Furthermore, in the behavior prediction device 100, the prediction process inputs the accumulated information of the user's behavior and the planned behavior information for the prediction target date into a trained machine learning model, and predicts the degree to which the user will perform each behavior type used in the behavior pattern in each of a plurality of time segments. In this way, the behavior prediction device 100 can predict behavior taking into account the user's internal state by using the accumulated information of the user's behavior.

[0115] Furthermore, in the behavior prediction device 100, the machine learning model sequentially executes predictions for a plurality of time segments, thereby enabling the behavior prediction device 100 to predict behavior at a fine time interval.

[0116] (system) The information, including the processing procedures, control procedures, specific names, various data, and parameters shown in the above documents and drawings, may be changed as desired unless otherwise specified. Furthermore, the specific examples, distributions, numerical values, etc. described in the embodiments are merely examples and may be changed as desired.

[0117] Furthermore, the specific form of distribution or integration of the components of each device is not limited to that shown in the drawings. In other words, all or some of the components may be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions of each device may be realized by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or may be realized as hardware using wired logic.

[0118] (Hardware) Fig. 19 is a diagram illustrating an example of the hardware configuration of the behavior prediction device 100. As shown in Fig. 19, the behavior prediction device 100 includes a communication interface 100a, a hard disk drive (HDD) 100b, a memory 100c, and a processor 100d. The components shown in Fig. 19 are connected to each other via a bus or the like.

[0119] The communication interface 100a is a network interface card or the like, and communicates with other servers. The HDD 100b stores, for example, programs and DBs that operate the functions shown in FIG.

[0120] The processor 100d is a hardware circuit that operates a process that executes each function described in FIG. 3 and the like by reading a program that executes processing similar to that of each processing unit shown in FIG. 3 from the HDD 100b or the like and expanding the program in the memory 100c. That is, this process executes functions similar to those of each processing unit included in the behavior prediction device 100. Specifically, the processor 100d reads a program having functions similar to those of the learning unit 100A and the prediction unit 100B from the HDD 100b or the like. Then, the processor 100d executes a process that executes processing similar to that of the learning unit 100A and the prediction unit 100B.

[0121] In this way, the behavior prediction device 100 operates as a device that executes behavior prediction processing by reading and executing a program that executes processing similar to that of each processing unit shown in Fig. 3, for example. The behavior prediction device 100 can also realize functions similar to those of the above-described embodiment by reading a program from a recording medium using a medium reading device and executing the read program. Note that the program in these other embodiments is not limited to being executed by the behavior prediction device 100. For example, this embodiment may also be applied in the same way to cases where another information processing device executes a program or where the behavior prediction device 100 and another information processing device cooperate to execute a program.

[0122] A program that executes the same processes as those of the processing units shown in Fig. 3 can be distributed via a network such as the Internet. This program can be recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, a magneto-optical disk (MO), or a digital versatile disk (DVD), and can be executed by being read out from the recording medium by a computer.

[0123] The behavior prediction process of the behavior prediction device 100 according to the above embodiment can be applied to a personalized recommendation service that predicts daily behavior for the next day and then recommends exercise or the like during free time, for example. [Explanation of symbols]

[0124] 1. Behavioral Prediction System 100 Behavior Prediction Device 100A Learning Department 100B Prediction Department 110 Cumulative calculation section 120 First feature extraction unit 130 Second feature extraction unit 140 Third feature extraction unit 150 Scheduled correlation feature extraction unit 160 Internal behavior factor feature extraction unit 170 Probability Prediction Unit 180 Update Department 190 Output section 210 Fourth feature extraction unit 220 Behavioral Factor Feature Extraction Unit 300 User terminal M0 prediction model

Claims

1. Obtaining information about the user's internal state; By analyzing the acquired information on the internal state of the user, a behavioral pattern of the user is predicted for at least one of a plurality of time segments obtained by dividing a preset period into predetermined time units; Identifying a time segment for recommending a predetermined behavior to the user from among the plurality of time segments based on the predicted behavior pattern. A behavior prediction program that causes a computer to execute processing.

2. The acquiring process acquires information about the user's behavioral record, The prediction process includes: By analyzing the acquired information on the user's behavioral record, a degree of psychological fatigue accumulated in the user is estimated; predicting the degree to which the user will perform each of the behavior types used in the behavior pattern in each of a plurality of time segments divided into predetermined time units based on the estimated degree of fatigue of the user; The process of specifying identifies a time segment for which exercise is to be recommended to the user from among a plurality of time segments, based on the predicted degree of performance of each behavior type in each of a plurality of time segments.

2. The behavior prediction program according to claim 1, wherein:

3. The analyzing process includes accumulating the number of times a behavior was performed for each behavior type used in the behavior pattern based on the behavior pattern for each of the plurality of time segments included in the information about the behavioral record of the user, and generating cumulative information about the behavior of the user.

3. The behavior prediction program according to claim 2.

4. The accumulated information on the user's behavior and the correct answer data of the user's behavior pattern for one day are input into a machine learning model that predicts the user's behavior pattern, thereby learning the user's internal state.

4. The behavior prediction program according to claim 3.

5. The machine learning model converts input data into features that indicate internal states based on deep learning.

5. The behavior prediction program according to claim 4.

6. The learning process involves updating the parameters of the machine learning model based on the error between a behavior pattern that indicates an output result when cumulative information on the user's behavior is input into the machine learning model and correct data for the user's behavior pattern, thereby learning the user's internal state.

5. The behavior prediction program according to claim 4.

7. The learning process includes: In addition to the accumulated information on the user's behavior, at least one of information on planned behavior on the same day as the correct answer data and information on the user's behavioral track record is input to the machine learning model to learn the user's internal state.

5. The behavior prediction program according to claim 4.

8. The prediction process inputs cumulative information of the user's behavior and information on the user's planned behavior on the prediction target date into the trained machine learning model, and predicts the degree to which the user will perform each behavior type used in the behavior pattern in each of a plurality of time segments.

5. The behavior prediction program according to claim 4.

9. The machine learning model performs predictions sequentially over multiple time intervals.

9. The behavior prediction program according to claim 8.

10. an acquisition unit that acquires information about the user's internal state; a prediction unit that predicts a behavior pattern of the user in at least one of a plurality of time segments obtained by dividing a preset period into predetermined time units by analyzing information about the user's inner state acquired by the acquisition unit; and an identification unit that identifies, from the plurality of time segments, a time segment in which a predetermined behavior is to be recommended to the user based on the behavior pattern predicted by the prediction unit; A behavior prediction device comprising:

11. Obtaining information about the user's internal state; By analyzing the acquired information on the internal state of the user, a behavioral pattern of the user is predicted for at least one of a plurality of time segments obtained by dividing a preset period into predetermined time units; Identifying a time segment for recommending a predetermined behavior to the user from among the plurality of time segments based on the predicted behavior pattern. A behavior prediction method characterized in that processing is executed by a computer.

Images