Information processing device, information processing method, and information processing program

Abstract

This information processing device is provided with: an acquisition unit that acquires behavioral data indicating the behavior of each of a plurality of users; a pre-processing unit that calculates a statistical value of the behavior data for each of the plurality of users over a prescribed period of time; a seasonal variation calculation unit that calculates the harmonic mean value of the maximum value and the minimum value of the statistical values calculated by the pre-processing unit, and calculates a seasonal variation amount, which is a variation amount common to the plurality of users and dependent on the prescribed period of time, on the basis of the harmonic mean value; and a profiling unit that performs user profiling of the behavioral data on the basis of the seasonal variation amount.

Description

Information processing device, information processing method, and information processing program

[0001] The present disclosure relates to techniques for user profiling.

[0002] Patent document 1 discloses a determination device that acquires behavioral information indicating a user's behavior, estimates the user's degree of interest in a trading object related to the behavior indicated by the behavioral information for each piece of behavioral information, and determines a trading object for which information will be provided to the user based on the behavioral information for which the estimated degree of interest satisfies predetermined conditions.

[0003] However, the technology disclosed in Patent Document 1 does not take into account the amount of seasonal variation inherent in behavioral information, and therefore further improvement is required to perform highly accurate user profiling.

[0004] Patent No. 7287926

[0005] The present disclosure has been made to solve such problems, and aims to provide a technology that can improve the accuracy of estimating a user's interests and concerns and perform highly accurate user profiling.

[0006] An information processing device in one aspect of the present disclosure includes an acquisition unit that acquires behavioral data indicating the behavior of each of a plurality of users; a pre-processing unit that calculates statistical values ​​of the behavioral data for each of the plurality of users over a predetermined period of time; a seasonal variation calculation unit that calculates a harmonic mean of the minimum and maximum values ​​of the statistical values ​​calculated by the pre-processing unit and calculates a seasonal variation amount, which is a variation amount common to the plurality of users that depends on the predetermined period, based on the harmonic mean value; and a profiling unit that performs user profiling based on the statistical values ​​and the seasonal variation amount.

[0007] According to the present disclosure, it is possible to improve the accuracy of estimating a user's interests and concerns, and perform highly accurate user profiling.

[0008] 1 is a block diagram illustrating an example of a configuration of an information processing system according to a first embodiment of the present disclosure; FIG. 2 is a graph illustrating a first loss function; FIG. 3 is a diagram illustrating an arrangement of line segments when a lower limit coefficient and an upper limit coefficient are in an initial state; ｖ，ｉ，ｊFIG. 10 is a diagram illustrating an arrangement of line segments sorted in ascending order of . FIG. 11 is a diagram illustrating an arrangement of line segments after adjustment. FIG. 12 is a flowchart illustrating processing of an information processing system when a first method is adopted. FIG. 13 is a flowchart illustrating processing of an information processing system when a second method is adopted. FIG. 14 is a block diagram illustrating an example of a configuration of an information processing system according to a second embodiment of the present disclosure.

[0009] (Findings underlying the present disclosure) User profiling is performed by acquiring log data from applications (apps) used by users as behavioral data indicating the user's behavior, and using this behavioral data to estimate the characteristics of individual users. Because behavioral data includes external factors such as seasonal variation, performing user profiling using the behavioral data itself presents a challenge: high accuracy cannot be achieved. Seasonal variation is a variation corresponding to a predetermined period (season) common to all users. For example, because winter is cold, all users tend to take fewer steps per day compared to spring or autumn. Seasonal variation is a value that indicates such a trend common to all users depending on the season. In a health promotion service, when encouraging users to change their behavior to increase their step count, it is impossible to accurately grasp the increasing trend in the user's step count unless seasonal variation is removed. For example, even if a user's step count in winter is lower than that in spring, taking seasonal variation into account, it is possible that the user's step count in winter is actually higher than that in spring. In this case, it would be inappropriate to encourage the user to change their behavior to increase the number of steps they take just because the number of steps measured in winter is less than the number of steps measured in spring.

[0010] In conventional technologies such as Patent Document 1, user profiling is performed without observing such seasonal fluctuations, so it is not possible to improve the accuracy of estimating user interests and concerns and achieve highly accurate user profiling.

[0011] Therefore, the inventors have come up with the idea of ​​the present disclosure based on the knowledge that by estimating the amount of seasonal variation inherent in behavioral data and performing user profiling taking that amount of seasonal variation into account, it is possible to improve the accuracy of estimating user interests and concerns and perform highly accurate user profiling.

[0012] (1) An information processing device according to one aspect of the present disclosure includes an acquisition unit that acquires behavioral data indicating the behavior of each of a plurality of users; a pre-processing unit that calculates statistical values ​​of the behavioral data for each of the plurality of users over a predetermined period of time; a seasonal variation calculation unit that calculates a harmonic mean of the minimum and maximum values ​​of the statistical values ​​calculated by the pre-processing unit and calculates, based on the harmonic mean, a seasonal variation amount that is a variation amount common to the plurality of users that depends on the predetermined period; and a profiling unit that performs user profiling based on the statistical values ​​and the seasonal variation amount.

[0013] According to this configuration, statistical values ​​of behavioral data for a predetermined period are calculated from the user's behavioral data for that period, and seasonal fluctuations dependent on the predetermined period are calculated based on the harmonic mean of the maximum and minimum values ​​of the calculated statistical values. User profiling is then performed based on the statistical values ​​and seasonal fluctuations. This improves the accuracy of estimating the user's interests and concerns, enabling highly accurate user profiling.

[0014] (2) In the information processing device described in (1) above, the seasonal variation calculation unit calculates the average value and standard deviation of the statistical values ​​for all of the multiple users, and the maximum value may be a value obtained by adding an offset value obtained by multiplying the standard deviation by a predetermined coefficient to the average value, and the minimum value may be a value obtained by subtracting the offset value from the average value.

[0015] According to this configuration, the maximum and minimum values ​​are calculated based on the average and variance of the statistical values ​​for all users, and the harmonic mean is calculated using these maximum and minimum values. Therefore, this configuration makes it possible to calculate the seasonal variation taking into account the statistical values ​​of all users, and to estimate the seasonal variation robustly against outliers.

[0016] (3) In the information processing device described in (2) above, the seasonal fluctuation calculation unit may calculate the amount of seasonal fluctuation as the value obtained by subtracting the square of the offset value from the square of the average value and dividing the result by twice the average value.

[0017] This configuration allows the amount of seasonal variation to be calculated more accurately.

[0018] (4) In the information processing device described in any one of (1) to (3) above, the profiling unit may calculate, for each of the plurality of users, a value obtained by subtracting the amount of seasonal variation from the statistical value as the result of the user profiling.

[0019] According to this configuration, the value obtained by subtracting the amount of seasonal variation from the statistical value is used as the result of user profiling, thereby achieving more accurate user profiling.

[0020] (5) The information processing device according to any one of (1) to (4) above may further include an output unit that outputs the results of the user profiling.

[0021] According to this configuration, for example, the results of user profiling can be presented to the user.

[0022] (6) In the information processing device according to any one of (1) to (5) above, the predetermined period may be one period when a year is divided into a plurality of periods.

[0023] This configuration makes it possible to realize user profiling that takes into account seasonal variations, for example, every spring, summer, fall, winter, every few months, or every few weeks.

[0024] (7) In the information processing device described in any one of (1) to (6) above, the behavioral data may be data indicating the number of steps taken by the multiple users, and the statistical value may be an average value of the number of steps taken by each of the multiple users over the specified period.

[0025] According to this configuration, it is possible to evaluate the actual increase or decrease in the number of steps taken by the user, and to encourage the user to make appropriate behavioral changes to improve their health.

[0026] (8) In the information processing device described in any of (1) to (7) above, a correction unit may be further provided that acquires schedule information for each of the plurality of users, identifies an event date on which a specified event occurred for each of the plurality of users from the schedule information, and corrects the statistical value or the result of the user profiling corresponding to the event date.

[0027] For example, on an event day such as a wedding anniversary or birthday, a user is likely to behave differently than usual, so if the statistical values ​​for that day were used as is, it may be impossible to perform accurate user profiling. With this configuration, the statistical values ​​or user profiling results are corrected on the event day, thereby further improving the accuracy of user profiling.

[0028] (9) An information processing method in another aspect of the present disclosure is an information processing method in a computer, including: acquiring behavioral data indicating the behavior of each of a plurality of users; calculating statistical values ​​of the behavioral data for each of the plurality of users over a predetermined period of time; calculating a harmonic mean of the minimum and maximum values ​​of the calculated statistical values; calculating a seasonal variation amount, which is a variation amount common to the plurality of users that depends on the predetermined period of time, based on the harmonic mean value; and performing user profiling based on the statistical values ​​and the seasonal variation amount.

[0029] According to this configuration, an information processing method capable of performing highly accurate user profiling is provided.

[0030] (10) In another aspect of the present disclosure, an information processing program causes a computer to acquire behavioral data indicating the behavior of each of a plurality of users, calculate statistical values ​​of the behavioral data for each of the plurality of users over a predetermined period of time, calculate a harmonic mean of the minimum and maximum values ​​of the calculated statistical values, and calculate a seasonal variation amount, which is a variation amount common to the plurality of users that depends on the predetermined period, based on the harmonic mean value, and perform user profiling based on the statistical values ​​and the seasonal variation amount.

[0031] According to this configuration, an information processing program capable of performing highly accurate user profiling is provided.

[0032] It goes without saying that the present disclosure allows such information processing programs to be distributed via a computer-readable non-transitory recording medium such as a CD-ROM or a communication network such as the Internet.

[0033] Note that each of the embodiments described below represents a specific example of the present disclosure. The numerical values, shapes, components, steps, and step orders shown in the following embodiments are merely examples and are not intended to limit the present disclosure. Furthermore, among the components in the following embodiments, components that are not described in the independent claims that represent the highest concept are described as optional components. Furthermore, in all of the embodiments, the respective contents can be combined.

[0034] 1 is a block diagram showing an example of a configuration of an information processing system 100 according to a first embodiment of the present disclosure. The information processing system 100 includes an information processing device 1 and a terminal 4.

[0035] The terminal 4 is a device that acquires information about a user who is the subject of information processing. The terminal 4 is, for example, a smartphone, a personal computer, a wearable terminal, a physical measurement device, etc. The wearable terminal is, for example, smart glasses, a smart watch, etc. The physical measurement device is, for example, a pedometer. Note that the form of the terminal 4 is not limited to these examples.

[0036] The number of terminals 4 may be one or more. The terminal 4 inputs the acquired log of the target user (hereinafter referred to as the user log) to the information processing device 1 at a predetermined timing. In the first embodiment, the user log is behavioral data indicating the user's behavior. The behavioral data includes, for example, the number of steps taken by the user, the number of app operations performed by the user on an app, the number of operations performed by the user on a home appliance, and user location information. The user location information is acquired, for example, by a GPS sensor.

[0037] For example, the app invites a user to go for a walk to improve the user's health. In detail, the app accepts posts from users about recommended places to visit and posts the places to visit on a bulletin board that can be viewed by all users. The app accepts comments from other users about the places to visit and posts the comments on the bulletin board in association with the places to visit. In this way, the app can motivate the user to visit the places to visit and take a walk. Furthermore, the user can get to know other users through the comments and build a personal network.

[0038] The user's location information is used to identify the user's activity range. Examples of home appliances include washing machines, televisions, vacuum cleaners, and refrigerators.

[0039] The behavioral data may include behavioral data recorded at regular intervals (for example, every hour or every 10 minutes) and behavioral data recorded irregularly. The behavioral data recorded at regular intervals is, for example, user location information acquired by a GPS sensor. On the other hand, the behavioral data recorded irregularly is app operations. App operations include, for example, screen switching within an app, posting, comments, stamps indicating "like," etc. The information processing device 1 uses at least one of the behavioral data recorded at regular intervals and the behavioral data recorded irregularly.

[0040] In the following description, the behavioral data indicates the number of steps. In this case, the terminal 4 is configured with a smartphone, a wearable terminal, and a physical measurement device that have a function of measuring the number of steps.

[0041] The information processing device 1 is a device for profiling a target user using a user log. The information processing device 1 is configured as, for example, a server. The server may be a cloud server or an edge server. The information processing device 1 and the terminal 4 may be configured as a single computer. When the information processing device 1 is configured as a server, the information processing device 1 and the terminal 4 are connected via a network. The network is, for example, the Internet.

[0042] The information processing device 1 is configured as a computer including a processor 2, a memory 3, and an interface circuit (not shown).

[0043] The processor 2 is hardware for performing user profiling based on behavioral data, and is, for example, a central processing unit. The processor 2 includes an acquisition unit 21, a preprocessing unit 22, a seasonal variation calculation unit 23, a profiling unit 24, and an output unit 25. The acquisition unit 21 to the output unit 25 may be realized by the central processing unit executing an information processing program, or may be configured by dedicated hardware circuits such as ASICs. The acquisition unit 21 to the output unit 25 may be distributed among multiple computers. Some of the acquisition unit 21 to the output unit 25 may be included in the terminal 4.

[0044] The memory 3 is configured by a rewritable non-volatile storage device such as a solid state drive (SSD) or a hard disk drive (HDD). The memory 3 includes a log storage unit 31. The log storage unit 31 stores the behavior data acquired from the terminal 4 by the acquisition unit 21.

[0045] The acquisition unit 21 acquires the behavioral data stored by the terminal 4, and stores the acquired behavioral data in the log storage unit 31. The log storage unit 31 stores the behavioral data input from the acquisition unit 21.

[0046] The preprocessing unit 22 reads out the behavioral data from the log storage unit 31 and performs statistical processing on the read out behavioral data so that it can be used for user profiling. Specifically, the preprocessing unit 22 calculates statistical values ​​of the behavioral data for a predetermined period. The statistical values ​​are, for example, the average number of steps per day for each of a plurality of users for a predetermined period. Hereinafter, any one of the plurality of users will be referred to as user i.

[0047] For example, the log storage unit 31 stores behavioral data indicating the number of steps taken per hour. In this case, the preprocessing unit 22 calculates the number of steps taken per day by the user i by adding up the number of steps taken per hour. The preprocessing unit 22 then calculates the average value of the number of steps taken per day by the user i over a predetermined period as a statistical value.

[0048] The statistical value is not limited to the average number of steps taken by user i per day. For example, the statistical value may be the total number of steps taken by user i per day for a predetermined period, the median number of steps taken by user i per day for a predetermined period, or the minimum or maximum number of steps taken by user i per day for a predetermined period.

[0049] The predetermined period is the period for which the amount of seasonal variation is calculated. The predetermined period is one period when a year is divided into multiple periods, such as three months, four months, or three weeks. Furthermore, the statistical value is the average number of steps per day, but it may also be the average number of steps over a unit period. The unit period is one period that constitutes the predetermined period, such as one hour, one day, one week, or one month.

[0050] The seasonal variation calculation unit 23 calculates the amount of seasonal variation based on the statistical values ​​calculated by the preprocessing unit 22. Specifically, the seasonal variation calculation unit 23 calculates the harmonic mean of the minimum and maximum values ​​of the statistical values ​​calculated by the preprocessing unit 22, and calculates the amount of behavioral variation based on the harmonic mean. The amount of behavioral variation is a variation common to multiple users that depends on a predetermined period, and is a scalar value. Specifically, the seasonal variation calculation unit 23 calculates the amount of seasonal variation using equation (10) described below.

[0051] Alternatively, the seasonal variation calculation unit 23 may calculate the amount of seasonal variation using equation (11) described below. In this case, the seasonal variation calculation unit 23 calculates the average value and variance value of the statistical values ​​for all of the multiple users. Then, the seasonal variation calculation unit 23 may calculate the amount of seasonal variation by subtracting the square of the offset value from the square of the average value and dividing the result by twice the average value.

[0052] The profiling unit 24 performs user profiling based on the statistical values ​​calculated by the preprocessing unit 22 and the seasonal variation calculated by the seasonal variation calculation unit 23. In particular, the profiling unit 24 calculates a user profiling value for each of a plurality of users by subtracting the seasonal variation from the statistical values. For example, if the behavioral data indicates the number of steps, the profiling unit 24 calculates the effective number of steps by subtracting the seasonal variation from the number of steps. This calculates the effective number of steps from which seasonal variation, such as the overall decrease in the number of steps in winter, has been removed. For example, a user who is interested in health and intends to increase the number of steps they take, will gradually increase the effective number of steps. In this case, the profiling unit 24 can infer that the user is interested in health.

[0053] The output unit 25 outputs the user profiling values ​​calculated by the profiling unit 24. The output destination is, for example, a user terminal used by the user or an administrator terminal of an administrator that provides services to the user. The user terminal or administrator terminal displays the user profiling values ​​on a display. The user terminal or administrator terminal may display a display screen on the display including, for example, today's observed step count, today's effective step count, and the amount of seasonal variation for a specified period to which today belongs. This display screen may include a message informing the user that the effective step count is on the rise if the effective step count is on the decline. Furthermore, this display screen may include a message encouraging the user to take a walk if the effective step count is on the decline.

[0054] The principle of calculating the amount of seasonal variation will be described below. The following describes an example of a process for calculating the amount of seasonal variation for a predetermined period from the number of steps taken by M users over the one year period from January 1, 2022 to December 31, 2022. The predetermined period is one of four periods: January to March, April to June, July to September, and October to December. Note that, although the number of steps is used as an example in the following description, the present disclosure is not limited thereto. For example, if the behavioral data includes the browsing time or the number of app operations for an SNS app, the amount of seasonal variation for the browsing time or the number of app operations is estimated.

[0055] In a predetermined period j, the observed value of the number of steps of user i on day t (hereinafter referred to as the observed number of steps) is defined as η ｉ，ｊ，ｔ The amount of seasonal fluctuation in a given period j is expressed as ω ｊ In this case, the seasonal variation ω ｊ The effective number of steps η' is the number of steps excluding ｉ，ｊ，ｔ is expressed by the formula (1).

[0056]

[0057] Observed steps η ｉ，ｊ，ｔ is the individual characteristics of user i, such as how much the user likes to go out, and the seasonal variation ω ｊ The average number of steps for user i is determined by the individual characteristics of user i, and the seasonal variation ω ｊ If we consider that the number of steps varies depending on ｉ，ｊ，ｔ is a normal distribution N(μ ｐ，ｊ，ｔ , σ ２ ｐ，ｊ，ｔ ) where μ ｐ，ｊ，ｔ is the effective number of steps η´ ｉ，ｊ，ｔ (hereinafter referred to as the average effective step count), and σ ２ ｐ，ｊ，ｔ is the effective number of steps η´ ｉ，ｊ，ｔ Therefore, the effective number of steps η' ｉ，ｊ，ｔ The standard deviation of is σ ｐ，ｊ，ｔ It is considered that the number of steps taken by user i on day t does not affect the number of steps taken the next day. Therefore, the effective number of steps η' ｉ，ｊ，ｔ is expressed as a normal distribution that is independent of the time series.

[0058] Seasonal variation in observed steps η ｉ，ｊ，ｔ The effect on is common to all users and is normally distributed N(μ ｓ，ｊ , σ ２ ｓ，ｊ ) In this case, due to the reproductive property of the probability distribution, the observed step count η ｉ，ｊ，ｔ is the effective number of steps η´ ｉ，ｊ，ｔ and seasonal variation ω ｊ The average value of the seasonal fluctuations μ ｓ，ｊ ) and the correlation coefficient ρ ｊ The resulting distribution follows the normal distribution shown in equation (2).

[0059]

[0060] In formula (2), μ ｖ，ｉ，ｊ is the observed number of steps of user i in a predetermined period j, η ｉ，ｊ，ｔ The mean value of the normal distribution N shown in Equation (2), that is, the mean observed step count μ ｖ，ｉ，ｊ is the average effective step count μ ｐ，ｉ，ｊ and the mean seasonal variation μ ｓ，ｊ The average observed step count μ ｖ，ｉ，ｊ The average seasonal variation μ ｓ，ｊ Let R be the proportion of user i. This proportion R will never be 0% or 100%. This is because if the proportion R is 0%, there will be no seasonal variation. Also, if the proportion R is 100%, there will be no characteristics of user i.

[0061] Therefore, it can be assumed that the ratio R has a lower limit coefficient a and an upper limit coefficient b (hereinafter referred to as assumption "1"). Furthermore, since the lower limit coefficient a and the upper limit coefficient b are coefficients of the ratio R, it can be assumed that 0<a, b<1 (hereinafter referred to as assumption "2").

[0062] In order to determine the lower limit coefficient a and the upper limit coefficient b that satisfy assumptions "1" and "2," a first loss function "Loss" is defined. Then, the lower limit coefficient a and the upper limit coefficient b that minimize this first loss function "Loss" are calculated.

[0063] FIG. 2 shows the first loss function Loss(μ ｖ，ｉ，ｊ , μ ｓ，ｊ 2, the y-axis represents the loss and the x-axis represents the ratio R (=μ ｓ，ｊ / μ ｖ，ｉ，ｊ ) is shown.

[0064] First loss function Loss(μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is expressed as 0 for the error in the range between the lower limit coefficient a and the upper limit coefficient b, and μ ｓ，ｊ / μ ｖ，ｉ，ｊ The error in the range where is smaller than the lower limit coefficient a is μ ｓ，ｊ / μ ｖ，ｉ，ｊ and the lower limit coefficient a, and μｓ，ｊ / μ ｖ，ｉ，ｊ The error in the range where is greater than the upper limit coefficient b is μ ｓ，ｊ / μ ｖ，ｉ，ｊ and the upper limit coefficient b, which is expressed by the square of the distance between them.

[0065]

[0066] However, μ ｓ，ｊ / μ ｖ，ｉ，ｊ The data that is 0 is excluded as missing data. ｖ，ｉ，ｊ , μ ｓ，ｊ ) to obtain μ ｓ，ｊ / μ ｖ，ｉ，ｊ It is possible to prevent the seasonal variation ω from becoming a value close to 0 or 1. ｊ can be estimated, and a highly accurate user profile can be realized.

[0067] First loss function Loss(μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is the sum of all users' values ​​as the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) The first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is expressed by equation (4).

[0068]

[0069] Here, the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) can be reduced by increasing the number of users for which the first loss function Loss is 0. Therefore, consider the conditions for the lower limit coefficient a and the upper limit coefficient b that make the first loss function Loss of all users 0. The shape of the first loss function Loss is shown in FIG. 2. Therefore, the conditions for the lower limit coefficient a and the upper limit coefficient b that make the first loss function Loss 0 are expressed by equation (5).

[0070]

[0071] From equation (5), the lower limit coefficient a and the upper limit coefficient b are aμ ｖ，ｉ，ｊ ≦μ ｓ，ｊ ≦bμ ｖ，ｉ，ｊIf all users satisfy this condition, the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) becomes 0 and is minimized. This condition is satisfied by the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is one of the best conditions for making the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is 0.

[0072] First objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is the sum of the first loss function Loss, but the first loss function Loss has a different shape from a general loss function. A general loss function has a loss that becomes 0 only at one point in the domain. On the other hand, the first loss function Loss becomes 0 in the range from the lower limit coefficient a to the upper limit coefficient b, as shown in Figure 2. That is, the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) can have multiple values ​​for the variable that are 0.

[0073] First objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ When applying a general nonlinear optimization (for example, the Levenberg-Marquardt method) to the first objective function L, there is a flat region in which the loss function is 0 in the range from the lower limit coefficient a to the upper limit coefficient b. １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is not stably optimized. This is because the existence of this flat region makes it difficult to optimize the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ This is because a saddle point occurs on the error surface of the first loss function Loss, and the saddle point is likely to be calculated as a local optimum solution rather than a global optimum solution. Note that the error surface refers to the surface when the first loss function Loss is regarded as a surface.

[0074] On the other hand, the present disclosure focuses on the fact that the first loss function Loss has a flat region, and １，ｊ (μｖ，ｉ，ｊ , μ ｓ，ｊ We have found that the conditions under which the equation (1) becomes 0 can be analytically calculated.

[0075] First objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is a nonlinear minimization problem, but analytically the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) to 0. Since there are multiple values ​​in the domain where the first loss function Loss is the minimum value, the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is minimized.

[0076] <First objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ 3 to 5 show the minimization of the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ 3 to 5, the first objective function L is minimized. １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) becomes 0. FIG. 3 is a layout diagram of a line segment 300 when the lower limit coefficient a and the upper limit coefficient b are in the initial state. The line segment 300 is ｖ，ｉ，ｊ and bμ ｖ，ｉ，ｊ In the initial state, the line segments 300 are arranged in the order of the user indexes.

[0077] Adjust the lower limit coefficient a and the upper limit coefficient b to obtain the average seasonal fluctuation μ ｓ，ｊ If the vertical line of the line 300 passes through all the line segments 300, the loss function becomes 0. Here, the lower limit coefficient a, the upper limit coefficient b, and the observed step count average value μ ｖ，ｉ，ｊ , the mean seasonal variation μ ｓ，ｊ are all positive, the set {aμ ｖ，１，ｊ , ..., aμ ｖ，M，ｊ , b μｖ，１，ｊ , ..., bμ ｖ，M，ｊ} the minimum value μ ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊThe element that satisfies the condition does not depend on the values ​​of the lower limit coefficient a and the upper limit coefficient b.

[0078] FIG. 4 shows a ｖ，ｉ，ｊ 4 is a layout diagram of the line segments 300 sorted in ascending order of the lower limit coefficient a, the upper limit coefficient b, and the observed step count average value μ ｖ，ｉ，ｊ , the mean seasonal variation μ ｓ，ｊ is obvious since all are positive.

[0079] First objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) to 0, the minimum value μ ｖ，ｍｉｎ，ｊ The corresponding user and maximum value μ ｖ，ｍａｘ，ｊ The lower limit coefficient a and the upper limit coefficient b are adjusted by focusing on the user corresponding to the minimum value μ ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊ is the number of users with two or more and the observed step count average μ ｖ，ｉ，ｊ Unless the values ​​of are the same, they will always be different.

[0080] 5 is a diagram showing the layout of the line segment 300 after adjustment. ｖ，ｍｉｎ，ｊ The right end bμ of the line segment 300 corresponding to ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊ The left end aμ of the line segment 300 corresponding to ｖ，ｍａｘ，ｊ The lower limit coefficient a and the upper limit coefficient b are selected so that the right end bμ ｖ，ｍｉｎ，ｊ = left end aμ ｖ，ｍａｘ，ｊ Here, the length of the line segment 300 is in ascending order. ｓ，ｊ passes through all the line segments 300, and the first objective function L １，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is 0. Therefore, the first objective function Ｌ１，ｊ (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is set to 0 as the condition (condition "3") of the lower limit coefficient a and the upper limit coefficient b. ｖ，ｍｉｎ，ｊ = aμ ｖ，ｍａｘ，ｊ is obtained.

[0081] <Minimization of the second objective function> bμ ｖ，ｍｉｎ，ｊ = aμ ｖ，ｍａｘ，ｊThere are an infinite number of combinations of the lower limit coefficient a and the upper limit coefficient b that satisfy the above equation.

[0082] Consider the case where there is no constraint other than 0<b<1 on the upper limit coefficient b of the ratio R. If the upper limit coefficient b is set to a value close to 1, as shown in FIG. 2, the first loss function Loss(μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is no longer a downward convex shape. In this case, the first loss function Loss(μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) approaches the shape of a straight line where the loss is zero in all intervals, and loses its properties as a loss function. ｊ The estimated value of bμ becomes meaningless. ｖ，ｍｉｎ，ｊ = aμ ｖ，ｍａｘ，ｊ It is necessary to determine an appropriate combination of the lower limit coefficient a and the upper limit coefficient b from among combinations of the lower limit coefficient a and the upper limit coefficient b that satisfy the above.

[0083] The first loss function Loss (μ ｖ，ｉ，ｊ , μ ｓ，ｊ ) is expected to have a downward convex shape, so a search is made for a pair of upper limit coefficient b and lower limit coefficient a such that upper limit coefficient b takes the smallest possible value and lower limit coefficient a takes the largest possible value.

[0084] In a general minimization problem, maximizing the lower limit coefficient a is defined as minimizing -a, and minimizing b - a is performed. To minimize b - a, it is sufficient to simply minimize the upper limit coefficient b. In this case, the upper limit coefficient b becomes 0, and the upper limit coefficient b = lower limit coefficient a = 0, so the lower limit coefficient a and the upper limit coefficient b become inappropriate values. Therefore, the second objective function shown below is defined.

[0085] <Explanation of the second objective function> Effective number of steps η´ ｉ，ｊ，ｔ and seasonal variation ω ｊ Both follow a normal distribution, and it is noted that a normal distribution has symmetry with respect to the mean value. In this case, it can be considered that symmetry also holds for the range from the lower limit coefficient a to the upper limit coefficient b. In other words, the lower limit coefficient a and the upper limit coefficient b satisfy the condition of formula (6).

[0086]

[0087] According to equation (6), the lower limit coefficient a and the upper limit coefficient b are linked, and for example, when the lower limit coefficient a is 0.3, the upper limit coefficient b is 0.7, and there is symmetry.

[0088] Furthermore, in addition to the symmetry regarding the lower limit coefficient a and the upper limit coefficient b, it is considered that the shape of the first loss function Loss is also symmetric. Focusing on FIG. 2, the range of the lower limit coefficient a and the upper limit coefficient b is 0<a, b<1, and when b<1 / 2, the first loss function Loss is clearly not symmetric in the range from 0 to 1. The second objective function L that makes the first loss function Loss symmetric ２，ｊ From FIG. 2, if the midpoint between the lower limit coefficient a and the upper limit coefficient b is close to 1 / 2, the symmetry of the first loss function Loss becomes high. Therefore, the second objective function L ２，ｊ is expressed by equation (7).

[0089]

[0090] bμ ｖ，ｍｉｎ，ｊ = aμ ｖ，ｍａｘ，ｊ Therefore, equation (7) becomes equation (8), and the second objective function L ２，ｊ The lower limit coefficient a and the upper limit coefficient b at which σ becomes 0 are expressed by the following equation (9).

[0091] Also, μ ｓ，ｊ = aμ ｖ，ｍａｘ，ｊ Therefore, the average seasonal variation μ ｓ，ｊ is expressed by equation (10).

[0092]

[0093] Here, HarmonicMean is the harmonic mean value. Therefore, the mean seasonal fluctuation value μ ｓ，ｊ can be calculated using equation (10). The mean seasonal variation μ ｓ，ｊ is the constant seasonal fluctuation ω in a given period j ｊ It is considered to be.

[0094] That is, the preprocessing unit 22 calculates the observed step count η for each of all users. ｉ，ｊ，ｔ From the observed step count average μ ｖ，ｉ，ｊ The seasonal variation calculation unit 23 calculates the average observed step counts μ ｖ，ｉ，ｊ The minimum value μ ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊThe seasonal variation calculation unit 23 determines the determined minimum value μ ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊ By substituting into equation (10), the mean seasonal fluctuation μ ｓ，ｊ The seasonal variation ω ｊ Hereinafter, the seasonal variation ω is calculated using equation (10). ｊ The method for calculating is called the first method.

[0095] Second objective function L ２，ｊ By using the above formula, an appropriate set of lower limit coefficient a and upper limit coefficient b can be determined from the sets of lower limit coefficient a and upper limit coefficient b, and seasonal fluctuations can be estimated with high accuracy, thereby realizing highly accurate user profiling.

[0096] <Robust Estimation> In the calculation of equation (10), the minimum value μ ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊ The seasonal variation ω ｊ Therefore, the seasonal variation ω ｊ We propose a more robust method to estimate the average observed step count for all users μ ｖ，ｉ，ｊ Consider a similar estimation from the distribution of the observed step count average μ ｖ，ｉ，ｊ is not guaranteed to be normally distributed, but the number of steps increases as a person walks, so there is a possibility that the observed step count average μ ｖ，ｉ，ｊ is not likely to occur frequently. In other words, the observed step count average μ ｖ，ｉ，ｊ It is believed that the distribution of is such that the central limit theorem, such as a long-tailed power-law distribution, holds. Therefore, if the number of users M is sufficiently large, the observed step count average value μ ｖ，ｉ，ｊ The distribution of can be approximated to a normal distribution by the central limit theorem. ｖ，ｉ，ｊ may be omitted.

[0097] Observed step count average μ ｖ，ｉ，ｊ The distribution of all users is normal distribution N(μ ｖ，ｊ , σ ２ ｖ，ｊ ) and the minimum and maximum values ​​of the harmonic mean are considered to follow the normal distribution N(μ ｖ，ｊ , σ ２ ｖ，ｊ) tail values ​​are used.

[0098] In detail, in the second method, the minimum value is μ ｖ，ｊ -cσ ｖ，ｊ is adopted, and μ ｖ，ｊ -cσ ｖ，ｊ will be adopted.

[0099] μ ｖ，ｊ is the average observed step count for all users μ ｖ，ｉ，ｊ is the average value of (hereinafter referred to as the overall average value). ｖ，ｊ is the average observed step count for all users μ ｖ，ｉ，ｊ is the standard deviation of the whole (hereinafter referred to as the overall standard deviation). c is a positive coefficient that determines the width of the tail, and is, for example, 1. However, the value of the coefficient c is not limited to 1, and an appropriate value such as 2 or 3 can be used. Furthermore, the coefficient c may be a decimal value.

[0100] In addition, the overall average value μ ｖ，ｊ is an example of the average statistical value for all users. Overall standard deviation σ ｖ，ｊ is an example of the standard deviation of the statistics across all users. ｖ，ｊ is an example of an offset value obtained by multiplying the standard deviation by a predetermined coefficient.

[0101] In the second method, equation (10) is expressed as equation (11).

[0102]

[0103] μ´ ｓ，ｊ is the average seasonal variation in the second method. For a set of users with a large variance, the average seasonal variation μ´ ｓ，ｊ Since the observed step count average value μ ｖ，ｉ，ｊ The average effective step count μ ｐ，ｉ，ｊ The seasonal fluctuation calculation unit 23 calculates the average seasonal fluctuation amount μ′ ｓ，ｊ The seasonal fluctuation amount ω that takes a constant value in a predetermined period j ｊ Adopted as.

[0104] In this way, the second method uses data from all users, so the seasonal fluctuation amount ω ｊ can be estimated.

[0105] 6 is a flowchart showing the processing of the information processing system 100 when the first method is adopted. This flowchart is executed, for example, once a day.

[0106] First, in step S1, the acquisition unit 21 acquires the observed number of steps η in a predetermined period j to which today belongs. ｉ，ｊ，ｔ is acquired from the log storage unit 31. ｉ，ｊ，ｔ is an example of behavioral data.

[0107] Next, in step S2, the preprocessing unit 22 calculates the observed number of steps η for each user i in a predetermined period j. ｉ，ｊ，ｔ By calculating the average value of ｖ，ｉ，ｊ Calculate the observed step count average value μ ｖ，ｉ，ｊ is an example of a statistic.

[0108] Next, in step S3, the seasonal variation calculation unit 23 calculates the observed step count average value μ ｖ，ｉ，ｊ The minimum value μ from the set ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊ Identify.

[0109] Next, in step S4, the seasonal variation calculation unit 23 calculates the minimum value μ ｖ，ｍｉｎ，ｊ and the maximum value μ ｖ，ｍａｘ，ｊ Substituting into equation (10), the mean seasonal variation μ ｓ，ｊ Calculate the seasonal variation ω ｊ get.

[0110] Next, in step S5, the profiling unit 24 calculates the observed number of steps η of each user today. ｉ，ｊ，ｔ From the seasonal variation ω ｊ (Equation (1)) to calculate the user profiling value. The profiling unit 24 may also calculate the effective number of steps for the past few days (for example, three days) including today. In this case, the profiling unit 24 calculates the average value of the observed number of steps for each user for the past three days including today as the observed number of steps η ｉ，ｊ，ｔ It should be adopted as such.

[0111] Next, the output unit 25 outputs the user profiling value to the user terminal or the administrator terminal.

[0112] 7 is a flowchart showing the processing of the information processing system 100 when the second method is adopted. Steps S11 and S12 are the same as steps S1 and S2 shown in FIG.

[0113] In step S13, the seasonal variation calculation unit 23 calculates the observed step count average value μ ｖ，ｉ，ｊ By calculating the average value of ｖ，ｊ Calculate.

[0114] Next, in step S14, the seasonal variation calculation unit 23 calculates the observed step count average value μ ｖ，ｉ，ｊ and the overall average value μ ｖ，ｊ and the overall standard deviation σ ｖ，ｊ Calculate.

[0115] Next, in step S15, the seasonal fluctuation calculation unit 23 calculates the overall average value μ ｖ，ｊ and the overall standard deviation σ ｖ，ｊ and are substituted into equation (11) to obtain the mean seasonal variation μ' ｓ，ｊ The seasonal variation ω ｊ It is calculated as follows.

[0116] The processes in steps S16 and S17 are the same as those in steps S5 and S6 shown in FIG.

[0117] As described above, according to the first embodiment, statistical values ​​of behavioral data for a predetermined period are calculated from the behavioral data of a user for that period, and seasonal fluctuations dependent on the predetermined period are calculated based on the harmonic mean of the maximum and minimum values ​​of the calculated statistical values. User profiling is then performed based on the statistical values ​​and seasonal fluctuations. This improves the accuracy of estimating user interests and concerns, enabling highly accurate user profiling.

[0118] Second Embodiment In a second embodiment, statistical values ​​or user profiling values ​​are corrected. Fig. 8 is a block diagram showing an example of the configuration of an information processing system 100A according to the second embodiment of the present disclosure.

[0119] The information processing system 100A differs from the information processing system 100 in that the processor 2A further includes a correction unit 26.

[0120] The correction unit 26 acquires schedule information for each of the multiple users and identifies an event date on which a predetermined event occurred for each of the multiple users from the schedule information. The correction unit 26 acquires a user profiling value from the profiling unit 24 and corrects the user profiling value if the acquired user profiling value corresponds to the event date. The correction unit 26 inputs the corrected user profiling value to the output unit 25.

[0121] The correction unit 26 may acquire user schedule information from, for example, an external server. Event days are special days on which users are likely to behave differently than usual, such as birthdays and wedding anniversaries. Schedule information is information indicating each user's daily plans.

[0122] The correction unit 26 may delete the user profiling value corresponding to the event date, or may reduce the user profiling value by multiplying the user profiling value by a predetermined coefficient d. An example of the coefficient d is 1 / 2. However, this is just an example, and the coefficient d can be any appropriate value less than 1, such as 1 / 4 or 3 / 4.

[0123] The correction unit 26 acquires statistical values ​​from the preprocessing unit 22, and if the acquired statistical values ​​correspond to an event day, corrects the statistical values. The correction unit 26 inputs the corrected statistical values ​​to the seasonal variation calculation unit 23. For example, the correction unit 26 may delete the statistical values ​​corresponding to the event day, or may multiply the statistical values ​​by a coefficient d. The correction unit 26 inputs the corrected statistical values ​​to the seasonal variation calculation unit 23. As a result, the seasonal variation amount ω ｊ When calculating the seasonal variation ω, the influence of the statistical values ​​on the event day is reduced, and the seasonal variation ω is calculated more accurately. ｊ can be calculated.

[0124] (Application Examples) An application example of user profiling will be described below. User behavior varies depending on the seasons, such as spring, summer, autumn, and winter, or months (January to December). As shown in the first embodiment, the average number of steps taken per day decreases in winter when temperatures are low and many users feel cold. In summer when temperatures are high and many users sweat more frequently, the average frequency of washing machine use per day increases. In a social network service (SNS) app, the number of posts in December, which has many events, is higher than the number of posts in June, which has no holidays and few events. In this way, seasonal variations also exist in app operations such as SNS apps.

[0125] An application example of the present disclosure will be described below using the number of steps taken by a user in winter as an example. The profiling unit 24 performs user profiling using the amount of seasonal variation calculated by the seasonal variation calculation unit 23. The profiling unit 24 calculates the effective number of steps as a user profiling value by subtracting the amount of seasonal variation from the observed number of steps. For example, for a user who is interested in health and intends to increase the number of steps taken, the effective number of steps will gradually increase, making user profiling possible from the effective number of steps. The effective number of steps is the number of steps from which seasonal variation due to cold winter weather, etc. has been removed. Therefore, the profiling unit 24 can infer that this user is a user who is interested in health. The output unit 25 outputs the user profiling value calculated by the profiling unit 24. The output destination is, for example, a display on a user terminal or an administrator terminal.

[0126] The above application example will be described in more detail below. The terminal 4 is a smartphone used by a resident of a certain community. A step count recording app designed to promote health is installed on the smartphone. The acquisition unit 21 acquires log data from the step count recording app as behavioral data and stores it in the log storage unit 31. The log storage unit 31 is, for example, a database built on the cloud, and stores the behavioral data acquired by the acquisition unit 21 from the smartphone. The behavioral data includes the number of steps taken by the user.

[0127] The processors 2 and 2A execute batch processing related to user profiling, for example, at 11:00 PM every day. The processors 2 and 2A calculate seasonal variations in step counts for the behavioral data acquired from the log storage unit 31 and perform user profiling based on the effective step counts. The processors 2 and 2A notify the administrator of an organization aiming to promote health (e.g., a department promoting health promotion in a local government) of the user profiling values. This allows the administrator to determine whether the step count recording app has increased the user's step count. The administrator can determine whether the user's effective step count has increased by comparing the user's effective step count immediately after installation of the step count recording app (the average number of steps taken over three days after installation) with the user's effective step count after three months of use. In the present disclosure, seasonal variations are removed from the observed step counts. This allows the administrator to confirm whether the step count recording app is effective in increasing residents' step counts. Thus, the present disclosure is useful for administrators of organizations aiming to promote health by enabling them to confirm whether the step count recording app is an effective app for increasing step counts.

[0128] The processor 2A also includes a correction unit 26. The correction unit 26 corrects the user profile values ​​or statistical values ​​based on the user's schedule information. For example, the correction unit 26 acquires the user's calendar information or schedule information. The correction unit 26 identifies an event date, such as a wedding anniversary, from the calendar information or schedule information.

[0129] On wedding anniversaries, users travel far away, take various photos, and post them to social media apps. Therefore, the number of steps taken and the number of app operations performed on social media apps on wedding anniversaries are likely to be so high that they can be considered outliers compared to non-event days. Therefore, the correction unit 26 performs a correction to exclude the statistical values ​​or user profiling values ​​on the event day. Alternatively, the correction unit 26 performs a correction to reduce the magnitude of the statistical values ​​or user profiling values ​​on the event day by half. This reduces the influence of data corresponding to outliers, enabling more accurate user profiling.

[0130] While the above provides examples of user profiling using the user's step count and the number of app operations, the step count is an example of user profiling using behavioral data in real space, and the number of app operations is an example of user profiling using behavioral data in cyberspace. User profiling using behavioral data in real space is likely to be able to estimate a user's interests and concerns with higher accuracy than user profiling using behavioral data in cyberspace.

[0131] The information device of the present disclosure is useful for user profiling.

Claims

1. An information processing device comprising: an acquisition unit that acquires behavioral data indicating the behavior of each of a plurality of users; a pre-processing unit that calculates statistical values ​​of the behavioral data for a predetermined period for each of the plurality of users; a seasonal variation calculation unit that calculates the harmonic mean of the minimum and maximum values ​​of the statistical values ​​calculated by the pre-processing unit and calculates a seasonal variation amount, which is a variation amount common to the plurality of users that depends on the predetermined period, based on the harmonic mean value; and a profiling unit that performs user profiling based on the statistical values ​​and the seasonal variation amount.

2. The information processing device of claim 1, wherein the seasonal variation calculation unit calculates the average value and standard deviation of the statistical values ​​for all of the multiple users, the maximum value is a value obtained by adding an offset value obtained by multiplying the standard deviation by a predetermined coefficient to the average value, and the minimum value is a value obtained by subtracting the offset value from the average value.

3. The information processing device according to claim 2, wherein the seasonal variation calculation unit calculates the amount of seasonal variation by subtracting the square of the offset value from the square of the average value and dividing the result by twice the average value.

4. The information processing device according to claim 1 or 2, wherein the profiling unit calculates, for each of the plurality of users, a value obtained by subtracting the amount of seasonal variation from the statistical value as the result of the user profiling.

5. The information processing device according to claim 1 or 2, further comprising an output unit that outputs the results of the user profiling.

6. The information processing device according to claim 1 or 2, wherein the predetermined period is one period when a year is divided into a plurality of periods.

7. An information processing device according to claim 1 or 2, wherein the behavioral data is data indicating the number of steps taken by the plurality of users, and the statistical value is an average value of the number of steps taken by each of the plurality of users during the specified period.

8. An information processing device as described in claim 1 or 2, further comprising a correction unit that acquires schedule information for each of the multiple users, identifies an event date on which a specified event occurred for each of the multiple users from the schedule information, and corrects the statistical value or the result of the user profiling corresponding to the event date.

9. An information processing method in a computer, comprising: acquiring behavioral data indicating the behavior of each of a plurality of users; calculating statistical values ​​of the behavioral data for each of the plurality of users over a predetermined period of time; calculating the harmonic mean of the minimum and maximum values ​​of the calculated statistical values, and calculating a seasonal variation amount, which is a variation amount common to the plurality of users that depends on the predetermined period, based on the harmonic mean value; and performing user profiling based on the statistical values ​​and the seasonal variation amount.

10. An information processing program that causes a computer to perform the following operations: acquire behavioral data indicating the behavior of each of a plurality of users; calculate statistical values ​​of the behavioral data for each of the plurality of users over a predetermined period; calculate the harmonic mean of the minimum and maximum values ​​of the calculated statistical values, and calculate a seasonal variation amount, which is a variation amount common to the plurality of users that depends on the predetermined period, based on the harmonic mean value; and perform user profiling based on the statistical values ​​and the seasonal variation amount.

Images