Learning device, stress estimation device, learning method, stress estimation method, and program

Abstract

This learning device 1X mainly has a classification means 14X and a learning means 17X. The classification means 14X classifies an observed feature quantity pertaining to a subject such that an index representing a correlation between the observed feature quantity and a correct stress value corresponding to the observed feature quantity becomes higher than before the classification. The learning means 17X trains a stress estimation model for estimating the relationship between observed feature quantity and stress value, for at least each class categorized by classification, on the basis of the observed feature quantity and the correct stress value.

Description

[Technical Field]

[0001] The present disclosure relates to the technical fields of a learning device, a stress estimation device, a learning method, a stress estimation method, and a storage medium that perform processing related to estimation of a stress state. [Background technology]

[0002] There are known devices or systems for determining the stress state of a subject based on data measured from the subject. For example, Patent Document 1 discloses a portable stress measurement device that determines the temporary stress level of a subject for each day based on the subject's biological data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2007-275287 Summary of the Invention [Problem to be solved by the invention]

[0004] When estimating a subject's stress level from their biometric data, there was a problem in that the estimation accuracy was not stable for unknown data.

[0005] In view of the above-mentioned problems, one of the objects of the present disclosure is to provide a learning device, a stress estimation device, a learning method, a stress estimation method, and a storage medium that perform processing to obtain stress estimation results with stable estimation accuracy. [Means for solving the problem]

[0006] One aspect of the learning device is After tentatively dividing the observed feature values ​​of the subject into a plurality of classes based on attribute information representing the attributes of the subject, each of the observed feature values ​​is moved between the plurality of classes, and if an index representing a correlation between the observed feature value after the movement and a correct stress value corresponding to the observed feature value becomes higher than before the movement, the movement is adopted, and the index is moved. Before a classification means for classifying the observed feature by returning the state before the movement when the observed feature becomes lower than the state before the movement; a learning means for learning a stress estimation model that estimates a relationship between the observed feature value and the stress value for at least each class divided by the classification, based on the observed feature value and the correct stress value; It is a learning device having the following.

[0007] One aspect of the stress estimation device is Subject to stress estimation Based on attribute information representing the attributes of the person to be estimated and a classification model, a classification score calculation means for calculating a classification score representing a degree of certainty that the observed feature value of the estimation target person belongs to each of a plurality of classes; a stress estimation means for acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; an integration means for calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; With death, the classification model is a model that classifies the observed feature values ​​for learning so that an index representing a correlation between the observed feature values ​​and a correct stress value corresponding to the observed feature values ​​becomes higher than that before classification. It is a stress estimation device.

[0008] The computer After tentatively dividing the observed feature values ​​of the subject into a plurality of classes based on attribute information representing the attributes of the subject, each of the observed feature values ​​is moved between the plurality of classes, and if an index representing a correlation between the observed feature value after the movement and a correct stress value corresponding to the observed feature value becomes higher than before the movement, the movement is adopted, and the index is moved. Before If the difference is smaller than the value of the observation feature, the state before the movement is restored, and the observed feature is classified. learning a stress estimation model that estimates a relationship between the observed feature value and the stress value for at least each class divided by the classification, based on the observed feature value and the correct stress value; It is a learning method. Note that the "computer" includes any electronic device (or a processor included in an electronic device), and may be configured from multiple electronic devices.

[0009] One aspect of the stress estimation method includes: The computer Subject to stress estimation Based on attribute information representing the attributes of the person to be estimated and a classification model, calculating a classification score representing a degree of certainty that the observed feature of the estimation subject belongs to each of a plurality of classes; acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; A stress estimation value is calculated by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score. death, the classification model is a model that classifies the observed feature values ​​for learning so that an index representing a correlation between the observed feature values ​​and a correct stress value corresponding to the observed feature values ​​becomes higher than that before classification. This is a stress estimation method.

[0010] One aspect of the program is After tentatively dividing the observed feature values ​​of the subject into a plurality of classes based on attribute information representing the attributes of the subject, each of the observed feature values ​​is moved between the plurality of classes, and if an index representing a correlation between the observed feature value after the movement and a correct stress value corresponding to the observed feature value becomes higher than before the movement, the movement is adopted, and the index is moved. Before If the difference is smaller than the value of the observation feature, the state before the movement is restored, and the observed feature is classified. This is a program that causes a computer to execute a process of learning a stress estimation model that estimates the relationship between the observed feature and the stress value, at least for each class divided by the classification, based on the observed feature and the correct stress value.

[0011] One aspect of the program is Subject to stress estimation Based on attribute information representing the attributes of the person to be estimated and a classification model, calculating a classification score representing a degree of certainty that the observed feature of the estimation subject belongs to each of a plurality of classes; acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; a computer to execute a process of calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; 、 the classification model is a model that classifies the observed feature values ​​for learning so that an index representing a correlation between the observed feature values ​​and a correct stress value corresponding to the observed feature values ​​becomes higher than that before classification. It is a program. [Effects of the Invention]

[0012] Stress estimation can be performed with stable estimation accuracy, or a stress estimation model for performing such stress estimation can be trained. [Brief explanation of the drawings]

[0013] [Figure 1] 1 shows a schematic configuration of a stress estimation system according to a first embodiment. [Figure 2] 1 shows an example of a hardware configuration of a stress estimation device common to each embodiment. [Figure 3] 3 is an example of a functional block in a learning phase of the information processing device according to the first embodiment. [Figure 4] FIG. 2 is a functional block diagram relating to a classification label generation unit and a classification model learning unit. [Figure 5] 1A and 1B are diagrams showing an outline of the first and second steps of the classification label generation process; [Figure 6] An example of applying the first and second steps to the refined classes is shown below. [Figure 7] 1 is an example of a functional block of a feature quantity selection unit. [Figure 8] 1 shows a histogram of correlations for a certain type of observed feature. [Figure 9] 10 is an example of a flowchart illustrating a procedure of a learning process executed by the information processing device in a learning phase in the first embodiment. [Figure 10] 3 is an example of a functional block in an estimation phase of the information processing device according to the first embodiment. [Figure 11] 10 is an example of a flowchart illustrating a procedure of a stress estimation process executed by the information processing device in the estimation phase in the first embodiment. [Figure 12] 10 shows a schematic configuration of a stress estimation system according to a second embodiment. [Figure 13] FIG. 10 is a block diagram of a learning device according to a third embodiment. [Figure 14] 11 is an example of a flowchart executed by the learning device in the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of a learning device, a stress estimation device, a learning method, a stress estimation method, and a storage medium will be described with reference to the drawings.

[0015] First Embodiment (1) System Configuration FIG. 1 shows a schematic configuration of a stress estimation system 100 according to a first embodiment. The stress estimation system 100 learns a model for estimating a person's stress (also referred to as a "stress estimation model") and performs stress estimation based on the learned stress estimation model. Hereinafter, a person who is the subject of stress estimation will be referred to as an "estimation subject," and a person who is measured in generating training data (learning samples) necessary for learning the stress estimation model will also be referred to as a "sample subject." When there is no particular distinction between an estimation subject and a sample subject, they will also be simply referred to as a "subject." Note that an "estimation subject" may be an athlete or employee whose stress state is managed by an organization, or an individual user.

[0016] The stress estimation system 100 mainly includes an information processing device 1, an input device 2, a display device 3, a storage device 4, and a sensor 5.

[0017] The information processing device 1 communicates data with the input device 2, the display device 3, and the sensor 5 via a communication network or by direct wireless or wired communication. Based on the input signal "S1" supplied from the input device 2 and the sensor signal "S3" supplied from the sensor 5, the information processing device 1 collects information necessary for learning a stress estimation model or for estimating the stress of the estimation subject using the stress estimation model, and stores the collected information in the storage device 4. The information processing device 1 also generates a display signal "S2" based on the estimation result of the stress state of the estimation subject (specifically, a stress value indicating the degree of stress), and supplies the generated display signal S2 to the display device 3. Note that in this embodiment, the stress estimated by the information processing device 1 is assumed to be chronic stress, which is stress from a long-term (chronic) perspective spanning several days to weeks or months.

[0018] The input device 2 is an interface that accepts user input (manual input) of information about each estimated subject. The user who inputs information using the input device 2 may be the estimated subject himself / herself, or a person who manages or supervises the estimated subject's activities. The input device 2 may be, for example, various user input interfaces such as a touch panel, buttons, a keyboard, a mouse, or a voice input device. The input device 2 supplies an input signal S1 generated based on the user's input to the information processing device 1. The display device 3 displays predetermined information based on a display signal S2 supplied from the information processing device 1. The display device 3 is, for example, a display or a projector.

[0019] The sensor 5 measures the biosignals and the like of the estimated subject and supplies the measured biosignals and the like to the information processing device 1 as a sensor signal S3. In this case, the sensor signal S3 may be any biosignal (including vital information) of the estimated subject, such as the heart rate, brain waves, sweat rate, hormone secretion level, cerebral blood flow, blood pressure, body temperature, electromyography, respiratory rate, pulse wave, acceleration, etc. The sensor 5 may also be a device that analyzes blood collected from the estimated subject and outputs a sensor signal S3 indicating the analysis results. The sensor 5 may also be a sensor provided in a wearable device worn by the estimated subject, a camera that captures images of the estimated subject, a microphone that generates an audio signal of the estimated subject's speech, or a sensor provided in a device such as a personal computer or smartphone operated by the estimated subject. For example, the wearable device described above may include a GNSS (global navigation satellite system) receiver, an acceleration sensor, or any other sensor that detects biosignals, and outputs the output signals of these sensors as the sensor signal S3. The sensor 5 may also supply information corresponding to the amount of operation of the personal computer, smartphone, or the like to the information processing device 1 as the sensor signal S3. The sensor 5 may also output a sensor signal S3 representing biometric data (including sleeping time) from the subject while the subject is sleeping. The sensor signal S3 is used to generate features (also referred to as "observed features") representing observed characteristics of the observed subject.

[0020] The storage device 4 is a memory that stores various information necessary for estimating a stress state, etc. The storage device 4 may be an external storage device such as a hard disk connected to or built into the information processing device 1, or may be a storage medium such as a flash memory. The storage device 4 may also be a server device that performs data communication with the information processing device 1. The storage device 4 may also be composed of multiple devices.

[0021] The storage device 4 functionally includes an attribute information storage unit 40, an observation data storage unit 41, a training data storage unit 42, an estimation model information storage unit 43, and a classification model information storage unit 44.

[0022] The attribute information storage unit 40 stores attribute information related to the attributes of the subject. Here, "attributes" refers to, for example, the subject's personality, stress tolerance, gender, occupation, age, cognitive tendency, or a combination thereof. The attribute information may be generated by the information processing device 1 and stored in the storage device 4, or may be generated in advance by a device other than the information processing device 1 and stored in the storage device 4. The attribute information may include information generated based on the subject's responses to a questionnaire. For example, the Big 5 personality test is a questionnaire used to measure the subject's personality. The attribute information is stored in the attribute information storage unit 40 in association with the subject's identification information.

[0023] The observation data storage unit 41 stores observation data generated based on the sensor signal S3 acquired by the information processing device 1 from the sensor 5, etc. In this embodiment, the observation data is information in which, for example, observation features, date and time information on when the observation was made, activity information indicating the subject's activity state at the time the observation was made (for example, physical exercise intensity, mental activity intensity such as mental workload, state such as sitting / walking / running, state such as awake / sleeping), and identification information of the subject are associated with each other. For ease of explanation, it is assumed that the observation data storage unit 41 stores the observation data of the estimated subject, and that the training data storage unit 42 stores the observation data of the sample subject.

[0024] The observed feature is an arbitrary index value representing the characteristics of data observed from the subject, or a vector (feature vector) having the index value as an element. The observed feature may be a feature based on a biological feature such as sweating, acceleration, skin temperature, or pulse wave, or a feature based on a behavioral feature related to the subject's behavior, such as the amount of operation of a device. The observed feature may also be a time-series feature (time-series data) representing the state of the subject at predetermined time intervals during the period in which the subject is observed. Here, the process of converting the sensor signal S3 into the observed feature may be performed by the information processing device 1 or by a device other than the information processing device 1. In this case, the observed feature may be generated from the sensor signal S3 based on any method for calculating the feature from a biological signal or any other arbitrary feature calculation method. The activity information is generated by the information processing device 1 or another device based on, for example, position information, acceleration, etc. included in the sensor signal S3.

[0025] The training data storage unit 42 stores training data used for training the stress estimation model. The training data is data generated for multiple sample subjects and includes multiple pairs of observed data of the sample subjects and correct stress values ​​(stress data) based on the sample subjects' responses to a questionnaire, etc. For example, the correct stress values ​​are PSS (Perceived Stress Scale) values. The PSS values ​​are calculated from the responses to a PSS questionnaire, which can measure dynamic stress that changes over time. Furthermore, as will be described later, during the learning stage, the information processing device 1 generates classification labels indicating the classes of observed features for each sample subject based on the attributes of the sample subjects, and stores the generated classification labels in the training data storage unit 42.

[0026] The estimation model information storage unit 43 stores parameters of the stress estimation model learned by the information processing device 1 (in other words, parameters necessary to configure the stress estimation model). The stress estimation model is a model that estimates the relationship between a subject's observed features and the subject's stress value. The stress estimation model is trained so as to output an estimated stress value of the subject when a combination of a specific subject's observed features (feature vector) is input. Here, the stress estimation model may be any machine learning model (including a statistical model), such as a neural network or a support vector machine.

[0027] As will be described later, the stress estimation model is trained for each class using training data divided into classes that are classified so as to increase the correlation between the observed data and the correct stress value. In this case, each stress estimation model may have an architecture appropriate for that class. Hereinafter, a "class" refers to a classification (group) to which the trained stress estimation model is uniquely associated. The number of classes corresponds to the number of trained stress estimation models. The estimation model information storage unit 43 stores information on parameters required to configure these stress estimation models. For example, if the stress estimation model is based on a neural network such as a convolutional neural network, the estimation model information storage unit 43 stores information on various parameters such as the layer structure, the neuron structure of each layer, the number and filter size of filters in each layer, and the weight of each element of each filter.

[0028] The classification model information storage unit 44 stores parameters of the classification model trained by the information processing device 1 (in other words, parameters necessary to configure the classification model). Here, the classification model is a model that estimates the relationship between the attributes of a subject and the class to which the observed features of the subject belong. The classification model is trained to output a score (also referred to as a "classification score") that represents the degree of certainty that the subject will be classified into each candidate class when attribute information of the subject is input. The architecture of the learning model for training such a classification model is, for example, a model based on a neural network such as a convolutional neural network. Note that the higher the degree of certainty for a certain class, the higher the classification score for that class. The classification model is trained based on attribute information of a sample subject and a classification label that represents the class of the observed features of the sample subject.

[0029] The configuration of the stress estimation system 100 shown in FIG. 1 is an example, and various modifications may be made to the configuration. For example, the input device 2 and the display device 3 may be configured as an integrated device. In this case, the input device 2 and the display device 3 may be configured as a tablet terminal that is integrated with or separate from the information processing device 1. In this case, the information processing device 1, the input device 2, the display device 3, and the sensor 5 (and may also include the storage device 4) may be configured as a single smartphone or wearable terminal used by the subject. The information processing device 1 may also be configured as a plurality of devices. In this case, the plurality of devices that make up the information processing device 1 exchange information required to execute pre-assigned processing between these plurality of devices. In this case, the information processing device 1 functions as an information processing system.

[0030] (2) Hardware configuration of information processing device 2 shows the hardware configuration of the information processing device 1. The information processing device 1 includes, as hardware, a processor 11, a memory 12, and an interface 13. The processor 11, the memory 12, and the interface 13 are connected via a data bus 90.

[0031] The processor 11 functions as a controller (arithmetic unit) that performs overall control of the information processing device 1 by executing programs stored in the memory 12. The processor 11 is, for example, a processor such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a TPU (Tensor Processing Unit). The processor 11 may be composed of multiple processors. The processor 11 is an example of a computer.

[0032] The memory 12 is configured by various types of volatile and non-volatile memories such as RAM (Random Access Memory), ROM (Read Only Memory), and flash memory. The memory 12 also stores programs for executing processes performed by the information processing device 1. Note that part of the information stored in the memory 12 may be stored in one or more external storage devices capable of communicating with the information processing device 1, or may be stored in a storage medium that is detachable from the information processing device 1.

[0033] The interface 13 is an interface for electrically connecting the information processing device 1 to other devices. These interfaces may be wireless interfaces such as network adapters for wirelessly transmitting and receiving data to and from other devices, or may be hardware interfaces for connecting to other devices via cables or the like.

[0034] The hardware configuration of the information processing device 1 is not limited to the configuration shown in Fig. 2. For example, the information processing device 1 may include at least one of an input device 2 or a display device 3. Furthermore, the information processing device 1 may be connected to or have a built-in sound output device such as a speaker.

[0035] (3) Learning Phase Next, we will explain the processing in the learning phase executed by the information processing device 1. In summary, the information processing device 1 classifies the observed features so that there is a high correlation between the observed features and the correct stress value, and learns a stress estimation model for each classified class. In this way, the information processing device 1 learns a stress estimation model specialized for each class with a bias in stress tendency, and acquires a stress estimation model that can perform high-accuracy stress estimation for unknown data that has not been used in learning.

[0036] (3-1) Functional Blocks FIG. 3 shows an example of functional blocks in the learning phase of the information processing device 1. In the learning phase, the processor 11 of the information processing device 1 functionally includes a first classification unit 14, "N" (N is an integer of 2 or more) second classification units 15 (151 to 15N), "M" (M is an integer of 2 or more) feature quantity selection units 16 (1611 to 16NM), and N estimation model learning units 17 (171 to 17N). The training data storage unit 42 functionally includes an observation data storage unit 421, a classification label storage unit 422, and a stress data storage unit 423. The estimation model information storage unit 43 functionally includes a first estimation model information storage unit 431 to an N-th estimation model information storage unit 43N, each storing parameters of the N stress estimation models to be learned. Note that in FIG. 3, blocks that exchange data are connected by solid lines, but the combination of blocks that exchange data is not limited to that shown. The same applies to other functional block diagrams described later.

[0037] The first classification unit 14 performs processing related to first classification, which classifies (clusters) the observed features used for learning into N classes so as to increase the correlation between the observed features and the correct stress value. Functionally, the first classification unit 14 has a classification label generation unit 141, a classification model learning unit 142, and a classification unit 143.

[0038] The classification label generation unit 141 references the attribute information storage unit 40, the observation data storage unit 421, and the stress data storage unit 423, and generates classification labels to be associated with each sample subject. As will be described later, in this embodiment, the classification label generation unit 141 adaptively determines "N", which corresponds to the number of classes (i.e., the number of stress estimation models) used in the classification labels, in a classification label generation process that will be described later. The classification label generation unit 141 stores the generated classification labels in the classification label storage unit 422. Details of the processing of the classification label generation unit 141 will be described later.

[0039] The classification model learning unit 142 references the attribute information storage unit 40 and the classification label storage unit 422 to learn the classification model. In this case, the classification model learning unit 142 sequentially extracts pairs of attribute information and classification labels corresponding to each sample subject and updates the parameters of the classification model. In this case, the classification model learning unit 142 determines the parameters of the classification model so that the error (loss) between the classification result output by the classification model when attribute information is input and the correct class indicated by the classification label is minimized. The attribute information input to the classification model is, for example, index values ​​indicating personality, gender, occupation, race, age, height, weight, muscle mass, lifestyle habits, and exercise habits, or combinations (vector values) of these index values. Furthermore, the algorithm for determining the parameters so as to minimize the loss may be any learning algorithm used in machine learning, such as gradient descent or backpropagation. The classification model learning unit 142 then stores the learned parameters of the classification model in the classification model information storage unit 44.

[0040] Classification unit 143 extracts observation features for learning from observation data storage unit 421, and classifies (clusters) the extracted observation features into N classes according to the classification labels stored in classification label storage unit 422. In this way, classification unit 143 forms a set of N observation features that are biased in stress, biological features, etc. Then, classification unit 143 supplies the observation features for each class to second classification units 151 to 15N corresponding to the relevant class.

[0041] Note that instead of classifying the training observation features into N classes according to the classification labels stored in the classification label storage unit 422, the classification unit 143 may classify the training observation features into N classes based on the classification results of the classification model trained by the classification model training unit 142. In this case, the classification unit 143 extracts attribute information associated with the sample subject from the attribute information storage unit 40, and classifies the training observation features of the sample subject into the class that has the highest classification score output when the extracted attribute information is input to the classification model. As a result, the first classification using the classification model is performed in the training phase as well, as in the estimation phase described below, and therefore it is expected that the estimation accuracy in the estimation phase will improve.

[0042] The second classification unit 15 (151 to 15N) performs second classification to classify the set of observation features for each class supplied from the first classification unit 14 into M subclasses based on the observed target of the observation features or the subject's activity state at the time of observation. In this way, the second classification unit 15 further classifies the observation features that should be treated differently in stress estimation. Then, each of the second classification units 151 to 15N supplies the set of observation features divided into M subclasses based on the second classification to the feature selection unit 16 (1611 to 16NM).

[0043] Here, the "observation target" refers to the observation target of the raw data used when the observation feature is calculated, and includes various biological features such as sweating, acceleration, skin temperature, and pulse wave. Therefore, in the case of observation features based on biological features, "classification based on the observation target" refers to classification such that the observation feature related to sweating, the observation feature related to acceleration, the observation feature related to skin temperature, and the observation feature related to pulse wave are each classified into different subclasses. Furthermore, "classification based on activity state" refers to classification into subclasses according to the level of exercise intensity (e.g., stationary state, walking state, running state) of the subject at the time of observation. Information indicating the observation target and activity state corresponding to each observation feature is stored, for example, in association with the observation feature in the observation data storage unit 421.

[0044] The feature selection unit 16 (1611 to 16NM) selects, from a set of N×M observation features classified based on the first and second classifications, observation features (also referred to as "stress estimation features") to be input to the stress estimation model based on correlation with the correct stress value. Here, the feature selection unit 16 selects "R" (R is an integer equal to or greater than 0) types of observation features as stress estimation features. Details of the processing by the feature selection unit 16 will be described later. Note that instead of uniformly providing M feature selection units 16 for each class, an appropriate number of feature selection units 16 may be provided for each class. Similarly, the value of R may also differ for each feature selection unit 16.

[0045] The estimation model learning units 17 (171 to 17N) learn a stress estimation model prepared for each class based on the first classification, based on the stress estimation feature quantities selected by the feature quantity selection units 16 and the correct stress values ​​referenced from the stress data storage unit 423. In this case, each estimation model learning unit 17 uses the M×R stress estimation feature quantities supplied from the M feature quantity selection units 16 as input data for the stress estimation model, and acquires multiple pairs of the corresponding stress values ​​referenced from the stress data storage unit 423 as correct data. Then, each estimation model learning unit 17 learns the corresponding stress estimation model based on the multiple pairs of the input data and correct data.

[0046] In training the stress estimation model, the estimation model training unit 17, for example, sequentially extracts pairs of the input data and correct answer data described above and updates the parameters of the stress estimation model. In this case, the parameters of the stress estimation model are determined so as to minimize the error (loss) between the estimation result output by the stress estimation model when input data is input and the stress value (here, the PSS value) that is the correct answer data. The algorithm for determining the parameters so as to minimize the loss may be any learning algorithm used in machine learning, such as gradient descent or backpropagation. Then, each estimation model training unit 17 stores the trained parameters of each stress estimation model in the first estimation model information storage unit 431 to the Nth estimation model information storage unit 43N, respectively.

[0047] The components of the first classification unit 14, the second classification unit 15, the feature selection unit 16, and the estimation model learning unit 17 described in FIG. 3 can be realized, for example, by the processor 11 executing a program. Alternatively, the necessary programs may be recorded in any non-volatile storage medium and installed as needed to realize the components. At least some of the components may not necessarily be realized by software programs, but may be realized by any combination of hardware, firmware, and software. At least some of the components may be realized using a user-programmable integrated circuit, such as an FPGA (Field-Programmable Gate Array) or a microcontroller. In this case, the integrated circuit may be used to realize a program consisting of the components. At least some of the components may be configured by an ASSP (Application Specific Standard Produce), an ASIC (Application Specific Integrated Circuit), or a quantum processor (quantum computer control chip). In this way, the components may be realized by various hardware. The same applies to other embodiments described below. Furthermore, each of these components may be realized by the cooperation of multiple computers, for example, using cloud computing technology.

[0048] (3-2) Details of classification label generation section 4 is a detailed functional block diagram of the classification label generation unit 141 and the classification model learning unit 142 included in the first classification unit 14. The classification label generation unit 141 generates classification labels based on the attribute information stored in the attribute information storage unit 40, the observation feature amounts stored in the observation data storage unit 421, and the correct stress values ​​stored in the stress data storage unit 423, and stores the generated classification labels in the classification label storage unit 422. Furthermore, the classification model learning unit 142 learns a classification model based on the classification labels stored in the classification label storage unit 422 and the attribute information stored in the attribute information storage unit 40, and stores parameters of the classification model obtained by learning in the classification model information storage unit 44.

[0049] The classification label generation unit 141 classifies the observed features for each sample subject into multiple classes based on attribute information, then randomly shuffles (moves) the observed features between the classes, and adopts the shuffle if the correlation between the observed features for each class and the correct stress value becomes higher than before the shuffle.The classification label generation unit 141 then repeats the shuffling to classify the observed features so that the correlation between the observed features for each class and the correct stress value becomes higher, and generates classification labels that represent the classification results.

[0050] Here, a specific example of a classification label generation method executed by classification label generation unit 141 will be described. In this specific example, in the first step, classes are provisionally subdivided based on attribute information, and in the second step, the above-mentioned shuffling is performed between the subdivided classes. Then, classification label generation unit 141 repeatedly executes the first and second steps until it determines that class subdivision is no longer necessary. Note that, as an example, here, classification label generation unit 141 subdivides classes hierarchically by repeating two-part division. Note that when classes are subdivided hierarchically, the number of divisions may be three or more.

[0051] 5(A) is a diagram showing an outline of the first step of the classification label generation process, in which a set of observed features for each sample subject is indicated by a circle.

[0052] First, the classification label generation unit 141 classifies all observed features for learning into two classes (class A and class B) based on the attribute type "X." Here, the classification label generation unit 141 provisionally classifies the observed features of a sample subject whose attribute type X is attribute Xa into class A, and provisionally classifies the observed features of a sample subject whose attribute type X is attribute Xb into class B. The attribute type X is, for example, personality, sex, occupation, race, age, height, weight, muscle mass, lifestyle, and exercise habit, and the attributes Xa and Xb are categories or ranges of index values ​​in the attribute type X. For example, if the attribute type X is sex, the attribute Xa is male and the attribute Xb is female.

[0053] Then, if the correlation between the observation feature and the correct stress value (also referred to as the "observation-stress correlation") increases due to the provisional classification based on attribute type X in the first step, the classification label generation unit 141 determines that subdivision of classes is necessary and proceeds to the second step. Specifically, first, the classification label generation unit 141 calculates the observation-stress correlation based on the overall observation feature before classification into class A and class B, the observation-stress correlation based on the observation feature classified into class A, and the observation-stress correlation based on the observation feature classified into class B. Then, if the observation-stress correlation of class A and the observation-stress correlation of class B are higher than the observation-stress correlation before classification, the classification label generation unit 141 determines that subdivision into class A and class B is necessary, and executes the second step targeting class A and class B. Note that the classification label generation unit 141 may also determine that subdivision is necessary if either the observation-stress correlation of class A or the observation-stress correlation of class B is higher than the overall observation-stress correlation before classification. In another example, the classification label generation unit 141 may determine that subdivision is necessary when the average of the observation-stress correlation of class A and the observation-stress correlation of class B is higher than the overall observation-stress correlation before classification.

[0054] Here, we will provide additional explanation about the "correlation" calculated in the first step. The classification label generation unit 141 may calculate a correlation coefficient as the correlation, or may calculate any other index value representing a correlation, such as mutual information. Furthermore, the classification label generation unit 141 may perform any normalization process to eliminate the influence of differences in the number of samples when calculating the above-mentioned index value.

[0055] FIG. 5B shows an overview of the second step for class A and class B. In the first step, the classification label generation unit 141 provisionally classifies the observation features into class A and class B based on attribute information. In the second step, the classification label generation unit 141 shuffles the observation features for each sample subject so as to improve the observation-stress correlation. In FIG. 5B, the classification label generation unit 141 calculates the observation-stress correlation of class A and the observation-stress correlation of class B when the observation features of sample subject s1, provisionally classified into class A based on attribute information, are provisionally moved to class B. If the movement increases the observation-stress correlation of class A and the observation-stress correlation of class B, the classification label generation unit 141 adopts the movement of the observation features of sample subject s1 from class A to class B. The classification label generation unit 141 then performs this movement and decision on the necessity of the movement for all sample subjects in class A and class B. This allows the classification label generating unit 141 to classify the observation feature quantities into class A and class B so that the observation-stress correlation is high.

[0056] Note that, if the movement causes the observation-stress correlation of the observation feature of the sample subject s1 in the class from which the sample subject s1 is moved to decrease and the observation-stress correlation in the class to which the sample subject s1 is moved to increase, the classification label generation unit 141 may cause the observation feature of the sample subject s1 to exist in both class A and class B. In this case, the classification label generation unit 141 classifies the observation feature of the sample subject s1 into both class A and class B. This allows the classification label generation unit 141 to improve the observation-stress correlation of both class A and class B.

[0057] Furthermore, if the observation-stress correlation of the observation feature of sample subject s1 in the class from which the sample subject s1 is moved increases and the observation-stress correlation in the class to which the sample subject s1 is moved decreases due to the movement, the classification label generation unit 141 does not have to classify the observation feature of sample subject s1 into either class A or class B. In other words, in this case, the observation feature of sample subject s1 is not used for learning. This also allows the classification label generation unit 141 to improve the observation-stress correlation of both class A and class B.

[0058] After executing the second step, the classification label generation unit 141 performs the first step on each of class A and class B, and then performs the second step on the classes subdivided in the first step. Fig. 6 shows an example in which the first step and the second step are applied to class A and class B, respectively.

[0059] Here, as an example, in the first step, the classification label generation unit 141 subdivides each of class A and class B into two classes based on the attribute type "Y." Specifically, the classification label generation unit 141 classifies the observed feature of class A corresponding to attribute "Ya" into class Aa, and the observed feature of class A corresponding to attribute "Yb" into class Ab. Furthermore, the classification label generation unit 141 classifies the observed feature of class B corresponding to attribute Ya into class Ba, and the observed feature of class B corresponding to attribute Yb into class Bb.

[0060] Note that instead of classifying based on an attribute type Y different from that in the first step of the first round, the classification label generation unit 141 may perform classification based on the same attribute type X as in the first step of the first round. In this case, for example, the classification label generation unit 141 classifies the observed features of class A into classes Aa and Ab based on attributes "Xaa" and "Xab" that are obtained by classifying (categorizing) attribute Xa in more detail, and classifies the observed features of class B into classes Ba and Bb based on attributes "Xba" and "Xbb" that are obtained by subdividing attribute Xb.

[0061] Then, when it is determined that the observation-stress correlation has increased as a result of the provisional classification based on the attribute information in the first step, the classification label generation unit 141 shuffles the observation feature quantities for each sample subject between the subdivided classes based on the second step. In the example shown in Fig. 6, since the observation-stress correlation increased when class A was subdivided into class Aa and class Ab, the classification label generation unit 141 shuffles the observation feature quantities belonging to class Aa and class Ab so as to improve the observation-stress correlation. Similarly, since the correlation increased when class B was subdivided into class Ba and class Bb, the classification label generation unit 141 shuffles the observation feature quantities belonging to class Ba and class Bb so as to improve the observation-stress correlation.

[0062] In this way, the classification label generation unit 141 increases the number of classes hierarchically by performing the first step and the second step on each generated class. Then, the classification label generation unit 141 ends the process when the observation-stress correlation does not increase due to the subdivision of the class in any of the classes. Then, the classification label generation unit 141 generates classification labels indicating the classes to which the observation features of each sample subject belong at the time of the end of the process, and stores the generated classification labels in the classification label storage unit 422.

[0063] As described above, classification label generation unit 141 hierarchically subdivides classes and determines the observation features that belong to each class based on changes in the observation-stress correlation caused by movement of observation features between classes. Classification label generation unit 141 then performs subdivision of existing classes into those for which the observation-stress correlation will increase as a result of the subdivision, until no classes exist for which the observation-stress correlation will increase as a result of the subdivision. This allows classification label generation unit 141 to adaptively determine the number of classes N and generate classification labels that will result in a high observation-stress correlation.

[0064] Note that instead of determining whether or not class subdivision is necessary based on the provisional classification results based on attribute information, the classification label generation unit 141 may determine whether or not class subdivision is ultimately necessary after shuffling in the second step is completed. For example, in the example of FIG. 6 , the classification label generation unit 141 determines to establish classes Aa and Ab when the observation-stress correlation of class Aa and the observation-stress correlation of class Ab after executing the second step are higher than the observation-stress correlation of class A as a whole. On the other hand, the classification label generation unit 141 determines that subdivision of class A into classes Aa and Ab is inappropriate when the observation-stress correlation of class Aa and the observation-stress correlation of class Ab after executing the second step are not higher than the observation-stress correlation of class A as a whole. Therefore, in this case, the classification label generation unit 141 generates classification labels that define both classes of the observation features classified into classes Aa and Ab as class A. This example makes it possible to more accurately determine whether or not class subdivision is necessary.

[0065] Furthermore, instead of adaptively determining the number of classes N, the classification label generation unit 141 may set the number of classes N to a fixed value. In this case, the classification label generation unit 141 classifies the observation feature for each sample subject into N classes based on the attribute information, and then performs processing corresponding to the second step to determine the observation feature that belongs to each of the N classes. In this case, in the second step, the classification label generation unit 141 may move the observation feature to the class that will result in the greatest increase in the observation-stress correlation of the destination class. Note that if moving to any class does not result in an increase in the observation-stress correlation of the destination class, the classification label generation unit 141 may not change the class of the target observation feature or may not classify the observation feature into any class (do not use it for learning).

[0066] In yet another example, classification label generation unit 141 may set multiple candidates for number of classes N (number of candidate classes), and determine the number of candidate classes with the highest observation-stress correlation as number of classes N. In this case, classification label generation unit 141 sets number of classes N as each number of candidate classes, performs classification based on the first step and the second step, and sets number of classes N as the number of candidate classes with the highest observation-stress correlation for each class after classification. For example, if the number of candidate classes is "2," "3," and "4," classification label generation unit 141 compares the average observation-stress correlation of each class after classification when the number of classes is fixed to two, the average observation-stress correlation of each class after classification when the number of classes is fixed to three, and the average observation-stress correlation of each class after classification when the number of classes is fixed to four. Then, classification label generation unit 141 determines the number of candidate classes with the highest average observation-stress correlation of each class as number of classes N, and generates classification labels based on the classification results when this number of candidate classes is used. Even in this example, the classification label generating unit 141 can determine the number of classes N and classification labels that maximize the observation-stress correlation.

[0067] (3-3) Details of the feature selection section Next, the details of the processing executed by the feature quantity selection unit 16 (1611 to 16NM) will be described. FIG. 7 shows an example of the functional blocks of a feature quantity selection unit 16nm (where "n" and "m" are integers that satisfy 1≦n≦N and 1≦m≦M). Functionally, the feature quantity selection unit 16nm has a group generation unit 50, a correlation calculation unit 51, a ranking unit 52, and a selection unit 53.

[0068] The feature selection unit 16nm selects the observed feature "F p,q ” is acquired from the stress data storage unit 423, and the observed feature F p,q The correct stress value (PSS value) corresponding to "S pHere, "p" indicates the index of the sample subject (1≦p≦P, where P is an integer equal to or greater than 2), and "q" indicates the index of the type of observed feature (1≦q≦Q, where Q is an integer satisfying "Q≧R"). Note that there are generally many types of observed feature (e.g., tens of thousands), and in the case of a feature related to sweating, for example, various indicators related to sweating such as the maximum, minimum, median, mean, and other arbitrary statistics are applicable.

[0069] The group generation unit 50 randomly selects a predetermined number of observed features F p,q is extracted L times (L is an integer equal to or greater than 1), and the extracted observation feature F p,q In this case, for example, the group generation unit 50 generates L groups by dividing the observed feature quantity F p,q If there are 100 sample subjects (i.e., P=100) corresponding to p,q Then, the group generation unit 50 forms the observed feature F p,q The groups are supplied to the correlation calculation units 511 to 51L, respectively.

[0070] The correlation calculation unit 51 (511 to 51L) calculates the observed feature F p,q Based on the group, the observed feature F p,q and stress value S p The correlation (correlation coefficient) with the observed feature F p,q The correlation coefficient may be any one of Pearson's product-moment correlation coefficient, Spearman's rank correlation coefficient, and Kendall's rank correlation coefficient, or a combination of multiple correlation coefficients, such as an average. In other words, the correlation calculation unit 51 calculates the correlation coefficient for each group generated by the group generation unit 50 and for each observed feature F p,q For each type q of p,q and stress value S p Calculate the correlation with

[0071] The ranking unit 52 ranks the observed feature F based on the calculation results of the L correlation calculation units 511 to 51L. p,q In this case, the ranking unit 52 ranks the types q of the observed feature F p,q For each type q, a score (also referred to as a "correlation score") is calculated based on the calculation results of the L correlation calculation units 511 to 51L, and the higher the correlation score, the higher the ranking is considered to be. In this case, the ranking unit 52 calculates the correlation score based on statistical values ​​such as the average of correlation between groups and the degree of sign reversal, which will be described later. The method of calculating the correlation score will be described later.

[0072] The selection unit 53 selects the observed feature values ​​F corresponding to the top R types in the ranking formed by the ranking unit 52. p,q as the stress estimation feature. In this case, the selection unit 53 stores information indicating the type of the observation feature selected as the stress estimation feature (also referred to as "feature selection information Ifs") in the estimation model information storage unit 43. As will be described later, the feature selection information Ifs is used in the process of selecting, from the observation feature, a stress estimation feature to be input to the stress estimation model in the estimation phase.

[0073] Here, a specific example of a method for calculating correlation scores by the ranking unit 52 will be described. Fig. 8 shows a histogram that aggregates correlations for type q, which is the target for calculating correlation scores, based on the calculation results of the correlation calculation units 511 to 51L. Note that, although a histogram is shown here for the sake of convenience, generation of a histogram is not an essential process for calculating correlation scores.

[0074] In this case, first, the correlation calculation unit 51 calculates the average (here, 0.15) of the correlations (correlation coefficients) calculated by the correlation calculation units 511 to 51L for the target type q based on the calculation results of the correlation calculation units 511 to 51L. Furthermore, the correlation calculation unit 51 calculates, as the sign reversal degree, the proportion of minority signs when the positive and negative signs of the calculated correlations are tallied. In the example of FIG. 8, since positive signs are the majority, the correlation calculation unit 51 recognizes the proportion of negative signs (0.3) as the sign reversal degree. The sign reversal degree has a value range from 0 to 0.5. Then, for example, the correlation calculation unit 51 determines the correlation score for the target type q to be a value obtained by multiplying the absolute value of the average of the correlations by a value obtained by subtracting the sign reversal degree from 1 (i.e., a value range from 0.5 to 1) as a weight, as follows: Correlation score = |average correlation| × (1 - degree of sign reversal) In the example of FIG. 8, the correlation score of type q is 0.105 (=|0.15|×0.7).

[0075] The method for calculating the correlation score is not limited to the above formula, and any formula or lookup table may be used that defines the correlation score so that it has a positive correlation with the average of the correlation and a negative correlation with the degree of sign reversal.

[0076] The feature selection unit 16nm has a functional configuration as shown in FIG. 7, and can therefore suitably select, as a stress estimation feature, an observed feature that has a stable correlation with the stress value regardless of individual differences.

[0077] (3-4) Processing flow FIG. 9 is an example of a flowchart showing the procedure of the learning process executed by the information processing device 1 in the learning phase in the first embodiment.

[0078] First, the first classification unit 14 of the information processing device 1 performs a process of generating classification labels (step S11). In this case, the information processing device 1 determines the number of classes N and generates classification labels based on the process described in the section "(3-2) Details of the classification label generation unit."

[0079] Next, the first classification unit 14 performs training of a classification model based on the classification labels, and performs first classification of the observation features for training stored in the observation data storage unit 421 (step S12). In this case, the information processing device 1 may perform the first classification based on the classification labels, or may perform the first classification based on the trained classification model and attribute information stored in the attribute information storage unit 40.

[0080] Next, the second classification unit 15 of the information processing device 1 performs a second classification of the observed features based on the observed target of the observed features and the activity state of the corresponding sample subject at the time of observation (step S13). In this case, for example, the second classification unit 15 further classifies the observed features into M subclasses for each of the sets of observed features divided into N classes based on the second classification according to the type of observed biometric feature, the exercise intensity of the sample subject, etc.

[0081] Next, the feature selection unit 16 of the information processing device 1 randomly generates groups for each set of observed features divided into N × M subclasses, and calculates the correlation between each type of observed feature and the correct stress value included in the training data in the generated groups (step S14). Furthermore, the feature selection unit 16 ranks the types of observed features according to the correlation and sign reversal degree for each set of observed features divided into subclasses, and selects the top R observed features corresponding to the types of observed features as stress estimation features (step S15).

[0082] Then, the estimation model learning unit 17 of the information processing device 1 learns a stress estimation model for each class divided by the first classification, based on the stress estimation feature and the corresponding ground truth stress value included in the training data (step S16). The information processing device 1 then outputs, as learning results, feature selection information Ifs related to the stress estimation feature selected in step S15 and parameters of the stress estimation model learned in step S16. Specifically, the information processing device 1 stores the feature selection information Ifs and the parameters of the stress estimation model in the storage device 4. This allows the information processing device 1 to store information required in the estimation phase in the storage device 4.

[0083] (4) Estimation Phase Next, a description will be given of the processing in the estimation phase executed by the information processing device 1. The information processing device 1 estimates the stress value of the estimation subject based on the classification model and stress estimation model learned in the learning phase.

[0084] 10 shows an example of functional blocks in the estimation phase of the information processing device 1. In the estimation phase, the processor 11 of the information processing device 1 functionally includes a classification score calculation unit 34, N feature quantity selection units 36 (361 to 36N), N stress estimation units 37 (371 to 37N), and an integration unit 38. Here, the first estimation model information storage unit 431 to the Nth estimation model information storage unit 43N included in the estimation model information storage unit 43 store parameters of N stress estimation models that have already been trained in the learning phase.

[0085] The classification score calculation unit 34 extracts attribute information of the estimation target person from the attribute information storage unit 40, and calculates a classification score for each class (i.e., N classes corresponding to the first to Nth estimation models) established in the first classification of the learning phase based on the extracted attribute information. In this case, the classification score calculation unit 34 obtains a classification score for each class output from the classification model by inputting the above-mentioned attribute information into a classification model based on the classification model information stored in the classification model information storage unit 44. Then, the classification score calculation unit 34 supplies the obtained classification score for each class to the integrating unit 38.

[0086] The feature selection units 36 (361 to 36N) select stress estimation features from the observation features of the estimation subject extracted from the observation data storage unit 41, based on the feature selection information Ifs stored in the estimation model information storage unit 43. In this case, the feature selection units 36n (n is an arbitrary integer from 1 to N) each extract, from the observation features of the estimation subject, observation features of the same type as the type of stress estimation feature indicated by the feature selection information Ifs generated by the feature selection units 16n1 to 16nM, as stress estimation features. Then, the feature selection units 36n supply the extracted stress estimation features to the corresponding stress estimation units 37n.

[0087] The stress estimation units 37 (371 to 37N) each estimate a stress value of an estimation subject based on the stress estimation feature values ​​and stress estimation models supplied from the feature value selection units 36 (361 to 36N). In this case, the stress estimation units 37n (n is an arbitrary integer from 1 to N) configure the corresponding nth estimation model by referring to the corresponding nth estimation model information storage unit 43n. The stress estimation units 37n then input the stress estimation feature values ​​supplied from the corresponding feature value selection unit 36n to the configured nth estimation model, thereby obtaining the stress value of the estimation subject output by the nth estimation model. The stress values ​​output by each stress estimation model correspond to candidate values ​​for the stress value of the estimation subject finally estimated by the integration unit 38. Each stress estimation unit 37 (371 to 37N) then supplies the stress value of the estimation subject output by the stress estimation model to the integration unit 38.

[0088] The integrating unit 38 integrates (i.e., performs weighted averaging on) the stress values ​​supplied from each stress estimation unit 37 (371 to 37N) by weighting them based on the classification score for each class supplied from the classification score calculation unit 34. The integrating unit 38 then outputs the integrated stress value as a final estimated value of stress for the estimation subject (also referred to as a "stress estimate value"). For example, the integrating unit 38 generates a display signal S2 for displaying information related to the integrated stress estimate value, and supplies the display signal S2 to the display device 3, thereby causing the display device 3 to display information related to the stress estimate value. In this case, the integrating unit 38 can integrate the stress values ​​of the estimation subject output by the stress estimation models through weighting processing based on the classification scores, thereby calculating a highly accurate stress estimate value.

[0089] Note that instead of or in addition to controlling the display of the stress estimate value itself, the integration unit 38 may control the display of information regarding the stress level determined based on a comparison between the stress estimate value and a predetermined threshold, and / or information regarding advice corresponding to the stress level. Note that in this case, the viewer of the display device 3 may be, for example, the person being estimated, or a person who manages or supervises the person being estimated. Furthermore, the integration unit 38 may output the information regarding the stress estimate value as audio using an audio output device (not shown).

[0090] 11 is an example of a flowchart showing the procedure of the stress estimation process executed in the estimation phase by the information processing device 1. The timing of the stress estimation process may be a timing requested by the user based on the input signal S1, or may be a predetermined timing.

[0091] First, the information processing device 1 acquires the observed feature of the person to be estimated and the attribute information of the person to be estimated (step S21). In this case, for example, the information processing device 1 acquires the observed feature from the observed data storage unit 41 and acquires the attribute information from the attribute information storage unit 40.

[0092] Next, the classification score calculation unit 34 of the information processing device 1 determines a classification score corresponding to each of the N stress estimation models based on the attribute information of the estimation subject and the classification model to which the parameters stored in the classification model information storage unit 44 are applied (step S22). In this case, the classification score calculation unit 34 acquires the classification score of each class that uniquely corresponds to each stress estimation model from the classification model to which the attribute information has been input.

[0093] Then, the feature selection unit 36 ​​of the information processing device 1 selects observation features to be input to the N stress estimation models provided for each class (step S23). In this case, the feature selection unit 36 ​​refers to the corresponding feature selection information Ifs from the observation features acquired in step S21 and selects stress estimation features that are observation features to be input to the stress estimation models. Note that the processing order of steps S22 and S23 is not limited to this order and may be executed simultaneously in parallel.

[0094] Then, the stress estimation unit 37 of the information processing device 1 calculates a stress value for each stress estimation model (i.e., for each class) based on the selected stress estimation feature and each stress estimation model configured based on the parameters stored in the estimation model information storage unit 43 (step S24). In this case, the stress estimation unit 37 inputs the stress estimation feature supplied from the feature selection unit 36 ​​to each stress estimation model configured with reference to the estimation model information storage unit 43, thereby calculating a stress value for each stress estimation model.

[0095] The integration unit 38 of the information processing device 1 then calculates an integrated stress estimation value by weighting the stress values ​​for each stress estimation model by the classification score for each class determined in step S22 (step S25), and outputs information related to the stress estimation values ​​(step S26).

[0096] (5) Variations Next, modifications applicable to the first embodiment will be described. (Variation 1) The stress estimation model may be provided for each subclass classified by the first and second classifications, instead of for each class classified by the first classification.

[0097] In this case, in the learning phase, the information processing device 1 provides stress estimation models corresponding to the N×M feature quantity selection units 1611-16NM, and performs learning on these stress estimation models using the stress estimation feature quantities output by the corresponding feature quantity selection units 16 as input data and the stress values ​​indicated by the corresponding stress data as correct answer data. Similarly to the feature quantity selection units 16 in the learning phase, there are N×M feature quantity selection units 36 in the estimation phase, and each of the stress estimation units 371-37N inputs the stress estimation feature quantities output by the corresponding M feature quantity selection units 36 to the corresponding M stress estimation models. The integration unit 38 then calculates a stress estimation value by performing a weighted average based on the stress values ​​output by the N×M stress estimation models and the classification scores set for each class.

[0098] In this way, even in this modified example, the information processing device 1 can accurately estimate the stress state of the person to be estimated from observed features not used in learning, based on the stress estimation model learned for each set with a bias in stress tendency.

[0099] (Variation 2) The stress estimated by the information processing device 1 is not limited to chronic stress, but may be short-term stress, which is stress that lasts for a relatively short period of time (several minutes to a day).

[0100] (Variation 3) The information processing device 1 may train the stress estimation model in the learning phase without performing at least one of the second classification by the second classification unit 15 and the feature selection by the feature selection unit 16. Even in this case, the information processing device 1 trains the stress estimation model for each class set based on the first classification so as to increase the correlation between the observed feature and the stress value, and can acquire a stress estimation model that can perform highly accurate stress estimation for unknown data that has not been used in training. Note that if feature selection is not performed by the feature selection unit 16, the information processing device 1 also does not perform feature selection by the feature selection unit 36 ​​in the estimation phase.

[0101] Second Embodiment 12 shows a schematic configuration of a stress estimation system 100A according to the second embodiment. The stress estimation system 100A according to the second embodiment includes a stress estimation device 1A that performs processing in the estimation phase of the information processing device 1 according to the first embodiment, a learning device 1B that performs processing in the learning phase of the information processing device 1 according to the first embodiment, a storage device 4, and a terminal device 8 and sensor 5 used by the person to be estimated. Hereinafter, the same components as those in the first embodiment will be appropriately designated by the same reference numerals, and their description will be omitted.

[0102] In the second embodiment, the stress estimation device 1A functions as a server, and the terminal device 8 functions as a client. The stress estimation device 1A and the terminal device 8 perform data communication via a network 7.

[0103] The learning device 1B has the same hardware configuration as the information processing device 1 shown in Fig. 2, and the processor 11 of the learning device 1B has the functional blocks shown in Fig. 3. Based on the information stored in the storage device 4, the learning device 1B performs processes such as learning a stress estimation model, learning a classification model, and generating feature selection information Ifs.

[0104] The terminal device 8 is a terminal used by a user who will be the estimation subject. The terminal device 8 has input, display, and communication functions and functions as the input device 2 and display device 3 shown in FIG. 1 . The terminal device 8 may be, for example, a personal computer, a tablet terminal such as a smartphone, or a personal digital assistant (PDA). The terminal device 8 is electrically connected to a sensor 5, such as a wearable sensor worn by the user, and transmits a biosignal of the estimation subject output by the sensor 5 (i.e., information corresponding to the sensor signal S3 in FIG. 1 ) to the stress estimation device 1A via the network 7. The terminal device 8 also accepts user inputs related to questionnaire responses and transmits information generated by the user input (information corresponding to the input signal S1 in FIG. 1 ) to the stress estimation device 1A. The sensor 5 may be built into the terminal device 8. The sensor 5 may also have the functions of the terminal device 8 and perform data communication with the stress estimation device 1A.

[0105] The stress estimation device 1A has the same hardware configuration as the information processing device 1 shown in FIG. 2, and the processor 11 of the stress estimation device 1A has the functional blocks shown in FIG. 10. The stress estimation device 1A receives information corresponding to the input signal S1 and the sensor signal S3 in FIG. 1 from the terminal device 8 via the network 7 and stores the received information in the storage device 4. The stress estimation device 1A then references the parameters of each stress estimation model learned by the learning device 1B, the parameters of the classification model, and the feature selection information Ifs, and executes stress estimation processing for the estimation subject. In addition, the stress estimation device 1A transmits an output signal for outputting the stress estimation result to the terminal device 8 via the network 7, based on a display request from the terminal device 8.

[0106] In this way, the stress estimation system 100A in the second embodiment has the learning phase and the estimation phase executed by separate devices, and can perform learning of the stress estimation model and stress estimation using the stress estimation model in the same way as in the first embodiment. Also, in the second embodiment, the stress estimation device 1A estimates the stress state of the estimation subject based on the biological signals of the estimation subject received from a terminal used by the estimation subject, and can suitably present the estimation result to the estimation subject on the terminal.

[0107] Third Embodiment 13 is a block diagram of a learning device 1X according to the third embodiment. The learning device 1X mainly includes a classification unit 14X and a learning unit 17X. Note that the learning device 1X may be configured by a plurality of devices.

[0108] The classification means 14X classifies the observed feature of the subject so that an index representing a correlation between the observed feature and a correct stress value corresponding to the observed feature becomes higher than before classification. The classification means 14X can be, for example, the first classification unit 14 in the first embodiment (including modified examples, the same applies hereinafter) or the second embodiment.

[0109] The learning means 17X learns a stress estimation model that estimates the relationship between the observed feature values ​​and the stress value for at least each class divided by the above classification, based on the observed feature values ​​and the correct stress value. In this case, there are as many stress estimation models as there are classes, and each stress estimation model is learned for each class. The learning means 17X can be, for example, the estimation model learning unit 17 in the first or second embodiment.

[0110] 14 is an example of a flowchart executed by the learning device 1X in the third embodiment. First, the classification means 14X classifies the observed features of the subject so that an index representing the correlation between the observed features and the correct stress value corresponding to the observed features is higher than before classification (step S31). Furthermore, the learning means 17X learns a stress estimation model that estimates the relationship between the observed features and the stress value, at least for each class divided by classification, based on the observed features and the correct stress value (step S32).

[0111] According to the third embodiment, the learning device 1X learns a stress estimation model for each group having a bias in stress tendency, and can learn a stress estimation model that can perform stress estimation with high accuracy.

[0112] In the above-described embodiments, the program can be stored using various types of non-transitory computer-readable media and supplied to a computer processor or the like. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic storage media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical storage media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). The program may also be supplied to a computer by various types of transitory computer-readable media. Examples of transitory computer-readable media include electrical signals, optical signals, and electromagnetic waves. The transitory computer-readable media can supply the program to a computer via a wired communication path such as an electric wire or optical fiber, or via a wireless communication path.

[0113] In addition, some or all of the above embodiments may be described as, but are not limited to, the following supplementary notes.

[0114] [Appendix 1] a classification means for classifying the observed feature of the subject so that an index representing a correlation between the observed feature and a correct stress value corresponding to the observed feature becomes higher than before classification; a learning means for learning a stress estimation model that estimates a relationship between the observed feature value and the stress value for at least each class divided by the classification, based on the observed feature value and the correct stress value; A learning device having the above configuration. [Appendix 2] The learning device according to claim 1, further comprising a classification model learning means for learning a classification model that estimates the relationship between the attributes and the classes based on attribute information representing the attributes of the subject and the results of the classification. [Appendix 3] 3. The learning device according to claim 2, wherein the classification means classifies the observed feature values ​​used for training each of the stress estimation models based on the attribute information and the classification model. [Appendix 4] The learning device according to any one of appendices 1 to 3, wherein the classification means classifies the observed features into a plurality of classes based on attribute information representing attributes of the subject, and then moves the observed features between the plurality of classes so that the index of each of the plurality of classes increases. [Appendix 5] The learning device according to any one of appendices 1 to 4, wherein the classification means classifies the observed features into a plurality of classes, and then subdivides each of the plurality of classes into classes whose indices become higher as a result of the subdivision. [Appendix 6] a second classification means for classifying the observed feature based on at least one of an observed object of the observed feature or an activity state of the subject; a feature selection means for selecting a stress estimation feature to be used for stress estimation from the observation feature classified based on the classification by the classification means and the classification by the second classification means, The learning device according to any one of appendices 1 to 5, wherein the learning means learns the stress estimation model for each class based on the stress estimation feature and a correct stress value corresponding to the stress estimation feature. [Appendix 7] The learning device according to claim 6, wherein the feature selection means selects the stress estimation feature based on a correlation between the observed feature classified based on the classification by the classification means and the classification by the second classification means and the stress value. [Appendix 8] a classification score calculation means for calculating a classification score representing a degree of certainty that an observed feature value of a subject to be estimated belongs to each of a plurality of classes; a stress estimation means for acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; an integration means for calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; A stress estimation device having the above structure. [Appendix 9] the classification score calculation means calculates the classification score based on attribute information representing attributes of the estimation target person and a classification model; The stress estimation device according to claim 8, wherein the classification model classifies the observed features for learning so that an index representing the correlation between the observed features and the correct stress value corresponding to the observed features becomes higher than before classification. [Appendix 10] The method further includes a feature selection means for selecting a stress estimation feature, which is a feature used for stress estimation, from the observed feature, The stress estimation device according to claim 8 or 9, wherein the stress estimation means calculates the stress estimation value by integrating the stress values ​​of the estimation subject estimated by stress estimation models corresponding to each of the plurality of classes based on the stress estimation features. [Appendix 11] The computer The observed feature of the subject is classified so that an index representing a correlation between the observed feature and a correct stress value corresponding to the observed feature becomes higher than before classification. learning a stress estimation model that estimates a relationship between the observed feature value and the stress value for at least each class divided by the classification, based on the observed feature value and the correct stress value; How to learn. [Appendix 12] The computer calculating a classification score representing a degree of certainty that the observed feature of the estimation subject for which stress estimation is to be performed belongs to each of a plurality of classes; acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; Stress estimation methods. [Appendix 13] The observed feature of the subject is classified so that an index representing a correlation between the observed feature and a correct stress value corresponding to the observed feature becomes higher than before classification. A storage medium storing a program that causes a computer to execute a process of learning a stress estimation model that estimates the relationship between the observed feature values ​​and the stress value, at least for each class divided by the classification, based on the observed feature values ​​and the correct stress value. [Appendix 14] calculating a classification score representing a degree of certainty that the observed feature of the estimation subject for which stress estimation is to be performed belongs to each of a plurality of classes; acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; A storage medium storing a program that causes a computer to execute a process of calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score.

[0115] Although the present invention has been described above with reference to the embodiments, the present invention is not limited to the above embodiments. Various modifications within the scope of the present invention that would be understood by those skilled in the art can be made to the configuration and details of the present invention. In other words, the present invention naturally includes various modifications and alterations that would be possible for those skilled in the art based on the entire disclosure, including the claims, and the technical ideas. Furthermore, the disclosures of the above-cited patent and non-patent documents are incorporated herein by reference. [Explanation of symbols]

[0116] 1. Information processing equipment 1A Stress estimation device 1B, 1X learning device 2 Input devices 3 Display device 4 Storage device 5 sensors 8 Terminal Equipment 100, 100A Stress Estimation System

Claims

1. a classification means for classifying the observed features by provisionally dividing the observed features of a subject into a plurality of classes based on attribute information representing the attributes of the subject, then moving each of the observed features between the plurality of classes, adopting the move if an index representing a correlation between the observed features after the move and a correct stress value corresponding to the observed feature becomes higher than before the move, and returning the index to the state before the move if the index becomes lower than before the move; a learning means for learning a stress estimation model that estimates a relationship between the observed feature value and the stress value for at least each class divided by the classification, based on the observed feature value and the correct stress value; A learning device having the above configuration.

2. 2. The learning device according to claim 1, further comprising a classification model learning means for learning a classification model that estimates a relationship between the attribute and the class based on the attribute information and the classification result.

3. The learning device according to claim 2 , wherein the classifying means classifies the observed feature quantities used for training each of the stress estimation models based on the attribute information and the classification model.

4. 4. The learning device according to claim 1, wherein the classification means classifies the observed features into a plurality of classes, and then subdivides each of the plurality of classes into a class whose index becomes higher as a result of the subdivision.

5. a second classification means for classifying the observed feature based on at least one of an observed object of the observed feature and an activity state of the subject; a feature selection means for selecting a stress estimation feature to be used for stress estimation from the observation feature classified based on the classification by the classification means and the classification by the second classification means, The learning device according to any one of claims 1 to 4, wherein the learning means learns the stress estimation model for each class based on the stress estimation feature and a correct stress value corresponding to the stress estimation feature.

6. 6. The learning device according to claim 5, wherein the feature selection means selects the stress estimation feature based on a correlation between the stress value and the observed feature classified based on the classification by the classification means and the classification by the second classification means.

7. a classification score calculation means for calculating a classification score representing a degree of certainty that an observed feature value of a subject to be estimated belongs to each of a plurality of classes based on attribute information representing attributes of the subject to be estimated and a classification model; a stress estimation means for acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; an integration means for calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; and The classification model is a stress estimation device that classifies observed features for learning so that an index representing the correlation between the observed features and the correct stress value corresponding to the observed features becomes higher than before classification.

8. The method further includes a feature selection means for selecting a stress estimation feature, which is a feature used for stress estimation, from the observed feature, The stress estimation device according to claim 7, wherein the stress estimation means calculates the stress estimation value by integrating the stress values ​​of the estimation subject estimated by stress estimation models corresponding to each of the plurality of classes based on the stress estimation feature.

9. The computer tentatively dividing the observed feature values ​​of a subject into a plurality of classes based on attribute information that represents the attributes of the subject, and then moving each of the observed feature values ​​between the plurality of classes; adopting the move if an index that represents a correlation between the observed feature value after the move and a correct stress value corresponding to the observed feature value is higher than before the move; and returning the index to the state before the move if the index is lower than before the move, thereby classifying the observed feature values; learning a stress estimation model that estimates a relationship between the observed feature value and the stress value for at least each class divided by the classification, based on the observed feature value and the correct stress value; How to learn.

10. The computer calculating a classification score representing a degree of certainty that the observed feature value of the estimation subject belongs to each of a plurality of classes based on attribute information representing the attributes of the estimation subject who is the subject of stress estimation and a classification model; acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; the classification model is a model that classifies the observed feature values ​​for learning so that an index representing a correlation between the observed feature values ​​and a correct stress value corresponding to the observed feature values ​​becomes higher than that before classification. Stress estimation methods.

11. tentatively dividing the observed feature values ​​of a subject into a plurality of classes based on attribute information that represents the attributes of the subject, and then moving each of the observed feature values ​​between the plurality of classes; adopting the move if an index that represents a correlation between the observed feature value after the move and a correct stress value corresponding to the observed feature value is higher than before the move; and returning the index to the state before the move if the index is lower than before the move, thereby classifying the observed feature values; A program that causes a computer to execute a process of learning a stress estimation model that estimates the relationship between the observed feature values ​​and the stress value, at least for each class divided by the classification, based on the observed feature values ​​and the correct stress value.

12. calculating a classification score representing a degree of certainty that the observed feature value of the estimation subject belongs to each of a plurality of classes based on attribute information representing the attributes of the estimation subject who is the subject of stress estimation and a classification model; acquiring a stress value of the estimation subject estimated by a stress estimation model corresponding to each of the plurality of classes based on the observed feature amount; causing a computer to execute a process of calculating a stress estimation value by integrating the stress values ​​of the estimation subject estimated by each of the stress estimation models using the classification score; The classification model is a model that classifies observed features for learning so that an index representing the correlation between the observed features and the correct stress value corresponding to the observed features becomes higher than before classification.

Images