WO2022249483A1 - 予測装置、学習装置、予測方法、学習方法およびプログラム - Google Patents
予測装置、学習装置、予測方法、学習方法およびプログラム Download PDFInfo
- Publication number
- WO2022249483A1 WO2022249483A1 PCT/JP2021/020539 JP2021020539W WO2022249483A1 WO 2022249483 A1 WO2022249483 A1 WO 2022249483A1 JP 2021020539 W JP2021020539 W JP 2021020539W WO 2022249483 A1 WO2022249483 A1 WO 2022249483A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- mood
- series
- data
- action
- prediction model
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 53
- 230000036651 mood Effects 0.000 claims abstract description 246
- 230000009471 action Effects 0.000 claims abstract description 184
- 238000000605 extraction Methods 0.000 claims abstract description 50
- 239000000284 extract Substances 0.000 claims abstract description 17
- 230000006399 behavior Effects 0.000 claims description 71
- 230000003542 behavioural effect Effects 0.000 claims description 41
- 238000010276 construction Methods 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 description 71
- 238000012545 processing Methods 0.000 description 38
- 230000008569 process Effects 0.000 description 29
- 239000013598 vector Substances 0.000 description 24
- 230000037007 arousal Effects 0.000 description 16
- 238000010586 diagram Methods 0.000 description 13
- 238000003860 storage Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 6
- 230000001174 ascending effect Effects 0.000 description 4
- 230000008451 emotion Effects 0.000 description 3
- 230000002996 emotional effect Effects 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 230000036642 wellbeing Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000012625 in-situ measurement Methods 0.000 description 1
- 230000003340 mental effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000035882 stress Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Definitions
- the present invention relates to a prediction device, a learning device, a prediction method, a learning method and a program.
- Non-Patent Document 1 as a technique for automatically estimating a user's psychological state from such behavior data, the obtained data is discretized and converted into a histogram, and a probabilistic generation model is used to indicate the degree of health and stress. A technique for estimating is disclosed.
- Non-Patent Document 2 discloses a technique for regressing the psychological state of the next day using daily series data of operation logs and screen time acquired from smartphones.
- Non-Patent Document 3 discloses a technique for predicting a future mood data series from a past behavior data series.
- Non-Patent Document 2 mood data posted many times in a day is digitized and converted into an average value, and then the average mood value for the next day is predicted.
- the human mood fluctuates throughout the day, it is not possible to present mood fluctuations to the user simply by predicting the average value.
- Non-Patent Document 3 cannot take into account the influence of one's actions in the future. For example, if the user refrains from taking an action (negative event, etc.) that is inconvenient for him/herself the next day, it is thought that some users will be affected by this and feel negative. Therefore, with the conventional technology, there is a problem that it is difficult to predict future changes in mood over time.
- the disclosed technology aims to predict chronological changes in the future mood of the prediction target.
- the disclosed technology uses a feature amount extraction unit that extracts feature amounts based on past action sequence data to be predicted and outputs the action feature data, and a trained action sequence prediction model for predicting the action sequence. using a behavioral sequence prediction unit that predicts the future behavioral sequence of the prediction target based on the behavioral characteristic data, and a learned mood sequence prediction model for predicting the mood sequence, the behavioral characteristic data and the prediction and a mood series prediction unit that predicts the future mood series of the prediction target based on the past mood series data of the target.
- FIG. 3 is a functional configuration diagram of a learning device; FIG. It is a functional block diagram of a prediction apparatus.
- 6 is a flowchart showing an example of the flow of learning processing; 6 is a flowchart showing an example of the flow of prediction processing; It is a figure which shows an example of action series data. It is a figure which shows an example of preprocessed action feature data. It is a figure which shows an example of mood series data.
- FIG. 4 is a diagram showing an example of preprocessed mood series data; It is a flowchart which shows an example of the flow of action data preprocessing.
- FIG. 6 is a flowchart showing an example of the flow of feature amount extraction processing
- 4 is a flowchart showing an example of the flow of mood data preprocessing
- It is a figure which shows an example of a behavior sequence prediction model.
- It is a flow chart which shows an example of the flow of action sequence prediction model learning processing.
- It is a figure which shows an example of action series prediction model DB and mood series prediction model DB.
- It is a figure which shows an example of a mood series prediction model.
- FIG. 11 is a flowchart showing an example of the flow of mood series prediction model learning processing
- FIG. It is a flowchart which shows an example of the flow of action sequence prediction processing.
- 4 is a flowchart showing an example of the flow of mood series prediction processing; It is a figure which shows the hardware configuration example of a computer.
- the prediction device extracts feature amounts from behavior data of a prediction target (person, animal, etc.), predicts the behavior sequence using a behavior sequence prediction model, and predicts the predicted behavior sequence and the past mood. Based on the data, a mood series prediction model is used to output future mood series data to be predicted.
- Behavior series data is data that indicates chronological changes in behavior.
- the mood series data is data indicating time-series changes in mood.
- the learning device uses machine learning based on learning data including behavior data and mood data of a learning target (person, animal, etc.) to determine the parameters of the behavior sequence prediction model and the mood sequence prediction model. Update.
- FIG. 1 is a functional configuration diagram of a learning device.
- Learning device 10 includes action series data DB 101, mood series data DB 102, action data preprocessing unit 103, mood data preprocessing unit 104, feature amount extraction unit 105, action series prediction model construction unit 106, action A series prediction model learning unit 107, a mood series prediction model learning unit 108, a mood series prediction model construction unit 109, an action series prediction model DB 110, and a mood series prediction model DB 111 are provided.
- the action series data DB 101 stores action series data.
- the action series data is data indicating past actions of the learning target in chronological order. A specific example of action series data will be described later.
- the mood series data DB 102 stores mood series data.
- the mood series data is data indicating the past mood of the learning target in time series. A specific example of the mood series data will be described later.
- the behavior series data DB 101 and the mood series data DB 102 are constructed in advance so that they can be associated with the date and time information when the data was recorded.
- the mood series data DB 102 also stores two types of numerical values, emotional valence and arousal valence, which represent the mood at the point in time when the learning object self-reported.
- the emotional valence indicates the degree of positivity or negativity of emotion at that time.
- the arousal valence represents the degree of emotional excitement.
- the mood series data configured in this way can be answered by the user using a known technique such as the Photographic Affect Meter (PAM) (reference [1]).
- PAM Photographic Affect Meter
- the target user may input the result of self-evaluation of mood, and the input result may be stored in the database.
- the action data preprocessing unit 103 preprocesses the action series data to be learned. Specifically, the action data preprocessing unit 103 extracts action series data of a type specified in advance from the action series data to be learned, scans each extracted action series data column, and extracts missing values and assumptions. If it contains a value that does not exist, replace it with another value. Note that when preprocessing is unnecessary, the action data preprocessing unit 103 does not need to perform preprocessing.
- the mood data preprocessing unit 104 preprocesses the mood series data to be learned. Specifically, the mood data preprocessing unit 104 aggregates the data for each data group obtained by dividing the learning target mood series data for each time range specified in advance. Note that if preprocessing is unnecessary, the mood data preprocessing unit 104 does not need to perform preprocessing.
- the feature amount extraction unit 105 aggregates data for each data group obtained by dividing the preprocessed action series data for each time range specified in advance, and uses the aggregated data as action feature data.
- the action series prediction model building unit 106 takes the action feature data as input and builds an action series prediction model for outputting the average and variance of the predicted action series data.
- the behavioral sequence prediction model learning unit 107 updates the parameters of the behavioral sequence prediction model by machine learning.
- the mood series prediction model learning unit 108 updates the parameters of the mood series prediction model by machine learning.
- the mood series prediction model construction unit 109 receives behavior feature data, predicted behavior series data, and preprocessed mood series data as input, and builds a mood series prediction model for outputting predicted mood series data. do.
- the action series prediction model DB 110 stores the action series prediction model together with the parameters.
- the behavior sequence prediction model DB 110 is an example of a behavior sequence prediction model storage unit that stores a behavior sequence prediction model.
- the mood series prediction model DB 111 stores mood series prediction models together with parameters.
- the mood series prediction model DB 111 is an example of a mood series prediction model storage unit that stores mood series prediction models.
- FIG. 2 is a functional configuration diagram of a prediction device.
- the prediction device 20 includes an action data preprocessing unit 201, a mood data preprocessing unit 202, a feature quantity extraction unit 203, an action sequence prediction model DB 204, an action sequence prediction unit 205, a mood sequence prediction model DB 206, and a mood sequence prediction model DB 206. and a sequence prediction unit 207 .
- the behavior data preprocessing unit 201 performs the same processing as the behavior data preprocessing unit 103 of the learning device 10 on past behavior series data to be predicted. That is, the action data preprocessing unit 201 extracts action series data of a type specified in advance from the past action series data to be predicted, scans each extracted action series data column, and extracts missing values and assumed values. If it contains a value that does not exist, replace it with another value.
- the mood data preprocessing unit 202 performs the same processing as the mood data preprocessing unit 104 of the learning device 10 on past mood series data to be predicted. That is, the mood data preprocessing unit 202 aggregates data for each data group obtained by dividing the past mood series data to be predicted for each time range specified in advance.
- the feature quantity extraction unit 203 performs the same processing as the feature quantity extraction unit 105 of the learning device 10. That is, the feature amount extraction unit 203 aggregates data for each data group obtained by dividing the preprocessed action series data for each time range specified in advance, and uses the aggregated data as action feature data.
- the action series prediction model DB 204 stores the action series prediction model built by the action series prediction model construction unit 106 of the learning device 10 together with the parameters.
- the stored parameters are parameters updated by the behavior sequence prediction model learning unit 107 of the learning device 10 .
- the action series prediction unit 205 predicts the future action series of the prediction target based on the action feature data using the action series prediction model.
- the mood series prediction model DB 206 stores the mood series prediction model constructed by the mood series prediction model construction unit 109 of the learning device 10 together with parameters.
- the stored parameters are parameters updated by mood series prediction model learning section 108 of learning device 10 .
- the mood series prediction unit 207 uses the mood series prediction model to predict the future mood series to be predicted based on the behavior feature data and preprocessed mood series data.
- the learning device 10 starts learning processing in response to a user's operation or the like, or periodically.
- FIG. 3 is a flowchart showing an example of the flow of learning processing.
- the action data preprocessing unit 103 receives and processes action series data from the action series data DB 101 (step S100).
- the feature amount extraction unit 105 receives and processes the preprocessed action series data from the action data preprocessing unit 103 (step S110).
- the mood data preprocessing unit 104 receives and processes the mood series data from the mood series data DB 102 (step S120).
- the action sequence prediction model construction unit 106 constructs an action sequence prediction model (step S130).
- the action sequence prediction model learning unit 107 receives the action feature data from the feature amount extraction unit 105, learns the action sequence prediction model, and outputs the learned action sequence prediction model to the action sequence prediction model DB 110 (step S140).
- the mood series prediction model construction unit 109 constructs a mood series prediction model (step S150).
- Mood series prediction model learning unit 108 receives action feature data from feature amount extraction unit 105, preprocessed mood series data from mood data preprocessing unit 104, and mood series prediction model from mood series prediction model construction unit 109, and performs mood series prediction.
- the model is learned, and the learned mood series prediction model is output to the mood series prediction model DB 111 (step S160).
- the prediction device 20 starts prediction processing in response to a user's operation or the like, or periodically.
- FIG. 4 is a flowchart showing an example of the flow of prediction processing.
- the action data preprocessing unit 201 receives and processes past action series data to be predicted as an input (step S200).
- the feature quantity extraction unit 203 receives and processes the preprocessed action series data from the action data preprocessing unit 201 (step S210).
- the action series prediction unit 205 receives and processes the action feature data from the feature quantity extraction unit 203 (step S220).
- the mood data preprocessing unit 202 receives and processes past mood series data to be predicted as an input (step S230).
- the mood series prediction unit 207 extracts behavior feature data from the feature amount extraction unit 203, future behavior series data predicted from the behavior series prediction unit 205, preprocessed mood series data from the mood data preprocessing unit 202, mood series A learned mood series prediction model is received from the prediction model DB, a future mood series prediction result is calculated, and output together with the future mood series prediction result (step S240).
- FIG. 5 is a diagram showing an example of action series data.
- the action series data 901 includes data ID, date and time, number of steps, exercise time, and conversation time as items.
- the value of the item “data ID” is an identifier for identifying each data record.
- the value of the item “date and time” is a value indicating the date and time when the action was performed.
- the item “number of steps”, the item “exercise time”, and the item “conversation time” are examples of items indicating the content of actions, and other items may be used.
- FIG. 6 is a diagram showing an example of preprocessed behavior feature data.
- the behavior feature data 902 includes data ID, date and time, number of steps, exercise time, and conversation time as items.
- the value of the item “data ID” is an identifier for identifying each data record.
- the value of the item “date and time” is a value indicating the starting point of the time range specified in advance by the system administrator.
- the item “number of steps”, the item “exercise time”, and the item “conversation time” are examples of items indicating the content of actions, and other items may be used.
- Each value of the item “number of steps”, the item “exercise time”, and the item “conversation time” is a feature quantity representing the characteristics of each behavior by statistical processing based on normalized values.
- FIG. 7 is a diagram showing an example of mood series data.
- the mood series data 903 includes data ID, date and time, Valence, and Arousal as items.
- the value of the item “data ID” is an identifier for identifying each data record.
- the value of the item “date and time” is a value indicating the date and time when the mood was declared.
- the item “Valence” and the item “Arousal” are examples of mood items, and other items may be used. Each value of the item “Valence” and the item “Arousal” is a score self-reported by a learning target or prediction target user or the like.
- FIG. 8 is a diagram showing an example of preprocessed mood series data.
- the preprocessed mood series data (mood data) 904 includes data ID, date and time, Valence, and Arousal as items.
- the value of the item “data ID” is an identifier for identifying each data record.
- the value of the item “date and time” is a value indicating the starting point of the time range specified in advance by the system administrator.
- the item “Valence” and the item “Arousal” are examples of mood items, and other items may be used. Each value of the item “Valence” and the item “Arousal” is the average value of each score in each time range.
- FIG. 9 is a flowchart showing an example of the flow of action data preprocessing.
- the action data preprocessing is processing by the action data preprocessing unit 103 in step S100 of the learning process or processing by the action data preprocessing unit 201 in step S200 of the prediction process.
- the action data preprocessing unit 103 receives action series data from the action series data DB 101 (step S300).
- the action data preprocessing unit 201 receives action series data as an input.
- the action data preprocessing unit 103 or the action data preprocessing unit 201 extracts the type of action series data specified in advance by the system administrator (step S310). For example, the action data preprocessing unit 103 or the action data preprocessing unit 201 defines column names of action series data to be extracted, and extracts column data that matches the column names.
- the action data preprocessing unit 103 or the action data preprocessing unit 201 scans the columns of each extracted action sequence data, and if missing values or unexpected values are included, replace them with other values. (Step S320). For example, the action data preprocessing unit 103 or the action data preprocessing unit 201 inserts the average value or 0 of the corresponding column in the case of numerical data, or inserts a character string representing a missing value in the case of character string type data. do.
- the action data preprocessing unit 103 or the action data preprocessing unit 201 transfers the converted preprocessed action series data and the corresponding date and time information to the feature amount extraction unit 105 or the feature amount extraction unit 203. (Step S330).
- FIG. 10 is a flowchart showing an example of the flow of feature quantity extraction processing.
- the feature amount extraction process is the process by the feature amount extraction unit 105 in step S110 of the learning process or the process by the feature amount extraction unit 203 in step S210 of the prediction process.
- the feature quantity extraction unit 105 or the feature quantity extraction unit 203 receives preprocessed action series data from the action data preprocessing unit 201 (step S400).
- the feature quantity extraction unit 105 or the feature quantity extraction unit 203 scans each column of action series data and normalizes the values (step S410). For example, the feature quantity extraction unit 105 or the feature quantity extraction unit 203 normalizes numerical data so that the average is 0 and the standard deviation is 1, and the character string type data is converted to one-hot Convert to a vector of representations.
- the feature quantity extraction unit 105 or the feature quantity extraction unit 203 scans the date and time information of the action series data, and divides the data into time ranges specified in advance by the system administrator (step S420). For example, when the time range is set to 6 hours, the feature amount extraction unit 105 or the feature amount extraction unit 203 divides the data every 6 hours.
- the feature quantity extraction unit 105 or the feature quantity extraction unit 203 aggregates data for each divided data group (step S430). For example, the feature quantity extraction unit 105 or the feature quantity extraction unit 203 calculates an average value or total value for each data group in the case of numerical data. Further, in the case of vector data of one-hot expression, the feature amount extraction unit 105 or feature amount extraction unit 203 counts the number of occurrences of each dimension to obtain new vector data.
- the feature amount extraction unit 105 transfers the obtained behavior feature data and the corresponding date and time information to the behavior sequence prediction model learning unit 107 and the mood sequence prediction model learning unit 108 (step S440). ).
- the feature amount extraction unit 203 passes the obtained behavior feature data and the corresponding date and time information to the behavior sequence prediction unit 205 and the mood sequence prediction unit 207.
- FIG. 11 is a flowchart showing an example of the flow of mood data preprocessing.
- the mood data preprocessing is processing by the mood data preprocessing unit 104 in step S120 of the learning process or processing by the mood data preprocessing unit 202 in step S230 of the prediction processing.
- the mood data preprocessing unit 104 receives mood series data from the mood series data DB 102 (step S500).
- the mood data preprocessing unit 202 receives mood series data as an input.
- the mood data preprocessing unit 104 or the mood data preprocessing unit 202 scans the date and time information of the mood series data, and divides the data into time ranges specified in advance by the system administrator (step S510). This time range is the same as the time range determined by the feature quantity extraction unit 105 or the feature quantity extraction unit 203 .
- the mood data preprocessing unit 104 or the mood data preprocessing unit 202 aggregates data for each divided data group (step S520). For example, mood data preprocessing unit 104 or mood data preprocessing unit 202 calculates and stores the average value of Valence and Arousal. If there is no data in the divided section, mood data preprocessing unit 104 or mood data preprocessing unit 202 stores 0 in both Valence and Arousal.
- the mood data preprocessing unit 104 delivers the obtained preprocessed mood series data and the corresponding date and time information to the mood series prediction model learning unit 108 (step S530).
- the mood data preprocessing unit 202 transfers the obtained preprocessed mood series data and corresponding date and time information to the mood series prediction unit 207 .
- FIG. 12 is a diagram showing an example of a behavior sequence prediction model.
- the action series prediction model 905 is, for example, a DNN (Deep Neural Network) having the following network.
- DNN Deep Neural Network
- the first network is a fully connected layer 1 that extracts more abstract features from behavioral feature data.
- the first network transforms the input in a fully connected layer, and then nonlinearly transforms the input feature amount using, for example, a sigmoid function or ReLu function to obtain a feature vector.
- the first network sequentially repeats this process for behavior feature data given as time-series data.
- the second network is RNN1 that further abstracts the abstracted feature vectors as series data, and is implemented with known technology such as LSTM (Long-Short term memory). Specifically, the second network sequentially receives series data, and repeatedly non-linearly transforms them while considering past abstracted information.
- LSTM Long-Short term memory
- the third network is a fully connected layer 2 that transforms the feature vectors obtained by RNN 1 into normal distribution parameters.
- the third network obtains parameters of mean value and variance corresponding to each input action data, and associates output vector data with the mean and variance of each action data. That is, the number of dimensions of the vector is the input number of dimensions ⁇ 2.
- the third network converts using a sigmoid function or the like so that the variance is non-negative.
- FIG. 13 is a flowchart showing an example of the flow of action series prediction model learning processing.
- the action series prediction model learning process is a process by the action series prediction model learning unit 107 in step S140 of the learning process.
- the behavior sequence prediction model learning unit 107 rearranges the received behavior feature data in ascending order with reference to date and time information (step S600). Next, the action series prediction model learning unit 107 constructs a pair of action feature data to be input and action series data expected to be output as learning data (step S610).
- the behavior sequence prediction model learning unit 107 receives the behavior sequence prediction model from the behavior sequence prediction model construction unit 106 (step S620).
- the action series prediction model learning unit 107 initializes the model parameters of each unit in the network (step S630). For example, the behavior sequence prediction model learning unit 107 initializes parameters with random numbers from 0 to 1.
- the action series prediction model learning unit 107 updates the parameters of the network using the learning data (step S640). Specifically, the action series prediction model learning unit 107 learns network parameters so that correct action series data can be predicted correctly from input action series data.
- the action series prediction model learning unit 107 calculates the negative logarithmic likelihood for the output action series data using the average and variance values obtained in the fully connected layer 2 of the network, and calculates the negative logarithmic likelihood
- the network may be learned by a known technique such as the error backpropagation method depending on the degrees. Note that a normal distribution may be used to calculate the logarithmic likelihood.
- the behavior sequence prediction model learning unit 107 outputs the behavior sequence prediction model, and stores the output result in the behavior sequence prediction model DB 110 (step S650).
- FIG. 14 is a diagram showing an example of a behavior series prediction model DB and a mood series prediction model DB.
- the behavior series prediction model DB and mood series prediction model DB store parameters in the form of data 906 .
- Data 906 includes a combination of parameters and parameter values.
- FIG. 15 is a diagram showing an example of a mood series prediction model.
- Mood series prediction model 907 is, for example, a DNN having the following network.
- the first network is a fully connected layer 3 that extracts more abstract features from behavioral feature data.
- the first network performs the same processing as the fully connected layer 1 and obtains feature vectors.
- the first network sequentially repeats this process for behavior feature data given as time-series data.
- the second network is RNN2, which further abstracts the abstracted feature vectors as series data, and performs processing using known technology in the same manner as RNN1 to obtain feature vectors.
- the third network is the Self-Attention Mechanism (ATT), which calculates feature vectors that consider the degree of importance of each feature vector abstracted by RNN1.
- a weight representing the importance of each feature vector is implemented by two fully connected layers.
- the third network transforms the feature vector into a context vector of arbitrary size in the first fully connected layer, and outputs the scalar value corresponding to the weight by inputting the context vector in the second fully connected layer.
- a context vector may be subjected to non-linear transformation processing.
- a third network converts the obtained set of importances into probability values, for example, with a softmax function.
- a third network calculates a weighted average of the obtained weight set and feature vector set to obtain a feature vector considering the importance of each feature vector.
- the fourth network is RNN3, which abstracts the preprocessed mood series data, and processes it with known technology in the same way as RNN1 to obtain feature vectors.
- the fifth network is a fully connected layer 4 that outputs Valence and Arousal at the time to be predicted from the feature vectors obtained by the self-attention mechanism and RNN3.
- the fifth network combines the feature vectors obtained by the self-attention mechanism and the RNN 3, and then inputs them to the fully connected layer 4 to obtain two-dimensional values corresponding to the values of Valence and Arousal.
- FIG. 16 is a flowchart showing an example of the flow of mood series prediction model learning processing.
- the mood series prediction model learning process is performed by the mood series prediction model learning unit 108 in step S160 of the learning process.
- the mood series prediction model learning unit 108 receives the behavior feature data from the feature amount extraction unit 105 and the preprocessed mood series data from the mood data preprocessing unit 104, and sorts the data in ascending order by referring to date and time information (step S700). .
- the mood series prediction model learning unit 108 builds learning data (step S710). For example, the mood series prediction model learning unit 108 extracts one piece of preprocessed mood series data and uses it as correct data. By referring to the date and time information of the preprocessed mood series data, N preprocessed mood data past that date and time, N behavior feature data past that date and time, and future actions past that date and time N pieces of feature data are extracted and used as learning data. The mood series prediction model learning unit 108 constructs these three data as one data for learning. Note that N is set in advance by the system administrator.
- the mood series prediction model learning unit 108 adds noise to the future action feature data obtained as learning data (step S720).
- the mood series prediction model learning unit 108 prepares a normal distribution in which each value of the action feature data is averaged and random numbers obtained from a gamma distribution having an arbitrary parameter are used as variances, and is obtained by sampling from the normal distribution. The obtained value is stored as new behavior feature data, and the random number obtained from the gamma distribution is stored as variance.
- the mood series prediction model learning unit 108 receives the model from the mood series prediction model learning unit (step S730).
- the mood series prediction model learning unit 108 initializes the model parameters of each unit in the network (step S740). For example, the mood series prediction model learning unit 108 initializes parameters with random numbers from 0 to 1.
- the mood series prediction model learning unit 108 updates the network parameters using the learning data (step S750). Specifically, the mood series prediction model learning unit 108 learns network parameters so as to correctly predict correct mood series data from input mood series data. For example, the mood series prediction model learning unit 108 uses a known error function to calculate the error with respect to Valence and Arousal given as correct data for the prediction results of Valence and Arousal obtained in the fully connected layer 4 of the network, The error may be used to train the network by a known technique such as error backpropagation.
- the mood series prediction model learning unit 108 outputs the mood series prediction model (network structure and model parameters), and stores the output result in the mood series prediction model DB 111 (step S760).
- FIG. 17 is a flowchart showing an example of the flow of action series prediction processing.
- the action series prediction process is a process by the action series prediction unit 205 in step S220 of the prediction process.
- the action series prediction unit 205 receives the action feature data from the feature amount extraction unit 203, refers to the date and time information, and rearranges the data in ascending order (step S800). Next, the action series prediction unit 205 receives the learned action series prediction model from the action series prediction model DB 204 (step S810).
- the action series prediction unit 205 calculates future action series data from the action feature data using the action series prediction model, and transfers it to the mood series prediction unit 207 (step S820).
- FIG. 18 is a flowchart showing an example of the flow of mood series prediction processing.
- the mood series prediction processing is processing by the mood series prediction unit 207 in step S240 of the prediction processing.
- the mood sequence prediction unit 207 receives the behavior feature data from the feature amount extraction unit 203, the behavior series data predicted from the behavior sequence prediction unit 205, and the preprocessed mood series data from the mood data preprocessing unit 202 (step S900).
- the mood series prediction unit 207 normalizes the behavior feature data (step S910). For example, the mood series prediction unit 207 uses 0 as the variance value because the average value and the variance value of each action correspond to the input data of the mood series prediction model, but the behavior feature data does not have the variance value. Store.
- the mood series prediction unit 207 combines the behavioral feature data and the predicted behavioral series data (step S920). For example, the mood series prediction unit 207 rearranges the action feature data in ascending order according to the date and time information, combines the action series data predicted from the last row in the order of output, and treats the obtained series data as action feature data.
- the mood series prediction unit 207 receives the learned mood series prediction model from the mood series prediction model DB 206 (step S930).
- the mood series prediction unit 207 uses the mood series prediction model to calculate and output the current mood (Valence, Arousal) from the behavior feature data and the preprocessed mood series data (step S940).
- Learning device 10 and prediction device 20 can be realized, for example, by causing a computer to execute a program describing the processing details described in the present embodiment.
- this "computer” may be a physical machine or a virtual machine on the cloud.
- the "hardware” described here is virtual hardware.
- the above program can be recorded on a computer-readable recording medium (portable memory, etc.), saved, or distributed. It is also possible to provide the above program through a network such as the Internet or e-mail.
- FIG. 19 is a diagram showing a hardware configuration example of the computer.
- the computer of FIG. 19 has a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., which are connected to each other via a bus B, respectively.
- a program that implements the processing in the computer is provided by a recording medium 1001 such as a CD-ROM or memory card, for example.
- a recording medium 1001 such as a CD-ROM or memory card
- the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000 .
- the program does not necessarily need to be installed from the recording medium 1001, and may be downloaded from another computer via the network.
- the auxiliary storage device 1002 stores installed programs, as well as necessary files and data.
- the memory device 1003 reads and stores the program from the auxiliary storage device 1002 when a program activation instruction is received.
- the CPU 1004 implements functions related to the device according to programs stored in the memory device 1003 .
- the interface device 1005 is used as an interface for connecting to the network.
- a display device 1006 displays a GUI (Graphical User Interface) or the like by a program.
- An input device 1007 is composed of a keyboard, a mouse, buttons, a touch panel, or the like, and is used to input various operational instructions.
- the output device 1008 outputs the calculation result.
- the computer may include a GPU (Graphics Processing Unit) or TPU (Tensor Processing Unit) instead of the CPU 1004, or may include a GPU or TPU in addition to the CPU 1004. In that case, the processing may be divided and executed such that the GPU or TPU executes processing that requires special computation, such as a neural network, and the CPU 1004 executes other processing.
- Learning device 10 includes preprocessed behavior data obtained by preprocessing mood sequence data, preprocessed behavior feature data obtained by preprocessing behavior sequence data, and behavior sequence data.
- a mood series prediction model is learned by combining the future behavior series obtained from the prediction model. As a result, the prediction device 20 can predict a prediction target mood series that could not be predicted conventionally.
- the learning device 10 predicts future behavior and learns a mood series prediction model using the prediction results as input.
- the prediction device 20 can predict the mood series to be predicted with high accuracy.
- the prediction device 20 predicts future behavior as an average and a variance, and expresses the reliability of the result as a variance. Then, the learning device 10 learns the mood series prediction model with the reliability taken into account. As a result, the prediction device 20 can predict the mood series by emphasizing actions with high reliability.
- This specification describes at least a prediction device, a learning device, a prediction method, a learning method, and a program described in each of the following items.
- (Section 1) a feature amount extraction unit that extracts feature amounts based on past behavioral series data to be predicted and outputs behavioral feature data; an action series prediction unit that predicts the future action series of the prediction target based on the action feature data using a learned action series prediction model for predicting the action series; a mood series prediction unit that predicts the future mood series of the prediction target based on the behavior feature data and the past mood series data of the prediction target, using a learned mood series prediction model for predicting the mood series; have a prediction device.
- (Section 2) a feature quantity extraction unit that extracts a feature quantity based on the action series data to be learned and outputs the action feature data; a mood series prediction model construction unit for constructing a mood series prediction model for predicting the mood series; a mood series prediction model learning unit that updates parameters of the constructed mood series prediction model based on the behavior feature data and the past mood series data to be learned; learning device.
- the mood sequence prediction model learning unit updates the parameters based on behavioral feature data obtained by adding noise to future behavioral feature data extracted as learning data, A learning device according to claim 2.
- the behavioral sequence prediction model learning unit calculates the negative logarithmic likelihood for the output behavioral sequence data using the average and variance values obtained by the network provided in the behavioral sequence prediction model, and based on the calculation result, update parameters, 5.
- the learning device according to item 4.
- (Section 6) A prediction method performed by a prediction device, a step of extracting feature amounts based on past action series data to be predicted and outputting action feature data; predicting a future behavioral sequence of the prediction target based on the behavioral feature data using a learned behavioral sequence prediction model for predicting the behavioral sequence; predicting the future mood series of the predictor based on the behavior feature data and the past mood series data of the predictor using a learned mood series prediction model for predicting mood series. , Forecast method.
- (Section 7) A learning method executed by a learning device, a step of extracting a feature amount based on the action series data to be learned and outputting the action feature data; building a mood sequence prediction model for predicting the mood sequence; updating the parameters of the constructed mood series prediction model based on the behavior feature data and the past mood series data of the learning target; learning method.
- (Section 8) A program for causing a computer to function as each unit in the prediction device according to item 1, or a program for causing a computer to function as each unit in the learning device according to any one of items 2 to 5. .
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Description
本実施の形態に係る予測装置は、予測対象(人物、動物等)の行動データから特徴量を抽出して行動系列予測モデルを用いて行動系列を予測し、予測された行動系列と過去の気分データに基づいて、気分系列予測モデルを用いて、予測対象の未来の気分系列データを出力する。
図1は、学習装置の機能構成図である。学習装置10は、行動系列データDB101と、気分系列データDB102と、行動データ前処理部103と、気分データ前処理部104と、特徴量抽出部105と、行動系列予測モデル構築部106と、行動系列予測モデル学習部107と、気分系列予測モデル学習部108と、気分系列予測モデル構築部109と、行動系列予測モデルDB110と、気分系列予測モデルDB111と、を備える。
図2は、予測装置の機能構成図である。予測装置20は、行動データ前処理部201と、気分データ前処理部202と、特徴量抽出部203と、行動系列予測モデルDB204と、行動系列予測部205と、気分系列予測モデルDB206と、気分系列予測部207と、を備える。
次に、学習装置10の動作例について、図面を参照して説明する。学習装置10は、ユーザの操作等を受けて、または定期的に、学習処理を開始する。
次に、予測装置20の動作例について、図面を参照して説明する。予測装置20は、ユーザの操作等を受けて、または定期的に、予測処理を開始する。
次に、本実施の形態に係る学習装置10および予測装置20が扱う各種データの形式例について説明する。
次に、本実施の形態に係る学習処理および予測処理の詳細について説明する。
学習装置10および予測装置20は、例えば、コンピュータに、本実施の形態で説明する処理内容を記述したプログラムを実行させることにより実現可能である。なお、この「コンピュータ」は、物理マシンであってもよいし、クラウド上の仮想マシンであってもよい。仮想マシンを使用する場合、ここで説明する「ハードウェア」は仮想的なハードウェアである。
[1]J.P. Pollak, P. Adams, G. Gay: PAM: A Photographic Affect Meter for Frequent, In Situ Measurement of Affect. In Proc. of CHI, 2011.
本実施の形態に係る学習装置10は、気分系列データを前処理して得られた前処理済み行動データと、行動系列データを前処理して得られた前処理済み行動特徴データと、行動系列予測モデルから得られた未来の行動系列を合わせて気分系列予測モデルを学習する。これによって、予測装置20は、従来予測できなかった予測対象の気分系列を予測可能となる。
本明細書には、少なくとも下記の各項に記載した予測装置、学習装置、予測方法、学習方法およびプログラムが記載されている。
(第1項)
予測対象の過去の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力する特徴量抽出部と、
行動系列を予測するための学習済みの行動系列予測モデルを用いて、前記行動特徴データに基づく前記予測対象の未来の行動系列を予測する行動系列予測部と、
気分系列を予測するための学習済みの気分系列予測モデルを用いて、前記行動特徴データおよび前記予測対象の過去の気分系列データに基づく前記予測対象の未来の気分系列を予測する気分系列予測部と、を備える、
予測装置。
(第2項)
学習対象の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力する特徴量抽出部と、
気分系列を予測するための気分系列予測モデルを構築する気分系列予測モデル構築部と、
構築された前記気分系列予測モデルのパラメータを、前記行動特徴データおよび前記学習対象の過去の気分系列データに基づいて更新する気分系列予測モデル学習部と、を備える、
学習装置。
(第3項)
前記気分系列予測モデル学習部は、学習用データとして抽出した未来の行動特徴データにノイズを加えた行動特徴データに基づいて、前記パラメータを更新する、
第2項に記載の学習装置。
(第4項)
行動系列を予測するための行動系列予測モデルを構築する行動系列予測モデル構築部と、
前記行動特徴データに基づいて、前記行動系列予測モデルのパラメータを更新する行動系列予測モデル学習部と、をさらに備える、
第2項または第3項に記載の学習装置。
(第5項)
前記行動系列予測モデル学習部は、前記行動系列予測モデルが備えるネットワークで得られる平均と分散の値を用いて、出力の行動系列データに対する負の対数尤度を計算し、計算結果に基づいて前記パラメータを更新する、
第4項に記載の学習装置。
(第6項)
予測装置が実行する予測方法であって、
予測対象の過去の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力するステップと、
行動系列を予測するための学習済みの行動系列予測モデルを用いて、前記行動特徴データに基づく前記予測対象の未来の行動系列を予測するステップと、
気分系列を予測するための学習済みの気分系列予測モデルを用いて、前記行動特徴データおよび前記予測対象の過去の気分系列データに基づく前記予測対象の未来の気分系列を予測するステップと、を備える、
予測方法。
(第7項)
学習装置が実行する学習方法であって、
学習対象の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力するステップと、
気分系列を予測するための気分系列予測モデルを構築するステップと、
構築された前記気分系列予測モデルのパラメータを、前記行動特徴データおよび前記学習対象の過去の気分系列データに基づいて更新するステップと、を備える、
学習方法。
(第8項)
コンピュータを、第1項に記載の予測装置における各部として機能させるためのプログラム、または、コンピュータを、第2項から第5項のいずれか1項に記載の学習装置における各部として機能させるためのプログラム。
20 予測装置
101 行動系列データDB
102 気分系列データDB
103 行動データ前処理部
104 気分データ前処理部
105 特徴量抽出部
106 行動系列予測モデル構築部
107 行動系列予測モデル学習部
108 気分系列予測モデル学習部
109 気分系列予測モデル構築部
110 行動系列予測モデルDB
111 気分系列予測モデルDB
201 行動データ前処理部
202 気分データ前処理部
203 特徴量抽出部
204 行動系列予測モデルDB
205 行動系列予測部
206 気分系列予測モデルDB
207 気分系列予測部
1000 ドライブ装置
1001 記録媒体
1002 補助記憶装置
1003 メモリ装置
1004 CPU
1005 インタフェース装置
1006 表示装置
1007 入力装置
1008 出力装置
Claims (8)
- 予測対象の過去の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力する特徴量抽出部と、
行動系列を予測するための学習済みの行動系列予測モデルを用いて、前記行動特徴データに基づく前記予測対象の未来の行動系列を予測する行動系列予測部と、
気分系列を予測するための学習済みの気分系列予測モデルを用いて、前記行動特徴データおよび前記予測対象の過去の気分系列データに基づく前記予測対象の未来の気分系列を予測する気分系列予測部と、を備える、
予測装置。 - 学習対象の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力する特徴量抽出部と、
気分系列を予測するための気分系列予測モデルを構築する気分系列予測モデル構築部と、
構築された前記気分系列予測モデルのパラメータを、前記行動特徴データおよび前記学習対象の過去の気分系列データに基づいて更新する気分系列予測モデル学習部と、を備える、
学習装置。 - 前記気分系列予測モデル学習部は、学習用データとして抽出した未来の行動特徴データにノイズを加えた行動特徴データに基づいて、前記パラメータを更新する、
請求項2に記載の学習装置。 - 行動系列を予測するための行動系列予測モデルを構築する行動系列予測モデル構築部と、
前記行動特徴データに基づいて、前記行動系列予測モデルのパラメータを更新する行動系列予測モデル学習部と、をさらに備える、
請求項2または3に記載の学習装置。
学習装置。 - 前記行動系列予測モデル学習部は、前記行動系列予測モデルが備えるネットワークで得られる平均と分散の値を用いて、出力の行動系列データに対する負の対数尤度を計算し、計算結果に基づいて前記パラメータを更新する、
請求項4に記載の学習装置。 - 予測装置が実行する予測方法であって、
予測対象の過去の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力するステップと、
行動系列を予測するための学習済みの行動系列予測モデルを用いて、前記行動特徴データに基づく前記予測対象の未来の行動系列を予測するステップと、
気分系列を予測するための学習済みの気分系列予測モデルを用いて、前記行動特徴データおよび前記予測対象の過去の気分系列データに基づく前記予測対象の未来の気分系列を予測するステップと、を備える、
予測方法。 - 学習装置が実行する学習方法であって、
学習対象の行動系列データに基づいて特徴量を抽出し、行動特徴データを出力するステップと、
気分系列を予測するための気分系列予測モデルを構築するステップと、
構築された前記気分系列予測モデルのパラメータを、前記行動特徴データおよび前記学習対象の過去の気分系列データに基づいて更新するステップと、を備える、
学習方法。 - コンピュータを、請求項1に記載の予測装置における各部として機能させるためのプログラム、または、コンピュータを、請求項2から5のいずれか1項に記載の学習装置における各部として機能させるためのプログラム。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/020539 WO2022249483A1 (ja) | 2021-05-28 | 2021-05-28 | 予測装置、学習装置、予測方法、学習方法およびプログラム |
JP2023523937A JPWO2022249483A1 (ja) | 2021-05-28 | 2021-05-28 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/020539 WO2022249483A1 (ja) | 2021-05-28 | 2021-05-28 | 予測装置、学習装置、予測方法、学習方法およびプログラム |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022249483A1 true WO2022249483A1 (ja) | 2022-12-01 |
Family
ID=84229582
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/020539 WO2022249483A1 (ja) | 2021-05-28 | 2021-05-28 | 予測装置、学習装置、予測方法、学習方法およびプログラム |
Country Status (2)
Country | Link |
---|---|
JP (1) | JPWO2022249483A1 (ja) |
WO (1) | WO2022249483A1 (ja) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH05346915A (ja) * | 1992-01-30 | 1993-12-27 | Ricoh Co Ltd | 学習機械並びにニューラルネットワークおよびデータ分析装置並びにデータ分析方法 |
JP2015184764A (ja) * | 2014-03-20 | 2015-10-22 | ヤフー株式会社 | 行動予測装置、行動予測方法、及び行動予測プログラム |
JP2019036116A (ja) * | 2017-08-15 | 2019-03-07 | ヤフー株式会社 | 決定装置、決定方法、及び決定プログラム |
-
2021
- 2021-05-28 JP JP2023523937A patent/JPWO2022249483A1/ja active Pending
- 2021-05-28 WO PCT/JP2021/020539 patent/WO2022249483A1/ja active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH05346915A (ja) * | 1992-01-30 | 1993-12-27 | Ricoh Co Ltd | 学習機械並びにニューラルネットワークおよびデータ分析装置並びにデータ分析方法 |
JP2015184764A (ja) * | 2014-03-20 | 2015-10-22 | ヤフー株式会社 | 行動予測装置、行動予測方法、及び行動予測プログラム |
JP2019036116A (ja) * | 2017-08-15 | 2019-03-07 | ヤフー株式会社 | 決定装置、決定方法、及び決定プログラム |
Also Published As
Publication number | Publication date |
---|---|
JPWO2022249483A1 (ja) | 2022-12-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7668826B2 (en) | Predicting apparatus, predicting method, and computer product | |
CN110291542B (zh) | 主要原因估计装置、主要原因估计系统以及主要原因估计方法 | |
WO2017010317A1 (ja) | 表示制御装置、表示制御方法、及び、プログラム | |
CN111737586B (zh) | 信息推荐方法、装置、设备及计算机可读存储介质 | |
CN112289442A (zh) | 预测疾病终点事件的方法、装置及电子设备 | |
US20180032674A1 (en) | Diagnostic system, diagnostic method, and storage medium | |
KR20160043777A (ko) | 질환 발병 예측 방법 및 그 장치 | |
CN111178986A (zh) | 用户-商品偏好的预测方法及系统 | |
US9875443B2 (en) | Unified attractiveness prediction framework based on content impact factor | |
JP6038727B2 (ja) | 分析システム及び分析方法 | |
WO2022249483A1 (ja) | 予測装置、学習装置、予測方法、学習方法およびプログラム | |
JP7400965B2 (ja) | 気分予測方法、気分予測装置及びプログラム | |
JP7424389B2 (ja) | 学習装置、推定装置、学習方法、推定方法、及びプログラム | |
CN111243738A (zh) | 生存分析模型的构建、生存率的预测方法、装置及设备 | |
US20230186092A1 (en) | Learning device, learning method, computer program product, and learning system | |
CN116344042A (zh) | 基于多模态分析的认知储备干预提升方法及系统 | |
CN113990506B (zh) | 一种健康状态评估方法、装置、存储介质及计算机系统 | |
US20210350260A1 (en) | Decision list learning device, decision list learning method, and decision list learning program | |
JP7435821B2 (ja) | 学習装置、心理状態系列予測装置、学習方法、心理状態系列予測方法、及びプログラム | |
JP7456503B2 (ja) | 心理状態分析方法、心理状態分析装置及びプログラム | |
EP3836069A1 (en) | Electronic device and control method therefor | |
CN112562861B (zh) | 一种训练传染病预测模型的方法和装置 | |
CN117409978B (zh) | 一种疾病预测模型构建方法、系统、装置及可读存储介质 | |
TonThat et al. | A Feature Subset Selection Approach For Predicting Smoking Behaviours | |
JP7504315B1 (ja) | 状態推定装置、プログラム、状態推定システム及び状態推定方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21943120 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18557158 Country of ref document: US Ref document number: 2023523937 Country of ref document: JP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21943120 Country of ref document: EP Kind code of ref document: A1 |