Feature extraction device and state estimation system

A state estimation system for estimating a state related to a terminal where an activity is executed or a user of the terminal, comprising circuitry configured to:

• acquire activity history information indicating an execution history of the activity in a specified period of time and including time-series activity logs;
• extract co-occurrence relationship information indicating a co-occurrence relationship of activities in a specified period of time based on the activity history information;
• aggregate activity feature quantities indicating a feature of each activity based on the co-occurrence relationship information, and generate an activity history feature quantity indicating a feature of the activity history information;
• estimate a state related to the terminal or a user of the terminal by inputting the activity history feature quantity generated by the circuitry to a state estimation model including a neural network built by machine learning based on learning data where the activity history feature quantity is an input feature quantity and identification information for identifying a state related to the terminal or a user of the terminal as a supervision label;
• and
• output state information indicating a state related to a terminal or a user of the terminal estimated by the circuitry,
• wherein the circuitry generates a feature quantity that aggregates the activity feature quantity of one activity and an activity feature quantity of one or more activities having the co-occurrence relationship with the one activity by a specified convolution method as a convolutional feature quantity of the one activity, and generates a feature quantity that aggregates convolutional feature quantities of all activities included in the activity history information by a specified aggregation method as the activity history feature quantity,
• wherein
• the circuitry performs hierarchical convolution of activity feature quantities in a specified number of layers from a first layer to a k-th layer (k is a predetermined integer of 1 or more) based on the co-occurrence relationship information,
• when each activity of a plurality of activities having a co-occurrence relationship as indicated by the co-occurrence relationship information is a first activity in an n-th layer (n is an integer of 0 or more), the circuitry defines one or more activities having a co-occurrence relationship with the first activity as a second activity,
• in convolution of the first layer, the circuitry generates the convolutional feature quantity of the first activity in the first layer by convolving an activity feature quantity of the first activity in a 0th layer and an activity feature quantity of one or more second activities having the co-occurrence relationship with the first activity by a specified convolution method,
• in convolution of an m-th layer (m is an integer of 2 or more), the circuitry generates the convolutional feature quantity of the first activity in the m-th layer by convolving an activity feature quantity of the first activity in an (m−1)th layer and the activity feature quantities of the one or more second activities having the co-occurrence relationship with the first activity by a specified convolution method,
• the circuitry generates the convolutional feature quantity of the first activity in the k-th layer in convolution of the k-th layer as the convolutional feature quantity of the one activity based on k-layer convolution, and
• the circuitry generates a feature quantity that aggregates the convolutional feature quantities based on the k-layer convolution of all activities included in the activity history information by a specified aggregation method as the activity history feature quantity.

A method implemented by circuitry of a feature extraction device for extracting a feature quantity indicating a feature related to an execution history of an activity based on the execution history, the feature quantity being input to a state estimation model for estimating a state related to a terminal where the activity is executed, or a user of the terminal, the method comprising:

• acquiring activity history information indicating an execution history of the activity in a specified period of time and including time-series activity logs;
• extracting co-occurrence relationship information indicating a co-occurrence relationship of activities in a specified period of time based on the activity history information;
• aggregating activity feature quantities indicating a feature of each activity based on the co-occurrence relationship information, and generate an activity history feature quantity indicating a feature of the activity history information;
• and
• outputting the activity history feature quantity as a feature quantity to be input to the state estimation model,
• wherein the circuitry generates a feature quantity that aggregates the activity feature quantity of one activity and an activity feature quantity of one or more activities having the co-occurrence relationship with the one activity by a specified convolution method as a convolutional feature quantity of the one activity, and generates a feature quantity that aggregates convolutional feature quantities of all activities included in the activity history information by a specified aggregation method as the activity history feature quantity,
• wherein
• the circuitry performs hierarchical convolution of activity feature quantities in a specified number of lavers from a first layer to a k-th layer (k is a predetermined integer of 1 or more) based on the co-occurrence relationship information,
• when each activity of a plurality of activities having a co-occurrence relationship as indicated by the co-occurrence relationship information is a first activity in an n-th layer (n is an integer of 0 or more), the circuitry defines one or more activities having a co-occurrence relationship with the first activity as a second activity,
• in convolution of the first layer, the circuitry generates the convolutional feature quantity of the first activity in the first layer by convolving an activity feature quantity of the first activity in a 0th layer and an activity feature quantity of one or more second activities having the co-occurrence relationship with the first activity by a specified convolution method,
• in convolution of an m-th layer (m is an integer of 2 or more), the circuitry generates the convolutional feature quantity of the first activity in the m-th layer by convolving an activity feature quantity of the first activity in an (m−1)th layer and the activity feature quantities of the one or more second activities having the co-occurrence relationship with the first activity by a specified convolution method,
• the circuitry generates the convolutional feature quantity of the first activity in the k-th layer in convolution of the k-th layer as the convolutional feature quantity of the one activity based on k-layer convolution, and
• the circuitry generates a feature quantity that aggregates the convolutional feature quantities based on the k-layer convolution of all activities included in the activity history information by a specified aggregation method as the activity history feature quantity.