Predicting Human Behavior Using User’s Contextual Embedding by Convolution of Action Graph

Predicting human behavior using logs that include user location information and categories of facilities visited is being actively researched. However, not enough research has focused on user behavioral embedding expressing user preferences. We have developed a behavior prediction model that uses an action graph with categories as nodes and transitions between categories as edges in order to capture the preference of transition on the basis of the context of the places visited by users. It uses the features of the action graph, which are extracted using a graph convolutional network. Experiments demonstrated that using user behavioral embedding extracted by graph convolution improves prediction accuracy. Quantitative and qualitative analyses demonstrated the effectiveness of action graph embedding representation.