Recommender system based on personal kansei evaluation tendency

Abstract

In the present paper, we propose a recommender system that considers the current user behavior tendencies to evaluate relevant contents. In, the proposed model, the user preference factors are estimated based on the collected kansei evaluations and image features of recommended contents. The features are acquired from the image of the content by using a deep neural network, and the user evaluation for each feature is estimated by applying the gradient boosting decision tree. In addition, we analyze the relationship between the user evaluation and the image feature, and search for the users with the similar evaluation tendency with regard to changes in the feature. Based on the obtained results, we recommend the individuals that are presumed to be highly rated by other users. In addition, we estimate the features that are the factors of the user preference corresponding to recommendation results and present the common features as images. The effectiveness of the proposed model is verified by performing a simulation using the user evaluation data.