Efficient Data Presentation Method for Building User Preference Model Using Interactive Evolutionary Computation

Abstract

In this paper, we propose data presentation methods for building user preference models efficiently in recommendation systems by interactive genetic algorithm. The user preference model of the recommender agent is represented by a three-layer neural network (NN). In order to generate training data for this NN, it is necessary to present some sample items to the user and obtain the user’s evaluation values. Based on the idea that the evaluation distribution of the presented data should not be biased, we propose two types of data presentation methods. One is the selection of three types of data such as like, dislike and neutral. The other is the Inverse Proportional Selection of sampling points based on the frequencies of already presented data. As the results of experiments using four types of pseudo-users, we found that the Inverse Proportional Selection was the most efficient method in learning the preferences of users.