Development of Serendipity-oriented Recommender System with Time Series Prediction of Unexpectedness

Abstract

Serendipity-oriented recommender systems have been proposed to solve the filter-bubble problem caused by the excessive pursuit of recommendation accuracy. Many previous studies of unexpectedness as a component of serendipity have used a user's browsing and rating history to calculate the degree of unexpectedness of items for the user, but all histories have been considered equal or only the most recent histories have been used. However, these methods cannot take into account changes in user preferences and trends over time. This study proposes a serendipity-oriented recommender system that predicts unexpectedness at the time of recommendation using the time series prediction model and evaluates the recommendation results using a benchmark dataset, MovieLens 100K. The experimental results show that the proposed system improves the performance of Recall, NDCG, and Serendipity by up to 0.84, 0.39, and 0.08 points respectively compared to the baseline systems.