User Profiling Based on Keyword Clusters for Improved Recommendations

Abstract

Recommender Systems (RS) have risen in popularity over the years, and their ability to ease decision-making for the user in various domains has made them ubiquitous. However, the sparsity of data continues to be one of the biggest shortcomings of the suggestions offered. Recommendation algorithms typically model user preferences in the form of a profile, which is then used to match user preferences to items of their interest. Consequently, the quality of recommendations is directly related to the level of detail contained in these profiles. Several attempts at enriching the user profiles leveraging both user preference data and item content details have been explored in the past. We propose a method of constructing a user profile, specifically for the movie domain, based on user preference for keyword clusters, which indirectly captures preferences for various narrative styles. These profiles are then utilized to perform both content-based (CB) filtering as well as collaborative filtering (CF). The proposed approach scores over the direct keyword-matching, genre-based user profiling and the traditional CF methods under sparse data scenarios as established by various experiments. It has the advantage of a compact user model representation, while at the same time capturing the essence of the styles or genres preferred by the user. The identification of implicit genres is captured effectively through clustering without requiring labeled data for training.