Combining Algorithms and User Experience: A Hybrid Personalized Movie Recommender Based on Perceived Similarity

Recommender systems, which filter information based on individual interests, represent a possible remedy for information overload. There are two major types of recommendation techniques—collaborative filtering and content-based. Although the content-based approach alleviates the “cold-start” problem faced by collaborative filtering, this approach generally produces lower accuracy. Thus, a hybrid strategy is often adopted. However, we identified that existing approaches are hampered by insufficient analysis of the unstructured content features of recommended products and a problematic assumption that ignores individual differences in the perception of similarity. Therefore, we propose a new recommendation framework that applies Latent Semantic Analysis to extract semantic features from unstructured text and uses Multiple Regression to identify a unique similarity weighting strategy for each user. By using a combined dataset from MovieLens and Microsoft Xbox, we developed a movie recommender as a proof-of-concept. The initial results represented a promising opportunity to combine behavioral studies and computer algorithms.