What makes a good movie recommendation? Feature selection for Content-Based Filtering
Maciej Gawinecki, Wojciech Szmyd, Urszula Żuchowicz and Marcin Walas
Nowadays, recommendation systems are becoming ubiquitous, especially in the entertainment industry, such as movie streaming services. In More-Like-This recommendation approach, movies are suggested based on attributes of a currently inspected movie. However, it is not obvious which features are the best predictors for similarity, as perceived by users. To address this problem, we developed and evaluated a recommendation system consisting of nine features and a variety of their representations. We crowdsourced relevance judgments for more than 5 thousand movie recommendations to evaluate the configurations of several dozen of movie features. From five embedding techniques for textual attributes, we selected Universal Sentence Encoder model as the best representation method for producing recommendations. Evaluation of movie features relevance showed that summary and categories extracted from Wikipedia led to the highest similarity on user perceptions in comparison to other analyzed features. We applied the feature weighting methods, commonly used in classification tasks, to determine optimal weights for a given feature set. Our results showed that we can reduce features to only genres, summary, plot, categories, and release year without losing the quality of recommendations.
