Explainable Recommender Systems
Recommenders systems add business value and help in better monetization for multiple stakeholders. The objectives of the recommender system vary from domain to domain and in this paper, the focus is on the movie recommender system.
The advent of powerful mobile devices has led to the proliferation of many mobile-based movie recommenders, whereas earlier they were confined to desktop or web-based applications. The future generation of users expects the recommender systems to explain themselves as they want to make informed choices rather than consuming just anything the AI offers to them. The veil of AI as magic is to be lowered as consumers seek meaningful explanations for the recommendations that are laid in front of them.
This paper builds a content-based recommender system based on the genre data of the MovieLens dataset by GroupLens. The clustering technique k-means is used to find hidden patterns and form clusters of similar movies. Surrogate models are used to provide explanations of k-means clustering using decision trees. The decision path, along with the rules, forms the explanation of cluster assignments and the reasoning behind recommended movies.
Traditionally, recommender systems are built on similarity measures based on user preferences, which leads to filter bubbles and echo chambers. Serendipitous recommendations can be generated by coming up with novel and unexpected movies using cosine similarity between centroid vectors of k-means clustering. These serendipitous movies can be explored by users based on the explanations given by the decision paths of the surrogate model.
Last but not least, the explanations can be delivered through multiple interfaces. The prevalent ones are textual and visual methods. This paper explores multimodal explainability through an additional audio medium.
Keywords: Recommender system, Content-based recommendations, explainable AI, Serendipitous recommendations, Explanation interfaces