RACE:GP – a Generic Approach to Automatically Creating and Evaluating Hybrid Recommender Systems

Abstract

In recent years, recommender systems have become a fundamental part of our online experience. Users rely on such systems in situations with many potential choices, such as watching a movie on a streaming service, reading a blog post, or listening to a song. Traditionally, these systems use techniques such as collaborative filtering and content-based recommendation. Both approaches have disadvantages, so to reduce those, recent research combines various techniques in different ways to create hybrid recommender systems. Creating a well-performing hybrid recommender system generally requires extensive knowledge of recommender systems, the domain on which one wants to provide recommendations, and trial and error. Automating this process makes recommender systems accessible for organizations that lack the resources to build these systems themselves. However, there is a lack of research regarding automating this process. This study aims to provide an initial exploration into this area by proposing RACE:GP, an end-to-end approach that automatically produces accurate non-trivial hybrid recommender systems with only a dataset and a definition of 'relevance'. RACE:GP automatically creates a programming language from a dataset in which any valid program is a recommender system on that dataset. By defining the relevant interaction, it can automatically evaluate the accuracy of these programs. It uses a search strategy based on genetic programming to find the best performing recommender systems in the language. To test our hypothesis, RACE:GP is used to produce recommender systems on three well-known datasets in recommender systems literature, and the results are compared to baselines based on collaborative filtering. Additionally, to verify its adaptability, we analyzed the produced recommender systems given different recommendation scenarios. The results showed that RACE:GP is able to produce recommender systems that outperform our chosen baselines by a significant margin. Furthermore, analysis of the produced recommender systems on different recommendation scenarios within a dataset shows that it can find systems that are especially accurate in situations with different densities or recommending specific interactions in datasets that contain different interactions. These results suggest that RACE:GP is a viable and generally applicable approach that makes creating hybrid recommender systems accessible for anyone with a dataset.