Adaptive user similarity measures for recommender systems: A genetic programming approach

Recommender systems signify the shift from the paradigm of “searching for items” to “discovering items” and have been employed by an increasing number of e-commerce sites for matching users to their preferences. Collaborative Filtering is a popular recommendation technique which exploits the past user-item interactions to determine user similarity. The preferences of such similar users are leveraged to offer suggestions to the active user. Even though several techniques for similarity assessment have been suggested in literature, no technique has been proven to be optimal under all contexts/data conditions. Hence, we propose a two-stage process to assess user similarity, the first is to learn the optimal transformation function to convert the raw ratings data to preference data by employing genetic programming, and the second is to utilize the preference values, so derived, to compute user similarity. The application of such learnt user bias gives rise to adaptive similarity measures, i.e. similarity estimates that are dataset dependent and hence expected to work best under any data environment. We demonstrate the superiority of our proposed technique by contrasting it to traditional similarity estimation techniques on four different datasets representing varied data environments.