The Effect of Combining Algorithms in Recommendation Systems

Abstract

Due to the increasing investment of industry and the scientific development, new recommendation systems are constantly emerging, seeking to increase the precision of item`s suggestions to consumers and to cover a greater number of application contexts. However, choosing the optimal algorithm for a given application is not always a trivial task. Recent papers have studied ways to accomplish this choice automatically, through meta-learning strategies. This work investigates the effects of the extension of this meta-learning process from the application context level to the user context. Some experiments were performed from four recommendation models, selecting, from two different criteria, the algorithm that presents better performance in the task of mapping the preferences of each user, verifying the effect of the customized application of the algorithms on the systems overall performance. Positive results were achieved when the algorithm selection was based on approaches with similar complexities.