Dynamic session-based music recommendation using information retrieval techniques

In this paper, we propose the use of information retrieval (IR) techniques in order to build dynamic and scalable, yet accurate, music recommender systems. We describe adaptations of a traditional text retrieval pipeline to tailor it to recommendation tasks, and demonstrate its use in the session-based music recommendation scenario. We propose three methods, two of them based on TF-IDF weighting (IR-TFIDF and IR-1NN), and a third method (IR-MC) that extends the first-order Markov chain method (MC) in order to consider longer past sequences than the original one. We evaluate the proposed methods against state-of-the-art recommender methods in two experiments; the first experiment compares the overall performance of the competitors, while the second explores their dynamic capabilities. The methods based on classic IR relevance weighting scheme has shown comparable performance results to the baselines, while the IR-MC method overcomes its competitors in different scenarios. We find that modeling the recommendation algorithms as IR problems not only expands the set of techniques available for handling the recommendation tasks, but also that the support of the traditional IR pipeline in the implementation of such algorithms plays an important role in the attempt of satisfying the specific requirements of dynamic recommendation scenarios, such as the capability of receiving online updates.