Natural language processing techniques for session-based recommendation.

Abstract

Recommender Systems is a field of research and application focused on identifying and retrieving relevant items given user preferences. There are many scenarios where a Recommender System can be applied, but its performance usually depends on the availability of user consumption historic data. In this work, we evaluate the performance of Recommender Systems in the Session-based scenario, in which the user cannot be identified, comparing the performance of naïve, matrix-based, sequential, and session-based models, also introducing an alternative implementation of one of these models, based on a specific type of Recurrent Neural Network called Gated Recurrent Units. We use Natural Language Processing techniques to create three different input strategies and create session embeddings, analyzing their performance in our model implementation, extracting insights, and applying fine tuning to achieve better results. This work was evaluated using a real-world database extracted from the Last.fm online radio platform.