Computing and Informatics

Day by day, music streaming services grow the volume of data on the internet. To help the users to find songs that fit their interests, music recommender systems can be used to filter a large number of songs according to the preference of the user. However, the context in which the users listen to songs must be taken into account, which justifies the usage of context-aware recommender systems. Although there are some works about context-aware music recommender systems, there is a lack of automatic techniques for extracting contextual information for these systems. Thus, the goal of this work is to propose two methods to acquire contextual information (represented by embeddings) for each song, given the sequence of songs that each user has listened to. The first method, called Metadata-Aware, uses tags and genres to enrich the embeddings with additional information. The second method, called Dual Recurrent Neural Network, uses such a network to improve the embeddings generated from long sequences of songs. The embeddings generated by both methods were evaluated with four context-aware music recommender systems in two datasets. The results showed that the embeddings, obtained by our proposals, present better results than the state-of-the-art method proposed in the literature (in some cases with gains of more than 100 %). Finally, the experiments also showed that our second method provides better results than the first one.