This paper describes the approach of the D2KLab team to the RecSys Challenge 2018 that focuses on the task of playlist completion. We propose an ensemble strategy of different recurrent neural networks leveraging pre-trained embeddings representing tracks, artists, albums, and titles as inputs. We also use lyrics from which we extract semantic and stylistic features that we fed into the network for the creative track. The RNN learns a probabilistic model from the sequences of items in the playlist, which is then used to predict the most likely tracks to be added to the playlist. Concerning the playlists without tracks, we implemented a fall-back strategy called Title2Rec that generates recommendations using only the playlist title. We optimized the RNN, Title2Rec, and the ensemble approach on a validation set, tuning hyper-parameters such as the optimizer algorithm, the learning rate, and the generation strategy. This approach is effective in predicting tracks for a playlist and flexible to include diverse types of inputs, but it is also computationally demanding in the training phase.
An ensemble approach of recurrent neural networks using pre-trained embeddings for playlist completion
RECSYS 2018, 12th ACM Conference on Recommender Systems, 2-7 October 2018, Vancouver, BC, Canada
© ACM, 2018. This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in RECSYS 2018, 12th ACM Conference on Recommender Systems, 2-7 October 2018, Vancouver, BC, Canada http://dx.doi.org/10.1145/3267471.3267484
PERMALINK : https://www.eurecom.fr/publication/5690