Content-aware Collaborative Music Recommendation Using Pre-trained Neural Networks

“Previous attempts on content-based music recommendation have achieved promising results. van den Oord et al. [13] utilize a neural network to map acoustic features to the song latent factors learned from the weighted matrix factorization [6]. As a result, given a new song that no one has ever listened to, a latent factor can still be predicted from the network and recommendation can be done in the same fashion as with a regular collaborative filtering model.

Our method is very similar to this approach, but we will point out two major differences:

  First, the neural network is used for different purposes. We use it as a content feature extractor, just like LDA in the collaborative topic model. The neural network in [13] maps content directly to the latent factors learned from pure collaborative filtering, and the resulting model is expected to operate similarly to collaborative filtering even when usage data is absent.

  Since the neural network is trained to map content to the latent factors learned from the weighted matrix factorization, the performance of [13] is unlikely to surpass that of the weighted matrix factorization. What we propose in this paper, on the other hand, uses content as an addition to the weighted matrix factorization, in a similar manner as the collaborative topic model described in Section 2.2. As we show in the experiment, we are able to achieve better result than the weighted matrix factorization when we only have limited amount of user feedback.

Other approaches that hybridize content and collaborative models include Yoshii et al. [17], McFee et al. [11], and Wang and Wang [15]. [17] train a three-way probabilistic model that joins user, item, and content by a latent “topic” variable; the model focuses on explicit feedback (user ratings). [11] take a similar approach to [13] and learn a content-based similarity function from collaborative filtering via metric learning. [15] also use a neural network to incorporate music content into the collaborative filtering model. The major difference is that in [15] the output of the neural network is treated as item factor and the neural network is trained to minimize a collaborative-filtering-based loss function. Therefore the content model itself does not have explicit musicological meaning.

3. PROPOSED APPROACH

Adopting the same structure as that of CTR, our system consists of two components: a content model which is based on a pre-trained neural network and a collaborative filtering model based on matrix factorization.”