Research On Attentive Autoencoder For Content-aware Music Recommendation

With the rapid development of mobile networks,users can access to online resources anytime and anywhere,but the overload information makes it difficult for users to select the most preferred music out of thousands of candidates.Nowadays,music recommendation has become a popular research area.With the tremendous increase of online music resources and users,personalized music recommendation still faces the following three challenges: first,user-music interaction data contains rich information,which is essential for predicting users’ preferences.However,the existing methods cannot model the complex non-linear relations in user behavior records with shallow linear model;second,user’s behavior can be affected by multiple factors,thus it is necessary to learn the fine-grained relations between user-user and user-music for predicting user preferences,which has not be taken into account by existing methods;third,audio features represent characteristics of music,which can be integrated for alleviating the problem of data sparsity.However,it is difficult to effectively combine the heterogeneous audio content.To address the aforementioned problems,the major contributions of this paper are as follows:(1)To overcome the limitation of conventional recommendation architectures,and capture the non-linear relations in user-music interaction data,we design the a stacked autoencoder as our basic model.With symmetrical structure and deep neural network,it can capture the non-linear and non-trivial information in user behavior data,and enables more abstract data representations in the latent space.In this way,we can model users’ preferences more accurately.(2)To fully leverage the complicated and fine-grained information in user-music interaction data,we propose a novel model called Attentive Auto-Encoder for Music Recommendation to model users’ preferences more accurately,which improves the vanilla stack autoencoder model with its encoder and decoder.In the encoder,we design a hierarchical attention module to learn the relations between user-user and user-music,and obtain a more accurate music hidden representation.In the decoder,we design a neighbor-aware module to model the influence of behavioral users exerted on unbehavioral users,which makes un-behavioral users who are more similar to behavioral users obtain higher prediction scores.(3)To alleviate the problem of data sparsity,music content is integrated from two aspects based on the Attentive Auto-Encoder for Music Recommendation model.On the one hand,we embed audio features to obtain content-based music hidden representation,and integrate them with user behavior-based music representation to obtain a more comprehensive understanding of music;On the other hand,we calculate similar music sets by content-aware clustering,and obtain the similar music hidden representations of target music.In the decoder,by modeling users’ preferences on similar music of the target music,the users’ preferences on target music can be indirectly reflected.Finally,we conduct extensive experiments on real-world dataset #nowplaying-RS,which includes comparative experiments and ablation experiments.In the comparative experiment,we compare the proposed method with many state-of-the-art methods,and the results demonstrate the effectiveness of our proposed model.In the ablation experiment,we carry out experiments for each component of the proposed model,and the experimental results verify that every component improves the performance of our model.