Dynamic Perspective Defined Graph Attention Network Towards Sequential Music Recommendation

Recommendation system is considered as effective approach to provide personalized recommendation to the user on the various data contents. Especially music recommendation is becoming crucial field for digital communication as it influenced and interacted by many users. Hence it becomes important to discover relationship between music data and user. However due to increase of digital contents streaming around the world, it leads to data complexity and data overloading challenges towards increasing user experiences. In order to mitigate those challenges, a new deep learning architecture entitled as Dynamic Perspective Defined Graph Attention Network towards Sequential Music Recommendation is proposed in this work. It associates graph neural network and attention mechanism to examine the diversity of user behaviours on their interaction with music contents. Initially user information and music information is converted into the feature vector which extract the attribute based on the collaboration information's from user collective behaviours. Those feature vector is used to exploit the high order relationship on processing the information on basis of the user-music graph in form of bipartite graph. Especially user graph facilitates in dynamic user preference modeling using regularization based approaches to capture long term and short term preference along latent factor of the user. Further Graph Neural Network discovers the complex sequential patterns on utilizing the user preference model to provide rich information. Meanwhile these models use weight functions to learn weight of each user through graph propagation such that attention weight of cascaded propagation provides importance of the target music to user on basis of higher order connectivity. In particular, embedding layers uses Self-attention mechanism to discover the different priorities of the user to the music items on the long term and short term structures of the user to generate the high order relationships for recommendations. Experimental analysis of the proposed model has been conducted on the real dataset and performance analysis of the proposed model achieves better prediction accuracy compared to the conventional approaches.