基于加权聚合图神经网络的音乐推荐模型

Music Recommendation Model Based on Graph Neural Network with Weighted Aggregation

摘要: 现有基于图神经网络的音乐推荐模型在聚合方式上忽略了节点自身与邻域的权重分配，多数模型只是简单的相加或拼接，这种聚合方式没有充分挖掘节点特征的差异性，针对以上问题，本文提出了基于加权聚合图神经网络的音乐推荐模型MGWA，首先根据用户项目交互的行为记录来构建知识图，然后图神经网络通过逐层聚合邻域信息来丰富项目表示和传播高阶信息，特别的，在聚合阶段，我们对节点自身和其邻域采用门控机制进行权重分配，从而更好地捕捉节点特征表示。此外，门控机制可以根据节点自身及其邻域的特征动态地分配权重，这使得聚合方式更加灵活和可学习，从而更好地适应不同的图结构和节点特征。最后，本文评估了所提出模型在真实数据集上的性能，与常见推荐模型进行对比，实验结果证明了我们的模型在AUC、F1等指标上均取得了提升。

Abstract: The existing music recommendation model based on graph neural network ignores the weight distribution between the node itself and the neighborhood, and most models simply add or splice, which does not fully exploit the differences of node characteristics. To solve the above problems, this paper proposes a music recommendation model (MGWA) based on graph neural network with weighted aggregation. Firstly, a knowledge graph is constructed according to the behavior records of user interaction. Then graph neural network enriches the project representation and spreads high-order information by aggregating neighborhood information layer by layer. Especially, in the aggregation stage, we use gating mechanism to assign weights to nodes themselves and their neighborhoods, so as to better capture the node feature representation. In addition, the gating mechanism can dynamically allocate weights according to the characteristics of nodes themselves and their neighbors, which makes the aggregation method more flexible and learnable, thus better adapting to different graph structures and node characteristics. Finally, this paper evaluates the performance of the proposed model on real data sets, and compares it with the common recommended models. The experimental results prove that our model has achieved improvement in AUC, F1 and other indicators.