Abstract: As a classic algorithm in recommender systems, two-tower neural networks have been widely used in large-scale information retrieval tasks. But because the two neural networks are independent of each other, the neural network cannot learn the interaction information between users and items. In addition, due to the lack of learning ability for graph connection information, its retrieval range is narrower than that of graph learning algorithms. To address the above problems, a new algorithm TSN(Two-stage siamese graph convolutional neural network recommendation algorithm) is proposed. First, build a heterogeneous graph based on user behavior; then, design a graph convolutional siamese network between the two-tower neural networks, so as to achieve information interaction while learning the connection information of the heterogeneous graph; finally, by designing a special structure of two-stage information sharing mechanism enables the neural networks on the user side and the item side to transmit information dynamically and bidirectionally during the training process, and effectively avoids neural network cascading. After comparative experiments based on MovieLens and Douban movie datasets, the average performance of HR@N, NDCG@N, and MRR@N is improved by 4.70% and 13.28% respectively compared with the optimal benchmark algorithm DAT. The results show that this method can alleviate the problem of lack of information interaction in the twin-tower neural network. Compared with the previous algorithm, the recommendation performance is significantly improved.

Key words: recommendation system, two-tower model, siamese network, deep learning, convolutional neural network

摘要： 作为推荐系统中的经典算法，双塔型神经网络已被广泛应用在大规模信息召回任务中。 但是由于两个神经网络之间相互独立，神经网络无法学习用户和商品之间的交互信息。此外，由于缺乏对图连接信息的学习能力，其召回范围较图学习算法更窄。针对上述问题，提出了一种二阶段孪生卷积神经网络推荐算法TSN(Two-stage siamese graph convolutional neural network recommendation algorithm)。首先，以用户行为构建异质图；然后，在双塔型神经网络之间设计图卷积孪生网络，从而实现在学习异质图连接信息的同时进行信息交互；最后，通过设计特殊结构的二阶段孪生信息共享机制，使得用户侧和商品侧的神经网络在训练过程中能够动态地、双向地传输信息，且有效避免神经网络串联。经过基于MovieLens和豆瓣电影数据集的对比实验，得到由HR@N、NDCG@N、MRR@N构成的平均性能相较于最优基准算法DAT分别提升了4.70%和13.28%。结果表明，该方法能够缓解双塔型神经网络缺乏信息交互的问题，相较于之前的算法，推荐性能提升显著。

关键词: 推荐系统, 双塔模型, 孪生网络, 深度学习, 卷积神经网络