DAS-GNN: Denoising autoencoder integrated with self-supervised learning in graph neural network-based recommendations

To enhance the recommendation performance, session-based recommendations typically model based on graph neural networks (GNN). These models use the most recently clicked item as the user's short-term interest, as well as the query vector in the attention mechanism. Based on it, the attention score is calculated with the remaining items to obtain the user's long-term interest. However, the obtained representation of long-term interest is one-sided. Furthermore, unlike other recommendation technology, such as collaborative filtering that includes the user's entire history information, the session-based recommendation is more vulnerable to data sparsity. Existing models primarily make predictions based on observable user-item interactions and ignore items not interacted with by users. To address the aforementioned issues, we propose the denoising autoencoder integrated with self-supervised learning (SSL) in graph neural networks (DAS-GNN). In DAS-GNN, the query extraction module based on denoising autoencoder can mine multiple user interests and assist long-term interest to express user needs more comprehensively. We propose an effective way of dividing positive and negative samples in the SSL module and use adaptive thresholds to mine negative hard samples, thereby improving training efficiency and alleviating data sparsity. Extensive experiments demonstrate that the proposed DAS-GNN outperforms state-of-the-art models on four benchmarks. The source code is available at: https://github.com/daijiuqian/DAS-GNN.