Graph neural networks (GNNs) have recently been widely leveraged for recommender systems based on graph representation learning advances. The current GNN-based models mostly neglect the valuable insights in the negative feedback of users and focus on positive feedback (high ratings). Therefore, negative feedback remains to be exploited to represent the preferences of users, considering useful information in negative feedback concerning the design of GNN-based recommender systems. The present work introduces SiGR to learn node representations within signed user-item interaction graphs. It exploits an embedding module to model the negative and positive preferences of high-order users on signed user-item interaction graphs before the rating graph is partitioned into two bipartite graphs via negative feedback and positive feedback. Then, the effective denoising of negative feedback is facilitated, the contrastive training of the design graph is performed, and non-graph information is extracted using a multilayer perceptron (MLP) for positive feedback. The final interest and disinterest embeddings are ultimately generated using an attention mechanism. The negative preferenceoptimized positive preferences are used for the final recommendations. A multi-task training scheme is also developed for the joint optimization of SiGR parameters. This study contributes to the exploration of negative feedback in recommender systems. It was experimentally demonstrated that SiGR outperforms the current benchmark techniques on three real-world datasets.
