Contrastive Learning (CL) has recently achieved significant progress in the field of recommender systems, as it leverages supervision signals from raw data to mitigate the issue of data sparsity. However, most existing methods rely on random or heuristic data augmentation strategies, which often disrupt the intrinsic structural relationships in the graph and introduce spurious connections, leading to suboptimal representations for recommendation tasks. To address these challenges, we propose a novel graph contrastive learning model based on Large Language Model (LLM), named LLMGCL, which leverages LLM for both embedding augmentation and graph augmentation, and incorporates inter-layer contrastive learning to improve recommendation performance. Specifically, the proposed method first designs an LLM-driven embedding augmentation strategy, where item embeddings pre-trained by LLM are aggregated to construct user embeddings, enabling contrastive views to retain richer semantic relationships. Furthermore, to enhance the structural integrity of the graph while minimizing noise, we develop two targeted graph augmentation strategies powered by LLM: one leverages LLM to identify high-confidence user-item interactions from historical behaviors to refine observed preferences, while the other infers potential but unobserved interactions to reinforce graph connectivity. Finally, we design an inter-layer contrastive learning module that aligns representations across different Graph Neural Network layers, effectively mitigating the over-smoothing effect and enhancing feature discrimination. These combined strategies significantly improve recommendation quality, as demonstrated by extensive experiments on public datasets.

Knowledge Graphs (KGs) have emerged as a critical technique to enhance recommendation performance by modeling complex relationships and semantics within heterogeneous networks. However, it faces issues such as longtail distribution, structural redundancy caused by semantically similar relations, and susceptibility to noise interference, which severely limit the effectiveness of graph-based recommendations. Aiming to tackle the challenges, we propose Relation-aware Contrastive Learning (RACL), a brand-new framework for knowledge-enhanced recommendations. Specifically, relationdriven subgraph construction is employed to cluster the KG into subgraphs with potential semantic associations, addressing the issue of structural redundancy while alleviating the long-tail effect through the integration of relationship types. Besides, we introduce a relation-aware aggregation module to inject relation-specific semantic features from KG into neighborhood propagation, effectively encoding multi-type relational contexts into user and item embeddings. Furthermore, a graph learner is established, which significantly improves the model's robustness in contexts with sparse and noisy data by integrating selfsupervised signals into model training. Comprehensive experiments on two publicly accessible datasets verify that our RACL surpasses the state-of-the-arts in terms of recommendation efficacy.

Graph unlearning involves removing certain elements, such as edges and nodes, from a trained graph neural network. This research has significant practical implications, such as enabling users to enforce data removal requests in compliance with their "right to be forgotten". Additionally, it can help models regain utility by eliminating the impact of poisoned data. Although some graph unlearning algorithms, such as GNNDelete, assign forgotten samples to random negative samples to negate their impact on the model, using random samples is not the optimal choice as it may lead to decreased performance of the model on downstream tasks. Moreover, GNNDelete requires additional space to mark the forgotten samples and their neighborhoods, making this method inconvenient to implement and difficult to continue training. In light of this, we propose a novel graph unlearning framework called Adversarial Example-Based Graph Contrastive Unlearning (AEGCU). AEGCU generates adversarial examples through a straightforward method and utilizes contrastive learning to compare the features of the forgotten samples with those of the adversarial examples, thereby aiming to thoroughly erase edge information. Our experiments demonstrate that AEGCU effectively forgets edges in the graph while maximizing the retention of predictive capabilities. Our model improves performance on both the link prediction as well as the node classification tasks, especially on the node classification task, where we improve performance by approximately 11% over GNNDelete with taking 30% less time on average compared to the GNNDelete method.