The core insight behind the work lies in how the model combines two complementary sources of information. Collaborative filtering, the workhorse of modern recommender systems, infers user preferences from patterns of interaction—users who behaved similarly in the past are assumed to share future tastes. Its weakness, however, has always been interpretability: the system learns that a user likes an item but cannot explain why in terms anyone can inspect. Knowledge graphs address this by encoding items, attributes, and their relationships in structured form, so that a movie can be linked to its director, genre, and era, and a product to its brand, category, and compatible accessories. Earlier systems such as Knowledge Graph Convolutional Networks, RippleNet, and the Knowledge Graph Attention Network demonstrated that propagating information across such graphs enriches the representations of both users and items, yielding more accurate and more explainable recommendations.

But graph neural networks, the deep learning architecture typically used to propagate information across knowledge graphs, carry a well-known pathology of their own. When these networks grow deeper, the representations of neighboring nodes converge toward one another, a phenomenon the field calls over-smoothing. In practical terms, after several rounds of message passing, the unique features that distinguish one item from another—the very features a recommender needs to personalize its output—get averaged away. Nodes end up embedded in nearly identical vectors, and the model’s discriminative power collapses. This problem grows worse with depth, which means the naive solution of simply stacking more graph convolution layers to capture longer-range dependencies in the knowledge graph can backfire, degrading rather than improving recommendation quality.

MCLRE’s answer to over-smoothing is its residual enhancement mechanism, an idea borrowed from the residual networks that revolutionized deep computer vision. In classical residual learning, each layer receives not only the transformed output of the previous layer but also a direct shortcut connection carrying the original input forward, ensuring that identity information survives even through many layers of transformation. The Jiangsu Normal University team adapted this principle to graph convolution on knowledge graphs: the node representation produced by the preceding layer is incorporated directly into the graph convolution operation of each subsequent layer. Formally, this means that when the model aggregates messages from an item’s neighbors in the knowledge graph—its attributes, related entities, and users who interacted with it—the aggregation result is combined with the node’s own representation from the previous hop rather than replacing it. Information that would otherwise be diluted or lost during propagation is preserved, and the model can stack multiple convolution layers to capture multi-hop relational structure without the representations collapsing into indistinguishable blobs.

The second major component, multi-level contrastive learning, is designed to strengthen the fusion of the two views of the data: the user–item interaction graph, which captures collaborative filtering signals, and the knowledge graph, which captures semantic structure. Contrastive learning is a self-supervised technique in which a model learns by comparing representations—pulling together views of the same entity while pushing apart representations of different entities, often using the InfoNCE-style objectives popularized in graph self-supervised learning frameworks such as Deep Graph Infomax and adaptive graph augmentation methods. What distinguishes MCLRE is that its contrastive strategy operates at multiple levels and across multiple structures. The model performs contrastive learning not only within a single graph—distinguishing node-level representations against perturbed or augmented views—but also between graphs, aligning the representation of the same item or user as seen through the collaborative filtering graph with its representation as seen through the knowledge graph.

This cross-graph alignment matters because the two views encode genuinely different information. The interaction graph tells the model what users actually did; the knowledge graph tells it why those co-occurrences might make sense in terms of item attributes and relations. By maximizing agreement between the two views of the same node while maintaining discrimination between different nodes, the model learns embeddings in which behavioral similarity and semantic similarity reinforce one another. The authors report that this strategy significantly enhances the fusion of collaborative filtering signals with knowledge graph data, improving both recommendation accuracy and robustness—an important consideration in real-world settings where interaction data is sparse, noisy, or affected by the cold-start problem that has plagued recommender systems since their inception.

The training objective in MCLRE combines the contrastive losses with a Bayesian personalized ranking loss, the standard criterion for implicit-feedback recommendation that encourages the model to score a user’s observed interactions higher than unobserved ones. The embeddings that feed the scoring function are produced by the residual-enhanced graph convolution encoder: user representations are learned from the interaction graph, item representations are iteratively refined by propagating along knowledge graph relations through multiple hops, and the final recommendation score is computed as the inner product between user and item embeddings. Optimization proceeds with the Adam stochastic gradient method, with model parameters initialized using standard schemes for deep feedforward networks.

Experimental validation was carried out on multiple public datasets, comparing MCLRE against state-of-the-art knowledge-aware baselines including knowledge graph attention networks, intent-learning models that trace user intents behind interactions, and multi-level cross-view contrastive approaches. Across all three reported metrics—AUC, the area under the receiver operating characteristic curve, which measures the model’s ability to rank a user’s preferred items above non-preferred ones; F1, the harmonic mean of precision and recall; and Recall@K, the fraction of truly relevant items captured among the top K recommendations—MCLRE achieved the best results. The consistent advantage across datasets and metrics suggests that the gains are not an artifact of any single benchmark or evaluation choice.

Perhaps the most instructive part of the evaluation is the ablation study, in which the researchers systematically removed individual components to measure their contribution. When the residual enhancement mechanism was stripped out, performance dropped, confirming that preserving node identity across convolution layers is not a technical nicety but a pivotal driver of accuracy—especially as the number of propagation layers increases and over-smoothing would otherwise set in. When the multi-level contrastive learning strategy was removed, performance degraded again, demonstrating that the cross-graph and intra-graph alignment objective is what welds the collaborative and semantic views into a coherent representation space. Together, the ablations indicate that the two innovations are complementary rather than redundant: the residual mechanism protects information during propagation, while contrastive learning shapes what that information should encode.

The work arrives amid a surge of interest in knowledge-aware recommendation. Recent efforts in the field have explored hypergraph representation learning combined with transformers, federated knowledge-graph recommenders that preserve user privacy, simplified GNN architectures such as LightKG for efficient recommendation, and curriculum-learning-based knowledge graph embedding. MCLRE’s contribution to this landscape is twofold. First, it offers a clean, architecturally simple remedy to over-smoothing in knowledge-aware graph convolution—one that requires no exotic machinery, only a well-placed residual connection at every layer. Second, it demonstrates that contrastive learning is most effective when applied hierarchically, capturing consistency both within each graph and across the collaborative–semantic boundary, rather than treating the two views as isolated islands of signal.

For the millions of users who encounter recommender systems daily, the practical consequences of such research are subtle but real: fewer irrelevant suggestions, better handling of new users and niche items with sparse interaction histories, and recommendations whose logic can be traced through the knowledge graph to the attributes and relations that motivated them. For the research community, the study reinforces a broader lesson that is emerging across deep learning: architectural safeguards that preserve information, such as residual connections, and self-supervised objectives that exploit structure, such as multi-level contrastive learning, are powerful allies when combined thoughtfully. The authors acknowledge support from the National Natural Science Foundation of China under grant No. 62277030 and the Jiangsu Normal University 2025 Graduate Research and Practical Innovation Program, and they note that all authors reviewed the final manuscript, with correspondence directed to Xia Wang at the School of Artificial Intelligence and Computer Science.