Multi-graph Fusion Cross-model Contrastive Learning for Recommendation

Shengjun Ma, Yuhai Zhao, Fenglong Ma, Baoyin Liu, Zhengkui Wang, Wen ShanAAAI 2026 · DOI

TL;DR

MFCCL (Multi-graph Fusion Cross-model Contrastive Learning) bridges the semantic gap between user-item interaction data and external knowledge graphs. By synthesizing auxiliary "interest" and "association" graphs and employing a data-augmentation-free contrastive learning strategy, it achieves state-of-the-art performance across major recommendation benchmarks.

What Problem Does This Paper Address?

Modern recommendation systems often supplement user-item interaction data (the Bipartite Graph or BG) with Knowledge Graphs (KG) containing item attributes (e.g., a movie's genre or director). However, two major hurdles remain:

  1. Semantic Misalignment: The BG is sparse and singular (user-item), while the KG is dense and diverse (head-relation-tail). Simply concatenating these representations often leads to "misalignment," where the model struggles to reconcile the two different views of the same item.
  2. Information Noise: KGs contain many relations that are irrelevant to a specific user's preference. Indiscriminately incorporating all KG relations introduces noise that can degrade performance.
  3. Augmentation Noise: Many current models use Contrastive Learning (CL) to handle data sparsity, but they rely on random graph perturbations (like dropping edges). This "manual" augmentation can accidentally destroy critical structural information.

Methods

The authors propose MFCCL, which operates through four main stages:

  • Multi-graph Generation: Instead of using the raw KG directly, they construct two specialized auxiliary graphs:
  • Item-association Graph (IaG): Captures how often items are co-purchased.
  • User-interest Graph (UiG): Connects users directly to item attributes from the KG, revealing specific feature preferences.
  • Virtual Relation Mechanism: To solve the KG noise problem, they replace complex real-world relations with a fixed set of learnable virtual relations. This distills the KG into a more manageable and relevant format for the recommendation task.
  • Multi-graph Representation Fusion: They use a two-tower GNN architecture to learn embeddings from the BG, KG, IaG, and UiG in parallel, using attention mechanisms to fuse these signals into "Multi-graph" embeddings ($e_{mg}$).
  • Cross-model Contrastive Learning (CMC): This is the "secret sauce." Instead of creating fake "augmented" views of a graph, they treat the output of a standard Collaborative Filtering (CF) module and the output of their Multi-graph module as two different views of the same user/item. They use the InfoNCE loss to maximize the similarity between these two views, effectively aligning semantic knowledge with structural collaborative signals without introducing augmentation noise.

Datasets

The model was evaluated on three widely used public benchmarks:

  • Alibaba-iFashion: A large-scale dataset for fashion recommendation.
  • Yelp2018: A dataset of business reviews and attributes.
  • Last-FM: A music recommendation dataset rich in artist and genre metadata.

Key Insights and Ideas

  • Alignment is Better than Augmentation: The most significant finding is that the "cross-model" contrastive approach—aligning a semantic-heavy model with a structure-heavy model—is more robust than traditional data augmentation (like edge dropping).
  • Virtual Relations as a Filter: By using a bottleneck of "virtual relations," the model learns to ignore the "messiness" of a raw KG and focuses only on the relational patterns that actually drive user behavior.
  • The Power of Latent Associations: Explicitly building the Item-association Graph (IaG) helps the model capture "beer-diaper" style relationships that are often buried in the raw interaction data.

Key Results

MFCCL consistently outperformed all baselines, including competitive KG-aware models like KGRec and KGIN.

  • Accuracy Gains: On the Last-FM dataset, MFCCL improved Recall@20 by 4.81% and NDCG@20 by 15.33% over the strongest baseline.
  • Ablation Criticality: The Cross-model Contrastive (CMC) module proved essential; removing it caused performance to drop by 13.85% to 22.29% across different metrics, proving that aligning the two model views is the primary driver of success.
  • Efficiency: Despite the multi-graph approach, the complexity remains comparable to existing SOTA models because the auxiliary graphs are constructed offline.

Limitations and Future Work

  • Static Graph Construction: The User-interest and Item-association graphs are constructed as a pre-processing step. Dynamic graph updates during training could potentially capture evolving user interests more accurately.
  • LLM Integration: The authors suggest that future work could involve using Large Language Models (LLMs) to generate even richer side information or to better interpret the semantic relations within the KG, potentially replacing the virtual relation mechanism with more "intelligent" feature extraction.

Summary updated

Abstract

Knowledge Graph (KG)-supported Graph Neural Network models are becoming crucial in recommendation systems due to their ability to mitigate the data sparsity challenge. However, these models remain suboptimal because they overlook the representation differences between the inherent user-item Bipartite Graph (BG) and the external head-relation-tail KG, leading to semantic misalignment. Moreover, they indiscriminately incorporate various types of relations from the KG, which may introduce noise information into the model, ultimately degrading recommendation performance. To address these challenges, we propose an end-to-end model named Multi-graph Fusion Cross-model Contrastive Learning (MFCCL). To uncover users' interest in items and explore the associations between items, we first construct a user-interest graph by integrating information from both the BG and KG, and an item-association graph derived from the BG. We devise a multi-graph representation learning module that incorporates rich semantics into user and item representations in parallel. Simultaneously, a classical collaborative filtering module is introduced to fully leverage user-item collaborative signals. Additionally, we design a novel free data-augmentation cross-model contrastive learning to facilitate the exchange of complementary information between different models. Empirical evaluations on three widely used benchmarks demonstrate that our MFCCL method achieved significant improvements over the baselines.
BibTeX
@inproceedings{ma2026multi,
  author = {Shengjun Ma and Yuhai Zhao and Fenglong Ma and Baoyin Liu and Zhengkui Wang and Wen Shan},
  title = {{Multi-graph Fusion Cross-model Contrastive Learning for Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i18.38580},
}
Figure