A multi-task heterogeneous graph learning with cross-attention fusion for robust recommendation

Recommender systems based on graph neural networks have demonstrated strong capability in modeling complex user-item interactions; however, their performance is often hindered by data sparsity, cold-start scenarios, and negative interference among heterogeneous learning objectives. In this paper, we propose a multi-task heterogeneous graph representation framework that jointly addresses these challenges at both the data and model levels. At the data level, we enrich the user-item interaction graph from MovieLens by integrating complementary semantic information from the IMDb knowledge graph, resulting in a multi-type heterogeneous graph that captures both behavioral interactions and item-level semantics. At the model level, we design a dual-branch architecture composed of two independently pre-trained, task-oriented graph encoders. The first branch employs a multi-task Graph Convolutional Network (GCN), with a primary focus on recommendation, to effectively model interaction intensity and structural patterns. The second branch utilizes a multi-task Heterogeneous Graph Transformer (HGT) that emphasizes node classification by explicitly capturing diverse node and relation types, thereby learning semantically coherent and discriminative representations. Each branch is optimized using task-weighted objectives to reduce negative task interference and preserve task-specific inductive biases. To integrate the complementary representations learned by the two branches, we introduce a bidirectional cross-attention mechanism that enables adaptive information exchange between interaction-driven and semantic-driven latent spaces. The fused representations are subsequently fine-tuned end-to-end to support recommendation and classification tasks jointly. Extensive experiments demonstrate that the proposed framework achieves more stable learning and superior performance compared to conventional GNN-based recommender models, highlighting the effectiveness of combining heterogeneous graph enrichment, task-oriented multi-task learning, and cross-attention-based feature fusion.