Adaptive Multi-modal Fusion and Noise-robust Graph for Product Recommendation

Product recommendation systems must efficiently process multi-modal information to enhance recommendation quality. However, existing methods often lack sufficient adaptability in multi-modal information fusion and overlook the impact of feature noise, leading to suboptimal model performance in complex environments. To address these issues, this paper proposes a product recommendation method based on Adaptive Multi-modal Fusion and Noise-robust Graph Construction (AMFNG). Firstly, we introduce a multi-head self-attention mechanism to enable the model to adaptively learn fusion weights for visual and textual features, thereby improving representation capability. Additionally, a feedforward neural network and residual connections are incorporated into the fusion process to enhance the model’s nonlinear modeling capacity. Secondly, we design a noise-robust graph construction method based on Gaussian noise model, which introduces noise perturbations in the feature space, allowing the model to better adapt to noise and effectively capture latent data relationships. Finally, extensive experiments on three public datasets validate the effectiveness of the proposed method.