A Study on Product Recommendation Based on Dynamically Adjusted Multimodal Weights

Product recommendation can provide users with product suggestions that meet their potential needs and interests, saving users' search time and energy. However, traditional recommendation methods usually rely only on structured product data, which makes it difficult to fully characterize the complex relationship between user preferences and product features. To this end, this paper proposes a study on product recommendation based on dynamically adjusted multimodal weights, which weighted fuses structured product data and unstructured text reviews to predict users' willingness to recommend purchased products. The model uses a dual-channel neural network (MLP+BERT) to extract features of product data and text reviews, and then uses MLP to achieve feature fusion and recommendation prediction. During the model training process, a confidence memory module is introduced to track the long-term performance of each modality on different samples, and the initial weights of different modalities are calculated based on historical confidence. Then, the SHAP feedback mechanism is used to analyze the impact of each modality on the decision, and the fusion weight of each modality is dynamically adjusted based on confidence. The comparative experimental results show that this method performs better than traditional methods on product recommendation, improves the accuracy and stability of the model, and can better recommend high-satisfaction products to consumers.