Leveraging Multimodal Shapley Values to Address Multimodal Collapse and Improve Fine-Grained E-Commerce Product Classification

Multimodal models can experience multimodal collapse, leading to sub-optimal performance on tasks like fine-grained e-commerce product classification. To address this, we introduce an approach that leverages multimodal Shapley values (MM-SHAP) to quantify the individual contributions of each modality to the model's predictions. By employing weighted stacked ensembles of unimodal and multimodal models, with weights derived from these Shapley values (MM-SHAP), we enhance the overall performance and mitigate the effects of multimodal collapse. Using this approach we improve previous results (F1-score) from 0.67 to 0.79.