Розробка методу пояснюваності мультимодальних нейронних моделей у задачах класифікації неоднорідних даних

Abstract

This research focuses on enhancing the explainability of multimodal neural networks for classification of heterogeneous data. We propose a method to quantify the contribution of each modality and individual features to the final prediction. The approach combines modality exclusion analysis with gradient based feature importance assessment. It generates structured explanations at both modality and feature levels without modifying the model architecture. The method is applicable to diverse multimodal neural architectures and integration strategies. Experimental evaluation demonstrates its effectiveness in improving transparency and interpretability of complex classification systems.