AURA: Adaptive Unified Routing Architecture with Cross-Attention and Expert Balancing for Multimodal Product Classification

Abstract

Accurate product classification is a critical task for modern e-commerce platforms, directly affecting user experience, search relevance, and revenue optimization. However, the rapidly growing number of products and categories — coupled with the inherent noise and ambiguity in real-world data — presents significant challenges for traditional models. In this paper, we propose AURA, a novel and efficient multimodal deep learning architecture that integrates product images and textual descriptions through a cross-modal fusion module, followed by a Mixture-of-Experts (MoE) classifier. Our fusion module captures fine-grained semantic interactions between modalities via bidirectional cross-attention, while the MoE component enables instance-level specialization by dynamically routing inputs to the most suitable expert networks. We further introduce a balanced training objective to ensure both classification performance and expert diversity. Extensive experiments on the large-scale GLAMI-1M dataset demonstrate that AURA significantly outperforms several strong multimodal baselines in accuracy. Our results highlight the effectiveness of combining lightweight expert networks with adaptive routing in addressing the scale and diversity of real-world e-commerce classification tasks.