Cross-Modal Attention Mechanisms for Inventory Optimization: Fusing IoT Sensor Data, News Sentiment, and Geospatial Information

Abstract

Modern supply chain management faces unprecedented challenges in maintaining optimal inventory levels amid volatile market conditions, disruption risks, and rapidly changing consumer demands. This paper proposes a novel cross-modal attention framework that integrates Internet of Things (IoT) sensor data, news sentiment analysis, and geospatial information to enhance inventory optimization decisions. The proposed architecture leverages a transformer-based encoder-decoder structure with multi-head attention mechanisms to dynamically weight and fuse heterogeneous data sources, enabling real-time adaptive inventory management. Through the implementation of global attention layers that compute context vectors across all input modalities, the framework captures complex inter-modal dependencies that traditional single-modality approaches fail to represent. Experimental validation demonstrates that the cross-modal attention approach achieves significant improvements in demand forecasting accuracy, reducing mean absolute percentage error by 18.7% compared to conventional methods, while simultaneously improving service levels from 92.3% to 96.8%. Visualization of learned attention alignment matrices reveals interpretable patterns where the model dynamically focuses on relevant information sources based on market conditions. The integration of news sentiment provides early warning signals for demand fluctuations, IoT sensors enable granular monitoring of inventory conditions, and geospatial analysis optimizes distribution network configurations. This work contributes to the emerging field of multimodal fusion in supply chain analytics by demonstrating how attention mechanisms can effectively integrate diverse data modalities for superior decision-making.