ACMBP: An Adaptive Graph-Based Meta-Learning Framework for Cross-Modal Behavior Prediction from IoT and Social Media Streams

Abstract

The growth of Internet of Things (IoT) devices and social media platforms generates large volumes of multimodal behavioral data that present both great opportunities and challenges to cross-domain behavior prediction. The current models do not deal with time misalignment, distributional drift, and personalization when only a limited amount of data is available. To address these issues, Adaptive Cross-Modal Behavioral Prediction (ACMBP), introduced in this study, is a new framework that merges the streams of IoT activity and social media characteristics via adaptive graph networks. ACMBP combines temporal-semantic offset attention to align activities by the IoT with social actions, drift-conscious dynamic graph rewiring to adapt to changing user relations, hierarchical cross-domain transfer to adapt across different platforms, and few-shot personalization with meta-learning and behavioral prototypes. ACMBP possibility in large-scale heterogeneous digital ecosystems is compared to a fused IoT social dataset (wearable sensors + Twitter) and two public multimodal benchmarks against 10 strong baselines (LSTM, GRU, BERT, GCN, GAT, HAN, GraphSAGE, DCRNN, DeepFM, Multimodal Transformer). ACMBP has been tested on a custom IoT–Social Fusion dataset (1,200 users, 90-day collection, 2.3M sensor records, 850K Twitter posts) and two public benchmarks (PAMAP2, USC-HAD). The results of all improvements are statistically significant (Wilcoxon signed-rank test, p < 0.001). ACMBP attains Behavioral Transition Accuracy (BTA) of 91.4%, Temporal Offset Prediction Error (TOPE) of 1.2 hours, Drift Detection Precision (DDP) of 84.7%, Cross-Platform Transfer Efficiency (CPTE) of 82.3%, and Few-Shot Adaptation only with five samples, which are improvements by 7.3%, 32%, 11.8%, 13.4%, and 50% respectively over the best baselines. Each module’s contribution is supported by ablation studies, and the results of scalability experiments indicate stable performance in large-scale heterogeneous IoT-social environments.