SenseMax: IoT-Based User–Device Interaction Prediction With Complementary Masked Modeling and Contrastive Learning

User–device interaction prediction for smart homes aims to accurately predict the next interaction between the user and the Internet of Things (IoT) devices. By analyzing the user’s historical interaction sequence, the system can self-adjust and provide personalized services, thereby enhancing automation and intelligence in smart home environments. However, efficiently capturing temporal dependencies while addressing data limitations and class imbalance remains a challenge. To address these issues, this article proposes SenseMax, which organically combines masked event–time modeling (METM) and event–time contrastive learning (ETCL) to provide targeted guidance for model training from both the feature and output perspectives, without relying on additional prior knowledge or external data. METM is designed to enable the model to jointly learn event–time associations and improve sample diversity through a complementary masking strategy that simultaneously masks both events and corresponding timestamps. As a complement to METM, ETCL is proposed to leverage the synergistic relationship between event sequences and their corresponding time series, providing task-oriented supervision to optimize sequence representations. Extensive experiments are conducted on publicly available real-world datasets from three different countries. The results show that SenseMax outperforms the state-of-the-art model, improving ACC@1 by 10.48%, 12.43%, and 7.36%, and enhancing Macro-<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$F1$ </tex-math></inline-formula> by 0.4066, 0.4559, and 0.3300 on three datasets, demonstrating its ability to capture complex daily behavior patterns in different regions.