A novel fine-tuning method for long short-term memory models on edge device to adapt to dynamic user habits in smart homes

Static machine learning models in smart home automation may struggle to adapt to dynamic user habits when deployed on resource-constrained edge devices. Although long short-term memory networks can capture temporal dependencies in user behavior, models trained on fixed historical data may lose accuracy as user patterns evolve. Periodically retraining the complete model is computationally demanding on an edge device. To address this problem, this study proposes an incremental fine-tuning method that updates a pre-trained long short-term memory model using small batches of newly labeled data derived from user feedback. Evaluation in a real-world smart home system comprised an offline chronological replay of the initial data stream, a subsequent 21-day continuous-operation stream, and separate model-update cost measurements on a Raspberry Pi 4B. In the initial replay, the incrementally fine-tuned model achieved the highest F1-score among the four evaluated models, reaching 0.9401 for air-conditioner prediction and 0.9430 for light prediction. During the subsequent 21-day air-conditioner evaluation, it maintained the highest daily F1-score after the initial transient period. Compared with full retraining, incremental fine-tuning reduced average processor utilization from 32.00% to 8.00% and peak memory use from 13.55 megabytes to 7.25 megabytes. Within the evaluated single-participant setting, these results indicate the potential of lightweight incremental model updates for adaptive smart home automation on resource-constrained edge devices; evaluation across more diverse users, prediction tasks, residential environments, and edge hardware is still required.