Hierarchical Deep Learning Framework for Daily Living Activity Recognition in Smart Homes: Addressing Class Imbalance Through Dual-Path Feature Fusion and Focal Loss Optimization

The ability to recognize users’ activities within their homes is essential for enabling assisted living and proactive health monitoring through sensor-based systems. However, in real-world smart home deployments, activity distributions are highly imbalanced: routine daily activities dominate the data, while critical yet infrequent activities remain under-represented. This imbalance significantly limits the generalization capability of conventional machine learning models. In this paper, we propose a deep learning architecture specifically designed to address class imbalance in streaming sensor data. The proposed framework employs a dual-path feature extraction mechanism that integrates hand-crafted statistical features (HCF) with high-level features (HLF) learned using Convolutional Neural Networks (CNNs). The resulting multimodal feature representations are then processed by an ensemble of temporal recurrent architectures, including (i) unidirectional Long Short-Term Memory (LSTM), (ii) Bidirectional LSTM (BiLSTM), and (iii) cascade LSTM layers. To further mitigate the effects of class imbalance, we introduce a specialized focal loss function, denoted as LDLA, which dynamically down-weights the contribution of majority classes during training. Extensive experimental evaluations conducted on five benchmark datasets from the Center for Advanced Sensors and Autonomous Systems (CASAS) demonstrate that the BiLSTM model combined with the proposed focal loss consistently outperforms traditional classifiers. Specifically, it achieves an improvement of 17% to 32% in overall accuracy while maintaining high precision and recall for minority activity classes.