Class imbalance in multi-resident activity recognition: an evaluative study on explainability of deep learning...

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Activity recognition in multi-resident homes is challenging due to unevenly distributed actions.
• Deep learning models struggle with class imbalance in sensor data for multi-occupant scenarios.
• Existing methods often focus on single-resident or balanced datasets.

Purpose of the Study:

• To survey class imbalance issues in multi-resident activity recognition.
• To evaluate Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) networks for this task.
• To investigate data-level and algorithmic strategies for mitigating class imbalance and enhancing model explainability.

Main Methods:

• Comprehensive literature survey on class imbalance in activity recognition.
• Experimental evaluation of LSTM and BiLSTM networks on imbalanced smart home datasets.
• Application of data-level and algorithmic techniques to address class imbalance.
• Analysis of model performance and explainability using various metrics.

Main Results:

• Deep learning models, particularly LSTM and BiLSTM, can effectively recognize multi-resident activities despite class imbalance.
• Data-level and algorithmic strategies significantly improve model performance on imbalanced datasets.
• Investigated strategies enhance the transparency and reliability of activity recognition systems.

Conclusions:

• Addressing class imbalance is crucial for accurate multi-resident activity recognition.
• LSTM and BiLSTM networks, with appropriate imbalance handling, offer a viable solution for active and assisted living technologies.
• This research contributes to developing more trustworthy and explainable AI systems for smart home environments.