Improved SMOTE Algorithm to Deal with Imbalanced Activity Classes in Smart Homes

Performance of resident-activity-recognition systems is an important measure in the evaluation of smart homes performance. An imbalanced distribution of activity classes, however, severely degrades this performance. Traditional approaches towards realization of activity recognition focus on the improvement of recognition algorithms rather than imbalanced-data adjusting. Even state-of-the-art recognition algorithms have been limited to exclusively improving activity-recognition performance. The proposed study focuses on imbalanced-data adjusting and presents an improved Synthetic Minority Oversampling Technique (SMOTE) algorithm to address issues concerning imbalanced activity classes. Instead of linear interpolation, the proposed algorithm uses the Euclidean distance of each minor activity to adjust the distribution of activity classes, thereby generating new synthetic minority activities in the neighborhood of remaining minority-class examples. Two public datasets were utilized in this study to validate the improved SMOTE algorithm. Results demonstrate that the proposed approach favorably outperforms traditional SMOTE algorithms.