Importance-weighted the imbalanced data for C-SVM classifier to human activity recognition

The class imbalance problem is one of the new problems that emerged in activity recognition and that caused suboptimal classification performance. To deal this problem, we propose an efficient way of choosing the suitable regularization parameter C of the Soft-Support Vector Machines (C-SVM) method to perform automatic recognition of activities in a smart home environment. We also discuss how they differ when not considering the weights in C-SVM formulation using cross validation and how it affects their performance. Then, we compare C-SVM with Conditional Random Fields (CRF) considered as the reference method. Our experimental results carried out on three real world imbalanced datasets show that C-SVM based our proposed criterion is capable of solving this class imbalance problem by improving the class accuracy of activity classification compared to other methods.