A Smoothed Naive Bayes-Based Classifier for Activity Recognition

Abstract A number of classifiers have been proposed by the researchers for activity recognition using binary and ubiquitous sensors. Many researchers have shown that the hidden Markov model (HMM) and the conditional random field (CRF)-based activity classifiers work well to classify activities in comparison with the widely used naüve Bayes-based activity classifier. However, it would not be an exact verdict if a naüve Bayes-based activity classifier is properly smoothed. Parameter estimation plays the central role in the performance of a naüve Bayes activity classifier. Data sparsity puts substantial challenges in parameter estimation because the sizes of the real-life activity datasets are relatively small. The distribution of the sensors may not be even among the activity classes. Additionally, some of the sensors would appear during testing but would not appear while training. This is called zero-frequency problems which assign zero probability of a sensor for a given activity. To prevent such estimation problems, we propose two smoothing techniques for adjusting the maximum likelihood to produce more precise probability of a sensor given an activity. We performed three experiments using three real-life activity datasets. It is observed that our proposed mechanism yields significant improvement in the accuracy of activity classification in comparison with its existing counterparts. We achieved the class accuracy ranging between 63% and 83%.