Analysis of sequential physiology data with weighted naive Bayes

In this project, I describe how I address the ICML 2004 Physiological Data Modeling Contest. For the gender prediction task, I compressed the large entry-based dataset to small session-based dataset and manually devised 90 features using a histogram method. Weighted naive Bayes (WNB) which is an extension of naive Bayes was applied and Markov Chain Monte Carlo was combined to solve the weight updating problem. The performance between the WNB and naive Bayes was assessed. Also explicit feature selection based on fisher’s discriminant ratio was applied for the newly devised features and the contribution for the subset of features on the accuracy was assessed. My experiment showed that WNB trained to produce much higher test accuracy up to 89.2% compared with 67.5% that naive Bayes can achieve.