Interview Data Analysis using Machine Learning Techniques to Predict Personality Traits

In this paper, we analyze the MIT Interview dataset and find the relation between the various prosodic features (such as intensity, pitch, frequency, etc.) and the likelihood of the person getting a virtuous assessment in the interview. These prosodic features help in rating a person on several grounds such as how engaging or excited or friendly the candidate was. We have demonstrated how selecting only a few of the prosodic features can give better prediction results. This was done by selecting the top features using the `recursive feature elimination' technique for five personality traits such as `Engaged', `Excited', `Friendly', `Calm' and `Speaking Rate'. It was found that for traits such as `Engaged' and `Excited', the prosodic features related to intensity play a major role. For personality trait `Friendly', prosodic features like pitch and duration of pause are more relevant. Similarly, for personality trait `Calm', prosodic features related to pitch play a major role and so on. Once the top features were selected, we applied three different regression models with tenfold cross validation to determine the best method for predicting these personality traits. These regression models were evaluated by calculating the negative mean squared error, coefficient of determination, etc. Based on the empirical results, decision tree proved to be the best method for predicting the personality traits based on the selected prosodic features.