Impact of transformed features in automated survey coding

Abstract

Survey coding is a process of transforming respondents' responses or description into a code in the process of data analysis. This is an expensive task and this is the reason for social scientists or other professionals in charge of designing and administering surveys tend to avoid the inclusion of many open-ended questions in their surveys. They tend to rely more on the less expensive multiple-choice questions, which by definition do not require a coding phase. However multiple-choice questions strictly limit the respondents’ possible answers. This study aims at automating the survey coding process using transformed features. Five intelligent coders were developed using k Nearest Neighbor algorithm, Support Vector Machine with linear function, Support Vector Machine with RBF function and Support Vector Machine with polynomial function. Different response features were applied to improve the coding performance. Techniques that were applied to origin response features include: Relative Frequency, Power Transformation, Relative Frequency Power Transformation and Term Frequency Weighted by Inverse Document Frequency. Furthermore the study proposed new features including: Normalized Relative Frequency, Normalized Relative Frequency with Power Transformation and Normalized Relative Frequency with Term Frequency Weighted by Inverse Document Frequency. The micro-averaged F-measure was used to evaluate the performance of each automated coder. Among all machine learning techniques used Support Vector Machine polynomial was the best when implemented with transformed features.