Comparison of Text Mining Feature Extraction Methods Using Moderated vs Non-Moderated Blogs
Abu Saleh Md Tayeen, Saleem Masadeh, Abderrahmen Mtibaa, Satyajayant Misra, Moumita Choudhury · 2019
Online social media is being widely used by social scientists to study human behavior. Researchers have explored different feature extraction (FE) and classification techniques to perform sentiment analysis, topic identification, etc. Most studies tend to evaluate FE and classification methods using only one particular class of datasets---well-defined with little/no noise or with well-defined noise. For instance, when the datasets under study have different noise characteristics, various FE and/or classification methods may fail to identify a given topic. In this paper, we fill this gap by quantitatively comparing multiple FE methods and classifiers using three different datasets (two moderator-controlled blogs and one single-authored personal blogs) related to Autism Spectrum Disorder (ASD). Our result shows that no particular combination of FE and classifier is the best overall, but choosing the right ones can improve accuracy by over 30%.