Predicting latent attributes by extracting lexical and sociolinguistics features from user tweets

Abstract

Twitter user profile information is very useful for various fields such as marketing, HRD, advertising, and personalization. Since user profile provided by Twitter is very limited, some latent attributes such as gender, age, work, or interest should be predicted. In this paper, we aim to predict those four latent attributes using her/his tweet and bio data by employing machine learning techniques. We conduct experiments in order to find the best algorithm, weighting method, minimal frequency number, preprocess, for each latent attribute we predicts. We also compare the accuracy of lexical feature and sociolinguistic feature classification models. Our experiment shows that SVM is the best performer and lexical feature models perform better than sociolinguistic feature models.