Fashioning data - A social media perspective on fast fashion brands

Abstract

In this paper, we study the performance of N-gram language models on classification tasks such as sentiment analysis and spam detection and evaluate the effect of prior probability estimates on the results. Our data is in the form of public online posts pertaining to fast fashion brands, from different social media channels (Twitter and Facebook). We propose a novel ensemble model based on the combination of different N-grams in order to deal with the heteroskedastic nature of data collected from these social media channels. This has been further extended to increase the efficacy of the classification results.