OpinionML: An Interpretable Machine Learning Framework for Opinion Mining on Social Networking Sites

Abstract

In recent years, the dramatic increase in the volume of user-generated content on social media, making effective opinion analysis systems more essential than ever. Tasks such as sentiment analysis, aspect extraction, topic modeling, and fake review detection are often addressed separately, although they are closely related. This isolating limits the potential for leveraging shared information. In this paper, we present OpinionML, a machine learning framework designed to unify tasks into a single framework. The framework uses a common feature engineering pipeline that combines several feature types: TF-IDF, lexical, syn- tactic, behavioral, and topic-based. Topic information is extracted using latent Dirichlet allocation (LDA) and augmented to obtain additional contextual cues. In modeling, different problems require different approaches—so we use support vector machines, random forests, and conditional random fields, depending on suitability. The proposed framework evaluated on standard datasets, (SemEval- 2016, Yelp, Amazon, and Sentiment140). The results show that our approach demonstrates competitive performance compared to traditional machine learn- ing methods while remaining interpretable and computationally feasible, and can make opinion analysis systems more effective and flexible in practice.

Keywords