Hybrid (G- Chi_square) Feature Selection for Sentiments Analysis of Social Data Using Deep Recurrent Neural Networks (HDRNN)

Abstract

Mobile-supported cloud-computing enables harnessing of data relating to natural people’s social interactions. These data provide valuable sources of intelligence about various domain-specific behaviors, and are useful for the detection, modeling, and analysis of such behaviors. This paper proposes a hybrid feature-selection framework that combines Chi-Square and Gini impurity criteria to extract the most informative text features, reducing dimensionality while preserving sentiment-discriminative power. The approach then applies Deep Recurrent Neural Networks (DeepRNNs) to detect feature patterns associated with sentimental behaviors as subsets of documents/records in categorical datasets. It implements a supervised learning technique that detects orderly occurrences of vector arrays to build appropriate data representations that can be inferred as context-aware text models. Method implementation and experiments on case studies test the algorithm using text-based social media inputs that contain context-preserving word features. The experiments reveal the role of investigating the hybrid (G-Chi-square) to enhance DRNN accuracy. Performance results indicate that our HDRNN approach can be more effective than related approaches in data mining and sentiment analysis.