Advanced Recurrent Neural Network Driven Social Media Sentiment Analysis Using Instagram Reviews

DOI

https://www.doi.org/10.59256/ijsreat.20260602021

Published In

Volume 6, Issue 2 (2026)

Published Online

March-April 2026

Pages

144-152

Article Type

Original Article

Check for updates

This study proposes a deep learning-based framework for performing sentiment analysis on Instagram user reviews, with the objective of extracting meaningful insights from large-scale user-generated textual data. The rapid expansion of social media platforms and mobile applications has led to a significant increase in user feedback in the form of reviews, which serve as an important source for understanding user satisfaction, application performance, and feature effectiveness. However, the unstructured, noisy, and high-volume nature of such data makes manual analysis inefficient and unreliable, thereby necessitating automated and scalable solutions. To address this challenge, a dataset comprising over 32,000 Instagram reviews is collected from platforms such as the Google Play Store and Apple App Store. The dataset captures real-world user interactions and includes informal linguistic patterns such as abbreviations, slang, and emojis. A comprehensive preprocessing pipeline is implemented to enhance data quality, including text cleaning, normalization, tokenization, and the removal of stop words and punctuation. The study employs and compares two sequence-based deep learning architectures: Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM). These models are designed to capture contextual dependencies within textual data, with BiLSTM offering enhanced capability by processing sequences in both forward and backward directions. The textual data is transformed into numerical representations using tokenization, sequence padding, and embedding techniques to facilitate effective model learning. The models are trained using the Adam optimizer and binary cross-entropy loss function, along with regularization techniques such as dropout to improve generalization and prevent overfitting. Experimental results demonstrate that both models achieve strong performance; however, the BiLSTM model outperforms LSTM, achieving an accuracy of 86.15% compared to 85.22%. Additionally, BiLSTM shows improvements in precision, recall, and F1-score, highlighting its effectiveness in capturing bidirectional contextual information, particularly in short and informal text.