Deep Learning Based Sentiment Analysis on Instagram Insights of Consumer Behavior for Improving Business Decision Making

Abstract

The increasing use of social media platforms such as Instagram has made them a significant source of consumer insights for businesses, highlighting the importance of automated sentiment analysis. This study aims to address the challenge of accurately classifying consumer sentiments in Instagram posts, where informal language, slang, and sarcasm often reduce the effectiveness of traditional models. To overcome this gap, two deep learning approaches were employed: a Bidirectional Long Short-Term Memory (BiLSTM) network as a classical recurrent baseline and transformer-based architectures (BERT and DistilBERT) as state-of-the-art models. A dataset of 184,010 Instagram posts was preprocessed, tokenized, and mapped into positive and negative sentiments, and the models were trained and evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices. The results demonstrated that BERT achieved the highest performance with an accuracy of 0.91 and an F1-score of 0.91, outperforming BiLSTM (accuracy 0.87, F1-score 0.86), while DistilBERT provided a competitive balance between accuracy (0.89) and efficiency. These findings confirm that transformer-based models, particularly BERT, are better suited for capturing nuanced sentiments in social media text. The study concludes that models can provide actionable insights into consumer behavior, enabling businesses to enhance brand monitoring and customer engagement.