Decoding the black box: LIME-assisted understanding of Convolutional Neural Network (CNN) in classification of social media tweets

• Published in: Social network analysis and mining 2024-07, Vol.14 (1), p.133
• Main Authors: Mazhar, Kashif, Dwivedi, Pragya
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial intelligence; Artificial neural networks; Classification; Complexity; Content analysis; Criminal investigations; Decision making; Decision trees; Decoding; Digital media; Evidence; Explainable artificial intelligence; Hate speech; Language; Literature reviews; Machine learning; Misinformation; Network reliability; Neural networks; Performance evaluation; Profanity; Public opinion; Reliability; Sentiment analysis; Social media; Social networks; Transparency; User behavior
• ISSN: 1869-5450; 1869-5469
• DOI: 10.1007/s13278-024-01297-8

The rise of social media has brought both opportunities and challenges to the digital age, including the proliferation of online trolls that have spread misinformation, hates, and disruptions. An automated classification system is crucial to mitigate the impact of trolls. This paper presents an innovative approach for classifying social media tweets into troll and non-troll categories using a machine learning (ML) approach and CNN. We also employed explainable artificial intelligence (XAI) to address the inherent opacity and complexity of the CNN model. This approach allowed us to provide a comprehensive explanation of the model’s behavior. We have achieved Accuracy = 91.45% with CNN2 model. The best results using ML methods were achieved by random forest classifier model, Accuracy = 86.57%. To enhance our trust in the CNN model, we leveraged the local interpretable model-agnostic explanation (LIME) technique within XAI. Algorithm correctly predicted troll tweets with a confidence of 93% and non-troll tweets with a confidence of 97%. This research lays the groundwork for better decision making in the ever-changing field of social media content analysis by bridging the gap between complex neural networks and insights that can be understood by humans. The transparency and reliability that LIME brings to public discussion are crucial tools for ensuring the responsible and efficient use of online content, as social media continues to influence public opinion.