Location
Innovation Lab Atrium
Start Date
4-3-2025 10:00 AM
End Date
4-3-2025 11:30 AM
Presentation Date
3 April 2025, 10:00am - 11:30am
Meeting Name
2025 - Miners Solving for Tomorrow Research Conference
Department(s)
Computer Science
Document Type
Poster
Document Version
Final Version
File Type
event
Language(s)
English
Rights
© 2025 The Authors, All rights reserved
Included in
Apr 3rd, 10:00 AM
Apr 3rd, 11:30 AM
Unsupervised Contrastive Learning Based Clustering for Robust Emotion Classification
Innovation Lab Atrium
Comments
Advisor: Sanjay Kumar Madria
Abstract:
Adversarial attacks exploit subtle perturbations in input data to mislead machine learning models. This becomes more effective in emotion classification tasks where neutral labels are often overrepresented, making models vulnerable to small changes in neutral texts. To address this, we propose a method to reassign potentially mislabeled neutral samples to more accurate emotion labels, improving model robustness. A preprocessing module using the Sim-Text algorithm identifies emotion-specific keywords in the text. Then, a Seq2Seq Autoencoder is jointly trained with the proposed Sim-EUCL algorithm to learn embeddings where semantically similar samples are closely aligned. UMAP is applied for dimensionality reduction, followed by HDBSCAN clustering to discover emotion-based clusters. Neutral samples are relabeled based on cluster context. Finally, the model is tested against adversarial attacks at the character, word, and sentence levels. Experimental results show a 14% improvement in classification accuracy, validating the effectiveness and robustness of the proposed approach.