CONTRASTIVE TF-IDF VECTORIZATION FOR IMPROVED CLASSIFICATION OF FASHION-RELATED TWEETS: A CROSS-DOMAIN TRANSFER LEARNING SEMI-SUPERVISED LEARNING APPROACH
International Research Journal of Modernization in Engineering Technology and Science · 2023
The proliferation of fashion-related discourse on social media, particularly on Twitter, presents a unique opportunity for real-time analysis of trends and preferences.To capitalize on this, we introduce a novel labeled dataset* of fashion-related tweets, paired with an innovative, resource-efficient machine learning model for classification.Drawing on principles of contrastive learning, we propose a modified TF-IDF vectorization technique, "Contrastive TF-IDF," which enhances the discriminative power of term vectors by emphasizing terms that are not only relevant within a document but also distinctively contrastive across different classes.This methodology enriches term representations by integrating contrastive learning directly into the TF-IDF scheme, thus pulling similar fashion topics closer while pushing disparate ones apart in the feature space.To validate our approach, we employ lightweight classifiers, such as Support Vector Machines (SVM) and Naive Bayes, which are known for their effectiveness in text classification tasks with less computational overhead than deep learning counterparts.These models are trained using our contrastive TF-IDF vectors, allowing us to demonstrate the efficacy of our methodology without the extensive resource requirements typically associated with deep learning.Our results show a marked improvement in classification accuracy, providing the fashion industry with a nimble and accurate tool for monitoring and analyzing fashion trends as they emerge on social media.