Dynamic Hyperparameter Tuning in Deep Co-Training for Semi-Supervised Sentiment Analysis on Social Media

The exponential growth of social media has generated vast volumes of emotionally expressive text, making sentiment analysis crucial for public opinion monitoring, marketing, and crisis detection.However, supervised deep learning models are severely constrained by the scarcity of labeled data.Semi-supervised learning via Co-Training offers a promising solution by leveraging unlabeled data through collaborative learning between multiple classifiers.While recent studies integrate deep architectures (CNN, LSTM) into Co-Training, most rely on static hyperparameters, leading to error propagation, suboptimal convergence, and performance degradation under noisy, real-world data.This paper introduces Co-TuneDL, a novel framework that embeds iterative, joint hyperparameter tuning directly into the Co-Training loop.We dynamically optimize three critical parameters, (1) confidence threshold () for pseudo-label selection, (2) learning rate () for model adaptation, and (3) batch size (b) for training stability, using Bayesian Optimization (BO) based on validation set performance after each iteration.Unlike prior approaches that tune hyperparameters once or independently, our method treats them as interdependent variables that evolve with the learning process.Evaluated on three benchmark datasets, including Indonesian Twitter, SemEval-2017, and a custom political sentiment corpus, Co-TuneDL achieves an average F1-score of 92.4%, outperforming static Co-Training baselines (87.1%) and conventional supervised CNN-LSTM models (89.3%).Statistical tests confirm significance (p < 0.01).Crucially, dynamic tuning reduces error propagation by 41%, accelerates convergence by 28%, and enables robust learning even with <2k labeled samples.This work establishes that hyperparameter tuning is not a preprocessing step but a core mechanism for enabling scalable, adaptive semi-supervised learning in Big Data environments.