Evolving optimal text clusters: A novel GA-driven framework for dynamic ensemble fusion of multi-model contextual...

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Artificial Intelligence

Background:

• Text clustering is vital for unsupervised NLP tasks like news classification and summarization.
• Contextual embedding models (SBERT, RoBERTa, DistilBERT) enhance clustering but have domain-specific performance variations.
• Existing ensemble methods use fixed weights, failing to leverage complementary model strengths dynamically.

Purpose of the Study:

• To introduce a novel Genetic Algorithm (GA)-based ensemble model for dynamic, label-free optimization of multi-model contextual embedding fusion.
• To address the gap in adaptive, domain-accommodative model combination for unsupervised text clustering.
• To hypothesize that evolutionary optimization can discover domain-adaptive weights outperforming standalone and fixed ensemble models.

Main Methods:

• Developed a GA-based ensemble framework to dynamically optimize fusion weights for SBERT, RoBERTa, and DistilBERT embeddings.
• Employed L2-normalization and a sum-to-one constraint for fair model integration.
• Utilized a composite fitness measure (Silhouette Score, Adjusted Rand Index, Topic Coherence) with tournament selection, uniform crossover, and Gaussian mutation for weight evolution.

Main Results:

• Achieved statistically significant improvements on heterogeneous benchmarks (AG News, 20 Newsgroups, Stack Overflow).
• Demonstrated Silhouette Score increases of +14-16% and Topic Coherence gains of +17% (p < 0.05).
• Evolved weights provide interpretable, domain-specific model roles, guiding future architectural choices.

Conclusions:

• The proposed GA-based ensemble framework effectively optimizes fusion weights for contextual embeddings in unsupervised clustering.
• The dynamic, adaptive approach surpasses standalone models and fixed ensemble baselines on diverse datasets.
• The system offers a scalable, modular solution for practical unsupervised NLP, extendable to new transformer and semi-supervised systems.