Biomedical Text Categorization Based on Ensemble Pruning and Optimized Topic Modelling

Area of Science:

• Computational linguistics
• Machine learning
• Bioinformatics

Background:

• Text mining, encompassing information retrieval, extraction, and categorization, is crucial in various research fields.
• Latent Dirichlet allocation (LDA) is a powerful tool for text categorization, but its performance is highly dependent on accurate parameter estimation.
• Existing methods for LDA parameter tuning and ensemble classifier construction can be further optimized for improved predictive performance.

Purpose of the Study:

• To propose an efficient multiple classifier system for text categorization using swarm-optimized topic modeling.
• To enhance the performance of Latent Dirichlet allocation (LDA) by optimizing its parameters using swarm intelligence.
• To develop a hybrid ensemble pruning approach that combines diversity measures and clustering for superior predictive accuracy.

Main Methods:

• Swarm optimization algorithms (e.g., genetic algorithms, particle swarm optimization) were used to estimate LDA parameters, including the number of topics.
• A hybrid ensemble pruning technique was developed, combining four diversity measures (disagreement, Q-statistics, correlation coefficient, double fault) and swarm intelligence-based clustering.
• Classifiers were clustered based on diversity, and the best-performing classifier from each cluster was selected for the final system.

Main Results:

• Swarm-optimized LDA demonstrated superior predictive performance over conventional LDA on five biomedical text benchmarks.
• The proposed multiple classifier system significantly outperformed traditional classification algorithms, ensemble learning, and existing ensemble pruning methods.
• Evaluation included comparisons of various metaheuristic algorithms for both swarm-optimized LDA and ensemble pruning.

Conclusions:

• Swarm-optimized topic modeling provides a robust enhancement for LDA in text categorization tasks.
• The proposed hybrid ensemble pruning method effectively constructs diverse and high-performing multiple classifier systems.
• This approach offers a significant advancement in the field of text mining, particularly for biomedical applications.