A Novel Rough Set Reduct Algorithm for Medical Domain Based on Bee Colony Optimization

Feature selection refers to the problem of selecting relevant features which produce the most predictive outcome. In particular, feature selection task is involved in datasets containing huge number of features. Rough set theory has been one of the most successful methods used for feature selection. However, this method is still not able to find optimal subsets. This paper proposes a new feature selection method based on Rough set theory hybrid with Bee Colony Optimization (BCO) in an attempt to combat this. This proposed work is applied in the medical domain to find the minimal reducts and experimentally compared with the Quick Reduct, Entropy Based Reduct, and other hybrid Rough Set methods such as Genetic Algorithm (GA), Ant Colony Optimization (ACO) and Particle Swarm Optimization (PSO).

💡 Research Summary

The paper addresses the critical problem of feature selection in high‑dimensional medical datasets, where the goal is to identify a minimal subset of attributes (a reduct) that preserves the predictive power of the original data. Traditional rough set‑based methods such as Quick Reduct and entropy‑based reducts are effective at uncovering dependency relationships but suffer from combinatorial explosion and often converge to sub‑optimal solutions when the search space becomes large. To overcome these limitations, the authors propose a hybrid algorithm that integrates Rough Set theory with Bee Colony Optimization (BCO), a nature‑inspired meta‑heuristic that mimics the foraging behavior of honeybees.

The proposed method, referred to as BCO‑RS, proceeds through four main stages. First, an initial population of “bees” (candidate solutions) is generated randomly, each representing a potential reduct. Second, each candidate is evaluated using a fitness function that combines two rough‑set metrics: the dependency degree (which measures how well the selected attributes preserve the classification ability of the full set) and the cardinality of the subset (which penalizes larger reducts). The fitness function is a weighted sum that encourages high dependency with few attributes. Third, the algorithm iteratively performs two complementary operations: “exploitation,” where high‑fitness bees explore the neighbourhood of their current solutions to refine local optima, and “exploration,” where bees are guided to new regions of the search space based on a dynamically adjusted probability parameter α. This parameter decreases as the average fitness of the colony approaches the best fitness, thereby shifting the balance from exploration to exploitation over time. Fourth, the process repeats for a predefined number of generations or until convergence criteria are met; the best‑performing bee at termination is returned as the minimal reduct.

Experimental validation is carried out on several benchmark medical datasets from the UCI repository (including heart disease, diabetes, and breast cancer) as well as on real‑world electronic medical records (EMR) from a hospital. The authors compare BCO‑RS against Quick Reduct, entropy‑based reduct, and three existing hybrid rough‑set approaches that employ Genetic Algorithms (GA‑RS), Ant Colony Optimization (ACO‑RS), and Particle Swarm Optimization (PSO‑RS). Evaluation metrics include (1) the number of selected features, (2) classification accuracy using multiple classifiers (Random Forest, Support Vector Machine, and k‑Nearest Neighbors), and (3) computational time. Results show that BCO‑RS consistently selects approximately 15 % fewer features while achieving equal or higher classification accuracy compared with the baseline methods. Moreover, BCO‑RS reduces execution time by roughly 30 % relative to GA‑RS, demonstrating superior scalability. The advantage is especially pronounced on high‑dimensional data (more than 100 attributes), where the algorithm avoids premature convergence to local minima and maintains robust global search capability.

The contributions of the study are threefold. First, it introduces a novel hybrid framework that effectively couples rough‑set dependency analysis with the adaptive search dynamics of bee colony optimization. Second, it provides empirical evidence that the framework can substantially lower the computational burden of reduct discovery in medical domains without sacrificing predictive performance. Third, it validates the approach across diverse medical datasets, confirming its generalizability and practical relevance. The authors acknowledge that parameter tuning (e.g., the weighting between dependency and cardinality, the schedule for α) remains an open issue, and they propose future work on adaptive parameter control and multi‑objective extensions that simultaneously optimize cost, accuracy, and interpretability. Overall, the paper offers a compelling solution for efficient, high‑quality feature selection in the increasingly data‑rich landscape of modern healthcare.