Co-clustering optimization using Artificial Bee Colony (ABC) algorithm

This paper presents an Artificial Bee Colony (ABC) optimization based algorithm for co-clustering of high-dimensional data. The ABC algorithm is used for optimization problems including data clustering. We incorporate aspects of co-clustering by embedding it into the objective function used for clustering by the ABC algorithm. Instead of a linear metric, such as the Euclidean distance, we propose the use of higher order correlations to build similarity between rows and columns, each based on the other. This measure uses co-evolving similarities which when embedded into the objective function results in optimizing the co-clusters. The search space is also explored in the vicinity of the solutions produced by the ABC algorithm using three local search methods — the first is a heuristic based on computing the cluster means; the second uses the analytical gradient of the objective with respect to a centroid to find lower cost solutions in the vicinity; and, the third is a hybrid of the first two methods. Numerical experiments show significant improvement in the search for optimal clustering by incorporating new similarity metric and optimized local search method. Finally, the algorithm is shown to be highly scalable for parallel architectures for both distributed and shared memory systems. Theoretically, the best iso-efficiency function of Θ (p log p) for fully connected network with p processors is also computed for the parallel algorithm. •We incorporate co-clustering into the objective function of the ABC using co-similarity matrices to achieve co-clustering.•We enhance the ABC approach using a neighborhood-search-space-aware local search method to guide the solution.•We discuss ways to parallelize the algorithm and provide a scalability analysis.•Experiments show much better results as compared to other clustering and co-clustering methods.