A novel adaptive density-based ACO algorithm with minimal encoding redundancy for clustering problems

In the so-called Big Data paradigm descriptive analytics are widely conceived as techniques and models aimed at discovering knowledge within unlabeled datasets (e.g. patterns, similarities, etc) of utmost help for subsequent predictive and prescriptive methods. One of these techniques is clustering, which hinges on different multi-dimensional measures of similarity between unsupervised data instances so as to blindly collect them in groups of clusters. Among the myriad of clustering approaches reported in the literature this manuscript focuses on those relying on bio-inspired meta-heuristics, which have been lately shown to outperform traditional clustering schemes in terms of convergence, adaptability and parallelization. Specifically this work presents a new clustering approach based on the processing fundamentals of the Ant Colony Optimization (ACO) algorithm, i.e. stigmergy via pheromone trails and progressive construction of solutions through a graph. The novelty of the proposed scheme beyond previous research on ACO-based clustering lies on a significantly pruned graph that not only minimizes the representation redundancy of the problem at hand, but also allows for an embedded estimation of the number of clusters within the data. However, this approach imposes a modified ant behavior so as to account for the optimality of entire paths rather than that of single steps within the graph. Simulation results over conventional datasets will evince the promising performance of our approach and motivate further research aimed at its applicability to real scenarios.