Evolving personalized modeling system for integrated feature, neighborhood and parameter optimization utilizing gravitational search algorithm

摘要

This paper introduces a new evolving personalized modeling method and system (evoPM) that integrates gravitational search inspired algorithm (GSA) for selecting informative features, optimizing neighbors and model parameters. For every individuals, evoPM creates a model that best predicts the outcome for this individual at the time of model creation. A comparative study is given for investigating the feasibility of the proposed system on several benchmark datasets using global, local and personalized modeling methods. The proposed evoPM system is capable of identifying a small group of the most informative features, optimizing the neighbors and model parameters relevant to the learning function (a classifier), which leads to improved classification performance. The experimental results show that evoPM not only outperforms several global and local modeling methods in terms of classification accuracy, but also finds the optimal or near-optimal solution to feature selection, and neighborhood, model parameters optimization in less number of iterations than many other evolutionary computational based optimizing algorithms.