A linear directional optimum weighting (LDOW) approach for parallel hybridization of classifiers

Abstract

Hybridization of classifiers can often yield outperformed performance compared to its best individual component and mostly have more generalization ability. The majority of combined classifiers reported in the literature benefit the parallel or ensemble topology. The performance of such parallel hybrid classifiers significantly depends on their applied weighting approaches and the accuracy strongly relies on the classifiers' weights. In the literature, several different weighting mechanisms have been developed to yield a higher classification rate, which can be generally categorized into three main categories individual classifiers, averaging-based, and optimization-based algorithms. Although these weighting mechanisms have been commonly and frequently used for parallel hybridization, none of them can guarantee that their obtained classification rate will be optimum. In addition, due to the use of iterative procedures, especially by meta-heuristic optimization-based algorithms, their computational time and cost are always unsatisfactory. In this paper, a linear direct optimal weighting (LDOW) approach is proposed in which it can be generally guaranteed that the optimum classification rate will be directly reached. In this way, the proposed approach can achieve the highest classification rate by the desired computational time and cost and both mentioned limitations of currently-used weighting algorithms can be simultaneously lifted. Empirical results of twenty-three different benchmark data sets from six dissimilar domains indicate that the proposed LDOW approach can yield more classification rate in all data sets and can averagely improve 5.36% the performance of its traditional linear direct non-optimal version with the same components. Furthermore, the proposed LDOW approach can averagely improve 6.87% the classification rate of the averaging weighting algorithms that are similar to the proposed algorithm are direct ones. In addition, the proposed model can even improve 2.80% the classification rate of nonlinear intelligent weighting algorithms, on average. Whereas, the proposed approach is a linear model and its complexity and computational cost are significantly lower than these nonlinear intelligent algorithms. Numerical results also indicate that the proposed model can averagely improve 6.08% the classification rate of meta-heuristic-based weighting algorithms. While the proposed model is a direct model and its computational cost is meaningfully lower than these iterative algorithms. Thus, in theory as well as in practice, it can be inferred that the proposed LDOW approach can be an efficient alternative weighting method for parallel hybridization in the classification field. This is particularly relevant when more accurate results are required or for big data situations where computational time and cost are critical factors to consider. © 2024 Elsevier B.V.