Traditional units of genome-wide association studies have serious defects such as low repeatability, difficulty to interpret, and epistasis analysis based on machine learning has troubles such as high computational complexity and insufficient prediction accuracy. This paper presented a new approach for the analysis of genome-wide epistatic. This method uses the framework of two-phase epistatic analysis method. It includes a filtering stage and an epistatic combinatorial optimization stage. The characteristics of the filtering stage presents a multicriteria fusion strategy for the evaluation of genetic loci from multiple perspectives to ensure that the weak effect of susceptibility loci can be retained, and then, this method uses the multiple criteria sorting fusion strategy to eliminate the low degree of genetic variation associated with disease states. Epistatic combinatorial optimization phase uses the greedy algorithm combination of heuristic search space in order to reduce the time complexity. Finally, a support vector machine was used as the epistatic evaluation model. Experiments with different parameters of linkage disequilibrium SNPruler with classical algorithms were compared with the performance of the ACO, and the experiment results show that the method can effectively keep weak effect locus and improve disease forecasting accuracy considerably.