Feature Weight Optimization Method Based on t-Memetic Algorithm

Feature weights have a significant effect on the accuracy of case-based reasoning (CBR) prediction models. Therefore, this paper proposes a feature weight optimization method combining a t-distribution mutation operator and a memetic algorithm (t-memetic). In this method, the mean absolute percentage error of the CBR prediction model is defined as a fitness function, and the memetic framework is used to realize the optimization process for case weights. A sparrow search algorithm and adaptive t-distribution mutation operator are used to realize the global search of the case feature weights, and the simulated annealing algorithm is used to perform a local search of the weighted individuals with the best fitness for the current population. Five UCI standard regression datasets are used respectively to test the effectiveness of the proposed method and compare it with classical feature weight optimization algorithms. The results show that the CBR prediction model has the highest accuracy after the weights are optimized by the t-memetic algorithm, which indicates that the proposed weight optimization method can be used effectively in CBR prediction models.