Multi-Dimensional Classification via NSGA-III Algorithm and Sparse Label Encoding

Traditional multi-dimensional classification (MDC) methods often assume that optimizing a single objective can improve overall performance, thus meeting the requirements of various applications. However, achieving optimal performance in MDC typically involves balancing trade-offs among multiple objectives or metrics. Consequently, this paper investigates the problem of multi-objective MDC. To overcome the limitations of random coding matrix A generation in the sparse labeled encoding-based MDC method (SLEM), a novel MDC method based on the NSGA-III algorithm, is proposed, named N3MDC. First, the NSGA-III algorithm is employed for iterative optimization to obtain an optimal set of non-dominated solutions for the targeted objectives. Then, based on preferences for unseen instances and the proposed model selection strategy, the optimal solution is selected and used as coding matrix A. Finally, using the generated coding matrix A, cascade operations of encoding, training, and decoding are performed to predict the class vectors of the unseen instances. To assess the effectiveness of the proposed method, extensive comparative experiments were conducted using 8 state-of-the-art MDC methods on 11 publicly available datasets, each representing a distinct real-world application problem. The results show that the N3MDC method delivers superior generalization performance compared to the benchmark methods.