Ensemble pruning of ELM via migratory binary glowworm swarm optimization and margin distance minimization

摘要

Ensemble pruning aims at attaining an ensemble composed of less size of leaners for improving classification ability. Extreme Learning Machine (ELM) is employed as a base learner in this work, in light of its salient features, an initial pool is constructed using ELM. An ensemble composed of ELMs with better performance and diversity can make it perform the best, but the average accuracy of the whole ELMs must be decreased as the increase of diversity among them. Hence there exists a balance between the diversity and the precision of ELMs. Existing works find it via diversity measures or heuristic algorithms, which cannot find the exact tradeoff. To solve the issue, ensemble pruning of ELM via migratory binary glowworm swarm optimization and margin distance minimization (EPEMBM) is proposed utilizing the integration of the proposed migratory binary glowworm swarm optimization (MBGSO) and margin distance minimization (MDM). First, the created ELMs in a pool can be pre-pruned by MDM, and it can markedly downsize the ELMs in the pool, and significantly alleviates its computation overhead. Second, the retaining ELMs are further pruned utilizing MBGSO, and the final ensemble is attained with a high efficiency. Experimental results on 21 UCI classification tasks indicate that EPEMBM outperforms techniques, and that its effectiveness and efficiency. It is a very useful tool for solving the selection problem of ELMs.