Ensemble regression kernel extreme learning machines for multi-instance multi-label learning

Abstract

The multi-instance multi-label learning (MIML) framework is an extension of the multi-label learning, where each object in MIML is represented by a multi-instance bag and associated with a multi-label vector. Recently, extreme learning machine (ELM) has been widely used in multi-instance multi-label classification due to its short runtime. Simultaneously, ELM also has good classification accuracy compared to other neural network models. However, this type of ELM-based MIML classification algorithms can easily lead to overfitting problems during training and the basic ELM algorithm with random initial weights and biases is not stable. In order to solve the above problems, ensemble learning is used to overcome overfitting problems and regression kernel extreme learning machine (RKELM) as classifier instead of the basic ELM effectively can solve the problem of instability of training. In this paper, Bagging-based RKELM (BRKELM) and AdaBoost-based RKELM (ARKELM) for MIML classifications are proposed. The comparison with other state-of-the-art multi-instance multi-label learning algorithms shows that the BRKELM and ARKELM are highly efficient, feasible and stable algorithms.