关键词: 多标记分类, 多示例学习, 权重, K近邻

Abstract: In most cases, the inherent ambiguity of each instance is explicitly expressed in the output space based on associations with multiple class labels. Recent studies indicate that the instance ambiguity can be expressed in the input space by transforming a single instance into a bag of instances and establishing the relations between sets of labels and bags of transformed instances. However, the bags of instances are generated by calculating the mean values of instances corresponding to the each label with equal weight for each instance. Because of the local distribution characteristics of data, taking the local distribution of data will generate more accurate instance bags. This paper fully considers the local distribution characteristics, and proposes a new multi-label classification algorithm. Experimental results show that it outperforms other proposed multi-label algorithms.

Key words: multi-label classification, multi-instance learning, weighting, K-Nearest Neighbour（KNN）