View article

Background

Abstract Multi-instance learning is regarded as a new learning framework. Previous researches mainly focus on multi-instance classification. Recently, multi-instance regression attracts the attention of the machine learning community. Manifold learning attempts to obtain the intrinsic structure of non-linearly distributed data, which can be used in non-linear dimensionality reduction (NLDR). In this paper, a manifold learning-based multi-instance regression algorithm, ManiMIL, is proposed. ManiMIL performs NLDR on the instances in training bags, selects the most diverse dimension that NLDR brings and builds a classifier only on this dimension and then makes the prediction. Experimental results show that the performance of ManiMIL outperforms that of existing multi-instance algorithms such as Citation-kNN.