A Novel Multiple Kernel Learning Approach for Semi-Supervised Clustering

Distance metrics are widely used in various machinelearning and pattern recognition algorithms. A main issue inthese algorithms is choosing the proper distance metric. In recentyears, learning an appropriate distance metric has become a veryactive research field. In the kernelised version of distance metriclearning algorithms, the data points are implicitly mapped into ahigher dimensional feature space and the learning process isperformed in the resulted feature space. The performance of thekernel-based methods heavily depends on the chosen kernelfunction. So, selecting an appropriate kernel function and/ortuning its parameter(s) impose significant challenges in suchmethods. The Multiple Kernel Learning theory (MKL) addressesthis problem by learning a linear combination of a number ofpredefined kernels. In this paper, we formulate the MKLproblem in a semi-supervised metric learning framework. In theproposed approach, pairwise similarity constraints are used toadjust the weights of the combined kernels and simultaneouslylearn the appropriate distance metric. Using both synthetic andreal-world datasets, we show that the proposed methodoutperforms some recently introduced semi-supervised metriclearning approaches.

کلیدواژه ها:

Distance Metric Learning ، Multiple Kernel Learning (MKL) ، Pairwise similarity constraints ، Semi-supervised clustering