A Multi-kernel Semi-supervised Metric Learning Using Multi-objective Optimization Approach
A kernel-matrix based distance measure is utilized for computing the similarities between the data points. The available few labeled data is used as constraints to project on initial kernel-matrix using Bregman projection. Since the projection of constraints onto the matrix is not orthogonal, we need to identify an appropriate subset of constraints subject to objective functions, measuring the quality of partitioning of the data. As the kernel-space is large in size, we have divided the original kernel space into multiple kernel sub-spaces so that each kernel can be processed independently and parallelly in advance GPU and kernel semi-supervised metric learning using multi-objective approach is applied on individual kernels parallelly. The multi-objective framework is used to select the best subset of constraints to optimize multiple objective functions for grouping the available data. Our approach outperforms the state of the art algorithms on the various datasets with respect to different validity indices.
KeywordsSemi-supervised Multi-objective optimization Classification Clustering Graphics processing unit (GPU)
- 1.Davis, J.V., Kulis, B., Jain, P., Sra, S., Dhillon, I.S.: Information-theoretic metric learning. In: Proceedings of the 24th International Conference on Machine Learning, pp. 209–216. ACM (2007)Google Scholar
- 2.Amid, E., Gionis, A., Ukkonen, A.: A kernel-learning approach to semi-supervised clustering with relative distance comparisons. In: Appice, A., Rodrigues, P.P., Santos Costa, V., Soares, C., Gama, J., Jorge, A. (eds.) ECML PKDD 2015. LNCS (LNAI), vol. 9284, pp. 219–234. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-23528-8_14CrossRefGoogle Scholar
- 7.Robert, V. Vasseur, Y.: Comparing high dimensional partitions, with the Coclustering Adjusted Rand Index. ArXiv Preprint ArXiv:1705.06760 (2017)
- 8.Rendón, E., Abundez, I., Arizmendi, A., Quiroz, E.M.: Internal versus external cluster validation indexes. Int. J. Comput. Commun. 5(1), 27–34 (2011)Google Scholar