Xuran ZHAO - Ph.D. student in the Multimedia Communication department Multimedia Communications
Date: - Location: Eurecom
A good distance metric is crucial for many real world applications involving high-dimensional data, such as image classification and clustering, text mining, and content-based image retrieval (CBIR). Most metric learning algorithms appeal to projecting the observed data into a lower dimensional subspace, where some geometric relationships, such as pairwise distances are preserved. In multi-media applications, multiple features can be extracted from the same data sample. For example, an image can be represented by its texture feature and color histogram, a person in a video can be represented by facial and vocal features. Therefore, we are facing an additional fusion task on top of the metric learning problem. We try to solve this multi-view metric learning (or dimensionality reduction) problem under a multi-view agreement assumption. For example, consider an audio-visual person clustering problem, given optimal distance metrics, we assume that the clustering results would be the same no matter facial or vocal features are used (both results will correspond to person identities). In this sense, we propose two algorithms for different application purposes. First, for a multi-view clustering purpose, we propose CoKmLDA algorithm which unifies unsupervised k-means clustering and supervised Linear Discriminant Analysis (LDA) within a co-training [1] framework. Second, for retrieval problems, we proposed an algorithm to learn distance metrics under which the similarity graph of different features agree with each other. In short, if two data samples are within the k nearest neighbor of each other in for one feature, we compel it to be so in the other feature. This objective is accomplished by co-training Locality Preserving Projections (LPP). In experiments of audio-visual person clustering and retrieval, the proposed algorithms out-perform recently proposed multi-view clustering/metrics fusion algorithms [2][3][4] with a significant margin. [1] Blum, A., Mitchell, T. Combining labeled and unlabeled data with co-training. COLT: Proceedings of the Workshop on Computational Learning Theory, Morgan Kaufmann, 1998, p. 92-100. [2] A. Kumar and H. D. III. A co-training approach for multi-view spectral clustering. In ICML, pages 393?400, 2011 [3] K. Chaudhuri, S. M. Kakade, K. Livescu, and K. Sridharan. Multi-view clustering via canonical correlation analysis. In Proceedings of the 26th Annual International Conference on Machine Learning, ICML ?09, 2009. [4] B. Wang, et al., Unsupervised Metric Fusion by Cross Diffusion, in CVPR 2012.
