An Iterative Fusion Approach to Graph-Based Semi-Supervised Learning from Multiple Views

Often, a data object described by many features can be naturally decomposed into multiple “views”, where each view consists of a subset of features. For example, a video clip may have a video view and an audio view. Given a set of training data objects with multiple views, where some objects are labeled and the others are not, semi-supervised learning with graphs from multi-views tries to learn a classifier by treating each view as a similarity graph on all objects, where edges are defined by the similarity on object pairs based on the view attributes. Labels and label relevance ranking scores of labeled objects can be propagated from labeled objects to unlabeled objects on the similarity graphs so that similar objects receive similar labels. The state-of-the-art, onecombo-fits-all methods linearly and independently combine either the metrics or the label propagation results from multi-views and then build a model based on the combined results. However, more often than not, the similarities between various objects may be manifested differently by different views. In such situations, the one-combo-fits-all methods may not perform well. To tackle the problem, we develop an iterative SemiSupervised Metric Fusion (SSMF) approach in this paper. SSMF fuses metrics and label propagation results from multi-views iteratively until the fused metric and label propagation results converge simultaneously. Views are weighted dynamically during the fusion process so that the adversary effect of irrelevant views, identified at each iteration of fusion process, can be reduced effectively. To evaluate the effectiveness of SSMF, we apply it on multi-view based and content based image retrieval and multi-view based multi-label image classification on real world data set, which demonstrates that our method outperforms the state-of-the-art methods.