Most of the approaches in multi-view categorization use early fusion, late fusion or co-training strategies. We propose here a novel classification method that is able to efficiently capture the interactions across the different modes. This method is a multi-modal extension of the Rocchio classification algorithm - very popular in the Information Retrieval community. The extension consists of simultaneously maintaining different "centroid" representations for each class, in particular "cross-media" centroids that correspond to pairs of modes. To classify new data points, different scores are derived from similarity measures between the new data point and these different centroids; a global classification score is finally obtained by suitably aggregating the individual scores. This method outperforms the multi-view logistic regression approach (using either the early fusion or the late fusion strategies) on a social media corpus - namely the ENRON email collection - on two very different categorization tasks (folder classification and recipient prediction).
展开▼
