Ali Fakeri-Tabrizi (1), Massih-Reza Amini (2), Cyril Goutte (3), Nicolas Usunier (4)
(1) Université Pierre et Marie Curie (2) Université Joseph Fourier (3) National Research Center (4) Université Technologie de Compiègne
In many applications, observations are available with different views. This is, for example, the case with image-text classification, multilingual document classification or document classification on the web. In addition, unlabeled multiview examples can be easily acquired, but assigning labels to these examples is usually a time consuming task. We describe a multiview self-learning strategy which trains different voting classifiers on different views. The margin distributions over the unlabeled training data, obtained with each view-specific classifier are then used to estimate an upper-bound on their transductive Bayes error. Minimizing this upper-bound provides an automatic margin-threshold which is used to assign pseudo-labels to unlabeled examples. Final class labels are then assigned to these examples, by taking a vote on the pool of the previous pseudo-labels. New view-specific classifiers are then trained using the labeled and pseudo-labeled training data. We consider applications to image-text classification and to multilingual document classification. We present experimental results on the NUS-WIDE collection and on Reuters RCV1-RCV2 which show that despite its simplicity, our approach is competitive with other state-of-the-art techniques.