Learning Using Multiple-Type Privileged Information and SVM+ThinkTank

In this paper, based on the extension to the standard Support Vector Machines Plus (SVM+) model for the Learning Using Privileged Information (LUPI) paradigm, a new SVM+ThinkTank (SVM+TT) model, is proposed for Learning Using Multiple-Type Privileged Information (LUMTPI). In cases that Multiple-Type Privileged Information (MTPI) from different perspectives is available for interpreting those training samples, such as in Big Data Analytics, it could be beneficial to leverage all these different types of Privileged Information collectively to construct multiple correcting spaces simultaneously for training the maximum margin based separating hyperplane of SVM model. In fact, from the practical point of view, organising Privileged Information from different perspectives might be a relatively easier task than finding a single type perfect Privileged Information for those hardest training samples. The MTPI collectively plays the role of a think tank as the single type Privileged Information plays the role of a single perfect master class teacher. The preliminary experimental results presented and analysed in this paper demonstrate that SVM+TT, as a new learning instrument for the proposed LUMTPI paradigm, is capable of improving the generalisation ability of the standard SVM+ in learning from a small amount of training samples but incorporating with the MTPI interpretations to these samples for improved learning performance.