Privileged Information Learning with Weak Labels

Privileged information learning is proposed to construct a better classifier by incorporating some privileged knowledge. At present, most of the privileged information learning methods assume that the instance is accurately labeled. However, in real-world applications, an instance may be annotated by a number of labelers, and the labels obtained by these labelers are used to calculate the final label. Since the weight of each labeler is unknown, people always give an equal or random weight to each labeler. It leads to the ambiguous labels for instances, which are known as weak labels. In this paper, we consider the weak labels problem in privileged information learning problem, where the training data and privileged data is associated with ambiguous labels, rather than accurate labels. To tackle this problem, we propose a novel privilege information learning method with weak labels (PLWB). We assign each labeler a weight and these labeler weights are then incorporated into a privileged information learning process. An heuristic framework is proposed to train the privileged information learning model and update the labeler weights alternately. Numerous experiments have shown that PLWB can get better classification performance on weakly labeled data compared with state-of-the-art privileged information learning methods.