RPL-SVM: Making SVM Robust Against Missing Values and Partial Labels

With increased data availability, data quality is the biggest problem in using AI models. The data may suffer from missing values and noisy values. Another major challenge is to get accurate labels for feature vectors in the training data. In contrast, in many applications, we can only get weak labels (for example, partial labels, positive-unlabeled data, bandit feedback, etc.). In this paper, we consider uncertainty in both features and labels. More specifically, we assume that feature vectors have missing attributes and are only given partial labels. We present a novel second-order cone programming framework to learn robust classifiers that can tolerate uncertainty in the observations of partially labeled multiclass classification problems. The proposed approach, RPL-SVM, is based on a chance-constrained framework. Experimental results show that RPL-SVM efficiently learns multiclass classifiers with missing values in a partial label setting. This demonstrates the remarkable resilience of RPL-SVM to real-world observational uncertainties.