Learning from Positive and Unlabeled Data

Abstract

Positive and Unlabeled (PU) data comprise positively labeled examples and a mixture of unlabeled data, potentially including both positive and negative examples. PU data is commonly encountered when we develop machine learning methods. In the real world, many applications rely solely on PU data, and extensive research has been conducted on this topic. Examples of applications that benefit from learning from PU data encompass anomaly detection, recommender systems, and knowledge graph completion. In particular, anomaly detection and recommender systems are representative applications for learning from PU data in the industry. This dissertation introduces two studies on anomaly detection and one recommender system study that deals with PU data. First, we introduce a metric learning-based anomaly detection method, utilizing a small set of positive anomalies along with unlabeled data. We mine and utilize positive and negative data from unlabeled data for training. Secondly, anomaly score functions for the self-supervised anomaly detection methods are introduced. The functions address the limitation of previous approaches, which lacked the capability for semantic-aware detection. Finally, a multi-domain recommender system is introduced that leverages data from other sources to address the data sparsity issue, which is one of the most challenging problems in recommender systems with PU data. The proposed methods exhibit exceptional performance in each application through the effective utilization of PU data. In summary, our study explores two distinct applications of learning from PU data and introduces state-of-the-art methods tailored for each.