Multiple Instance Learning With Instance-Level Positive-Unlabeled Learning in Anomaly Detection

Abstract

We propose a method for learning a classifier that accurately predicts both instance and bag classes in multiple instance learning (MIL) for anomaly detection, achieving significant performance improvement. MIL, a form of weakly supervised learning, represents datasets as sets of bags labeled as either positive or negative, where each bag contains multiple instances. A positive bag includes at least one positive instance, whereas a negative bag contains only negative instances. Taking into account the lack of instance-level labels, a classifier must learn to predict instance-level labels from only bag labels during training. Whereas bag label annotation is relatively straightforward in many applications, annotating individual instances is often expensive. Consequently, MIL settings often arise in various tasks. Recent studies in MIL for anomaly detection integrate bag-level training with instance-level training focused on feature reconstruction. However, these methods use only instances in normal bags for instance-level training, and some methods also make it difficult to determine a suitable decision boundary due to the nature of training based on the area under the receiver operating characteristic curve. Our proposed method uses positive and unlabeled learning for enhancing instance-level classification performance and weighted-noisy-OR for training bag-level classification. We validate the proposed method in terms of the F1 score and area under the precision-recall curve (PR-AUC) on 35 anomaly-detection datasets. The experimental results indicate that our method outperforms conventional methods in both F1 score and PR-AUC, achieving an F1 score/ PR-AUC of 0.706/0.713 for instance classification and 0.789/0.845 for bag classification.