Multi-Instance Learning And Anomaly Detection In Open Environment

Machine learning is one of the core research areas of artificial intelligence,and is the fundamental way to achieve intelligence.Traditional machine learning always assumes that the training and test data follow the same distribution and a certain amount of labeled data is needed to build an effective learner.With the expansion of machine learning techniques to more application fields,we are facing more and more real tasks under open environments,specifically,non-i.i.d.data and lacking of labeled data may always occur.This thesis focuses on the problem of multi-instance learning and semi-supervised anomaly detection,and achieves the following innovations:1.An embedding-based method MIKI is proposed,which is insensitive to key in-stance shift.Traditional MIL studies are mostly based on the i.i.d.assumption,and has rarely considered the distribution shift problem.In this thesis,we propose the MIKI method.We first do instance prototypes learning,and transform the original bag to a single vector representation,so that the transformed bag vectors keep both bag-level information and the key instance shift information.In addition,we learns the impor-tance weights for the training data to narrow the distribution gap between training and test data.Experimental results show that the proposed method performs significantly better than the existing MIL algorithms.2.A two-stage method ADOA is proposed for the task of semi-supervised anomaly detection.Previous studies of anomaly detection mainly focus on the supervised and unsupervised setting,and few studies focus on the setting in which only a handful of labeled anomalies and a large amount of unlabeled data can be obtained.In this thesis,ADOA first do clustering for the observed anomalies,and find potential anomalies and reliable normal examples from the unlabeled data according to their isolation degree and the similarity degree to the observed anomalies.Furthermore,a weighted model is trained.Experiments demonstrate the effectiveness of the proposed method in semi-supervised anomaly detection.