Bag-level active multi-instance learning

Multi-Instance Learning (MIL) is a special scheme in machine learning. In recent research it is successfully applied in text classification problem. However, MIL is naturally semi-supervised since the instances labels are unknown for positive bags, which would cut down the accuracy of predictors, or require more computational cost to reduce uncertainty, or to guess such labels at a high probability. In this paper, we attempt to tackle MIL problem by introducing active learning, which is another learning scheme attracted much research interests. Active learning relies on an oracle that can give ground truth labels as required. The proposed method is based on query for bags and it adopts a Fisher Information Matrix (FIM) based method to construct the criteria of query for oracle. We launch experiment on a famous text classification data set - 20 group news. Compared to the randomly selected query strategy as a baseline method and recent methods, the proposed method is of higher accuracy and outperforms others.