[4L3-GS-4] Web intelligence

[4L3-GS-4-01] Active Learning from the Web

〇Ryoma Sato^1,2 (1. Kyoto University, 2. RIKEN AIP)

Keywords:Web Mining, Active Learning

Active learning is a technique that aims to reduce the cost of labeling by repeatedly selecting data to be labeled from a pool of unlabeled data. Many methods have been proposed for criteria to select data from the pool. However, how to construct a pool has been less explored, and most methods assume that task-specific pools are available free of charge. In this paper, we advocate that task-specific pools are not always available and propose using the myriad of unlabeled data on the Web for active learning pools. The problem is that the pool is so large that it is not possible to exhaustively compute the acquisition function for all the data. We propose a method to efficiently acquire useful data from the web in terms of active learning using a user-side search algorithm. In our experiments, we use the online Flickr environment as an active learning pool. This pool contains more than 10 billion images, which is several orders of magnitude larger than existing active learning pools. The proposed method shows better performance than existing methods using smaller pools.