HIGCALS: a hierarchical graph-theoretic clustering active learning system

Active learning aims at automatically selecting highly informative unseen data for humans to label under the condition that only few labeled samples are ready for use. Most of previous researches in active learning take advantage of supervised learning. Due to semantic gap problem, these existing active learning algorithms are not very effective on lessening human labeling efforts, especially in multiclass applications. In this paper, we propose a novel active learning framework based on unsupervised learning, and implement a hierarchical graph-theoretic clustering active learning system (HIGCALS). HIGCALS outperforms the existing active learning systems in several aspects, such as flexibility in system architecture and simpleness in system upgrade. Experiments on KDDCUP99 data set has demonstrated that HIGCALS can effectively reduce the workload of manual labeling without losing much accuracy.