Selective Sampling via Similarity Factor for Sequential Imbalanced Data

Considering labeling problems in some online processes, sampling strategies with limited budget would save both time and cost for the operator; while keeping results in an acceptable range. Selective sampling has made supervised learning possible for some expensive or inapplicable online processes. A good sampling strategy give this opportunity to label prediction system to have labels that will improve system’s accuracy. An ideal sampling strategywould query labels of data points that have been classified in wrong class. Sampling problem has common characteristics with anomaly detection problem. This paper proposes a selective sampling method based on principal curves. Proposed algorithm returns a similarity factor between input stream and existing structures of predictor system in each step. No randomness is involved in the proposed algorithm to make it applicable in sensitive real-world problems.

Keywords:

selective sampling , query budget , principal curves