Feature Selection in Large Scale Data Stream for Credit Card Fraud Detection

Ise Masayuki
Niimi Ayahiko
Konishi Osamu
There is increased interest in accurate model acquisition from large scale data streams. In this paper, because we have focused attention on time-oriented variation, we propose a method contracting time-series data for data stream. Additionally, our proposal method employs the combination of plural simple contraction method and original features. In this experiment, we treat a real data stream in credit card transactions because it is large scale and difficult to classify. This experiment yields that this proposal method improves classification performance according to training data. However, this proposal method needs more generality. Hence, we'll improve generality with employing the suitable combination of a contraction method and a feature for the feature in our proposal method.