Abstract
In this paper, we propose an unsupervised learning algorithm to predict an anomaly detection which loyal customers lose their loyalty. A flow of our method is that finds characteristic items for loyal customers, enumerates sequential item patterns from their historical purchasing data, and expresses normal or abnormal state for the sequential patterns as a score. Then for each customer, total score is calculated as a difference between aggregated normal and abnormal scores which the customer has. When the total score is minus, we detect an anomaly. From some computational experiments using a practical POS data, we show that our method is suitable for practical use.
