Efficient Discovery of Generalized Sentinel Rules

This paper proposes the concept of generalized sentinel rules (sentinels) and presents an algorithm for their discovery. Sentinels represent schema level relationships between changes over time in certain measures in a multi-dimensional data cube. Sentinels notify users based on previous observations, e.g., that revenue might drop within two months if an increase in customer problems combined with a decrease in website traffic is observed. If the vice versa also holds, we have a bi-directional sentinel, which has a higher chance of being causal rather than coincidental. We significantly extend prior work to combine multiple measures into better sentinels as well as auto-fitting the best warning period. We introduce two novel quality measures, Balance and Score, that are used for selecting the best sentinels. We introduce an efficient algorithm incorporating novel optimization techniques. The algorithm is efficient and scales to very large datasets, which is verified by extensive experiments on both real and synthetic data. Moreover, we are able to discover strong and useful sentinels that could not be found when using sequential pattern mining or correlation techniques.