Bounding Entities within Dense Subtensors.

Group-based fraud detection is a promising methodology to catch frauds on the Internet because 1) it does not require a long activity history for a single user; and 2) it is difficult for fraudsters to avoid due to their economic constraints. Unfortunately, existing work does not cover the entire picture of a fraud group: they either focus on the grouping feature based on graph features like edge density, or probability-based features, but not both. To our knowledge, we are the first to combine these features into a single set of metrics: the complicity score and fraud density score. Both scores allow customization to accommodate different data types and data distributions. Even better, algorithms built around these metrics only use localized graph features, and thus scale easily on modern big data frameworks. We have applied our algorithms to a real production dataset and achieve state-of-the-art results comparing to other existing approaches.