ABFlow: Alert Bursting Flow Query in Streaming Temporal Flow Networks

Flow analyses on temporal flow networks have recently been found to be an appealing tool for a wide range of applications. For example, in financial fraud detection applications, suspicious activities can be alerted by bursty, substantial transfers and require continuous monitoring. Although determining bursting flows in static temporal networks has recently been proposed, its high complexity limits its applications to the emerging streaming scenarios. Motivated by these, we study the novel bursting flow query in the streaming scenario and propose the A lert B ursting Flow (ABFlow) query. Specifically, given two groups of nodes, S and T , and a stream of temporal flow networks, our goal is to find the flow with maximum burstiness from S to T within the stream being monitored, where the burstiness of a flow is defined as the ratio of the flow value to the flow duration. To solve this query, we propose a novel suffix flow problem, which leads to a practical incremental solution. Based on this solution, we further propose i) a novel constraint that enables a reduced-complexity recursive solution, and ii) optimizations for streaming, yielding a solution called SuffixFlow str . Our experiments verify that SuffixFlow str is up to two orders of magnitude faster than a baseline. Two case studies on real-world datasets showcase the anomaly detection applications of the ABFlow query.