GPU-Accelerated Maximal Bicliques Mining Framework for Large E-commerce Networks

Many of Taobaos important daily data mining tasks, such as anomaly attack detection and interest group detection, require efficient algorithmic solutions for mining specific graph patterns. The most common graph pattern is biclique which has a very dense structure and often contains rich implicit information. An important question to address is whether and how we can efficiently find all the interesting bicliques in large e-commerce networks, which is coined as the Maximal Biclique Enumeration problem (MBE). MBE involves enumerating all the maximal bicliques in the given graph, which is rather computationally expensive for large networks. However, recent research works on MBE haven't made good use of GPU, a very widely used high-speed computing resource. In this paper, we propose GMBE, a novel framework that achieves an efficient utilization of the power of GPUs to parallelize the MBE algorithm to find all maximal bicliques. We design a programmable API for data analysts to meet different business needs, enabling GMBE to become the middleware to effectively support various graph mining applications in e-commerce domain. Extensive experiments show that GMBE achieves significant (12X) speedup on average over the state-of-the-art MBE algorithms.