REGULAR PATTERN MINING (WITH JITTER) ON WEIGHTED-DIRECTED DYNAMIC GRAPHS

• Main Authors: A. GUPTA, H. K. THAKUR, T. GUPTA, S. YADAV
• Format: Article
• Language: English
• Published: Taylor's University 2017-02-01
• Series: Journal of Engineering Science and Technology
• Subjects: Regular patterns; Jitter; Quasi-regular patterns; Centrality; Impact factor
• Online Access: http://jestec.taylors.edu.my/Vol%2012%20issue%202%20February%202017/12_2_5.pdf

Real world graphs are mostly dynamic in nature, exhibiting time-varying behaviour in structure of the graph, weight on the edges and direction of the edges. Mining regular patterns in the occurrence of edge parameters gives an insight into the consumer trends over time in ecommerce co-purchasing networks. But such patterns need not necessarily be precise as in the case when some product goes out of stock or a group of customers becomes unavailable for a short period of time. Ignoring them may lead to loss of useful information and thus taking jitter into account becomes vital. To the best of our knowledge, no work has been yet reported to extract regular patterns considering a jitter of length greater than unity. In this article, we propose a novel method to find quasi regular patterns on weight and direction sequences of such graphs. The method involves analysing the dynamic network considering the inconsistencies in the occurrence of edges. It utilizes the relation between the occurrence sequence and the corresponding weight and direction sequences to speed up this process. Further, these patterns are used to determine the most central nodes (such as the most profit yielding products). To accomplish this we introduce the concept of dynamic closeness centrality and dynamic betweenness centrality. Experiments on Enron e-mail dataset and a synthetic dynamic network show that the presented approach is efficient, so it can be used to find patterns in large scale networks consisting of many timestamps.