Hardeo Kumar Thakur

Mining Regular Patterns in Weighted-Directed Networks

Mining of regular patterns in dynamic networks finds immense application in characterizing the local properties of the networks, like behaviour (friendship relation), event occurrence (football matches). They in then are used to predict their future trends. But if they do not entail weight and direction aspects of the dynamic network, there can be loss of several significant details, such as strength of a relationship or event, specification of the person responsible for it in a relationship, winning or losing in case of events. To the best of our knowledge, no work has been reported yet to extract regular patterns that take into account weight and direction aspects of dynamic networks. We thus propose a novel method to mine regular patterns in weighted and directed networks. In the proposed method, different snapshots of the dynamic network are taken, and through the concept of Regular Expression, we obtain repetition rule for each of: occurrence sequence, weight sequence, direction sequence and weight-direction sequence. For each of these four categories, edges having same rule are grouped to obtain evolution patterns. To ensure the practical feasibility of the approach, experimental evaluation is done on the real world dataset of Enron emails. The results obtained show that, 2.39%, 6.92%, 9.96% and 1.81% of the edges are found to be regular on weight, direction, occurrence and weight-direction respectively.

Mining regular pattern in edge labeled dynamic graph

Static Graphs consist of a fixed sequence of nodes and edges which does not change over time, hence lack in providing the information regarding evolution of the network. In contrast, Dynamic Graphs to a greater extent relate to real-life events and so provide complete information about the network evolution. That is why many researchers [1, 2, 3, 5, 6, 7, 8, 9 and 10] have developed interest in mining of Dynamic Graphs. We feel, that the topic can be further sub-divided structurally into four major categories, which are mining of Labeled, Edge Unlabeled, Directed and Undirected Dynamic Graphs. However, the main focus of research till now is on the mining of Edge Unlabeled Dynamic Graphs. But the limitation is that it does not provide the complete insights of graphs where edge strengths i.e. weights are also changing with time. For example in case of Coauthor network mining in Unlabeled Dynamic Graphs gives information only about the occurrence of relation whereas that in Labeled Dynamic Graphs provides more detailed information like the number of paper published jointly at different instants of time. To address this problem, the present paper proposes a novel method to find out Weighted Regular Patterns in Edge Labeled Dynamic Graphs. The proposed method consists of creating a summary graph to find weight occurrence sequence of edges enabling to determine weighted regular patterns. The method is applied to real world dataset, PACS networks, to ensure its practical feasibility and to understand how Weighted Dynamic Graphs behave regularly over time.

MaDGM: Multi-aspect Dynamic Graph Miner

Dynamic graphs model time-varying interactions between related entities in a network. Extensive studies have been carried out on the mining of frequent, regular, and periodic behavior of such interactions. Some of the previous research focused on providing users the platform to mine periodic patterns on a single aspect (structure, weight, or direction) at a time. However, the designed tool needs to be run multiple times, to mine significant information that is gained by the study of multiple aspects simultaneously in a network. Moreover, it lacks capability of mining the regular patterns in a network, and the applicability of which lies in wide-ranging domains. In the present research, a tool, the Multi-aspect Dynamic Graph Miner, is proposed that fills up these gaps by providing users an integrated platform for mining regular and periodic patterns on multiple aspects. Besides the primary features, it facilitates easy visualization of tool output and provides a converter for the users of previous works that some portion of our software is based on. We further discuss its applicability by testing on real-world and synthetic datasets.

Evolution of regular directed patterns in dynamic social networks

Existing Dynamic Graph mining algorithms focus typically on finding patterns in undirected, unweighted and weighted dynamic networks ignoring the fact that some of them could be directed also. In this paper, we focus on finding regular evolution patterns in edges, outdegree and indegree of all the nodes, featuring consecutively at fixed time intervals during the growth of an unweighted and directed dynamic graph. Such regular patterns will help in finding the characteristics (such as popularity, inactiveness) of the nodes. A methodology of occurrence rule is proposed in order to determine regular evolution patterns, which are considered to be regular if they follow the same occurrence rule. The methodology is also used to find the patterns in the outdegree and indegree of all the nodes. These patterns are used to describe exhaustively the neighbourhood properties of dynamic graphs as in the social networks. To ensure its practical feasibility, the method has been applied to real world dataset on Facebook-like Forum network, and the results have shown that 37.6% of the edges are directed regular edges, and 59% of the total users, not having their indegree patterns are unpopular users.