Many Heads are Better than One: Local Community Detection by the Multi-walker Chain

Shared-Memory Branch-and-Reduce for Multiterminal Cuts

“…There is a multitude of approaches adding and removing vertices greedily [2,10,27,30]. PageRank [35] is reported to be well suited for the problem [25] and there are multiple approaches that aim to make PageRank perform even better [1,4,26]. Another approach is to use machine learning methods such as geometric [43] or relational [28] neighborhood classifiers.…”

Graph analysis techniques and applications to bitcoin forensics

“…The process is reapplied on the created clusters (instead of the nodes) to e↵ectuate a bottom-up, scalable, hierarchical clustering. Our clustering approach inherits the flexibility of the underlying centrality model, and is applicable to all variations of graphs (un/directed and un/weighted graph), in contrast to most existing graph clustering algorithms that are designed and applicable to only one variant each [7,[34][35][36][37][38][39][40][41][42]. We present these ideas in Chapter 4.…”

“…These include [5] which used edge removal algorithm to cluster an undirected graph by recomputing total Markov entropy for every graph instance created by removal of each possible edge, to determine which edge to remove, so that the aggregate entropy decreases the most; [38] which studies multiple chained walker model to allow multiple walkers to explore the network, so that confinement probability of random walks is increased within a local community which contains the query (start of the walk) node; and [39] which was an extension of [38] to support multiple query nodes. A variant of community detection using Markov chain is [41] where a random walk starts from a link and transition probability of Markov chain is used as similarity between a link pair.…”

“…A variant of community detection using Markov chain is [41] where a random walk starts from a link and transition probability of Markov chain is used as similarity between a link pair. Our flow confinement based clustering algorithm (discussed in Chapter 4) uses similar idea of exploiting the probability distribution gradient as explored in [38] [92] to classify unlabelled data from labelled data by combining data class distribution and random walk. Later, the same authors extend their work [92] by using Gaussian random field [93] model through random walk in [94].…”

“…In the studies [38,162], local clusters are detected using multi-agent random walks, using the intuition that multiple walkers are more likely to stay confined within local communities is such structures exist in a network, than escape the local communities simultaneously. These studies inspire us to explore whether and how the local communities are inferred from the gradient observed in the evolution of the entropic centralities but with the use of a single random walker, by initiating the random walks at nodes with low entropic centrality since nodes at the interface of multiple local communities are expected to have higher centrality.…”