Improving fuzzy C-mean-based community detection in social networks using dynamic parallelism

Abstract

In Social Network Analysis (SNA), a common algorithm for community detection iteratively applies three phases: spectral mapping, clustering (using either the Fuzzy C-Means or the K-Means algorithms) and modularity computation. Despite its effectiveness, this method is not very efficient. A feasible solution to this problem is to use Graphics Processing Units. Moreover, due to the iterative nature of this algorithm, the emerging dynamic parallelism technology lends itself as a very appealing solution. In this work, we present different novel GPU implementations of both versions of the algorithm: Hybrid CPU-GPU, Dynamic Parallel and Hybrid Nested Parallel. These novel implementations differ in how much they rely on CPU and whether they utilize dynamic parallelism or not. We perform an extensive set of experiments to compare these implementations under different settings. The results show that the Hybrid Nested Parallel implementation provide about two orders of magnitude of speedup.

Introduction

Due to their importance and popularity, the attention in Social Networks (SNs) has significantly increased recently. A SN is defined as a set of vertices or nodes (representing individuals, organizations, countries, etc.) connected by links or edges representing relationships between them such as friendship, ownership, alliances, etc. With their billions of users, studying Online SNs (OSNs), such as Facebook, Twitter, etc., is of great importance to many parties in the academia as well as the industry, which gave a great boost to the field of SN Analysis (SNA).

Community detection is one of the very basic and fundamental tasks in SNA and it is concerned with finding sub-groups such that the nodes in each group share certain common characteristics. This is very useful in many practical applications such as in recommendations and advertisement [1]. Many algorithms were proposed for community detection. One of the popular ones was originally proposed by Newman and Girvan in 2004 (henceforth, referred to as NG) and was later extended by Zhang et al. in 2007 (henceforth, referred to as ZWZ). Despite their effectiveness, such algorithms do not scale well.

The SNs being studied these days have millions or even billions of nodes and edges [2]. This means that, without efficient implementations, SNA algorithms (such as community detection algorithms) will be impractical [3]. Graphics Processing Units (GPUs) represent a good option to address this problem. It is now affordable to obtain a laptop equipped with a very powerful GPU consisting of hundreds or thousands of cores.

In this work, we aim at speeding up the process of extracting communities in SNs by using the parallelization capabilities of the GPUs to. To be more specific, our focus will be on the ZWZ algorithm of [1], which comprises of three stages: spectral mapping, clustering and modularity computation. We exploit the GPU capabilities to improve the performance of the algorithm by speeding up the both the modularity and the clustering operations. We employ different parallelization techniques (some of which have already been presented in our earlier work, while the others are novel) and conduct various experiments.

The remainder of this paper is organized as follows. The following section summarizes the related works on community detection algorithms and GPU-based implementations. Section 3 presents our sequential implementations of the different versions of the algorithm under consideration. Section 4 discusses our parallel implementations. Section 5 presents the experimental setup and the results. Finally, the paper is concluded in Section 6.