Limited random walk algorithm for big graph data clustering
 Honglei Zhang^{1}Email authorView ORCID ID profile,
 Jenni Raitoharju^{1},
 Serkan Kiranyaz^{2} and
 Moncef Gabbouj^{1}
DOI: 10.1186/s4053701600605
© The Author(s) 2016
Received: 23 September 2016
Accepted: 15 November 2016
Published: 1 December 2016
Abstract
Graph clustering is an important technique to understand the relationships between the vertices in a big graph. In this paper, we propose a novel randomwalkbased graph clustering method. The proposed method restricts the reach of the walking agent using an inflation function and a normalization function. We analyze the behavior of the limited random walk procedure and propose a novel algorithm for both global and local graph clustering problems. Previous randomwalkbased algorithms depend on the chosen fitness function to find the clusters around a seed vertex. The proposed algorithm tackles the problem in an entirely different manner. We use the limited random walk procedure to find attractor vertices in a graph and use them as features to cluster the vertices. According to the experimental results on the simulated graph data and the realworld big graph data, the proposed method is superior to the stateoftheart methods in solving graph clustering problems. Since the proposed method uses the embarrassingly parallel paradigm, it can be efficiently implemented and embedded in any parallel computing environment such as a MapReduce framework. Given enough computing resources, we are capable of clustering graphs with millions of vertices and hundreds millions of edges in a reasonable time.
Keywords
Graph clustering Random walk Big data Community findingBackground
Graph data are important data types in many scientific areas, such as social network analysis, bioinformatics, and computer and information network analysis [1]. In recent years, the size of graph data has grown dramatically. For example, a typical social network graph may contains millions of vertices and hundreds of million of edges. Further more, these graphs may continuously evolve over time. Processing these dynamic big graph data is very challenging and time consuming. In general, big graphs are normally heterogeneous. They have such nonuniform structures that edges between vertices in a group are much denser than edges connecting vertices in different groups. Graph clustering (also named as “community detection” in the literature) algorithms aim to reveal the heterogeneity and find the underlying relations between vertices [2]. This technique is critical for understanding the properties, predicting dynamic behavior and improving visualization of big graph data.
Graph clustering is a computationally challenging and difficult task, especially for big graph data. Many algorithms have been proposed over the last decades [2–4]. The criteriabased approaches try to optimize clustering fitness functions using different optimization techniques. Newman defined a modularity measurement based on the probability of the link between any two vertices. He applied a greedy search method to minimize this modularity fitness function in order to partition a graph into clusters [5]. Blondel et al. used the same fitness function but combined it with other optimization techniques [6–8]. Spielman and Teng opted the graph conductance measurement as the fitness function [9]. Other than criteriabased methods, spectral analysis has also been widely adapted in this area [10, 11]. Randomwalkbased methods tackle the problem from a different angle [12–14]. These methods use the Markov chain model to analyze the graph. Each vertex represents a state and the edges indicate transitions between the states. The probability values that are distributed among the states (vertices) reveal the graph structure.
For big graph data, the problem becomes more challenging or even intractable. Very often, people are only interested in finding the cluster for a given seed vertex. This problem is called local clustering problem [9, 15, 16]. For example, from an end user’s perspective, finding the closely connected friends around him or her is more important than revealing the global user clusters of a large social network. It is unnecessary to explore the whole graph structure for this problem. Recently, random walk methods have gained great attention on this local graph clustering problem, since a walk started from the seed vertex is more likely to stay in the cluster where the seed vertex belongs. Comparing to the criteriabased methods, the randomwalkbased methods are capable of extracting local information from a big graph without the knowledge of the whole graph data. In [17–19], a random walk is first applied to find important vertices around the seed vertex. Then a sweep stage is involved to select the vertices that minimize the conductance of the candidate clusters.
The accuracy of any criteriabased clustering method (or those combined with the random walk procedures) is greatly affected by the chosen clustering fitness function. Furthermore, most local clustering algorithms use the criteria that are more suitable for the global graph clustering problem. These choices greatly degrade the performance of these algorithms when the graph is big and highly uneven. Also the majority of the graph clustering algorithms are designed in sequential computing paradigm. Therefore, they do not take advantage of modern highperformance computing systems.
In this paper, we propose a novel randomwalkbased graph clustering algorithm—the limited random walk (LRW) algorithm. First of all, the LRW algorithm does not rely on any clustering fitness function. Furthermore, the proposed method can efficiently tackle the computational complexity using a parallel programming paradigm. Finally, as a unique property among many graph clustering methods, LRW can be adapted to both global and local graph clustering in an efficient way.
The rest of the paper is organized as follows: basics of random walk procedure and the proposed LRW algorithm are explained in "Methodology" section; an extensive set of experiments on the simulated and real graph data, along with both numerical and visual evaluations are given in "Experiments" section; finally, the conclusions and future work are discussed in "Conclusions" section.
Methodology
Basic definitions and the random walk procedure
Let G(V, E) denote a graph of n vertices and m edges, where \(V=\left\{ v_{i}i=1,\ldots n\right\}\) is the set of vertices and \(E=\left\{ e_{i}i=1,\ldots m\right\}\) is the set of edges. Let \(A\in \mathcal {R}^{n\times n}\) be the adjacency matrix of the graph G and \(A_{ij}\) are the elements in the matrix A. Let \(D\in \mathcal {R}^{n\times n}\) be the degree matrix, which is a diagonal matrix whose elements on the diagonal are the degrees of each vertex. In this paper, we assume the graph is undirected, unweighted and does not contain selfloops.
Clustering phenomenon is very common in big graph data. A cluster in a graph is a vertex set where the density of the edges inside the cluster is much higher than the density of edges that link the inside vertices and the outside vertices.
Random walk on a graph is a simple stochastic procedure. At the initial state, an agent stays on a chosen vertex (seed vertex). At each step, the agent randomly picks a neighboring vertex and moves to it. The agent repeats this movement and there is certain probability that the agent lands on a vertex after each movement.
Let \(x_{i}^{(t)}\) denote the probability that the agent is on vertex \(v_{i}\) after step t, where \(i=1,2,\ldots n\). \(x_{i}^{(0)}\) is the probability of the initial state. Let s be the seed vertex. We have \(x_{s}^{(0)}=1\), and \(x_{i}^{(0)}=0\) for \(i\ne s\). Let \(x^{(t)}=\left[ x_{1}^{(t)},x_{2}^{(t)},\ldots ,x_{n}^{(t)}\right] ^{T}\) be the probability vector, where the superscript T denotes the transpose of a matrix or a vector. By the definition of the probability, it is easy to see that \(\sum _{i=1}^{n}x_{i}^{(t)}=1\) or \(\left\ x^{(t)}\right\ _{1}=1\).
For an undirected, connected and aperiodic graph, there exists an equilibrium state \(\pi\), such that \(\pi =P\pi\). This state is unique and irrelevant to the starting point. By iterating Eq. 3, \(x^{(t)}\) converges to \(\pi\). More details about the Markov chain process and the equilibrium state can be found from [20].
Limited random walk procedure
Definitions
At each walking step, the probability vector \(x^{(t)}\) is computed using Eq. 3. Note that, in general, elements in \(x^{(t)}\) that are around the seed vertex are nonzeros and the rest are zeros. So we do not need the full transition matrix to calculate the probability vector for the next step.
We call the aforementioned procedure the limited random walk (LRW) procedure. Comparing to the normal random walk procedure defined in "Basic definitions and the random walk procedure" section, LRW involves inflation and normalization operations in each walking step. These nonlinear operations limit the agent to walk around the neighborhood of the seed vertex, especially if there is a clear graph cluster boundary.
The MCL algorithm simulates flow within a graph. It uses the inflation and normalization operation to enhance the flow within a cluster and reduce the flow between clusters. The MCL procedure is a timeinhomogeneous Markov Chain in which the transition matrix varies over time. The MCL algorithm starts the random walk from all vertices simultaneously—there are n agents walking on the graph at the same time. The walking can only continue after all agents have completed a walking step and the result probability matrix has been inflated and normalized. Unlike in the MCL algorithm, the LRW procedure is a timehomogeneous Markov Chain. We initiate random walk from a single seed vertex, and do the inflation on the probability values of this walking agent. This design has many advantages. First, it avoids unnecessary walks since the graph structure around the seed vertex may be exposed by a single walk. Second, the procedure is suitable for the local clustering problems because it does not require the whole graph data. Third, if multiple walks are required, each walk procedure can be executed independently. Thus the algorithm is fully parallelizable.
The LRW procedure involves a nonlinear operation, thus it is difficult to analyze its properties on a general graph model. Next we study the equilibrium of the LRW procedure.
Equilibrium of the LRW procedure
Theorem 1
There exists a fixedpoint \(x^{*}\) such that \(\mathcal {L}\left( x^{*}\right) =x^{*}\).
Proof
We use the Brouwer fixedpoint theorem to prove this statement.
Given \(0\le x_{1},x_{2},\cdots ,x_{n}\le 1\), the set X is clearly bounded and closed. Thus X is a compact set.
Let \(u,v\in X\) and \(w=\lambda u+(1\lambda )v\), where \(\lambda \in \mathcal {R}\) and \(0\le \lambda \le 1\). So \(w_{i}=\lambda u_{i}+(1\lambda )v_{i}\) for \(i=1,2,\cdots n\). Obviously \(0\le w_{i}\le 1\).
Since function f(x) is continuous over the set X and function g(x) is continuous over the codomain of function f(x), function \(\mathcal {L}\) is continuous over the set X.
Given \(\mathcal {L}\) is a continuous function that maps a convex set to itself, according to the Brouwer fixedpoint theorem, there is a point \(x^{*}\) such that \(\mathcal {L}\left( x^{*}\right) =x^{*}\). \(\square\)
Theorem 1 shows the existence of fixedpoint of the LRW procedure, i.e., the LRW procedure will not escape from a fixedpoint whenever the point is reached. Since the LRW procedure is a nonlinear discrete dynamic system, it is difficult to analytically investigate the system behavior. However, when \(r=1\), the LRW procedure is simply a Markov chain process, in which the fixedpoint \(x^{*}\) is the unique equilibrium state \(\pi\) and the global attractor. In another extreme case when \(r\rightarrow \infty\), a fixedpoint can be an unstable equilibrium and the LRW procedure may have limit cycles that oscillate around a star structure in the graph. In one state of the oscillation, the probability value of the center of a star structure is close to one. In practice, we chose r from (1, 2]. This makes the LRW procedure close to a linear system and oscillations are extremely rare. In this case, the fixedpoints of the LRW procedure are stable equilibriums.
Limited random walk on general graphs
Without any prior knowledge of the cluster formation, we normally start the LRW procedure from an initial state where \(x_{s}=1\), \(x_{i}=0\) for \(i\ne s\) and s is the seed vertex. During the LRW procedure, there are two simultaneous processes—the spreading process and the contracting process. When the two processes can balance each other, a stationary state is reached.
During the spreading process, the probability values spread as the walking agent visits new vertices. The number of visited vertices increases exponentially at first. The growth rate depends on the average degree of the graph. The newly visited vertices will always receive the smallest probability values. If the graph has an average degree of d, it is not difficult to see that the expected probability value of a newly visited vertex at step t is \(\left( {1}/{d}\right) ^{t}\). As the walking continues, the probability values tend to be distributed more evenly among all visited vertices.
The other ongoing process during the LRW procedure is the contracting process. During this process, the probability values of the visited vertices contract to some vertices. Since the graph is usually heterogeneous, some vertices (and groups of vertices) will receive higher probability values as the procedure continues. The inflation operation further enhances this contracting effect. The degrees of a vertex and its surrounding vertices determine whether the probability values concentrate to or diffuse from these vertices. Some vertices, normally the center of a star structure, receive larger probability values than others. We call these vertices attractor vertices and they can be used to represent the structure of a graph.
Because the density of edges inside a cluster is higher than that of linking the vertices inside and outside the cluster, the probability that a walking agent visits vertices outside the cluster is small. Thus, the LRW procedure will find attractor vertices that the seed vertex is associated. We can use these vertices as features to cluster the vertices.
The larger the inflation exponent r is, the faster the algorithm converges to the attractor vertices. The LRW procedure tries to find the attractor vertices that are near the seed vertex. However, if r is too large, the probability values concentrate to the nearest attractor vertex (or the seed vertex itself) before the graph is sufficiently explored. If r is too small, the probability values will concentrate to the attractor vertices that may belong to other clusters. The performance of the LRW algorithm depends on choosing a proper inflation exponent r. From this aspect, it is similar to the MCL algorithm. In practice, r is normally chosen between 1 and 2 and the value 2 was found to be suitable for most graphs.
LRW for global graph clustering problems
In this section, we propose how to LRW in global graph clustering problems. Our algorithm is divided into two phases—graph exploring phase and cluster merging phase. To improve the performance on big graph data, we also propose a multistage strategy.
Graph exploring phase
In the graph exploring phase, the LRW procedure is started from several seed vertices. At each iteration, the agent moves one step as defined in Eq. 3. Then the probability vector x is inflated by Eq. 6 and normalized by Eq. 7. The iteration stops when the probability vector x converges or the predefined maximum number of iterations is reached. Let \(x^{(*,i)}\) denote the final probability vector of a random walk that was started from the seed vertex \(v_{i}\). As described in the previous section, the LRW procedure explores the vertices that are close to the seed vertex. Thus, the vector \(x^{(*,i)}\) has nonzero elements only on these neighboring vertices.
Algorithm 1 illustrates the graph exploring from a seed vertex set Q. Note that for small graph data, we can set the seed vertex set \(Q=V\) (i.e. the whole graph). In such case, the LRW procedure is executed on every vertex of the graph and the multistage strategy is not used.
Note that the threshold \(\epsilon\) limits the number of nonzero elements in the probability vector x. It is easy to prove that the number of nonzero elements in \(x^{(t,i)}\) is less than \({1}/{\epsilon }\). A larger \(\epsilon\) eliminates very small values in \(x^{(t,i)}\) and prevent unnecessary computing efforts. However, \(\epsilon\) does not impose a limit on the largest cluster we can find. Further, the choice of \(\epsilon\) has little impact on the final clustering results because either the LRW procedure finds the most dominant attractor vertices in a cluster or the small clusters are merged in the cluster merging phase.
Cluster merging phase
After the graph has been explored, we will find the clusters in the cluster merging phase. We treat each \(x^{(*,i)}\) as the attractor vector for the vertex \(v_{i}\). Vertices belonging to the same cluster have attractor vectors that are close to each other. Any unsupervised clustering algorithm, such as kmeans or single linkage clustering method, can be applied to find the desired number of (k) clusters. Because of the computational complexity of these clustering algorithms, we design a fast merging algorithm that can efficiently cluster vertices according to their attractor vectors.
Each element \(x_{j}^{(*,i)}\) in \(x^{(*,i)}\) is the probability value of the stationary state that the walking agent hits the vertex \(v_{j}\) when the seed vertex is \(v_{i}\). The attractor vector \(x^{(*,i)}\) is determined by the graph structure of the cluster that the initial vertex \(v_{i}\) has. Thus, vertices in the same cluster should have very similar attractor vectors. We first find the vertex that has the largest value in the vector \(x^{(*,i)}\). Suppose \(m=\arg \max _{j}\left( x_{j}^{(*,i)}\right)\), we call \(v_{m}\) the attractor vertex of vertex \(v_{i}\). Grouping vertices by their attractor vertex can be done in a fast way (complexity of \(O\text{(1) }\)) using a dictionary data structure. After the grouping, each vertex is assigned to a cluster that is identified by the attractor vertex. However, it is possible that some vertices in one cluster do not have the same attractor vertex. This may happen when the cluster is large and the edge density in the cluster is low. We then apply the following cluster merging algorithm to handle this overclustering problem.
The vertices that have large values, which are determined by a threshold relative to \(x_{m}^{(*,i)}\), in \(x^{(*,i)}\) are called significant vertices for vertex \(v_{i}\). If two vertices have large enough overlaps of their significant vertices, they should be grouped into the same cluster. From this observation, we first collect significant vertices for the found clusters. Then we merge clusters if their significant vertices overlap more than a half. Note that the attractor vertex and the significant vertices are always in the same cluster as the seed vertex. This is very useful when we use the multistage graph strategy.
Algorithm 2 shows the details of the merging phase of the LRW algorithm. Note that, for small graph data, we set the seed vertex set Q = V and the initial clustering dictionary \(\mathcal {D}\) to be empty.
Multistage strategy
For small graph data, we can do the LRW procedure on every vertex of the graph. So the seed vertex set \(Q=V\). The graph clustering is completed after a graph exploring phase and a cluster merging phase. However, when the graph data is large, it is timeconsuming to perform the LRW procedure from every vertex of the graph. A multistage strategy can be used to greatly reduce the number of required walkings. First, we start the LRW procedure from a randomly selected vertex set. After the first round of the graph exploring, some clusters can be found after the cluster merging phase. Next we generate a new seed vertex set by randomly selecting vertices from those vertices that have not been clustered. Then we do the graph exploration from the new seed vertex set. We repeat this procedure until all vertices are clustered.
Algorithm 3 shows the global graph clustering algorithm using the multistage strategy.
LRW for local graph clustering problems
For the local graph clustering problems, the LRW procedure can efficiently find the cluster from a given seed vertex. To achieve this, we first perform graph exploring from the seed vertex in the same way as described in "Graph exploring phase" section. Let \(x^{(*)}\) be the probability vector after the graph exploration. If a probability value in \(x^{(*)}\) is large enough, the corresponding vertex is assigned to the local cluster without further computation. Similar to the global graph clustering algorithm, we use a relative threshold \(\eta\) that is related to the maximum value in \(x^{(*)}\). Vertices whose probability values are greater than \(\eta \cdot \max \left( x_{j}^{(*)}\right)\) are called significant vertices. The significant vertices are assigned to the local cluster directly. A small value of \(\eta\) will reduce the computational complexity, but may decrease the accuracy of the algorithm. Suitable values of \(\eta\) were experimentally found to be between 0.3 and 0.5.
The vertices with low probability values can either be outside the cluster or inside the cluster but with relatively low significance. Unlike [9, 15, 16], which involve a sweep operation and a cluster fitness function, we do another round of graph exploring from these insignificant vertices. After the second graph exploring is completed, we apply the cluster merging algorithm described in "Cluster merging phase" section.
Algorithm 4 presents the LRW local clustering algorithm.
Computational complexity
We first analyze the computational complexity of the LRW algorithm for the global graph clustering problem. We assume the graph G(V, E) has clusters. Let \(\bar{n}_{c}\) be the average cluster size—the number of vertices in the cluster, and C is the number of clusters. We have \(\bar{n}_{c}\cdot C=n\). Note \(C\ll n\). The most timeconsuming part of the algorithm is the graph exploring phase. For each vertex, every iteration involves a multiplication of the transition matrix P and the probability vector x. The LRW procedure visits not only the vertices in the cluster but also a certain amount of vertices close to the cluster. Let \(\gamma\) be the coefficient that indicates how far the LRW procedure explores the graph before it converges. Notice the maximum number of nonzero elements in a probability vector is \({1}/{\epsilon }\). Let J denote the number of vertices that the LRW procedure visits in each iteration, thus \(J=\min \left( \gamma \bar{n}_{c},{1}/{\epsilon }\right)\). Thus the transition step at each iteration has complexity of \(O\left( J\bar{n}_{c}\right)\). The inflation and normalization steps, which operate on the probability vector x, have the complexity of \(O\left( J\right)\). Let K be the number of iterations for the LRW procedure to converge. So, the computational complexity for a complete LRW procedure on each vertex is \(O(KJ\bar{n}_{c})\). For a global clustering problem when performing the LRW procedure on every vertex, the graph exploration phase has a complexity of \(O\left( KJ\bar{n}_{c}n\right)\). In the worst case, the algorithm has a complexity of \(O\left( n^{3}\right)\). This is an extremely rare case and it only happens when the graph is small; does not have a cluster structure; and the edge density is high. This worst case scenario is identical to the MCL algorithm [12]. Notice that the variables J and K have upper bounds and \(\bar{n}_{c}\) is determined by the graph structure, the algorithm has a complexity of O(n) for big graph data.
The computational complexity of the cluster merging phase involves merging clusters that were found using the attractor vertices. This merging requires \({C \atopwithdelims ()2}\) times of set comparison operations, where C is the number of clusters found by the attractor vertices. The complexity of this phase is roughly \(O(C^{2})\). This does not impose a significant impact to the overall complexity of the algorithm, since \(C\ll n\). The time spent in this phase is often negligible. Experiments show that the clusters found using the attractor vertices are close to the final results. For applications where speed is more important than accuracy, the cluster merging phase can be left out.
When the LRW algorithm is used in local graph clustering problems, the first graph exploration (started from the seed vertex) has a complexity of \(O(KJ\bar{n}_{c})\). After the first graph exploration, there are LJ vertices to be further explored, where L is related to the threshold \(\eta\) and \(L<1\). The overall complexity of the LRW local clustering algorithm is thus \(O(LKJ^{2}\bar{n}_{c})\).
The LRW algorithm is a typical example of embarrassingly parallel paradigm. In the graph exploring phase, each random walk can be executed independently. Therefore it can be entirely implemented in a parallel computing environment such as a highperformance computing system. The time spent for graph exploring phase decreases roughly linearly with respect to the number of available computing resources. The twophase design also fits the MapReduce programming model and can easily be adapted into any MapReduce framework [23].
Experiments
The LRW algorithm uses the following parameters: inflation exponent r, maximum number of iterations \(T_{max}\), small value \(\epsilon\), merging threshold \(\tau\) and local clustering threshold \(\eta\). In practice, except the inflation exponent r, the values of the other parameters have little impact to the final results. The inflation power r should be chosen according to the density of the graph. A sparse graph should use a smaller value of r, though \(r=2\) is suitable for most real world graphs. In our experiments, we chose \(r=2\) unless otherwise specified. The other parameters have been set as: \(T_{max}=100,\) \(\epsilon =10^{5}\) and \(\tau =\eta =0.3\). We will show the impact of some parameters in "The sensitivity analysis of the parameters" section.
Simulated data for global graph clustering problem
We first show the performance of the LRW algorithm using simulated graph data. The simulated graph is generated using the ErdosRenyi model [24] with some modifications to generate clusters. Using the ground truth of the cluster structure, we can evaluate the performance of graph clustering algorithms. This kind of simulated data are widely used in the literature [5, 25–27].
The graphs are generated by the model G(n, p, c, q) where c is the number of clusters, n is the number of vertices, p is the probability of the link between two vertices, and \(q=d_{in}/d_{out}\) is the parameter that indicates the strength of the cluster structure, where \(d_{in}\) is the expected number of edges linking one vertex to other vertices inside the same cluster, and \(d_{out}\) is the expected number of edges linking a given vertex to other vertices in other clusters. Larger q indicates stronger cluster structure. When \(q=1\), each vertex has equal probability that it links to vertices that are inside and outside the cluster—the graph has a very weak cluster structure. Let d be the expected of degree of a vertex. So, \(d=d_{in}+d_{out}=p(n1)\). We use this model to generate graphs that consist of c clusters and each cluster has the same number of vertices. For each pair of vertices, we link them with the probability \(\frac{qpc(n1)}{(q+1)(nc)}\) if they belong to the same cluster, and the probability of \(\frac{pc(n1)}{n(q+1)(c1)}\) if they belong to different clusters.
Two simulated graphs are shown in Fig. 1, where the clusters are colored differently and the graphs are visualized by forcedirected algorithms.
The NMI values and the numbers of clusters of the clustering results on simulated graph data
q  GN  Louvain  Infomap  MCL  LRW 

4.0  0.975 (4)  1.0 (4)  1 (4)  1 (4)  1.0 (4) 
3.0  1 (4)  1.0 (4)  1 (4)  1 (4)  1.0 (4) 
2.33  0.950 (4)  1.0 (4)  1 (4)  0.860 (7)  1.0 (4) 
1.86  0.900 (4)  1.0 (4)  1 (4)  0.478 (95)  1.0 (4) 
1.5  0.890 (4)  1.0 (4)  0 (1)  0.453 (119)  0.975 (4) 
1.22  0.593 (4)  0.771 (5)  0 (1)  0.444 (128)  0 (1) 
1  0.232 (4)  0.304 (7)  0 (1)  0.444 (128)  0 (1) 
From the results, Louvain is the best performing algorithm and the LRW algorithm comes as the second. It can be seen that the LRW algorithm can find the correct structure if the graph has a strong cluster structure. When the cluster structure diminishes as q decreases, the walking agents quickly spread to the whole graph before the contraction dominates. Thus, the LRW algorithm returns the whole graph as one cluster. This behavior is beneficial when we need to find the true clusters in a big graph. The GN and Louvain algorithms are more like graph partition algorithms. They optimize certain cluster fitness functions using the whole graph data. They tend to partition the graph into clusters even though the cluster structure is weak. That explains their better NMI scores in Table 1 when q is small.
We use real graph data to evaluate the performance of the LRW global clustering algorithm on heterogeneous graphs. Details of the experiments and the results are given in "Real world data" section.
Simulated data for local graph clustering problems
In this section, we compare the LRW algorithm with other local clustering algorithms. The test graphs are generated using the protocol defined in [26]. To simulate the data that are close to real world graphs, the vertex degree and the cluster size are chosen to follow the power law. Each test graph contains 2048 vertices. The vertex degree has the minimum value of 16 and the maximum value of 128. The minimum and maximum cluster sizes are 16 and 256, respectively. Similar to the previous section, the inboundoutbound ratio q defines the strength of the cluster structure.
Jaccard index of local graph clustering results on the simulated graphs
q  CheG  CheS  NCutG  NCutS  IRDG  IRDS  LRW 

4.0  0.753  0.840  0.752  0.820  0.753  0.830  0.945 
3.0  0.671  0.812  0.671  0.801  0.671  0.798  0.927 
2.33  0.668  0.774  0.668  0.758  0.668  0.776  0.880 
1.86  0.593  0.650  0.593  0.681  0.593  0.684  0.823 
1.5  0.492  0.630  0.493  0.629  0.492  0.609  0.660 
1.22  0.444  0.544  0.437  0.549  0.444  0.529  0.504 
1  0.298  0.410  0.296  0.452  0.298  0.424  0.295 
From the results, it is obvious that the LRW algorithm clearly outperforms other methods when the graph has a clear cluster structure. For the same reason explained in "Simulated data for global graph clustering problem" section, it does not give good result if the cluster structure is weak. This is the main difference between the LRW algorithm and graph partition algorithms.
Real world data
In this section, we evaluate the performance of the LRW algorithm on some real world graph data.
Zachary’s karate club
We first do clustering analysis on the Zachary’s karate club graph data [31]. This graph is a social network of friendship in a karate club in 1970. Each vertex represents a club member and each edge represents the social interaction between the two members. During the study, the club split into two smaller ones due to the conflicts between the administrator and the coach. The graph data have been regularly used to evaluate the performance of the graph clustering algorithms [27, 30, 32]. The graph contains 34 vertices and 78 edges. We applied the LRW algorithm on this graph and the result shows two clusters that are naturally formed. Figure 2 shows the clustering result, where clusters are illustrated using different colors.
As the figure shows, the LRW algorithm finds the two clusters of the Zachary’s karate club. Actually the two clusters perfectly match the ground truth—how the club was split in 1970.
Clustering results of the GN, Louvain, Infomap and MCL algorithms are given in Additional file 1.
EgoFacebook graph data
The second data we used is the egoFacebook graph data [33]. The social network website Facebook allows users to organize their friends into “circles” or “friend lists” (for example, friends who share common interests). This data was collected from volunteer Facebook users for researchers to develop automatic circle finding algorithms. Egonetwork is the network of an end user’s friends. The egoFacebook graph is a combination of egonetworks from 10 volunteer Facebook users. There are 4039 vertices and 88234 edges in the graph.
We applied the LRW, GN [27], Louvain [6], Infomap [30] and MCL [12] graph clustering algorithms to this data. To compare the results, we generated the ground truth clustering by combining the vertices in the “circles” of each volunteer user. So, the ground truth contains 10 clusters. If a vertex appears in the circles of more than one volunteer, we assign the vertex to all of these ground truth clusters. We evaluated the number of clusters and the NMI values of the results that each competing algorithm generated.
We also calculated the mean conductance (MC) value of the clustering results. The conductance value of a cluster is calculated using Eq. 14. We then took the mean of all the conductance values of the clusters that an algorithm finds. Smaller MC values indicate better clustering results. Note that MC tends to favor smaller numbers of clusters in general. If the numbers of clusters are roughly the same, MC values give good evaluation of the clustering results. It is also worth noting that MC value is capable of evaluating clustering algorithms without the ground truth. We shall use this metric in later experiments where the ground truth is not available.
Global graph clustering results on the egoFacebook graph
GN  Louvain  Infomap  MCL  LRW  

Mean conductance  0.156  0.133  0.397  0.0882  0.0770 
NMI  0.778  0.796  0.723  0.908  0.910 
Number of clusters  16  19  76  10  10 
The clustering result of the LRW algorithms is shown in Fig. 3.
Clustering results of other algorithms are shown in the Additional file 1.
Heterogeneous graph data
Since there is no ground truth available for these test data, we evaluated each clustering algorithm by the mean conductance (MC) values. The results are in Table 5. The best MC scores are shown in a italic font. The numbers of clusters found by each algorithm are placed between parentheses. We also plot the clustered graphs in which the vertices are located using a forcedirected algorithm and colored according to their associated clusters. These clustered graphs are given in the Additional file 1 for subjective evaluation.
Global graph clustering results on the real heterogeneous graph data
GN  Louvain  Infomap  MCL  LRW  

Dolphins  0.425 (4)  0.440 (5)  0.487 (6)  0.675 (12)  0.347 (4) 
Arenesjazz  0.485 (4)  0.455 (4)  0.577 (7)  0.529 (5)  0.364 (4) 
Infectious  0.162 (5)  0.214 (6)  0.465 (17)  0.673 (40)  0.175 (5) 
Polblogs  0.524 (12)  0.501 (11)  0.727 (36)  0.777 (45)  0.427 (11) 
Reactome  0.108 (110)  0.099 (114)  0.315 (248)  0.478 (352)  0.221 (191) 
Based on the MC scores and the visualized clustering results, the LRW algorithm achieves a superior clustering performance in most of the cases. Note that the reactome data has a weak cluster structure, thus the LRW algorithm has difficulty to find a good partition for it.
The sensitivity analysis of the parameters
The proposed LRW algorithm depends on a number of parameters to perform global and local graph clustering. In this section, we perform the sensitivity analysis on the parameters.
We use both simulated and realworld graphs in our experiments. Test graph G1, G2 and G3 are similar to those used in "Simulated data for global graph clustering problem" section except that we vary the density of each graph. The expected degree d, which is a measure of the graph density, of graph G1, G2 and G3 are 12, 16 and 20 respectively and q is set to be 1.86 for all test graphs. Test graph G4 is generated in the same way as described in "Simulated data for local graph clustering problems" section. The egoFacebook graph data in "EgoFacebook graph data" section is used as an example of realworld graphs. We performed global graph clustering on these test graphs using the LRW algorithm with different parameters. the NMI scores are used to evaluate the performance of the algorithm. The experiments using simulated graph data were repeated 10 times and the average NMI scores and the average number of clusters are reported.
The NMI scores and the number of clusters by the different inflation exponent values
r  G1  G2  G3  G4  

1.2  0 (1)  0 (1)  0 (1)  0.155 (5)  0.902 (10) 
1.4  0 (1)  0 (1)  0 (1)  0.927 (22.7)  0.906 (10) 
1.6  0 (1)  0 (1)  0 (1)  1.0 (23)  0.908 (11) 
1.8  1 (4)  1 (4)  1 (4)  1.0 (20.8)  0.910 (10) 
2  0.971 (4.6)  1 (4)  1 (4)  1.0 (25)  0.910 (10) 
2.4  0.868 (7.0)  0.990 (4.2)  1 (4)  1.0 (21.8)  0.910 (10) 
3  0.822 (6.3)  0.962 (4.8)  0.988 (4.2)  1.0 (24.3)  0.910 (10) 
The test results show the relationship among the inflation exponent r, the density of the test graphs and the performance the LRW algorithm. A large inflation exponent r may overcluster the data as the results on graph G1 and G2 shows. It can be easily noticed that the LRW algorithm is not sensitive to the choice of r for test graph G4 and the egoFacebook graph. These graphs, and almost all realworld graphs, are more heterogeneous than the simulated graphs G1, G2 and G3. The LRW algorithm performs better on this type of graph since the attractor vertices and significant vertices are more stable on these graphs.
The parameter \(T_{max}\) sets a limit on the number of iterations for the LRW procedure to converge. According to our experiments, value 100 is large enough to ensure the convergence of almost all cases. For example, only 3 out of 88,234 LRW procedures do not converge within 100 iterations on the egoFacebook graph. A few exceptional cases has no impact on the final clustering results. The parameter \(\epsilon\) is used to remove small values in the probability vector thus decrease the computational complexity. It has no impact on the final clustering result as long as the value is small enough, for example \(\epsilon <10^{4}\).
We also conducted the experiments by varying the threshold value \(\tau\) from 0.1 to 0.5. The NMI scores and the number of clusters found by the LRW algorithm with different \(\tau\) values are almost identical to the values in the corresponding cells in Table 6. This indicates that the choice of \(\tau\) has very little impact on the clustering performance.
According to these results, one only needs to choose a proper inflation exponent r to use the LRW algorithms. Other parameters can be chosen freely from a wide range of reasonable values. \(r=2\) is suitable for most of graphs and is preferable because of the computational advantage.
Big graph data
Clustering performance of realworld big graph data
comAmazon  comYoutube  comLiveJournal  comOrkut  

Vertices  334,863  1,134,890  3,997,962  3,072,441 
Edges  925,872  2,987,624  34,681,189  117,185,083 
CPU cores (graph exploration)  12  96  96  96 
Memory per CPU core  4G  4G  8G  8G 
Graph exploration (in hours)  0.83  3.08  17.4  24.0 
Cluster merging (in hours)  0.20  1.34  9.44  2.53 
Clusters  37,473  170,569  381,246  165,624 
Rand index  0.908  0.755  0.951  0.751 
Table 7 shows that the LRW algorithm is able to find clusters from large graph data with a reasonable computing time and memory usage. The Rand index values indicate that the clusters returned by the LRW algorithm match well the ground truth. The time spent on the graph exploration phase is inversely proportional to the number of CPU cores. Computational time can be further reduced if more computing resources are available. The proposed algorithm can efficiently handle graphs with millions of vertices and hundreds of millions of edges. For even larger graphs that exceed the memory limit for each computing process, a mechanism that retrieves part of the graph from a central storage can be used. Since the LRW procedure is capable of exploring a limited number of vertices that are near a seed vertex, the algorithm can cluster much larger graphs if such a mechanism is implemented.
Conclusions
In this paper, we proposed a novel randomwalkbased graph clustering algorithm, the socalled LRW. We studied the behavior of the LRW procedure and developed the LRW algorithms for both global and local graph clustering problems. The proposed algorithm is fundamentally different from previous randomwalkbased algorithms. We use the LRW procedure to find attracting vertices and use them as features to cluster vertices in a graph. The performance of the LRW algorithm was evaluated using simulated graphs and realworld big graph data. According to the results, the proposed algorithm is superior to other wellknown methods.
The LRW algorithm can be efficiently used in both global and local graph clustering problems. It finds clusters from a big graph data by only locally exploring the graph. This is important for extreme large data that may not even fit in a single computer memory. The algorithm contains two phases—the graph exploring phase and the cluster merging phase. The graph exploring phase is the most critical part and also the most timeconsuming part of the algorithm. This phase can be implemented in embarrassingly parallel paradigm. The algorithm can easily be adapted to any MapReduce framework.
From our experiments, we also noticed the limitations of the LRW algorithm. First, when used as a global clustering algorithm, the computational complexity can be high, especially when the graph cluster structure is weak. This is due to the fact that the graph may be analyzed multiple times during the graph exploration phase, if we perform the LRW procedure from every vertex of the graph. However, using the multistage strategy can dramatically reduce the computation time. Second, if the cluster structure is weak, the LRW algorithm may return the whole graph as one cluster—though this behavior is desired in many cases.
The experiments show that the performance of the proposed LRW graph clustering algorithm is not sensitive to any parameter except the inflation exponent r, especially when the graph is not heterogeneous. For future research, we will further improve the LRW algorithm so that it can optimally select the inflation function that best suits the problem at hand.
Abbreviations
 GN:

GirvanNewman
 LRW:

limited random walk
 MCL:

Markov clustering algorithm
 MC:

mean conductance
 NMI:

normalized mutual information
Declarations
Authors' contributions
HZ carried out the conception and design of the study, participated in the analysis and interpretation of data, and was involved in drafting and revising the manuscript. JR, SK and MG made substantial contributions to the design of the study, the analysis and interpretation of the data, and were involved in critically reviewing the manuscript. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 Benson AR, Gleich DF, Leskovec J. Higherorder organization of complex networks. Science. 2016;353(6295):163–6.View ArticleGoogle Scholar
 Schaeffer SE. Graph clustering. Comput Sci Rev. 2007;1(1):27–64.MathSciNetView ArticleMATHGoogle Scholar
 Lambiotte R, Delvenne JC, Barahona M. Random walks, Markov processes and the multiscale modular organization of complex networks. IEEE Trans Netw Sci Eng. 2014;1(2):76–90.MathSciNetView ArticleGoogle Scholar
 He P, Xu X, Hu K, Chen L. Semisupervised clustering via multilevel random walk. Pattern Recognit. 2014;47(2):820–32.View ArticleMATHGoogle Scholar
 Newman ME. Fast algorithm for detecting community structure in networks. Phys Rev E. 2004;69(6):066133.View ArticleGoogle Scholar
 Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E. Fast unfolding of communities in large networks. J Stat Mech Theory Exp. 2008;2008(10):10008.View ArticleGoogle Scholar
 Clauset A, Newman ME, Moore C. Finding community structure in very large networks. Phys Rev E. 2004;70(6):066111.View ArticleGoogle Scholar
 Waltman L, van Eck NJ. A smart local moving algorithm for largescale modularitybased community detection. Eur Phys J B. 2013;86(11):1–14.MathSciNetView ArticleGoogle Scholar
 Spielman DA, Teng SH. A local clustering algorithm for massive graphs and its application to nearlylinear time graph partitioning. arXiv:0809.3232; 2008.
 Qiu H, Hancock ER. Graph matching and clustering using spectral partitions. Pattern Recognit. 2006;39(1):22–34.View ArticleGoogle Scholar
 Spielman D, Teng S. Spectral sparsification of graphs. SIAM J Comput. 2011;40(4):981–1025.MathSciNetView ArticleMATHGoogle Scholar
 Dongen S. Graph clustering by flow simulation. PhD thesis, Universiteit Utrecht, Utrecht, The Netherlands; 2000.
 Macropol K, Can T, Singh AK. RRW: repeated random walks on genomescale protein networks for local cluster discovery. BMC Bioinform. 2009;10(1):283.View ArticleGoogle Scholar
 Xin Y, Xie ZQ, Yang J. The adaptive dynamic community detection algorithm based on the nonhomogeneous random walking. Phys A Stat Mech Appl. 2016;450:241–52.View ArticleGoogle Scholar
 Chung F, Kempton M. A local clustering algorithm for connection graphs. In: Algorithms and models for the web graph; 2013. p. 26–43.
 Macko P, Margo D, Seltzer M. Local clustering in provenance graphs. In: Proceedings of the 22nd ACM international conference on conference on information & knowledge management; 2013. p. 835–840.
 Andersen R, Chung F, Lang K. Using pagerank to locally partition a graph. Internet Math. 2007;4(1):35–64.MathSciNetView ArticleMATHGoogle Scholar
 Buhler T, Rangapuram SS, Setzer S, Hein M. Constrained fractional set programs and their application in local clustering and community detection. arXiv:1306.3409; 2013.
 Zhu ZA, Lattanzi S, Mirrokni V. A local algorithm for finding wellconnected clusters. In: Proceedings of the 30th international conference on machine learning (ICML13); 2013. p. 396–404.
 Norris JR. Markov chains applied probability and stochastic networks; 1998)
 Harel D, Koren Y. On clustering using random walks. In: Hariharan R, Vinay V, Mukund M, et al., editors. Theoretical computer science., Lecture notes in computer scienceBerlin: Springer; 2001. p. 18–41.Google Scholar
 Cai B, Wang H, Zheng H, Wang H. An improved random walk based clustering algorithm for community detection in complex networks. In: 2011 IEEE international conference on systems, man, and cybernetics (SMC); 2011. p. 2162–2167.
 Dean J, Ghemawat S. MapReduce: simplified data processing on large clusters. Commun ACM. 2008;51(1):107–13.View ArticleGoogle Scholar
 Newman M. Networks: an introduction. 1st ed. New York: Oxford; 2010.View ArticleMATHGoogle Scholar
 Danon L, DíazGuilera A, Arenas A. The effect of size heterogeneity on community identification in complex networks. J Stat Mech Theory Exp. 2006;2006(11):11010.View ArticleGoogle Scholar
 Lancichinetti A, Fortunato S, Radicchi F. Benchmark graphs for testing community detection algorithms. Phys Rev E. 2008;78(4):046110.View ArticleGoogle Scholar
 Newman ME, Girvan M. Finding and evaluating community structure in networks. Phys Rev E. 2004;69(2):026113.View ArticleGoogle Scholar
 Ana LNF, Jain AK. Robust data clustering. In: Proceedings 2003 IEEE computer society conference on computer vision and pattern recognition, vol. 2; 2003. p. 128–1332.
 Danon L, DiazGuilera A, Duch J, Arenas A. Comparing community structure identification. J Stat Mech Theory Exp. 2005;2005(09):09008.View ArticleGoogle Scholar
 Rosvall M, Bergstrom CT. Maps of random walks on complex networks reveal community structure. Proc Natl Acad Sci. 2008;105(4):1118–23.View ArticleGoogle Scholar
 Zachary WW. An information flow model for conflict and fission in small groups. J Anthropol Res. 1977;1:452–73.View ArticleGoogle Scholar
 Sahai T, Speranzon A, Banaszuk A. Hearing the clusters of a graph: a distributed algorithm. Automatica. 2012;48(1):15–24.MathSciNetView ArticleMATHGoogle Scholar
 Leskovec J, Mcauley JJ. Learning to discover social circles in ego networks. In: Pereira F, Burges CJC, Bottou L, Weinberger KQ, editors. Advances in neural information processing systems 25; 2012. p. 539–547.
 Kunegis J. Konect—the Koblenz network collection. In: Proceedings of international conference on World Wide Web companion; 2013. p. 1343–1350.
 Lusseau D, Schneider K, Boisseau OJ, Haase P, Slooten E, Dawson SM. The bottlenose dolphin community of doubtful sound features a large proportion of longlasting associations. Behav Ecol Sociobiol. 2003;54(4):396–405.View ArticleGoogle Scholar
 Gleiser P, Danon L. Community structure in jazz. Adv Complex Syst. 2003;06(04):565–73.View ArticleGoogle Scholar
 Isella L, Stehlé J, Barrat A, Cattuto C, Pinton JE, van den Broeck W. What’s in a crowd? Analysis of facetoface behavioral networks. J Theor Biol. 2011;271(1):166–80.View ArticleGoogle Scholar
 Adamic LA, Glance N. The political blogosphere and the 2004 US election: divided they blog. Proceedings of the 3rd international workshop on link discovery., LinkKDD ’05New York: NY, USA; 2005. p. 36–43.
 Croft D, Mundo AF, Haw R, Milacic M, Weiser J, Wu G, Caudy M, Garapati P, Gillespie M, Kamdar MR, Jassal B, Jupe S, Matthews L, May B, Palatnik S, Rothfels K, Shamovsky V, Song H, Williams M, Birney E, Hermjakob H, Stein L, D’Eustachio P. The reactome pathway knowledgebase. Nucleic Acids Res. 2013;1102:472–7.Google Scholar
 Leskovec J. Stanford large network dataset collection.
 Yang J, Leskovec J. Defining and evaluating network communities based on groundtruth. arXiv:1205.6233; 2012.