Error-Correcting Decoders for Communities in Networks

TL;DR

This paper introduces a novel community detection method inspired by error-correcting decoders, simplifying the algorithm to achieve linear time complexity while maintaining satisfactory performance on artificial and real networks.

Contribution

The paper presents a new community detection algorithm based on error-correcting decoding principles, with a simplified version that reduces complexity to linear time.

Findings

The simplified algorithm maintains competitive accuracy.

The method performs well on artificial benchmarks.

Effective on real-world network data.

Abstract

As recent work demonstrated, the task of identifying communities in networks can be considered analogous to the classical problem of decoding messages transmitted along a noisy channel. We leverage this analogy to develop a community detection method directly inspired by a standard and widely-used decoding technique. We further simplify the algorithm to reduce the time complexity from quadratic to linear. We test the performance of the original and reduced versions of the algorithm on artificial benchmarks with pre-imposed community structure, and on real networks with annotated community structure. Results of our systematic analysis indicate that the proposed techniques are able to provide satisfactory results.

Tables

Table 1. Table 1: The conditional probabilities between for the variables A i , j subscript 𝐴 𝑖 𝑗 A_{i,j} and θ i , j subscript 𝜃 𝑖 𝑗 \theta_{i,j} . The last column was calculated using Bayes’ rule with an assumption on the prior of P ​ ( θ i , j = 1 ) = 1 / 2 𝑃 subscript 𝜃 𝑖 𝑗 1 1 2 P(\theta_{i,j}=1)=1/2 .

$A_{i,j}$	${\theta }_{i,j}$	$P\left(A_{i,j}|{\theta }_{i,j}\right)$	$P\left({\theta }_{i,j}|A_{i,j}\right)$
1	0	$p_{in}$	$\frac{p_{in}}{p_{in}+p_{out}}$
1	1	$p_{out}$	$\frac{p_{out}}{p_{in}+p_{out}}$
0	0	1 - $p_{in}$	$\frac{1-p_{in}}{2-\left(p_{in}+p_{out}\right)}$
0	1	1 - $p_{out}$	$\frac{1-p_{out}}{2-\left(p_{in}+p_{out}\right)}$

Table 2. Table 2: NMI of the metadata communities and the communities detected by several algorithms applied to five network datasets: Zachary Karate Club  zachary1977information , 2000 US College football  girvan2002community , the 2004 US Political Books from Amazon  krebs2008network , the US Political Blogs  adamic2005political , and a small portion of the Facebook social network  traud2012social . For Facebook we consider different metadata to define communities. The first four columns report respectively name of the network (and eventual metadata used to define communities), number of nodes N 𝑁 N in the network, number of edges M 𝑀 M , and number of communities C 𝐶 C according to the metadata classification. All other columns refer to results obtained using community detection algorithms. First, we considered our proposed algorithms (original and reduced) and the two different starting conditions (random and regular). Given the stochasticity of the outcome, we report median values and the 90 % percent 90 90\% confidence interval for NMI values. Due to its high computational complexity, we couldn’t use the original version of the our proposed algorithm to analyze large networks. We performed the same analysis using Louvain and Infomap. In this case, the outcome of the community detection algorithm is deterministic, so we report a single NMI value.

				Original Random			Original Regular			Reduced Random			Reduced Regular			Louvain	Infomap
Graphs	$N$	$M$	$C$	5%	50%	95%	5%	50%	95%	5%	50%	95%	5%	50%	95%	50%	50%
karate	34	78	2	0.23	0.43	0.49	0.38	0.47	0.51	0.21	0.68	0.84	0.00	0.00	0.84	0.52	0.58
football	115	615	13	0.72	0.78	0.81	0.86	0.89	0.90	0.86	0.94	0.98	0.34	0.83	0.98	0.21	0.49
polbooks	105	441	3	0.42	0.45	0.48	0.43	0.45	0.47	0.27	0.62	0.66	0.60	0.62	0.70	0.04	0.11
polblogs	1222	16714	2	-	-	-	-	-		0.53	0.59	0.67	0.04	0.78	0.79	0.01	0.02
fb-dorm	10001	362892	112	-	-	-	-	-	-	0.01	0.11	0.12	0.01	0.06	0.13	0.21	0.13
fb-gender	10001	362892	7	-	-	-	-	-	-	0.00	0.01	0.03	0.01	0.01	0.02	0.00	0.01
fb-high school	10001	362892	691	-	-	-	-	-	-	0.13	0.16	0.17	0.02	0.08	0.16	0.10	0.29
fb-major	10001	362892	180	-	-	-	-	-	-	0.01	0.08	0.09	0.01	0.06	0.07	0.03	0.12

Equations

$${\sigma }_i+{\sigma }_j+{\theta }_{i,j}=0,$$

$$\begin{array}{rl}{\zeta }_{i\to j}^t& =\{\begin{array}{ll}{\ell }_i& t=0\\ {\ell }_i+{\sum }_{s\ne i,j}F[\tanh \frac{{\ell }_{i,s}}{2},{\zeta }_{s\to i}^{t-1}]& t>0\end{array}\end{array}$$

$$F[a,x]=\log \frac{1+a\tanh \frac{x}{2}}{1-a\tanh \frac{x}{2}},$$

$$\begin{array}{rl}{\hat{\ell }}_i^t& ={\ell }_i+\sum _{s\ne i}F[\tanh \frac{{\ell }_{i,s}}{2},{\zeta }_{s\to i}^{t-1}]\\ {\hat{\ell }}_{i,j}^t& ={\ell }_{i,j}+F[\tanh \frac{{\zeta }_{i\to j}^{t-1}}{2},{\zeta }_{j\to i}^{t-1}]\end{array}.$$

$$\displaystyle\begin{array}[]{ll}\zeta_{i\rightarrow j}^{t}=&\ell_{i}+(N-k_{i}-1)\;F\left[\tanh{\frac{\ell_{non}}{2}},\mathcal{Z}^{t-1}\right]+\\ &+\sum_{s\in\mathcal{N}_{i}\backslash j}\;F\left[\tanh{\frac{\ell_{con}}{2}},\zeta_{s\rightarrow i}^{t-1}\right]\end{array}\;$$

$$\begin{array}{rl}{\ell }_{non}& =\log P({\theta }_{i,j}=0|A_{i,j}=0)-\log P({\theta }_{i,j}=1|A_{i,j}=0)\\ {\ell }_{con}& =\log P({\theta }_{i,j}=0|A_{i,j}=1)-\log P({\theta }_{i,j}=1|A_{i,j}=1)\end{array}.$$

$$\displaystyle\mathcal{Z}^{t=0}=\frac{\sum_{i=1}^{N}(N-k_{i}-1)\ell_{i}}{N(N-1)-2M}\$$

$$\displaystyle\begin{array}[]{ll}\mathcal{Z}^{t}=&\mathcal{Z}^{t=0}+\;F\left[\tanh{\frac{\ell_{non}}{2}},\mathcal{Z}^{t-1}\right]+\\ &+\frac{\sum_{i=1}^{N}\sum_{j\in\mathcal{N}_{i}}\;F\left[\tanh{\frac{\ell_{con}}{2}},\zeta_{i\rightarrow j}^{t-1}\right]}{N(N-1)-2M}\end{array}$$

$$\begin{array}{rl}{\hat{\ell }}_i& ={\ell }_i+\sum _{s\in {\mathcal{N}}_i}F\left[\tanh \frac{{\ell }_{con}}{2},{\zeta }_{s\to i}^{t-1}\right]+(N-k_i)F\left[\tanh \frac{{\ell }_{non}}{2},{\mathcal{Z}}^{t-1}\right]\\ {\hat{\ell }}_{i,j}& =\log \frac{p_{in}}{p_{out}}+F\left[\tanh \frac{{\zeta }_{i\to j}^{t-1}}{2},{\zeta }_{j\to i}^{t-1}\right].\end{array}$$

$$NMI=\frac{I(True,Predicted)}{\sqrt{H(True)H(Predicted)}}.$$

$$\begin{array}{rl}{p}_{in}+p_{out}& =\frac{2⟨k⟩}{N}\\ p_{in}-p_{out}& >\frac{2\sqrt{⟨k⟩}}{N}\end{array}.$$

$$\begin{array}{rl}{p}_{in}& =\alpha \frac{⟨k⟩+\sqrt{⟨k⟩}}{N}\\ p_{out}& =\max \left\{0,\frac{2⟨k⟩}{N}-p_{in}\right\},\end{array}$$

Full text

I Introduction

Real networks often exhibit organization in communities, intuitively defined as groups of nodes with a higher density of edges within rather than between groups girvan2002community ; fortunato2010community . Most of the research on this topic has focused on the development of algorithms for community identification. Proposed approaches vary widely, including hierarchical clustering algorithms friedman2001elements , modularity-based methods newman2004finding ; newman2004fast ; clauset2004finding ; guimera2007module ; duch2005community ; newman2006finding ; newman2006modularity , random walk based algorithms zhou2003distance ; rosvall2008maps , and statistical inference methods newman2007mixture ; hastings2006community ; decelle2011inference ; karrer2011stochastic ; peixoto2014hierarchical ; peixoto2013parsimonious ; peixoto2017bayesian , to mention a few of them. Whereas algorithms differ much in spirit, they all share two intrinsic limitations. First, as described by the No Free Lunch Theorem peel2017ground , there is no community detection algorithm that works best for all networks and community structures; an algorithm good for one class of networks may be equally bad for another class of networks. A second type of limitation arises from self-consistency tests, where community detection methods are applied to instances of the stochastic block model to uncover the community structure pre-imposed in the model. Algorithms can recover a non-vanishing portion of the true community structure of the graph only if the amount of fuzziness in the network is below the detectability threshold decelle2011inference ; nadakuditi2012graph ; krzakala2013spectral ; radicchi2013detectability ; radicchi2014paradox ; abbe2015community ; abbe2017community . Also, exact detection of the true cluster structure is subjected to a threshold phenomenon  abbe2016exact ; abbe2017community ; mossel2013proof . This phenomenon can be understood through the lens of coding theory by interpreting the problem of defining and identifying communities in networks as a classical communication task over a noisy channel, analogous to the one originally considered by Shannon shannon2001mathematical . The value of the exact recovery threshold can be estimated in the limit of infinitely large graphs abbe2016exact ; abbe2015community ; abbe2017community ; mossel2013proof . A bound on the value of the threshold for finite-size graphs can be obtained as an application of the Shannon’s noisy-channel coding theorem radicchi2017decoding .

In this paper, we exploit the analogy between coding theory and community structure in networks, and develop a novel class of algorithms for community detection based on a state-of-the-art decoding technique gallager1962low ; mackay1996near . The idea has been already considered in Ref. radicchi2017decoding for the simplest case of network bipartitions. Here, we expand the method to find multiple communities by iterating the bipartition method in a way similar to what already considered in Refs. newman2013community ; kernighan1970efficient ; fiduccia1982linear . As the decoding method considered in Ref. radicchi2017decoding has computational complexity that scales quadratically with the number of nodes in the network, we further propose an approximation of the algorithm that makes the method complexity scale linearly with the number of edges, thus making it linearly dependent with system size in sparse networks. We perform systematic tests of the both algorithm versions on synthetic and real-world graphs. Performances appear satisfactory in all cases.

II Methods

II.1 Community detection as a communication process

For sake of clarity, we repeat the same description already provided in Refs. abbe2016exact ; abbe2015community ; abbe2017community ; mossel2013proof ; radicchi2017decoding of how the definition and detection of communities in a network can be framed as a communication process (see Figure 1).

We assume that there are $N$ nodes in the network and that each node $i$ has associated a single information bit $\sigma_{i}=0,1$. The value of the bit identifies the group of node $i$. The message is encoded by adding $N(N-1)/2$ parity bits $\theta$, each for every pair of nodes. The parity bit $\theta_{i,j}=0$ if $\sigma_{i}=\sigma_{j}$, or $\theta_{i,j}=1$, otherwise. The parity bits are essentially added to the original message according to the rule

[TABLE]

where the sum is performed in modulo-$2$ arithmetic. The set of $N(N-1)/2$ equations defines the code used in the communication process. In the jargon of coding theory, Eqs. (1) defines a low-density parity-check (LDPC) code. These type of codes are often used in practical communication tasks, given their effectiveness gallager1962low ; mackay1996near ; mackay2003information . In graphical terms, the encoded message can be seen as a network composed of two disconnected cliques, where each identifies a community of nodes.

Once encoded, the message is transmitted trough a communication channel. There, noise alters the bit values. Information bits $\sigma$ are deleted so that there is no longer information about node memberships; some parity bits $\theta$ are flipped giving rise to the observed network. The goal of the decoder is to use information from the observed network together with a hypothesis on the noise characterizing the channel to infer the original message about group memberships.

II.2 Stochastic block model as a noisy channel

As already done in Refs. abbe2016exact ; abbe2015community ; abbe2017community ; mossel2013proof ; radicchi2017decoding , we make a strong hypothesis on the noisy channel. We assume that the observed network is given by a stochastic block model, where pairs of nodes within the same group are connected with probability $p_{in}$, and pairs of nodes belonging to different groups are connected with probability $p_{out}$. This corresponds to assuming that the noisy channel is given by an asymmetric binary channel, and that the parity bits $\theta$ are flipped with probabilities defined in Table 1. Further, it allows us to use Bayes’ theorem to derive the conditional probability $P(\theta_{i,j}|A_{i,j})$ for the parity check bit $\theta_{i,j}$ depending on whether nodes $i$ and $j$ are connected in the observed network, i.e., $A_{i,j}=1$ or $A_{i,j}=0$. Please note that, since there is no prior knowledge of the true parity bits values, we assume $P(\theta_{i,j}=1)=1/2$ radicchi2017decoding . This represents a strong assumption in the model, and the resulting algorithm is biased towards the detection of homogenous communities.

II.3 Gallager community decoder

To find the community structure of an observed network, we take advantage of a widely-used decoding technique for LDPC codes. The technique consists in iteratively solving the system of parity-check equations that defines the code, given the knowledge of the noisy channel gallager1962low ; mackay1996near . The application of the method to community detection was considered in Ref. radicchi2017decoding . Specifically, the technique is used to solve Eqs. (1) using properties of the channel from Table 1. The $t$-th iteration of the algorithm is based on

[TABLE]

for all ordered pairs of nodes $i\to j$. The function $F$ is defined as

[TABLE]

where $\tanh(\cdot)$ is the hyperbolic tangent function. In the algorithm, the quantity $\ell_{i}$ is the log-likehood ratio (LLR) $\ell_{i}=\log P(\sigma_{i}=0)-\log P(\sigma_{i}=1)$ associated with node $i$, that is the natural logarithm of the ratio between the probabilities that the parity bit $\sigma_{i}$ equals zero or one. $\ell_{i,j}=\log P(\theta_{i,j}=0|A_{i,j})-\log P(\theta_{i,j}=1|A_{i,j})$ is instead the LLR associated with the parity bit $\theta_{i,j}$ given the hypothesis on the noisy channel and the evidence from the observed network. The variable $\zeta_{i\rightarrow j}^{t}$ is still a LLR. It is defined for all pairs of nodes $i$ and $j$, irrespective of whether they are connected or not. $\zeta_{i\rightarrow j}^{t}$ may be interpreted as a message that node $i$ sends to node $j$ regarding the value that the information bit $\sigma_{i}$ should assume based on the knowledge of the code, the noisy channel, and the evidence collected by observing the network. Please note that two distinct messages are exchanged for every pair of nodes $i$ and $j$, depending on the direction of the message, either $i\to j$ or $j\to i$. At every iteration $t$, convergence of the algorithm is tested by first calculating the best estimates of the LLRs as

[TABLE]

Then, one evaluates the best estimates of the information bits, according to $\hat{\sigma}_{i}=0$ if $\hat{\ell}_{i}^{t}>0$, and $\hat{\sigma}_{i}=1$, otherwise. A similar rule is used for the best estimate of the parity bit $\hat{\theta}_{i,j}$. Finally, the best estimates of the bits are plugged in the system of Eqs. (1). If the equations are all satisfied, the algorithm has converged. Otherwise, one continues iterating for a maximum number of iterations $T$. In our calculations, we set $T=100$.

We remark three important facts. First, possible solutions of the algorithm are classifications of nodes in either one or two groups. In the first case, the algorithm indicates absence of block structure in the network. Second, knowledge of the noisy channel and evidence of the observed network is used in the definition of the initial LLRs $\ell_{i,j}$. For the choice of the initial values of the LLRs for individual nodes $\ell_{i}$ there is not a specific rule. If the community structure is strong enough, initial conditions for the iterative algorithm are not very important. However, in regimes where community structure is less neat, they may determine the basis of attraction for the iterated map. In this paper, we will consider two different choices for the initial values of the nodes’ LLRs. Finally, we stress that the algorithm is the ad-literam adaptation of the Gallager decoding algorithm to the detection of two communities. As such, the algorithm iterates over all possible pairs of nodes, irrespective of whether they are connected or not. Each iteration of the algorithm requires a number of operations that scales with the network size $N$ as $\mathcal{O}(N^{2})$, thus making the algorithm applicable only to small/medium sized networks.

II.4 Reducing the computational complexity of the community decoder

We leverage network sparsity to reduce the computational complexity of the algorithm without significantly deteriorating algorithm performance. The way we decrease the complexity is rather intuitive. In the original implementation, a node sends a message to all other nodes, even if there is not an edge connecting them. In the reduced algorithm, we instead assume that (i) messages are delivered only along existing edges, (ii) the message passed from a node to any unconnected node is the same regardless of the actual pair of nodes considered. This reduces the total number of messages to twice the number of edges in the network, and thus the complexity from $\mathcal{O}(N^{2})$ to $\mathcal{O}(N\langle k\rangle^{2})$, where $\langle k\rangle$ is the average degree of the network. Our proposed reduction makes the algorithm linearly dependent on the number of edges in the network, which corresponds to a linear dependence with the system size if the network is sparse.

Specifically, the equations that define the algorithm are as follows. For connected pairs of nodes $i$ and $j$, we define the initial message $\zeta_{i\rightarrow j}^{t=0}=\ell_{i}$, and

[TABLE]

for iteration $t\geq 1$. In the equation above, $\ell_{non}$ stands for the LLR of non-connected node, and $\ell_{con}$ is the LLR for connected nodes. These quantities are defined as

[TABLE]

Further, in Eq. (7), $k_{i}$ is the degree of node $i$, and $\mathcal{N}_{i}$ indicates the set of neighbors of node $i$. Non-existing edges deliver the single message $\mathcal{Z}$. This corresponds to the average value of all messages among non-connected pairs of nodes in the original version of the algorithm. The equations that define the iterations for $\mathcal{Z}$ are

[TABLE]

and

[TABLE]

for iteration $t\geq 1$. We used $2M=\sum_{i}k_{i}$, i.e., the sum of the degrees of all the nodes in the network.

Convergence of the equations above is tested using the same procedure described in the original algorithm. In particular, the best estimates of the LLRs are computed using

[TABLE]

These values are used to find the best estimates of the bits $\sigma$s and $\theta$s and, in turn, are plugged into the parity-check Eqs. (1). To keep the computational complexity linear, only parity-check equations corresponding to existing edges are actually tested. The maximum number of iterations $T$ that we considered before stopping the algorithm for lack of convergence is $T=1,000$.

II.5 Initial conditions

As we mentioned above, the initial value $\ell_{i}$ of the LLR for every node $i$ requires initialization. The initialization is potentially a very important decision for the performance of the algorithm as it determines the basin of attraction of the iterative system of equations. In this paper, we consider two different strategies for the determination of the starting conditions:

Regular

A random node $i$ is chosen such that $\ell_{i}=1$ and $\ell_{j}=0$, $\forall j\neq i$.

Random

For every node $i=1,\ldots,N$, $\ell_{i}$ is a random variable extracted from the uniform distribution with support $[-1,1]$.

II.6 Multiple communities

Up to now, we have described how to find a bipartition in a network according to our procedure. We remark that the output of the algorithm may also indicate no division of the network. Our goal, however, is to detect an arbitrary number of communities in our graph. To this end, we adopt a simple iterative procedure (see Figure 2). The procedure is identical to the one already adopted in Refs.  newman2013community ; kernighan1970efficient ; fiduccia1982linear , and it may be summarized as follows. At the beginning, we define a list of subgraphs $L$ to be analyzed, and a list of detected communities $C$. The list $L$ contains only one element, the entire graph $G$, while $C$ is empty. We then apply the following steps:

• 1

Take a graph $g$ from the list $L$. Remove the graph from the list.

• 2

Apply the bipartition algorithm to the graph $g$.

• a

If the algorithm finds a split of $g$ in two sets of nodes, namely $g_{1}$ and $g_{2}$, reconstruct each set as a graph using only nodes within the set, and only edges between pairs of nodes within the set. Place $g_{1}$ and $g_{2}$ into the list $L$.

• b

If the algorithm finds only a set of nodes, so that no actual split was detected, $g$ is considered as a community and placed in the list $C$.

• 3

Go back to point 2 until $L$ is empty. The list of detected communities is given by $C$.

II.7 Learning the parameters of the noisy channel

So far, we tacitly assumed to know the values of the probabilities $p_{in}$ and $p_{out}$. The assumption has been used in the bipartition algorithm of Ref. radicchi2017decoding when applied to instances of the stochastic block model with two communities. In practical situations, however, prior knowledge of the probabilities $p_{in}$ and $p_{out}$ is not available. These parameters should instead be learned in a self-consistent way by the algorithm relying only on information from the observed network. Here, we simultaneously propose and validate a simple learning strategy. To this end, we generate instances of the so-called Girvan-Newman (GN) benchmark graph girvan2002community , a variant of the stochastic model with $N=128$ and $Q=4$ communities. Different from the original version of the GN model we allow nodes to have average degree $\langle k\rangle\neq 16$. The average connectivity of the model is set by fixing the sum of the true parameter values $\hat{p}_{in}$ and $\hat{p}_{out}$, while the strength of the community structure is instead determined by their difference. We consider four different combinations $(\hat{p}_{in},\hat{p}_{out})$ for the true values of the model parameters to generate four instances of the model. To each of the four instances, we apply the original algorithm with the regular starting conditions to the network using the parameters values $\tilde{p}_{in}$ and $\tilde{p}_{out}$. We measure the performance of the algorithm to recover the pre-imposed community structure of the graph, using normalized mutual information (NMI)

[TABLE]

NMI is defined as the mutual information $I$ between the predicted and true clusters normalized by the square root of the product of the individual entropies $H$ strehl2002cluster ; danon2005comparing .

In Fig. 3, we display the outcome of our tests when the community detection algorithm is applied relying on prior information given by $\tilde{p}_{in}$ and $\tilde{p}_{out}$. We consider only combinations $(\tilde{p}_{in},\tilde{p}_{out})$ that lay in the regime of detectability decelle2011inference . The figure shows that our algorithm reproduces accurately the community structure of the graph for several combinations $(\tilde{p}_{in},\tilde{p}_{out})$. This fact happens as long as $(\tilde{p}_{in},\tilde{p}_{out})$ is not too far from the ground truth $(\hat{p}_{in},\hat{p}_{out})$. The finding tells us that knowing the exact value is not a necessary requirement for the correct detection of the modules; we need only a good guess of the values of the parameters. In particular, the analysis suggests a simple criterion for the choice of the parameter values $p_{in}$ and $p_{out}$ that can be used in the algorithm. We can use any combination that satisfy the equations

[TABLE]

where $\langle k\rangle$ is the average degree observed in the network. The first equation imposes that the parameters $p_{in}$ and $p_{out}$ are compatible with the average degree of the observed network. The inequality appearing in the bottom of Eq. (15) is instead restricting our possibilities only in the regime of detectability decelle2011asymptotic . As any point in the segment determined by Eqs. (15) is equivalent in terms of performance, the values of the parameters $p_{in}$ and $p_{out}$ used by our algorithm are obtained with

[TABLE]

where $\alpha>0$ is a tunable parameter, whose value is chosen appropriately such that $p_{in}>p_{out}\geq 0$. In our numerical results, we set $\alpha=1.2$. However, we verified that the performance of the algorithm doesn’t change if we choose small $\alpha$ values at random.

III Results

III.1 Artificial graphs

First, we perform tests of the original and reduced versions of the algorithm on synthetic graphs with pre-imposed community structure. These are compared with 100 realizations from both the well-established methods Louvain blondel2008fast and Infomap rosvall2008maps . In our numerical tests, we used the implementations of the two algorithms provided by the Python library igraph igraph . In particular, we use the best partition found by Louvain, the community structure obtained looking at the lowest level of the multiresolution method lancichinetti2009community . We consider two different variants of the stochastic block model: the Girvan-Newman (GN) benchmark graph girvan2002community and the Lancichinetti-Fortunato-Radicchi (LFR) benchmark graph lancichinetti2008benchmark . We measure the performance of the algorithms using NMI as a function of the community strength of the model, determined by the value of the mixing parameter $\mu=\frac{k_{out}}{k_{out}+k_{in}}$, i.e., the ratio between external and total degree of the nodes. This parametrization allows for a direct comparison between our results on those reported in Ref. lancichinetti2009community .

In Fig. 4A, we show the performance of the algorithms on the Girvan-Newman (GN) graph. The original algorithm is tested on $100$ instances for each $\mu$ value. We compare results using both starting conditions. Similarly, Fig. 4B shows the results of the reduced algorithm on $100$ instances of the GN graph. In the original implementation, at around $\mu=0.5$, the performances of both algorithm reduce to 0. Both tend to outperform Infomap for large values of $\mu$ but perform worse than Louvain. In the reduced version of the algorithm, the performance of the regular implementation reduces to 0 when $\mu\geq 0.3$. The random implementation is similar to Infomap and both start to drop around $\mu=0.4$. As before, both perform worse than Louvain. The values of $\mu$ where we see a drop in performance are tantamount with the level of fuzziness where most of the algorithms start to systematically fail on the GN benchmark lancichinetti2009community . In most of the cases, either perfect communities or one large community was predicted. An interesting finding is that the reduced version of the algorithm is able to perform just as well as the original version with the regular conditions and just slightly worse with the random conditions for low values of $\mu$.

Tests on the LFR graphs are reported in Fig. 5. Similar to Ref. lancichinetti2009community , our tests were performed on networks with size either $N=1000$ or $N=5000$, generated under condition S, i.e., small communities with size in the range $[10,50]$ nodes per community, or under condition $B$, i.e., large communities with size in the range $[20,100]$. In the generation of graph instances, community sizes are chosen at random according to power-law functions with exponent $-1$ defined over the aforementioned ranges. Node degrees are random variates extracted from a power-law degree distribution with exponent $-2$, such that the average degree of the nodes is $20$ and maximum degree equals $50$. We tested the performance of our algorithms over $100$ instances of the model for each $\mu$ value. Given the high complexity of the original version of the algorithm, we could test in a systematic fashion only the performance of the reduced algorithm. The algorithm was started from both initial conditions. The results of Fig. 5 provide evidence that the algorithm is able to achieve good performance, although the ability to recover the right community structure of the model decreases to zero for a level of noise slightly smaller than those of other algorithms lancichinetti2009community .

III.2 Real networks

Recently, community detection algorithms have been focusing on incorporating edge and node metadata into community formation newman_structure_2016 . An interesting point in this context is understanding how much the community structure of a network is actually representative for exogenous classifications of nodes obtainable from metadata hric_community_2014 .

We run both versions of the algorithms 100 times on 5 well-known datasets with metadata. For each dataset, we applied three filters; splitting communities into connected components, removing duplicates, and removing singletons hric_community_2014 . The Zachary Karate Club network is a social network of 34 nodes and 78 edges of self reported friends zachary1977information . A disagreement between the two leaders led to the splitting of the club into two groups. The US College football network is a network of college football teams in which edges represent a scheduled game in the Fall of 2000 girvan2002community . The communities are the 12 conferences each of the teams belong to. The US Political Book network represents all books co-purchased on Amazon.com around the 2004 election in which edges are Amazon recommendations indicating co-purchases from other users while the groups represent the political leanings of the book (Liberal, Neutral, or Conservative) found by human ratings krebs2008network . The US Political Blog dataset is a network of hyperlinks between blogs with the groups being Conservative or Liberal adamic2005political . Finally, the Facebook social networks are undirected friendship networks from 97 different colleges across the US traud2012social . We specifically use network 82 with dorms, gender, high school, and major as the communities. Due to the size, we only ran 5 iterations on the Facebook network.

Table 2 shows the performance of algorithms, under both initial conditions, on the various datasets. Performance is still measured in terms of NMI between the community structure recovered by the algorithms and the one given by the metadata. Best matches between topological communities and metadata were observed for the US College Football network, similar to Hric et al. hric_community_2014 . The result is expected as college football teams play more against teams within their conference rather than teams outside their conference. Interestingly, the communities found by our algorithm seem to provide significantly higher NMI values than those obtained via Louvain and Infomap on the US Political Book and US Political Blog networks.

IV Conclusion

In this paper, we exploited the interpretation of the problem of defining and identifying communities in networks as a classical communication task over a noisy channel, and made use of a widely-used decoding technique to generate a novel algorithm for community detection. Although the primitive version of the algorithm was introduced in Ref. radicchi2017decoding , we extended the idea in three respects. First, we generalized the algorithm, originally designed for the detection of two communities only, to the detection of an arbitrary number of communities. The generalization consists of iterating the binary version of the algorithm till convergence. Second, we accounted for the sparsity of graphs which community detection methods are usually applied to, and reduced the complexity of the algorithm from quadratic to linear. The simplification allowed us to generate a method able to deal with potentially large networks without renouncing too much to the basic principles of the original version of the algorithm. Third, we systematically tested the performance of the new algorithm on both synthetic networks and real-world graphs. These tests provided results that are consistent with what already observed in the literature for other well-established algorithms for community detection. In particular, the algorithm outperformed top community detection algorithms in tests based on the standard SBM, i.e., involving the detection of equally sized communities in graphs with homogenous degree distributions. On the basis of the performance results obtained here, we believe that our algorithm may represent an effective and efficient alternative to other methods that rely on the SBM ansatz to infer network community structure.

Availability of data and material

The code and the datasets used in this analysis are available in a Github repository at https://github.com/kbathina/Error-decoding-community-detection.

Acknowledgements.

FR acknowledges support from the National Science Foundation (CMMI-1552487) and from the US Army Research Office (W911NF-16-1-0104).