A social network graph (sometimes called an “activity graph”) [1] contains nodes representing actors (generally people or organizations), and edges representing relationships or communications between the actors. In graph representation, a community structure consists of several nodes which shows dense internal connection compared to the rest of the network. It is one of the well-known graph patterns observed in large social networks, and it reflects the fact the human society is organized into groups, families, filial groups, tribes, villages, friendship circles, professional circles, etc. The graph structure of social networks is influenced by the underlying human social behavior, and there is a strong connection between association and similarity in human society. Those with similar interests or of the same kind tend to form groups, which are shown by the age-old adage: “birds of a feather flock together.” This principle, the homophily principle [2] , implies the nodes that belong to a cluster or community in a social network are expected to have homogeneous characteristics. The identification of communities hidden within the structure of large complex network is a challenging problem which has attracted a considerable amount of interest. It has been widely studied in the literature, and there have been significant advancements with contributions from different fields, such as statistical physics. The fact that there is no agreed-upon definition for a community adds to the challenge. Different community detection methods are developed from various applications of specific needs which establish, explicitly or implicitly, its own definition of community. This means the definition of a community depends on the application domain and the properties of the graph under consideration.

Community detection is analogous to the clustering problem, and it can be viewed as a data mining analysis on graphs: an unsupervised classification of its nodes. However, classical data mining clustering algorithms [3] like K-means, Expectation-Maximization do not deal with graph data where the entities are nodes connected to each other through edges. On the other hand, the well-known graph clustering techniques, like clustering based on normalized cut [4] , modularity [5] or structural density [6] , utilize the relationships of the network to partition the graph into several densely connected components, but do not take the properties of the nodes into account.

Most community detection techniques focus mainly on the topological structure of the graph. However, online social networks are a rich source of information that can help profiling the users because they register their patterns of social interaction, dialogue, exchange and collaboration. Hence, in social network graph, nodes have attributes, and can be represented as attributed graphs, [7] [8] where an attribute vector associated with a node represent properties of the node like demographic features, personal preferences, and features characterizing the user's utilization of specific social network functions or applications. Since the nodes forming a community are expected to have homogeneous characteristics because of the homophily principle, it can be useful to take into account the node properties in the clustering process for increasing the accuracy of the partitions. Under this assumption, we can argue that, considering both structural data and attribute data simultaneously, instead of separately, can inform us the properties of topological structure and the impact and the role of the attributes on the topological structure which will be helpful for more meaningful and accurate community detection.

A community is intuitively understood as a set of entities where each entity is closer, in the network sense, to the other entities within the community than to the entities outside it [9] . One of the widely-used definition of communities is based on the number of edges within a group (density) compared to the number of edges between different groups. A community is thought of as a group of nodes that has more and/or better-connected edges between its members (intra-cluster edges) than between its members and the remainder of the network (inter- clusters edges) [10] .

Community detection methods can be broadly classified [11] into four categories: 1) node-centric community detection, 2) group-centric community detection, 3) network-centric community detection, and 4) hierarchy- centric community detection. Community detection based on node-centric criteria requires each node in a group to satisfy certain properties like mutuality, reachability, or degrees. On the other hand, community detection based on group-centric criteria means each group has to satisfy certain requirements like density, and the connections inside a group are considered as a whole. Network-centric community detection has to consider the connections of the whole network. It aims to partition the nodes into a number of disjoint sets. A group in this case is not defined independently. Typically, some quantitative criterion of the network partition is optimized. Hierarchy-centric community detection constructs a hierarchical structure of communities which facilitates the examination of communities at different granularity. There are mainly three types of hierarchical clustering: agglomerative, divisive, and structure search. Newman-Girvan algorithm [5] , is a divisive hierarchical clustering algorithm which used Modularity to measure the quality of the community.

Community detection in social network graph can be viewed as an unsupervised classification of its nodes, i.e., graphs clustering. While most graph clustering methods focus only on the topological structure of a graph, there are a few methods that aims to partition the graph depending solely on the similarity of the attributes, i.e., the node properties. Some recent methods attempt to consider both the graph topological structure and the node properties. We describe below some of the representative methods which demonstrate that utilizing both the structural data and the attribute data simultaneously is promising.

Steinhaeuser et al. [12] utilize the node attributes for weighting the edges of the graph. They define a novel metric, node attribute similarity (NAS), to assign meaningful weights to the edges. The weight of an edge depends on the attribute-similarity of two nodes connected by it. Communities are identified using a thresholding method. Any pair of nodes whose normalized edge weight exceeds the threshold is placed in the same community. The proposed community detection method is relatively simpler and highly scalable which is able to produce extremely high modularity scores (values over 0.9), as shown by experimental evaluation on a large graph with 1.3 million nodes and 1.2 million edges that represents a real-world social network constructed from cellular phone records. It is reported that the attribute values alone contain some extremely valuable information about the community structure.

In [13] , the authors study the evolution of Google+ social network by augmenting the social network structure with four users attributes (School, Major, Employer and City). Measurements of the network structure of the resultant Google+ Social-Attribute Network (SAN) is done by means of several canonical network metrics such as the reciprocity, density, clustering coefficient, and degree distribution [11] [14] [15] . Then, by generalizing the definition of the network metrics, a set of attribute metrics such as attribute density, attribute distance and diameter, attribute clustering coefficient is obtained. By measuring the network structure as well as the attribute structure by means of the defined metrics, it is possible to gain a more fine-grained understanding of a social network. In particular, how the attributes influence the social structure of the Google+ SAN is studied in [13] . It is observed that the nodes sharing common attributes are likely to have higher reciprocity, and that some attributes have much stronger influence than others.

Elhadi et al. [16] propose a novel clustering method, termed Selection method, which does not use a modified objective function to combine the structure data, and the attribute data, rather makes the choice to utilize either the structure data, or the attribute data depending on the type of graph and the level of information in the attributed graph.

Gunnemann et al. [17] propose a combined clustering model, and the algorithm GAMER (Graph & Attribute Miner) for detection of densely connected subgraph (clusters) that exhibit attribute similarity in subsets of the dimensions (subspace clusters). It is empirically shown for synthetic and real world datasets that clustering quality is improved by simultaneous use of both graph data and attributes data.

We have designed four schemes, which use structural and attribute similarity of the nodes for detection of community structure. Each of the schemes will deliver separate outputs. To find the set of community which represent the real community structure of the social network, outputs of different schemes are combined by using Consensus clustering [18] . Brief descriptions of the schemes are given below.

The community detection schemes will be quantitatively evaluated on datasets representing large real-world social networks. For this reason, we are going to need a collection of large real-world social network data in graph representation, where group memberships (ground-truth information) of nodes are explicitly stated, and the properties or attributes of the nodes are known.

A large collection of benchmark datasets, with ground-truth information, and also with both structural and attribute data do not yet exist, to the best of our knowledge. To solve this problem, a collection of benchmark datasets with both structural and attribute data will be gathered by crawling [14] online social networks using the public web interface provided by the sites. Also, we would utilize existing datasets of real-world social networks gathered by researchers, like the dataset used in [7] .

The absence of a universal definition of a community has implications for community structure detection and evaluation. It is hard to evaluate the identified communities and obtain the comparison of different community detection methods. In this paper, four schemes are designed which utilize structural and attribute data for community detection. These schemes will improve the quality of community detection.

NasifMuslim, (2016) A Combination Approach to Community Detection in Social Networks by Utilizing Structural and Attribute Data. Social Networking,05,11-15. doi: 10.4236/sn.2016.51002