Attribute based Weighted Mechanism for Community Detection in Social Internet of Things

A new paradigm of Internet of Things (IoT) is emerging rapidly by socializing the smarter physical devices called as Social Internet of Things (SIoT). Social relationships established between these objects make them autonomously connected for services, without any human intervention. Since SIoT is a large-scale network with huge data involved, the content spreading behaviour need to be exploited. In order to ensure the growth of the content spread, the large-scale SIoT network is divided into several communities based on the social attributes in this work. We first divided the SIoT network into high quality Sociality based Weighted Communities (SWC). Social attributes like user preferences, social similarities, and mutual friends’ degrees are main metrics for achieving the best rate function. The weighted method based on these social attributes determine the nodes to be present in their respective communities. Also, the controlling of the local community augmentation using cluster concepts is done in our approach. Finally, a Credential Acclaimed Information Spreading (CAIS) mechanism is proposed which selects the best node with the maximum credential to surge the content spreading behaviour in the detected communities of SIoT network. The proposed social-driven attribute based weighted mechanism for community detection is validated using


Introduction
With the massive growth in technology and by incorporating smartness in the physical devices, Internet of Things (IoT) marches the real-world communication. The ultimate aim of IoT is to deliver livelihood for mankind by performing similar services. Progression of IoT intends to be a boon to the broad variety of various application fields by enabling smarter device to device communication [1]. Implementing the social perceptions in the IoT paradigm, unveils a new era called Social Internet of Things (SIoT) [2]. The SIoT model indicates an environment which consents humans and smart devices to interact within a society through numerous varieties of relationships. SIoT uplifts the socialization of the smarter devices connected with one another without any human intervention [3]. SIoT devices are proficient of constructing any number of friends, which in turn to form several communities based on the interested attributes of the devices involved. Therefore, detecting and characterizing the large-scale SIoT network into various communities is imperative for the better service discovery. The nodes are classified and grouped as several communities which serve as the elementary component of SIoT networks. The leaders of these communities effectively influence the other nodes present. 3 The main two steps in the community detection is the identification of promising leaders in the SIoT network and examination of nodes likenesses to construct various communities.
Usually, community detection in networks is based on the several conventional community detection algorithms like Louvain, Girvan Newman and Bron Kerbosch. Louvain algorithm detects the disjoint communities in a directed social network using greedy optimization of modularity. Louvain algorithm tends to be one of the fastest community detection algorithms with a high modularity score. But this method is not suited for the smaller communities which limits the resolution [4]. Bron Kerbosch algorithm works goods for unweighted undirected graph for discovering the overlapping communities. This algorithm computes the utmost cliques by hunting for the perfectly linked nodes of a network. The major problem with Bron Kerbosch algorithm is that it does not hold good for output sensitive problems [5]. Girvan Newman algorithm eliminates the devices with the highest number of the shortest routes between the devices. The edges of the devices that are joining the other devices in a community is maintained to have maximum betweenness. The major problem with Girvan Newman algorithm is that, it is not suited for detecting the communities of huge and complex network structures [6]. Rosvall et al proposed the Infomap algorithm which detects the communities by engaging arbitrary strides to evaluate the content spreading behaviour in the networks. This algorithm does the encoding the content in the network as an encoded graph via a restricted channel.
Finally, the original graph is decoded by constructing the set of probable participants. Lesser the number of participants, larger the content about the network, therefor, nor suited for several participants in a network [7].
Paper is planned as follows: section 2 precises the related works; the proposed SWC detection model is discoursed in section 3; CAIS mechanism is explained in section 4; experimental evaluation is discussed in section 5; the results and discussions are elaborated in section 6 and finally the conclusion and the proposed work is offered in section 7. proposed a community detection technique to divide the SIoT network into many superiority communities and then the content spreading is maximized via two phases such as candidate and greedy phases to select the best candidate for the maximum content sharing. But the influence of node content on the information spread is ignored [11]. Incorporation of Louvain algorithm with the fuzzy network is proposed for finding the communities in a SIoT network.
Shapely index is used as the primary degree for obtaining the fuzzy measures [12]. Liu et al presented the progresses in the community detection via deep learning networks. Deep learning 5 models learn the pattern of nodes, neighborhoods, and subgraphs present in their respective communities of the real-world scenario. Currently convolutional neural networks (CNN), autoencoders and generative adversarial networks (GAN) are mostly used for the community detection but the following are the gap between deep learning networks and community detection: detection and recognition of the spatial variations among various communities are not done, combination of temporal based information and spatial content-based information are yet to be learned in these deep networks [13] Though these algorithms are suited for social networks, but really not appropriate for SIoT networks. In this work, we aim to divide the SIoT network into several smaller communities and maximize the content spread among these communities. The major contributions presented in our work are the following: 1. A Sociality based Weighted Community detection (SWC) algorithm for dividing the SIoT network into high quality smaller communities is developed.
2. An effective mechanism for maximizing the content spreading behaviour among the detected communities is proposed via a Credential Acclaimed Information Spreading (CAIS) strategy.
3. The suggested model is estimated on three different datasets like ARAS, MIT and CASAS datasets.
4. To end, the performance of the proposed attribute-based community detection is compared with various available approaches.

Social Attribute based Community Detection
From the research studies, a SIoT network is a random large-scale network containing several nodes of diverse relationships. Generally, SIoT networks are represented interms of weighted graphs by including the social belongings of the links between nodes. In our work, we used a weighted method which is based on the social attributes such as user preferences, 6 social similarities, degree of the mutual friends are for achieving the best rate function. If two nodes are connected to a node with a lesser degree, then those nodes behave with the linked characteristics. This is same like, when few persons discuss on an uncommon theme, then those persons are with the similar interests. Therefore, it is clear that in a SIoT network, content spreading behaviour can be maximized only if it is divided into several smaller communities.
Our aim is to divide a large SIoT network into several small communities based on the communication relationships between the nodes present in each community.

Sociality based Weighted Community (SWC) Division Mechanism
Let us consider that our SIoT network possesses only local structures with a sub graph containing few nodes in it. Hence, we initially choose local clusters and then the size of these clusters is increased consistently by choosing the nodes with the best rate function. Let regulates the number of communities to be formed and it is always a positive integer. When = 0 ≥ 0.5, it forms only one community and when > 2, it forms several communities.
Since, we pursue to form several smaller communities, we chose = 1.
Our algorithm uses the social attributes such as user preferences, social similarities, highest degree and maximum mutual friends are for achieving the best rate function. Here, the social similarities which obtained from the user preferences are considered to be Direct Intimacy ( ) and degree of the mutual friends are considered to be Indirect Intimacy ( ). Let us consider two nodes, and . The is measured as the summation of the weights associated to node multiplied by the summation of the weights associated to node . When there is no social similarity between the nodes and , then = 1. The is measured as the reciprocal of the Consider a node to sub graph , the difference in the best rate function to with and without is equal to the best rate function of with node added with the best rate functions of without node . If the difference in the best rate function plus the summation of the weights to with and without is greater than zero, this specifies that node improved the rate function of its community fitness and therefore it is then added to sub graph . If the difference in the best rate function plus the summation of the weights to with and without is lesser than zero, this specifies that node reduced the rate function of its community fitness and therefore it is not joined to sub graph . Thus, only if a node makes some improvement to the community fitness function, it joins that respective community. Thus, this method is suitable for exploring the overlapping community detection by means of controlling the number of communities via adjusting the value of.

Local Community Augmentation Control
SIoT network contains huge amount of triangular assemblies, which emerges several negatively influences the community and that node is not added to . Therefore, if the community can be more reliable, the node is added to . Else, even if the node is with the best fitness function, it is not added to .

SWC Algorithm
Our 3. Select a random node , which does not belong to any of the local communities.
4. Use the Direct Intimacy ( ) and Indirect Intimacy ( ) functions to compute the rate function of the neighboring. 9 5. Choose a node with the best rate function. If the best rate function is nonnegative, estimate the reliability of node comparative to community for sub graph .
6. If the best rate function is negative, then repeat from step 3.
7. If the reliability of node comparative to the local community is superior than -0.05, then will be added in , creating a larger local community; else, repeat from step 3.
8. Recompute the best rate function and reliability of each node.
9. If any of a node possess a negative rate function and its reliability of the same does not satisfy the constraint, then discard that node from the larger local community, which in turn generates new sub graph.
11. If all nodes fulfil the constraint of rate function and its reliability, then return to step 4.

Credential Acclaim based Information Spreading (CAIS) Mechanism
After discovering several communities and regularizing it from a SIoT environment, next our aim is to increase the information spreading quality of then communities formed. It can be achieved only by electing a leader, who is capable of spreading the maximum information among the nodes of their respective community. In this work, we used Credential Acclaim (1) Then the fitness of the sail fishes and sardines are estimated using the equations (3) and (4)

11
The position of the best sail fish and injured sardine with the best fitness value is saved in each iteration and considered as the elite, the position of sail fishes and sardines are then updated towards the best solution as given in the equations (5) and (6).
In the equations (5) and (6) Then estimate the fitness of the sardine fishes, if there is a better fitness solution in sardine, then that injured sardine is replaced with the elite sail fish as given in the equation (7) = ; ( ) > ( ) Such best fitted sail fishes are given higher credentials and are considered as the leader of the other nodes in its respective community. The content of the information is spread using the elected leader via SFO algorithm. The entire CAIS algorithm is described in the following steps of Algorithm 2

2.
Calculate the fitness of the sail fishes and sardines using the equations (3) and (4) 3.

7.
Update the position of all the sardine fishes using the equation (6), if > 0.5

9.
Update the position of selected sail fish using the equation (5) 10. Calculate the fitness of all the sardines using the equation (4) 11. Replace elite sail fish by injured sardine using the equation (7) 12. Remove the hunted sardine fish from the population 13. Update the best sail fish and sardine 14. Give high credentials to such best sail fish and sardine fishes. 13 15. Choose the highest credentialed fish/node as the leader for its own community 16. Spread the content of the information via such an elected leader 17. End

Experimental Evaluation
To prove the versatility of our proposed model, we validated it using three different real world datasets are Center for Advanced Studies in Adaptive Systems (CASAS) [15], Massachusetts Institute of Technology (MIT) [16] and Activity Recognition with Ambient Sensing (ARAS) [17] for recognising the actions using machine learning. We tested our model on 16 subjects. Totally 15791 number of actions are collected from 427 sensors which are preinstalled in 11 flats. The details of the datasets used in our work is shown in the Table 1.

Results and Discussions
The most commonly used metrics in the evaluation of community detection algorithms are influence speed, Normalized Mutual Information (NMI), modularity, F-measure, precision, recall and computation time [17]. Influence spread is maximized based on the credentials acquired by the nodes which are elected as the leader via SFO algorithm. Fig. 1. depicts the influence spread plot for different community detection algorithms. It is evident that the information spread is maximum for our proposed attribute-based method with the increase in the credentials. NMI is estimated via a confusion matrix, each row corresponds to the number of originally existing community and each column corresponds to the number of detected communities in a SIoT network.

Conclusion
Proposed an attribute-based weighted mechanism for the detection of communities in a SIoT environment. We focussed our work in dividing the SIoT network into high quality Sociality based Weighted Communities (SWC). We exploited the important social attributes like user preferences, social similarities, mutual friends' degrees are main metrics for achieving the best rate function. The weighted method based on these social attributes determine the nodes to be present in their respective communities. We also presented a mechanism for controlling the local community augmentation. A Credential Acclaimed Information Spreading (CAIS) mechanism is implemented for selecting the best node with the maximum credential to surge the content spreading behaviour in the detected communities of SIoT network. Experimental results prove that the proposed social-driven attribute based weighted mechanism for