An Edge-based Clustering Algorithm to Detect Social Circles in Ego Networks

Transcription

1 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER An Edge-based Clustering Algorithm to Detet Soial Cirles in Ego Networks Yu Wang Shool of Computer Siene and Tehnology, Xidian University Xi an,77, China Shool of Eonomis and Management, Xidian University Xi an 77, China Lin Gao Shool of Computer Siene and Tehnology, Xidian University Xi an,77, China Abstrat Organizing users friends in personal soial networks, i.e., ego networks, into irles is an important task for online soial networks. Soial networking sites allow users to manually ategorize their friends into soial irles. However, it is time onsuming and does not update automatially as a user adds more friends. In this paper, we propose an edge-based lustering algorithm to detet soial irles in ego networks automatially. Firstly, we reonstrut ego networks by prediting the missing links. Then, we define the similarity of adjaent edges and luster edges by single-linkage hierarhial lustering algorithm. Finally, we label eah irle by abstrating its ommon properties to explain why this irle forms. The experimental results demonstrate it is a better way to haraterize soial irles from the respet of edges. Our algorithm outperforms the link ommunity algorithm and low-rank embedding algorithm in terms of auray, and is more effiient than the probabilisti model algorithm. Our proposed method is validated as an effetive algorithm in identifying soial irles. They are serving as a ommuniation and information-sharing platform for their users. Users in suh online soial networks usually have hundreds of their friends and aquaintanes. So it is neessary to organize their friends into what we refer to as soial irles. Deteting soial irles for a speifi user an be formulated as an ego network lustering problem. The ego network of a user is the network of friendships among his friends. An illustration of ego network is shown in Fig.. In this example, user s friends are organized into four irles. The irle of olleagues is overlapped with the irle of ollege friends. Index Terms graph lustering, ego networks, soial irles, link ommunity I. INTRODUCTION Understanding and modeling network strutures have been a fous of attention in a number of diverse fields, inluding physis, biology, omputer siene, statistis, and soial sienes. Soial network analysis started in the 93's and has beome one of the most important topis in soiology [][2]. A fundamental problem in the study of soial networks is ommunity detetion and some algorithms are proposed to solve this problem[3][4][5][6]. Communities an have onrete appliations. For instane, identifying lusters of ustomers with similar interests in the network of purhase relationships between ustomers and produts of online retailers (like, e. g., enables to set up effiient reommendation systems [7], that better guide ustomers through the list of items of the retailer and enhane the business opportunities. In reent years, Soial Networking Servies (SNS) suh as Faebook and Twitter have beome a major part of people s daily life. Fig.. An illustration of ego networks. Previous ommunity detetion methods have two main disadvantages in solving this problem. One flaw is they find dense subgraphs as ommunities in the observed network. However, there are hidden relations in ego networks. For example, some people are friends and form a irle but they have not built onnetions on website yet. If we only onsider the density of subgraphs, we will lose these irles. Another limitation is they don t give explanations about why we organize these people as a irle: they are families, olleagues, living in the same towns or something else. To overome these problems, 23 ACADEMY PUBLISHER doi:34/jp

2 2576 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER 23 additional information suh as user profiles is used. Tetsuya Yoshida [8] proposed low-rank embedding approah to find hidden ommunities based on user profiles. Nevertheless, this approah needs users to speify the number of irles in advane and annot identify the overlapped ommunities. Stephan Günnemann et al. [9] developed the algorithm DB-CSC to find density-based subspae in graphs with feature vetors. This algorithm misses finding irles with small size. These two methods still annot interpret what makes eah irle emerge. Julian MAuley et al. [] suggested a probabilisti model algorithm to disover irles. This approah an not only identify irles, but also expliate what dimensions of the profile aused the irle to form. However, this approah never identifies more than irles per user. What s more, the effiieny of this approah is very low. It needs hours of time to exeute when the size of the ego network is bigger than. Considering the fat that a irle forms beause people in that irle have the same relationship, it is a natural thought to identify irles by finding similar edges, sine edges represent relationships in graph model. In this paper, we detet irles in ego networks by an improved link ommunity algorithm. Cirles we identified are not only densely onneted but its members also share ommon properties. We first reonstrut ego networks to reover the missing relations. Then we give a new similarity definition of the edges and luster edges by link ommunity algorithm. Finally, we label irles to expliate what features aused these irles emergene. The experimental results show that ompared with link ommunity algorithm and low-rank embedding algorithm, our method ahieves a higher auray between the predited irles and the known irles. Our algorithm outperforms the probabilisti model algorithm in terms of running time. Our proposed method is validated as an effetive algorithm in disovering soial irles in ego networks. II. METHODS A. Link Community Algorithm Ahn et al. [] suggested an unorthodox approah whih reinvented ommunities as groups of edges rather than verties. They defined the inlusive neighbors of a vertex i as: n () i { x d(, i x) }, where dix (, ) is + the length of the shortest path between verties i and x. The set simply ontains the vertex itself and its neighbors. From this, they defined the similarity between eah pair of adjaent edges. The similarity between edges e and e ik jk is n () i n ( j) + + Se (, e ) = ik jk n () i n ( j) + +. Single-linkage hierarhial lustering is employed to build a dendrogram where eah leaf is an edge from the original network and branhes represent link ommunities. To solve the problem of utting this dendrogram at whih level, oneptions of link density and partition density are given. For a network with M edges and N verties, P = { P,..., P } C is a partition of the edges into C subsets. The number of edges in subset P is m = P. The number of indued verties, all verties that those edges touh, is n = {, i j}. eij P The link density D of ommunity is m ( n ) D =. n ( n )/2 ( n ) The partition density D is 2 m ( n ) D = m, whih is the average of M ( n 2)( n ) D. The dendrogram ould be ut at the level with maximum partition density, but it doesn t mean it is the only level worth exploring. For example, students at university may organize at multiple sales, from shools, down to the departmental level and then further down to researh groups and small-sale ollaborations. The most modular struture may form at the departmental level, but the strutures of both smaller researh groups and larger shool-wide organization are still relevant. B. Network Reonstrution Ego networks are dynami, and new relationships build over time. A group of friends with relatively sparse relationships now ould be found as a irle later when the network grows with more edges in future. To identify suh irles, we predit those missing links and reonstrut ego networks. We predit there should be an edge between two verties if these two verties are similar in both topologial struture and feature profile. In our datasets, the profile information of all people in an ego network an be represented as a tree. An example is shown in Fig. 2. Users features are lassified. A leaf of the tree represents a speifi feature that at least one person in the ego network has. Note that user profile trees are defined per ego network: while many thousands of olleges (for example) exist among all soial network users, only a small number appear among any partiular user s friends. We use the profile vetor to haraterize a user s feature. If a person x whose hometown is Boston is working as a programmer and has studied in MIT and Stanford, then his profile vetor, profile( x ), is (,,,,,,,). We use Jaard oeffiient to measure the topologial similarity, sine it onsumes little time and performs relatively good among many loal indies [2]. We use osine whih is ommonly used to measure the similarity of two profile vetors. Algorithm is used to reonstrut ego network. For eah pair of diretly linking verties, we ompute the topologial similarity and profile similarity of them in line 2-5. We label the median of all the topologial similarities as M _ VTS and the median of all the profile similarities as M _ VPS in line 6-7. We predit whether there should be a missing link in line 23 ACADEMY PUBLISHER

3 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER For eah pair of verties who has no diret onnetion, if their topologial similarity is bigger than M _ VTS and profile similarity is bigger than M _ VPS, we link these two verties. Fig. 2. An illustration of user profile tree. Algorithm. Ego network reonstruting algorithm. Input: The adjaent matrix A of an ego network G = ( V, E) ; Users profile information. Output: The adjaent matrix RA of the reonstruted ego network RG = ( V, E R ). Steps: : RA = A ; VTS =Φ; VPS = Φ ; //Initialization //Set VTS stores the topologial similarities of verties. //Set VPS stores the profile similarities of verties. 2: for eah edge e E do 3: alulate the topologial similarity of the two verties linked by edge e and insert it in set VTS ; 4: alulate the profile similarity of the two verties linked by edge e and insert it in set VPS ; 5: endfor 6: ompute the median of set VTS, M _ VTS ; 7: ompute the median of set VPS, M _ VTS ; 8: for eah vertex i V do 9: for eah vertex j V do : if vts( i, j) > M_VTS and vps( i, j) > M_VPS do // vts( i, j ) is the topologial similarity of vertex i and // vertex j. // vps( i, j ) is the profile similarity of vertex i and // vertex j. : set RA( i, j ) = ; 2: endif 3: endfor 4: endfor 5: return RA C. Cirle Detetion Our irle detetion algorithm ontains three steps. Firstly, we ompute the similarity of eah pair of edges who have ommon vertex. The similarity of edges onsists of two parts: topologial similarity and profile similarity. The definition of the topologial similarity of two adjaent edges, lts( e, e ), is the same as what in ik jk link ommunity algorithm. The profile similarity of edges e and e is defined as ik jk lps( e, e ) = os( profile( i), profile( j)). We use ik jk parameter α to balane the effet of topologial struture and profile information. The final similarity of edges e and e is ik jk S( e, e ) = α lts( e, e ) + ( α) lps( e, e ). Then ik jk ik jk ik jk we employ single-linkage hierarhial lustering algorithm to partition these edges. The indued verties of similar edges onstitute a irle. To simplify the lustering proess, we utilize similarity threshold parameter τ to deide the luster granularity. Finally, we abstrat ommon features from the edges in a irle to explain why this irle emerges. Algorithm 2 is applied to detet the irles. For eah edge in the reonstruted network RG = ( V, E R ), if it isn t in a known irle, we insert it into an empty set named by edges _ irle in line 5-9. Then we find the edges similar to any of the edge in set edges _ irle, and insert these edges into set edges _ irle in line -2 (When we talk about two edges are similar, we mean the similarity between them is bigger than the parameter τ ). This proess will stop until we ouldn t insert any edge into set edges _ irle. The indued verties of the edges _ irle form a irle and we insert this irle into set CS in line 7. Finally, we obtain the ommon features of the edges in set edges _ irle by the irle feature abstrat proedure in line 8. Algorithm 2. Cirle deteting algorithm Input: The adjaent matrix RA of the reonstruted ego network RG = ( V, E R ); Users profile information; Similarity threshold parameter τ. Output: The set of soial irles, SC ; The set of irle feature vetors, CFV. // Eah irle feature vetor is orresponding to a irle // to interpret the reason this irle exists. Steps: : SC = Φ ; 2: CFV = Φ ; 3: Searhed _ Edges = Φ ; // Set Searhed _ Edges stores the edges in known // ommunities 23 ACADEMY PUBLISHER

4 2578 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER 23 4: ompute the similarity of eah pair of edges in graph RG = ( V, E R ); 5: for eah edge i E do R 6: if i Searhed _ Edges do 7: insert i into Searhed _ Edges ; 8: edges _ irle =Φ; 9: insert i into edges _ irle ; : for eah edge j edges _ irle do : if edge k E and R k Searhed _ Edges and Sk (, j) > τ do 2: insert k into edges _ irle ; 3: insert k into Searhed _ Edges ; 4: endif 5: endfor 6: endif 7: insert the indued verties of edges _ irle into SC ; 8: irle _ feature _ vetor =irle feature abstrat( edges _ irle ); 9: insert irle _ feature _ vetor into CFV ; 2: endfor 2: return SC, CFV A irle forms beause people in this irle have the same relationship. We an explain the meaning of a irle by abstrating the ommon features of the edges in the irle. The feature vetor of an edge e i, j is the ommon features of vertex i and vertex j, whih is defended as edge _ feature _ vetor( e ) = profile () i profile ( j) ij,. In a irle, if there are some features appearing in almost every edge feature vetor, and then it is reasonable to say these features ause the existene of the irle. Algorithm 3 is a subprogram of algorithm 2. It omputes whih dimensions of users profile information lead to densely linked irles. Algorithm 3 returns a vetor, alled irle feature vetor. The value of element i in the vetor is a weight, and it represents the ontribution of feature i to forming the irle. If it is big enough, say,, then we an onsider that feature i should be responsible for the emergene of the irle. The detail of algorithm 3 is as follows: Algorithm 3. Cirle feature abstrat ( ES ) Input: Edge set ES of irle C. Output: The irle feature vetor of irle C. Steps: : sum _ vetor = ; //initialization 2: for eah edge e ES, do 3: sum _ vetor = sum _ vetor + edge _ feature _ vetor( e); 4: endfor sum_ vetor 5: irle _ feature _ vetor = ; ES 6: return irle _ feature _ vetor D. Our Algorithm Individuals an partiipate in multiple irles, but relationships often exist for one dominant reason. It is appropriate to onsider irles as sets of similar onnetion. Based on the analysis above, our method identifies irles by partitioning edges.we give the overall framework of our method in algorithm 4. Algorithm 4. Our algorithm Input: The adjaent matrix A of an ego network G = ( V, E) ; Users profile information; Similarity threshold parameter τ. Output: The set of soial irles; The set of irle feature vetors. Steps:. Run algorithm to get a reonstruted ego network; 2. Run algorithm 2 to detet soial irles and the meaning of eah irle. 3. Return the set of soial irles and the set of irle feature vetors. Our method an detet overlapping and hierarhial 2 irles simultaneously. The time omplexity is Om ( ), whih is relatively high. Here m is the number of edges in the ego network. Fortunately, sine the sale of an ego network is usually not so big, the running time of our algorithm is aeptable. III. EXPERIMENTS AND RESULTS A. Experimental Data We obtained ego networks and ground-truth from Faebook. The ground-truth data work as benhmark. We obtained profile and network data from ego networks, onsisting of 93 irles and 4,39 users. On average, users identified 9 irles in their ego networks, with an average irle size of 22 friends. Examples of suh irles inlude students of ommon universities, sports teams, relatives, et. The same dataset are used in paper 6. B. Evaluation Metris We validate the auray of our algorithm by omparing the set of predited irles P with the set of benhmark irles B. Two evaluation metris are employed, -BER and F-measure. ) -BER To measure the alignment between a predited irle p and a benhmark irle b, we ompute the Balaned Error Rate (BER) between these two irles [3], p\ b b\ p BER( b, p) = +. 2 p b This measure assigns equal importane to false 23 ACADEMY PUBLISHER

5 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER positives and false negatives, so that trivial or random preditions inur an error of on average. Sine we do not know the orrespondene between irles in B and P, we ompute the optimal math via linear assignment by maximizing: max ( BER( p, f ( p))), f : P B f p dom( f ) where f is a (partial) orrespondene between P and B. That is, if the number of predited irles P is less than the number of benhmark irles B, then every irle p P must have a math b B, but if P > B,we do not inur a penalty for additional preditions that ould have been irles but were not inluded in the ground-truth. 2) F-measure To measure whether a predited irle p and a ground-truth irle b are mathed or not, we ompute neighborhood affinity sore (NA) between the two irles, 2 b p NA( p, b) =. If NA( p, b) ω, then they are b p onsidered to be mathing, and ω is usually set as.2. We use N to represent the number of predited irles p whih math at least a real irle. The mathemati N = p p P, b B, NA( p, b) ω. expression is { } p N is the number of real irles whih math at least a b predited one, the mathemati expression is N = { b b B, p P, NA( p, b) ω}. Preision and b reall are defined as [4]: N p Preision = Reall = P B F-measure whih is the harmoni mean of preision and reall is defined as: F measure = 2 Preision Reall /( Preision + Reall ). It is used to evaluate the overall performane of the predit results. C. Our Results ) Parameter analysis Our algorithm employs two parameters, inluding similarity threshold parameter τ used to ontrol lustering granularity and balane parameter α used to balane the effet of topologial struture and profile information. Sine the preferene τ is inherited from link ommunity algorithm, we didn t analyze it in this paper and just set it as the average similarity of edge pairs. To study how the parameter α affets the results, we use. as the step length and inrease α from. to.9 step by step. To explore the relationship between α and the number of predited irles, we randomly hoose 5 ego networks to observe the hanging trends. The experimental results displayed in Fig. 3 illustrate that with the inrement of α, the number of predited irles N b inreases too. This is beause of the way we define the topologial similarity and profile similarity. We sort topologial similarities and profile similarities of all the adjaent edges in an ego network, and find that profile similarity inreases slowly. An example is shown in Fig. 4. We utilize the average topologial similarity and the average profile similarity of edges as the thresholds to ut topologial similarity urve and profile similarity urve respetively. Muh more edge pairs are onsidered dissimilar in the topologial similarity urve, and this may ause relatively small irles. Parameter α is the weight of topologial similarity, so when we set similarity threshold parameter τ as the average similarity of edges, the bigger α is, the more dissimilar edge pairs are. An illustration is displayed in Fig. 5. Then we study how the α affets -BER and F-measure. The experimental results are displayed in Fig. 6 and Fig. 7. In Fig. 6, hanging regulations between α and -BER are not very lear. The same situation is observed in Fig. 7. This is beause the profile and topologial information are not the only effetive fators for users to organize their irles. Many irles are reated using information that users do not expliitly make available on soial networks, suh as a irle of best friends. So we an t give a preise relationship between α and the auray evaluation metris. However, we notie that in our data set, when α is small, say.2 or.3, we get the higher -BER and F-measure. So we set α =.3 in our experiments. number of predited irles ego network ego network 2 ego network 3 ego network 4 ego network parameter alpha Fig. 3. The relationship between parameter α and the number of predited irles. 5 ego networks are employed to illustrate the hanging trend. similarity topologial similarity profile similarity x 4 edge pairs with ommon vertex Fig. 4. The inrement urves of topologial similarity and profile similarity in an ego network. Vertial lines represent utting these two urves with their average similarity. The average similarity of topologial similarity urve is.25, and the similarities of 44 edge 23 ACADEMY PUBLISHER

6 258 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER 23 pairs are lower than.25. The average similarity of profile similarity urve is.36, and the similarities of 383 edge pairs are lower than.36. shool and having the same level of eduation and living together. similarity.9.7 alpha=. alpha=.3 alpha= alpha=.7 alpha=.9 weight same gender same first name Fig edge pairs with ommom vertex x 4 The inrement urves of similarity with different parameter α in an ego network. Vertial lines represent utting these urves with their average similarity respetively. With the inrement of α, there are more and more edge pairs whose similarity is lower than average similarity appear. -BER.9.7 ego network ego network 2 ego network 3 ego network 4 ego network 5 weight same shool feature vetor for a irle (A) work for the same employer feature vetor for a irle (B) same loal parameter alpha.7 same language Fig. 6. The relationship between parameter α and -BER. weight F-measure Fig ego network ego network 2 ego network 3.3 ego network 4 ego network parameter alpha The relationship between parameter α and F-measure. 2) Cirles deteted by our algorithm On average, our algorithm identified 2 irles in ego networks, with an average irle size of 23. It is very lose to the real situation. Our -BER sore is.76 and f-measure sore is 2. Fig. 8 shows the irle feature vetors of 4 irles we deteted. People in Fig. 8 (A) have the same gender and they are relatives. People in Fig. 8(B) are olleagues and they live in the same town. People in Fig. 8(C) are friends who speak the same language. People in Fig. 8(D) are shoolmate. They are onentrating on the same speialty, learning in the same. same onentration weight Fig. 8. feature vetor for a irle (C) same shool same level fo eduation same loal feature vetor for a irle (D) Cirle feature vetors of four irles. D. Experimental Results Comparative Evaluation In this setion, we ompare the performane, inluding auray and running time, of our algorithm with link ommunity, low-rank embedding and probabilisti model algorithm. Fig. 9 shows the results of the omparison of our 23 ACADEMY PUBLISHER

7 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER algorithm and three ompetitive approahes on -BER and F-measure. The -BER sore of our algorithm is.76, whih is 26.7% and 2% higher than link ommunity and low-rank embedding. The F-measure of our algorithm is 2, whih is 25.9%, 88.9% higher than link ommunity and low-rank embedding. In the respet of auray, the probabilisti model is better than us. Our -BER sore and F-measure are 9.5% and.9% lower than that of probabilisti model..2 -BER low-rank embedding probabilisti model our algorithm link ommunity F-meature Fig. 9. The omparisons between link ommunity, probability mode, low-rank embedding and our algorithm on -BER and F-measure. Fig. displays the running time of 4 algorithms. We hoose 3 ego networks with different sales. When the vertex number is 66, the running time of link ommunity, low-rank embedding and our approah is less than seond, while probabilisti mode needs 5 seonds. When the vertex number is 59, our algorithm needs seonds, whih is only a little more than link ommunity. The running time of probabilisti mode is 79 seonds at this sale. When the sale of the network is 46, the running time of our algorithm is 832 seonds. It is almost twie longer than that of low-rank embedding. Probabilisti model approah needs almost 3 hours to exeute the omputing proess. running time (se) our algorithm probabilisti model Low-Rank Embedding link ommunity vertex numbers Fig.. The omparisons between link ommunity, probability mode, low-rank embedding and our algorithm on running time. IV. CONCLUSION AND DISCUSSION One important feature of personal soial networks is the ability to group friends and ontats into irles. Creating and maintaining these irles is a time onsuming business. In this paper, we propose an effiient algorithm to analyze a user s entire list of ontats and automatially divide them up into useful irles. We improve link ommunity algorithm by adding users profile information to detet irles. And alulate what features ause the emergene of a irle. The experimental results show that our algorithm is more aurate than link ommunity and low-rank embedding method. Although our algorithm is less aurate than probabilisti mode approah, it is omparable. And the running time we need is far less than that of probabilisti mode approah. Our method gets a good balane between auray and effiieny. Experimental results demonstrate that haraterizing soial irles from the respet of edge is reasonable. Our proposed method is an effetive algorithm in identifying soial irles. ACKNOWLEDGMENT This work was supported by the National Natural Siene Foundation of China (Nos , 936, 657 and 6246), Provinial Soial Siene Foundation of Shaanxi (No. M6) and the Fundamental Researh Funds for the Central Universities (No. K5564). REFERENCE [] J. Sott, Soial Network Analysis: A Handbook, 2, SAGE Publiations, London, UK. [2] S. Wasserman and K. Faust, Soial network analysis,994, Cambridge University Press, Cambridge, UK. [3] K. P. Reddy, M. Kitsuregawa, P. Sreekanth, and S. S. Rao, A Graph Based Approah to Extrat a Neighborhood Customer Community for Collaborative Filtering, In Proeedings of the Seond International Workshop on Databases in Networked Information Systems, 22, pp [4] M. Girvan and M. Newman, Community struture in soial and biologial networks, Pro Natl Aad Si, Vol. 99, 22, pp [5] Xianghua Fu, Chao Wang, Zhiqiang Wang and Zhong Ming, Threshold Random Walkers for Community Struture Detetion in Complex Networks, Journal of Software, 23, Vol. 8, No. 2, pp [6] Yueping Li, Yunming Ye and Xiaolin Du, A New Vertex Similarity Metri for Community Disovery: a Loal Flow Model, Journal of Software, 2, Vol. 6, No. 8, pp [7] Chengying Mao, A Heuristi Algorithm for Bipartite Community Detetion in Soial Networks, Journal of Software, 22, Vol. 7, No., 24-2 [8] T. Yoshida, Toward finding hidden ommunities based on user profiles, In Proeeding of the IEEE International Conferene on Data Mining Workshops, 2. pp [9] Stephan Günnemann, Brigitte Boden, Thomas Seidl, Finding density-based subspae lusters in graphs with feature vetors, Data Mining and Knowledge Disovery, 22, vol.25, pp [] Julian MAuley and Jure Leskove, Learning to disover soial irles in ego networks, Neural Information Proessing Systems, 22, pp [] Y. Y. Ahn, J. P. Bagrow and S. Lehmann, Link ommunities reveal multisale omplexity in networks, Nature, 2, 466:76 [2] D. Liben-Nowell and J. Kleinberg, The link-predition problem for soial networks, Journal of the Amerian 23 ACADEMY PUBLISHER

8 2582 JOURNAL OF COMPUTERS, VOL. 8, NO., OCTOBER 23 Soiety for Information Siene and Tehnology, 27, 58 (7): pp. 9-3 [3] Y. Chen and C. Lin, Combining SVMs with various feature seletion strategies, Studies in Fuzziness and Soft Computing, 26, Vol. 27, pp [4] C. J. van Rijsbergen, Information Retrieval, Butterworth 979 Yu Wang is urrently a PhD student at the Shool of Computer Siene and Tehnology of Xidian University Xi an, China. She reeived her Master of Management from the Shool of Eonomis and Management of Xidian University (26) and her Bahelor of Management from Shool of Eonomis and Management of Xidian University (23). Her researh interests inlude data mining and omplex network. Lin Gao reeived the B.S and M.S. in Computational Mathematis from Xi an Jiaotong University and Northwest University in 987 and 99, respetively, and the. Ph. D. degree inciruit and System from Eletronis Engineering Institute from Xidian University in 23. She was a visiting sholar at University of Guelph, Canada from 24 to 25. At present, she is a professor in the Department of Computer Siene and Tehnology, Xidian University. Her researh interests inlude bioinformatis, data mining in biologial data, graph theory and intelligene omputation. 23 ACADEMY PUBLISHER