CN104008165A - Club detecting method based on network topology and node attribute - Google Patents

Abstract

The invention discloses a club detecting method based on network topology and node attribute. The club detecting method based on the network topology and the node attribute comprises the steps of: firstly, obtain significance of each node in a network based on global topology by utilizing a PageRank algorithm and calculating a structural linkage closeness between nodes by utilizing the neighbor structure of the nodes; then adopting different extraction ways and attribute similarity calculation ways based on different forms of the node attribute; then adjusting the balance by using a weight adjustment factor; finally regarding the similarity of the node neighbor and the category center node neighbor as the degree of the node which belongs to the category in a clustering process. According to the club detecting method based on the network topology and the node attribute disclosed by the invention, the obtained club is relatively high in linkage closeness and relatively good in homogeneity.

Description

Corporations' detection method of a kind of topological structure Network Based and nodal community

Technical field

The present invention relates to similarity measurement technology between PageRank technology, analysis of networks topology technology, topic model technology, attribute extraction technology, attribute, corporations' detection and clustering algorithm technical field, specifically a kind of corporations' detection technique based on complex network topologies and network node attribute information.

Background technology

Complex network more and more receives publicity in recent years, such as social networks, and scientist's coauthorship network, electronic mail network etc.In these networks, node can be expressed as a people or one piece of article etc., and attribute information in network (such as social user on the network's sex, hobby; Author's research field on paper coauthorship network), the attribute information that the link between nodes and node have has formed attributed graph.The target that corporations based on attributed graph are detected is to divide this attributed graph, make the node of same corporations tightr than the node connection between different corporations, and same corporations built-in attribute is similar as far as possible.

Traditional corporations' detection method major part is the topological structure (being the linking relationship between node) based on figure, based on attribute similarity.Method based on topological structure mainly can be divided into three types: figure division, figure density, hierarchical structure cluster.Figure divides by edged in network chart or the mode of deleting limit and reaches corporations' testing goal; The link density of the topological structure of figure density based on figure is carried out corporations' detection; The intensity of hierarchical structure cluster based on linking between node is divided into some corporations by network.Also have in addition the method based on genetic optimization, utilize the survival of the fittest of biological evolution thought, from objective function, select optimum individual to be final corporations and divide.Method common practices based on attribute similarity is that by the same corporations of being divided into of attribute similarity, the attribute information that the corporations that finally obtain can meet same corporations is similar by certain determination methods.

The method of independent topological structure Network Based, does not consider the impact that attribute forms testing process above, thereby the corporations' internal node link tight ness rating detecting is higher, but between node, attributes similarity is not high; Method based on attribute similarity, does not consider network topology structure, thereby the corporations that detect often have attribute similarity, but the link tight ness rating of corporations inside is not high.Consider above-mentioned deficiency, scholar combines consideration network topology structure and attribute information in recent years, mainly contains: the method based on distance and the method based on model.Method based on distance is in conjunction with the attribute information of topological structure and node, utilize the distance between node in Random Walk Algorithm computational grid (saying it is the similarity between node in a certain meaning), thereby obtain the distance between node, then by certain clustering method, node is carried out to cluster.Method based on model is mainly to utilize probability model, and the affiliated corporations of node are probability distribution, and by weighing the probability connecting between node, recycling clustering algorithm, or maximal possibility estimation algorithm, carry out corporations' detection.Yet the distance algorithm based on random walk relates to matrix multiple, there is higher time complexity; Current most of method is not considered the overall importance of network node in network topology structure in addition, the overall importance of node and local topology (being this characteristic of neighbours of node) is not combined.

Summary of the invention

The topological structure a kind of Network Based that provides for the deficiencies in the prior art and node attribute information combine, corporations' detection method more is efficiently provided, the method has proposed node for digraph and has combined in the importance of network topology structure and local neighbours' structure of node, obtain the link strength between node, in conjunction with the attributes similarity between node, finally by clustering method, node rendezvous is arrived to corresponding corporations again.

The concrete technical scheme that realizes the object of the invention is:

Corporations' detection method of topological structure Network Based and nodal community, the method comprises the following steps:

A) node topology structure link intensive analysis

The importance of node based on topological structure in computational grid, utilizes the local neighbours of node, again the link strength between node metric;

B) nodal community extracts and similarity measurement

Extract attribute, then according to different attribute types, the similarity of each attribute between computing node, the similarity of all properties between last computing node;

C) topological structure combines with attribute

After the link strength and nodal community similarity of trying to achieve between node, the heavy regulatory factor of exploitation right is in conjunction with the two, as the similarity based on topological structure and attribute between node.

D) node clustering

First initialization classification Centroid, then carries out node category division according to the similarity at each node and classification center, then upgrades classification Centroid, and last calculating target function value judges whether convergence.

Described step is the middle importance of nodes based on topological structure of calculating a), specifically comprises:

First build the adjacency matrix A of node, read after whole network structure, build adjacency matrix, if there is node v _ito v _jlink, corresponding position is set to 1, otherwise is 0;

Adopt the PageRank value of PageRank algorithm computing node based on topological structure, this value is the importance degree based on overall network topology structure as node again, obtains the importance b of each node after algorithm operation _i;

Described step a) in the link strength between node metric again, specifically comprise:

Node v _iand v _jbetween link strength be node v _ito v _jlink strength and node v _jto v _ilink strength sum; And v _ito v _jlink strength be v _joverall importance b _jdivided by node v _iall chains go out the overall importance sum of neighbor node; In like manner calculate v _jto v _ilink strength; Node v _ito v _jlink strength TS (i, j) computing formula as follows:

Wherein, NO (i) represents node v _ichain go out neighbours collection, according to above-mentioned computing formula, obtain node v _iand v _jbetween link strength T _ss (i, j), is calculated as follows:

T _sS(i,j)＝TS(i,j)+TS(j,i)。

Described step b) in, extract attribute, specifically comprise:

If the descriptor of node is content of text, the theme of using potential Di Li Cray (LDA) topic model to extract text information distributes, and then this theme is distributed as the attribute of this node; If the descriptor of node is discrete or continuous, the direct attribute using them as node.

Described step b) similarity of each attribute between computing node in, specifically comprises:

For discrete type attribute, directly compare the whether identical of attribute; For continuous type attribute, utilize the similarity of Euclid's formula computation attribute; For text-type attribute, after obtaining the distribution of text theme, utilize cosine similarity based method metric attribute similarity.

Described step b) in, between node, all properties similarity is calculated, and specifically comprises:

After obtaining the similarity of each attribute between node, these attributes similarities are unifiedly calculated to attributes similarity XS (i, j) all as between node; Be calculated as follows:

$\mathrm{XS}(i,j)=\frac{{\mathrm{\Σ}}_{r=1}^m{\mathrm{\ω}}_{a_{\mathrm{ijr}}}}{{\mathrm{\Σ}}_{r=1}^m{w}_{a_r}\·w_{a_r}}$

Wherein represent node v _iand v _jat single attribute a _ron similarity, wa _rrepresent attribute a _rweight, m represents the attribute number of node.

Described step c) the similarity PaS (i, j) based on topological structure and attribute between node in, is specifically calculated as follows:

PaS(i,j)＝(1-λ)·T _sS(i,j)+λ·XS(i,j),λ∈(0,1)。

Wherein λ is that the weight of link strength and attributes similarity between node regulates parameter.

Described steps d) in, initialization classification Centroid is: the similarity in conjunction with node based on structure and attribute, first by introducing function, the similarity between node is mapped to [0,1) scope, and using the node similarity of this value after node normalization; Then calculate the similarity of other nodes of each node and whole network, obtain node in the importance based on topological structure and attribute of global network, this importance is as the global impact power of node, by the size of node influence power, carry out descending sort, before selecting, the node of K influence power maximum is as the Centroid of an initialization K classification.

Described steps d) in, node category division specifically comprises:

Calculate each node v _icentroid c with each classification _jsimilarity, try to achieve first respectively node v _iwith Centroid c _jneighbor node collection, computing node v then _iand c _jneighbor node between similarity S (i, j), comprising chain, enter neighbours and chain goes out neighbours, using this similarity as node v _ibe assigned to such other possibility, by calculating v _iwith each classification center c _jsimilarity, select wherein maximum one as v _ithis assigned classification of iteration; Be calculated as follows:

$S(i,j)=\frac{\mathrm{PaS}(\mathrm{NIO}\left(i\right),\mathrm{NIO}\left(j\right))}{\left|\mathrm{NIO}\left(i\right)\right|+\left|\mathrm{NIO}\left(j\right)\right|}$

Wherein PaS (NIO (i), NIO (j)) represents node v _ineighbours and Centroid c _jneighbours' similarity, | NIO (i) | represent node v _ineighbours' number, | NIO (j) | represent classification Centroid c _jneighbours' number.

Described steps d) in, upgrading classification Centroid is: after distributing the classification ownership of each node, calculate the average similarity between each classification internal node, immediate as Centroid with the average similarity of this classification for each one of classification selection, consider that the Centroid that chooses may be similar larger on other nodal communitys of this classification, and link is not strong in structure, in the process of selecting Centroid, increase a restrictive condition, the number of degrees of the Centroid of electing (all neighbours' numbers) are greater than the average number of degrees of this classification node.

Described steps d) in, calculate target function value and judge whether that convergence is: calculate whole network and be divided in the value on objective function, by with last iteration after the comparison of target function value, if the two differs in certain threshold value, iteration stops, otherwise continues next iteration.

The present invention can effectively and efficiently carry out corporations' detection to complex network, link between corporations' internal node that can make to detect is tight compared with the node link between corporations, the node of corporations inside has good attribute similarity simultaneously, has more embodied the homogeney of corporations.The present invention can provide strong support for the researchs such as analysis of complex network.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is embodiment of the present invention schematic flow sheet.

Embodiment

Embodiment

In order to further illustrate the present invention, now 2 describe the present invention by reference to the accompanying drawings.Concrete steps are as follows:

Step 1: obtain network topology structure, build node adjacency matrix

First from the node link relational file of data set, read the linking relationship between network node, this linking relationship is directive, build adjacency matrix, matrix intermediate value is that node corresponding to 1 expression row subscript exists node corresponding to sensing row subscript, and value is that the node of 0 expression correspondence does not exist linking relationship.The chain of the node that the capable subscript of the every line display of adjacency matrix is corresponding goes out neighbours' collection, and list shows that the chain of the node that row subscript is corresponding enters neighbours' collection.

Step 2: node importance in global Topological Structure is calculated

After obtaining the adjacency matrix of whole network, utilize the importance of PageRank algorithm computing node topological structure Network Based, computing method are as follows:

$\mathrm{PageRank}\left(v_i\right)=\frac{1-d}{\left|V\right|}+d\underset{v_j\∈\mathrm{NI}\left(v_i\right)}{\mathrm{\Σ}}\frac{\mathrm{PageRank}\left(v_j\right)}{\left|\mathrm{NO}\left(v_j\right)\right|}$

NI (v _i) be v _jchain enter neighbours collection, in whole network, there is v _ithe set of node of link, NO (v _j) be v _jchain go out neighbours collection, exist link from v _ito other node in network, the set that these nodes form, V is nodes sum, d is ratio of damping, is set to 0.5.After algorithm operation, can obtain the importance b of each node _i.

Step 3: according to link strength between neighbours' computing node of node

Link in digraph between node is directive, it is considered herein that arbitrary node v _iand v _jbetween link strength be: v _ito v _jlink strength and v _jto v _ilink strength sum.And v _ito v _jlink strength not only whether have with two nodes that to link (can from adjacency matrix, obtain) relevant, also with node v _jat v _ichain to go out the concentrated significance level of neighbours relevant, this significance level is v _jthe importance b based on global Topological Structure _jwith node v _iall chains business of going out neighbours' importance sum.Node v _ito v _jlink strength TS (i, j) computing formula as follows, if v _iand v _jdo not have link, link strength is 0.

Therefore node v _iand v _jlink strength T _ss (i, j) is TS (i, j)+TS (j, i), wherein b _jfor the PageRank value calculating, NO (i) represents node v _ichain go out neighbours collection.

Step 4: for each node of network, extract node attribute information

The attribute of node has three kinds of forms: discrete type, continuous type, text-type.The attribute extraction of its Chinese version adopts potential Di Li Cray (LDA) topic model to carry out subject extraction to it, and each theme extracting is the probability distribution about word, and every piece of text is the probability distribution about theme.Discrete type and attribute continuous type all obtain than being easier to, below main explanation how from text-type, to obtain information as the attribute of node.

LDA topic model can carry out subject extraction to text, and each text is finally expressed as theme probability distribution, and the probability distribution that each theme expression is word.For example at paper, collaborate in net, node represents Authors of Science Articles, and the relation between node represents that these authors collaborateed paper, every piece of title that paper text message is paper, summary, key word.Each author may have many pieces of papers, the text message of these papers is integrated into a larger text as the text of Authors of Science Articles, from these texts, set the number of topics of whole network, as following table, number of topics is set as 100, has enumerated wherein the probability distribution of five authors on five themes:

Enumerated the probability distribution of the word of Topic1 below:

Step 5: calculate in different ways similarity according to the form of expression of attribute

For multi-form attribute, adopt different similarity calculating methods, for the attribute of discrete type, directly relatively whether property value equates; For the attribute of continuous type, adopt Euclid's formula to calculate, similarity simA (i, j) formula is as follows:

$\mathrm{simA}(i,j)=\frac{1}{1+\sqrt{{\mathrm{\Σ}}_d^D{(v_i^d-v_j^d)}^2}}$

The dimension that wherein D is this attribute, for node v _ithe value of d dimension on this attribute.For the attribute of text-type, the theme that utilizes LDA topic model to try to achieve text distributes, then adopts cosine similarity calculating method.

Step 6: utilize all properties similarity between cosine similarity computing node

The attribute of node may contain various ways simultaneously, the similarity calculating method of every kind of form is different, thereby after obtaining the similarity of each attribute, adopts the computing method of cosine similarity, the value of trying to achieve is as the attribute similarity XS (i, j) between node.

Step 7: Analysis of Topological Structure is combined with attribute similarity

The exploitation right joint factor lambda of resetting is come the link strength of adjustment node based on topological structure and the attribute similarity between node.PaS (i, j) represent node between similarity.

PaS(i,j)＝(1-λ)·T _sS(i,j)+λ·XS(i,j),λ∈(0,1)

Step 8: classification center initialization

Similarity PaS (i, j) between any two nodes that the method for utilizing the similarity of above-mentioned node based on topological structure and attribute to combine is tried to achieve, node can be weighed with following formula in the importance of whole network:

$f_V\left(v_i\right)=\underset{v_j\∈V}{\mathrm{\Σ}}(1-e^{-\mathrm{PaS}(v_i,v_j)})$

Wherein V is the set of node in whole network, obtains, after the importance of each node, by importance, carrying out descending sort, and before selecting, the node of K maximum is as the initial center node of K classification.

Step 9: node classification ownership is differentiated

The present invention makes full use of the local neighbours of node, the neighbours that the classification ownership degree of node to be measured is node to be measured and the neighbours' of classification Centroid similarity, consider node neighbours number (neighbours are more on similarity calculating impact simultaneously, node to be measured and classification Centroid similarity are higher), the neighbours' number to it divided by two nodes.Similarity S (i, j) is higher shows that the probability that this node to be measured belongs to current classification is larger, has calculated after the ownership degree of node to be measured and all categories, selects maximum belonging to as node classification.

$S(i,j)=\frac{\mathrm{PaS}(\mathrm{NIO}\left(i\right),\mathrm{NIO}\left(j\right))}{\left|\mathrm{NIO}\left(i\right)\right|+\left|\mathrm{NIO}\left(j\right)\right|}$

Wherein NIO (i) represents node v _iall neighbours.| NIO (i) | be node v _iall neighbours' numbers.

Step 10: classification Centroid upgrades

Each iterative process all can be carried out classification distribution by node again, and upgrades the Centroid of each classification.First calculate the similarity between the node under this classification, then average, the similarity of Centroid under this classification approaches this mean value as far as possible, in this difference of the present invention way in the past, the present invention also in addition adds a restrictive condition, and the number of degrees of the Centroid of electing (i.e. the neighbor node of this node sum) are greater than the mean value of the number of degrees of such all nodes.Every iteration is once just upgraded the Centroid of each classification.After this iteration finishes, the value after recalculating whole network and dividing on objective function, by with last iteration after the comparison of target function value, if the two differs in certain threshold value, iteration stops, otherwise continues next iteration.

Claims (10)

1. corporations' detection method of topological structure Network Based and nodal community, is characterized in that, described method comprises the following steps:

A) node topology structure link intensive analysis

The importance of node based on topological structure in computational grid, utilizes the local neighbours of node, again the link strength between node metric;

B) nodal community extracts and similarity measurement

Extract attribute, then according to different attribute types, the similarity of each attribute between computing node, the similarity of all properties between last computing node;

C) topological structure combines with attribute

After the link strength and nodal community similarity of trying to achieve between node, the heavy regulatory factor of exploitation right is in conjunction with the two, as the similarity based on topological structure and attribute between node.

D) node clustering

First initialization classification Centroid, then carries out node category division according to the similarity at each node and classification center, then upgrades classification Centroid, and last calculating target function value judges whether convergence.

2. method according to claim 1, is characterized in that calculating the importance of nodes based on topological structure during described step is a), specifically comprises:

First build the adjacency matrix A of node, read after whole network structure, build adjacency matrix, if there is node v _ito v _jlink, corresponding position is set to 1, otherwise is 0;

Adopt the PageRank value of PageRank algorithm computing node based on topological structure, this value is the importance degree based on overall network topology structure as node again, obtains the importance b of each node after algorithm operation _i;

Described step a) in the link strength between node metric again, specifically comprise:

Node v _iand v _jbetween link strength be node v _ito v _jlink strength and node v _jto v _ilink strength sum; And v _ito v _jlink strength be v _joverall importance b _jdivided by node v _iall chains go out the overall importance sum of neighbor node; In like manner calculate v _jto v _ilink strength; Node v _ito v _jlink strength TS (i, j) computing formula as follows:

Wherein, NO (i) represents node v _ichain go out neighbours collection, according to above-mentioned computing formula, obtain node v _iand v _jbetween link strength T _ss (i, j), is calculated as follows:

T _sS(i,j)＝TS(i,j)+TS(j,i)。

3. method according to claim 1, is characterized in that described step b) in extract attribute, specifically comprise:

If the descriptor of node is content of text, with potential Di Li Cray LDA topic model, extract the theme distribution of text information, then this theme is distributed as the attribute of this node; If the descriptor of node is discrete or continuous, the direct attribute using them as node.

4. method according to claim 1, is characterized in that described step b) in the similarity of each attribute between computing node, specifically comprise:

For discrete type attribute, directly compare the whether identical of attribute; For continuous type attribute, utilize the similarity of Euclid's formula computation attribute; For text-type attribute, after obtaining the distribution of text theme, utilize cosine similarity based method metric attribute similarity.

5. method according to claim 1, is characterized in that described step b) between node all properties similarity calculate, specifically comprise:

After obtaining the similarity of each attribute between node, these attributes similarities are unifiedly calculated to attributes similarity XS (i, j) all as between node; Be calculated as follows:

$\mathrm{XS}(i,j)=\frac{{\mathrm{\Σ}}_{r=1}^m{\mathrm{\ω}}_{a_{\mathrm{ijr}}}}{{\mathrm{\Σ}}_{r=1}^m{w}_{a_r}\·w_{a_r}}$

Wherein represent node v _iand v _jat single attribute a _ron similarity, represent attribute a _rweight, m represents the attribute number of node.

6. method according to claim 1, is characterized in that described step c) in the similarity PaS (i, j) based on topological structure and attribute between node, be specifically calculated as follows:

PaS(i,j)＝(1-λ)·T _sS(i,j)+λ·XS(i,j),λ∈(0,1)

Wherein λ is that the weight of link strength and attributes similarity between node regulates parameter.

7. method according to claim 1, it is characterized in that described steps d) in initialization classification Centroid be: the similarity in conjunction with node based on structure and attribute, first by introducing function, the similarity between node is mapped to [0,1) scope, and using the node similarity of this value after node normalization; Then calculate the similarity of other nodes of each node and whole network, obtain node in the importance based on topological structure and attribute of global network, this importance is as the global impact power of node, by the size of node influence power, carry out descending sort, before selecting, the node of K influence power maximum is as the Centroid of an initialization K classification.

8. method according to claim 1, is characterized in that described steps d) in node category division specifically comprise:

Calculate each node v _icentroid c with each classification _jsimilarity, try to achieve first respectively node v _iwith Centroid c _jneighbor node collection, computing node v then _iand c _jneighbor node between similarity S (i, j), comprising chain, enter neighbours and chain goes out neighbours, using this similarity as node v _ibe assigned to such other possibility, by calculating v _iwith each classification center c _jsimilarity, select wherein maximum one as v _ithis assigned classification of iteration; Be calculated as follows:

$S(i,j)=\frac{\mathrm{PaS}(\mathrm{NIO}\left(i\right),\mathrm{NIO}\left(j\right))}{\left|\mathrm{NIO}\left(i\right)\right|+\left|\mathrm{NIO}\left(j\right)\right|}$

Wherein PaS (NIO (i), NIO (j)) represents node v _ineighbours and Centroid c _jneighbours' similarity, | NIO (i) | represent node v _ineighbours' number, | NIO (j) | represent classification Centroid c _jneighbours' number.

9. method according to claim 1, it is characterized in that described steps d) in upgrade classification Centroid and be: after distributing the classification ownership of each node, calculate the average similarity between each classification internal node, immediate as Centroid with the average similarity of this classification for each one of classification selection, in the process of selecting Centroid, increase a restrictive condition, the number of degrees of the Centroid of electing are greater than the average number of degrees of this classification node simultaneously.

10. method according to claim 1, it is characterized in that described steps d) in calculate target function value and judge whether that convergence is: calculate whole network and be divided in the value on objective function, by with last iteration after the comparison of target function value, if the two differs in certain threshold value, iteration stops, otherwise continues next iteration.