CN108491449A - A kind of community discovery method based on neighbour's feature propagation label - Google Patents
A kind of community discovery method based on neighbour's feature propagation label Download PDFInfo
- Publication number
- CN108491449A CN108491449A CN201810158349.6A CN201810158349A CN108491449A CN 108491449 A CN108491449 A CN 108491449A CN 201810158349 A CN201810158349 A CN 201810158349A CN 108491449 A CN108491449 A CN 108491449A
- Authority
- CN
- China
- Prior art keywords
- label
- node
- influence power
- iteration
- neighboring
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Abstract
The present invention proposes that a kind of community discovery method based on neighbour's feature propagation label, this method introduce label and pre-allocate mechanism, and local optimum is carried out to the label of random initializtion；In communication process, the label of each node is updated according to the sequence of influence power, has been abandoned the way for updating node label in original method by random sequence, can have been shown better stability and higher accuracy.The method of the present invention introduces label predistribution mechanism, three factors of tight ness rating between the influence power and node of the number, neighboring node that are occurred in neighboring node according to each label optimize initialization result, it is propagated into row label on the basis of optimization, has reduced or eliminated influence of the result to community division result of random initializtion.In addition, this method calculates according to the sequence of influence power from high to low in communication process and update the label of each node, reduce the update repeatedly to each node label.
Description
Technical field
The invention belongs to network fields, and in particular, to a kind of community discovery method based on neighbour's feature propagation label.
Background technology
Most complication system can carry out abstract processing with complex network modeling in reality.Studies have shown that removing
Outside with small world effects and scaleless property, also in the prevalence of community structure in complex network, community structure is currently without system
One definition, it is considered that community structure shows as community's interior nodes contact closely, is connected between different community's nodes sparse
Feature.This structure is very common in actual life, such as：Attract constituted good friend by different interest in social networks
Circle；The industry circle being made of different industries in economy and trade network；Interaction shape in protein network between the protein of different structure
At different units.It was found that the community structure in complication system not only has higher research value, but also we are understood
The function and mutual relationship that the composition of complication system, apparent each component part play also have important reality meaning
Justice.
GN algorithms were proposed in 2002, community discovery field starts by the extensive of researcher from Girvan and Newman
Concern.Existing method improves the performance of label propagation algorithm to a certain extent at present, or but these methods disappear completely
In addition to randomness, the characteristics of label propagation algorithm is by network topology discovery community cannot be embodied；Possess the higher time
Complexity cannot carry out the community discovery of large scale network.
Description of the drawings
Fig. 1 is 5 meshed network figures.
Fig. 2 is 8 meshed network figures.
Invention content
The present invention proposes that a kind of community discovery method based on neighbour's feature propagation label, this method introduce label predistribution
Mechanism carries out local optimum to the label of random initializtion；In communication process, each section is updated according to the sequence of influence power
The label of point has abandoned the way for updating node label in original method by random sequence, can show better stability and more
High accuracy.
A kind of community discovery method based on neighbour's feature propagation label, it is characterised in that：This approach includes the following steps：
The label for obtaining random initializtion carries out local optimum to above-mentioned label, and each node can obtain a part after optimization
Optimal label；Finally based on the local optimum label that each node obtains, starts label and propagate, specially
The first step initializes the label of all nodes in network, assigns each node one globally unique label；
Second step, label are propagated in advance；
Third walks, and note iterations are t, and the initial value of t is 1；
All nodes in network are formed one by the 4th step by the descending sequence of the neighboring node influence power of each node
A ordered set；
The most label of the occurrence number in its all neighbor node is made each node in set by the 5th step
The label obtained in current iteration for the node；
6th step, for the arbitrary node in set, if node is in current iteration and the mark obtained in last iteration
Sign it is identical, then label propagation terminate；Otherwise, it enables t=t+1 and returns to the 4th step.
Particularly, in the pre- communication process of label,
1st step,Then viAnd vjFor neighbor relationships, and it is denoted asWhereinIndicate the neighbor relationships between node；
2nd step, ifThen viAnd vjNeighboring node each other；
3rd step,viNeighbour set
4th step, calculate node tight ness ratingIndicate any two node viAnd vjIt
Between tightness degree；
5th step, node viNeighboring node influence power be denoted asIt is obtained by iterating to calculate.Remember finf(vi, k) and indicate vi
Influence power at the kth iteration, if when kth time iteration all nodes finfThe sum of with (k-1) secondary iteration when all nodes
FinfThe sum of difference be less than particular value (ε), then restrain, the influence power that each node is obtained at the kth iteration is as the node
Final influence power, that is, have
6th step, it is assumed that node vjIt is viNeighboring node, by vjLabel by viThe degree of receiving is denoted as vjLabel for
viThe decision value of label distribution, is expressed asAnd
7th step, the set that constitutes of the corresponding label of the powerful node bigger than present node influence power be denoted as L, vi
Candidate tag setWhereinIndicate node vjLabel；
8th step, if
9th step, ifAnd
10th step, ifAnd
The method of the present invention introduces label predistribution mechanism, the number occurred in neighboring node according to each label, close
Three factors of tight ness rating between the influence power and node of neighbors optimize initialization result, into rower on the basis of optimization
Label are propagated, and influence of the result to community division result of random initializtion has been reduced or eliminated.In addition, this method is in communication process
In the label of each node is calculated and updated according to the sequence of influence power from high to low, reduce to each node label repeatedly more
Newly.
Specific implementation mode
The Research Thinking that the present invention uses is as follows：
The label for first obtaining random initializtion carries out local optimum to these labels, and each node can obtain after optimization
To the label of a local optimum, finally based on the local optimum label that each node obtains, starts label and propagate.
The first step initializes the label of all nodes in network, assigns each node one globally unique label；
Second step, label are propagated in advance
Third walks, and note iterations are t, and the initial value of t is 1；
All nodes in network are formed one by the 4th step by the descending sequence of the neighboring node influence power of each node
A ordered set；
The most label of the occurrence number in its all neighbor node is made each node in set by the 5th step
The label obtained in current iteration for the node；
6th step, for the arbitrary node in set, if node is in local iteration and the mark obtained in last iteration
Sign it is identical, then label propagation terminate；Otherwise, it enables t=t+1 and returns to the 4th step.
Wherein in the pre- communication process of second step, following steps are specifically included：
1st step describes the neighbor relationships between node；
2nd step, description meet the node of neighbor relationships, i.e. neighboring node；
3rd step constitutes neighboring node set by neighboring node；
4th step, according to the tightness degree contacted between the attribute calculate node of neighboring node set；
5th step calculates the influence power that some node has under neighboring node effect；
6th step describes the labeling assignments of some node to acceptance level of another node as its local optimum label；
7th step, candidate tag set describe a set being made of special tag, each label in these set
Some node may be all assigned to as its local optimum label, specific method：In the neighboring node set for checking the node
All nodes, the label that influence power in set is more than to all nodes of this node form a set, are exactly the time of this node
Select tag set；
8th step, if the candidate tag set of some node is sky, the existing label of this node is exactly that it is being propagated in advance
The local optimum label obtained in the process；
9th step exists if the candidate tag set of some node only includes a label using this label as this node
The local optimum label obtained in pre- communication process；
If the candidate tag set of the 10th step some node includes multiple labels, by the maximum label of label decision value
The local optimum label obtained in pre- communication process as this node.
Network, wherein V={ v are indicated with figure G=(V, E)i| i=1,2 ..., n } indicate network in all nodes set,
E={ (vi,vj)|vi,vj∈ V } indicate network in all sides set, n indicate network node total number, m indicate network it is total
Number of edges.Neighbor relationships and neighboring node attribute are defined as follows：
Define 1 neighbor relationships.IfThen viAnd vjFor neighbor relationships, and it is denoted asWhereinIndicate the neighbor relationships between node.
Define 2 neighboring nodes.IfThen viAnd vjNeighboring node each other.
Define 3 neighboring node set.viNeighbour set
Define 4 node tight ness ratings.Arbitrary node v is describediAnd vjBetween tightness degree, be denoted asIt can be according to formula (1)
It is calculated：
The bigger expression node v of valueiAnd vjIt is closer.
In complex network, the effect that each node plays is different, and significance level in a network is also different.Node influence power
It is a kind of method that can quantify this significance level, present invention introduces the influence powers of neighboring node to describe the important of neighboring node
Degree.The influence power of neighboring node is defined as follows：
Define 5 neighboring node influence powers.The influence that certain node generates under the collective effect of its all neighboring node is described
Power is denoted as Inf.Node viNeighboring node influence power beIt need to be obtained by iterating to calculate.Remember finf(vi, k) and indicate vi
Influence power when kth time iteration.If the f of all nodes when kth time iterationinfThe sum of with all nodes when (k-1) secondary iteration
finfThe sum of difference be less than particular value ε, then restrain, the influence power that each node is obtained at the kth iteration as the node most
Whole influence power.With viFor,finf(vi, k) and it is calculated by formula (2), (3)：
By taking 5 meshed networks as shown in Figure 1 as an example, the influence power for initializing each node in network is 1, then according to section
Point numbers the influence power that ascending sequence calculates each node successively.Due to v1Neighboring node there was only v2, v1Influence power it is complete
Entirely by v2It determines, i.e.,v2Influence power by its neighboring node v11 and v3It determines,
I.e.Similarly, f can be obtainedinf(v3,2)finf(v3,2)finf(v3, 2) and it is respectively 5/
3,19/18 and 39/36.It is iterated calculating as procedure described above until meeting the condition of convergence.
When initializing node label using random method, community division result will appear fluctuation by a relatively large margin and make to draw
Divide result highly unstable.To overcome the problems, such as this, the present invention introduces label before label propagation and pre-allocates step, by right
The local directed complete set of network node label reduces influence of the mode to final result because of random initializtion node label to improve society
Area finds stability and the accuracy of result.
It is obtained it is considered that the formation of the final community of each in community network is gradually extended by the small community in part,
And local community is collectively constituted by having certain similar features and contacting more close node.In these nodes, shadow
Ring the larger node of power has large effect to the distribution of neighboring node label.But if only according to the influence power of neighboring node come
Determine the distribution of node label, then the label spread scope of the larger node of individual effect power can become very wide, easily shape
At super community.To avoid such case, present invention introduces label Decision of Allocation values, are distributed into row label according to the value.
Define 6 label Decision of Allocation values.The description influence degree of neighboring node to present node when distributing label.It is false
If node vjIt is viNeighboring node, then vjTo viLabel Decision of Allocation value be denoted asIt can be calculated according to formula (4).
If the Inf neighboring nodes that certain node has more than one influence power larger, these nodes are likely to influence
Label to present node distributes, i.e., the label of some node can be pre-assigned to present node in these nodes.The present invention draws
The label of some node may be pre-assigned to by entering candidate tag set description.
Define 7 candidate tag sets.By the powerful neighboring node corresponding label bigger than present node influence power
The set of composition, is denoted as L, with viFor, viCandidate tag setIt is indicated by formula (5).
WhereinIndicate node vjLabel.
The policy depiction that label pre-allocation process is taken is as follows：
If strategy 1
If strategy 2And
If strategy 3And
According to strategy 1, if viCandidate tag set be sky, the v in pre- propagateiLabel remain unchanged.According to strategy
2, if viCandidate tag set only includes a label, then in pre- propagate by viTag update is the label in candidate collection.
According to strategy 3, if candidate tag set includes multiple labels, the label point of each candidate label corresponding node is calculated first
With decision value, then using the label corresponding to maximum value as distributing to v in pre- communication processiLabel.
Below with the node v in Fig. 25For care label pre- communication process.It is calculated according to influence power shown in the present invention
Method calculates v5The influence power of each neighboring node, and the sequence of influence power from high to low is pressed to these node sequencings, obtained section
Point sequence is v4,v8,v6,v5,v7,v2,v3,v1.The v known to ranking results4,v8And v6Influence power ratio v5Greatly, thus v5Candidate
Tag set should be made of the label of these three nodes, i.e., It is not sky, it should be according to strategy 3
To v5Pre-allocate label.It is computed, v4、v6And v8Label decision valueWithRespectively 0.27263,
0.34102 and 0.34104.Due toMaximum, therefore should be by v8Corresponding label is as v5The mark obtained in pre- communication process
Label.
Strategically 1-3 continues label and propagates in advance, the label until having updated all nodes in Fig. 2.It is propagated through in advance
Journey has carried out local optimum according to certain strategy to the label that each node obtains in initialization procedure, with the result of optimization
The initial labels of each node when as formal communication reduce the influence for being randomly assigned label to final community division result with this.
The pre- communication process of label reduces random point to being optimized to a label for node distribution in initialization procedure
Influence with label to community division result.The label obtained in pre- communication process with each node is by label communication process
Basis is iterated propagation according to network topology structure to label, and using the result of iterative diffusion as final division community
Foundation.
The problem of slow method convergence that label may be brought, unstable result being propagated by random node sequence.To avoid
This problem, it is necessary to simultaneously be propagated in any suitable order into row label since suitable node.Context of methods changes label
For the start node in communication process, selected the maximum node of influence power and propagated as label, and by node influence power by greatly to
Small sequence be updated to the label of each node, until convergence.The label of each node, divides society when according to convergence
Area.
Claims (2)
1. a kind of community discovery method based on neighbour's feature propagation label, it is characterised in that：This approach includes the following steps：It obtains
The label for obtaining random initializtion carries out local optimum to above-mentioned label, and each node can obtain a part most after optimization
Excellent label；Finally based on the local optimum label that each node obtains, starts label and propagate, specially
The first step initializes the label of all nodes in network, assigns each node one globally unique label；
Second step, label are propagated in advance；
Third walks, and note iterations are t, and the initial value of t is 1；
All nodes in network are formed an ordered set by the 4th step by the descending sequence of neighboring node influence power；
5th step, for each node in set, using the most label of the occurrence number in its all neighbor node as this
The label that node is obtained in current iteration；
6th step, for the arbitrary node in set, if node is in local iteration and the label phase obtained in last iteration
Together, then label propagation terminates；Otherwise, it enables t=t+1 and returns to the 4th step.
2. according to a kind of community discovery method based on neighbour's feature propagation label described in claim 1, it is characterised in that：
In the pre- communication process of label,
1st step,Then viAnd vjFor neighbor relationships, and it is denoted asWherein
Indicate the neighbor relationships between node；
2nd step, ifThen viAnd vjNeighboring node each other；
3rd step,viNeighbour setAnd
4th step, calculate node tight ness ratingIndicate any two node viAnd vjBetween
Tightness degree；
5th step, node viNeighboring node influence power be denoted asIt is obtained by iterating to calculate.finf(vi, k) and indicate viIn kth
Influence power when secondary iteration；If the f of all nodes when kth time iterationinfThe sum of f with all nodes when (k-1) secondary iterationinf
The sum of difference be less than particular value (ε), then restrain, the influence power that each node is obtained at the kth iteration as the node most
Eventually, that is, have
6th step, it is assumed that node vjIt is viNeighboring node, by vjLabel by viThe degree of receiving is denoted as vjLabel for viMark
The decision value for signing distribution, is expressed asAnd
7th step, the set that constitutes of the corresponding label of the powerful node bigger than present node influence power be denoted as L, viCandidate
Tag setAndWhereinIndicate node vjLabel；
8th step, if
9th step, ifAnd
10th step, ifAnd
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810158349.6A CN108491449A (en) | 2018-02-25 | 2018-02-25 | A kind of community discovery method based on neighbour's feature propagation label |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810158349.6A CN108491449A (en) | 2018-02-25 | 2018-02-25 | A kind of community discovery method based on neighbour's feature propagation label |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108491449A true CN108491449A (en) | 2018-09-04 |
Family
ID=63340711
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810158349.6A Pending CN108491449A (en) | 2018-02-25 | 2018-02-25 | A kind of community discovery method based on neighbour's feature propagation label |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108491449A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109379282A (en) * | 2018-10-25 | 2019-02-22 | 浙江工业大学 | The network community detection method propagated based on multi-tag |
| CN109410078A (en) * | 2018-09-12 | 2019-03-01 | 河南理工大学 | A kind of information propagation prediction method for the mobile social networking shared suitable for object oriented file |
| CN111221875A (en) * | 2020-01-06 | 2020-06-02 | 河南理工大学 | Constraint-based seed node data mining system |
| CN112257865A (en) * | 2020-09-09 | 2021-01-22 | 中国科学院信息工程研究所 | Belief propagation method based on coloring optimization on GPU |
-
2018
- 2018-02-25 CN CN201810158349.6A patent/CN108491449A/en active Pending
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109410078A (en) * | 2018-09-12 | 2019-03-01 | 河南理工大学 | A kind of information propagation prediction method for the mobile social networking shared suitable for object oriented file |
| CN109410078B (en) * | 2018-09-12 | 2021-09-28 | 河南理工大学 | Information propagation prediction method suitable for mobile social network facing file sharing |
| CN109379282A (en) * | 2018-10-25 | 2019-02-22 | 浙江工业大学 | The network community detection method propagated based on multi-tag |
| CN109379282B (en) * | 2018-10-25 | 2020-11-13 | 浙江工业大学 | Network community detection method based on multi-label propagation |
| CN111221875A (en) * | 2020-01-06 | 2020-06-02 | 河南理工大学 | Constraint-based seed node data mining system |
| CN111221875B (en) * | 2020-01-06 | 2022-11-04 | 河南理工大学 | Constraint-based seed node data mining system |
| CN112257865A (en) * | 2020-09-09 | 2021-01-22 | 中国科学院信息工程研究所 | Belief propagation method based on coloring optimization on GPU |
| CN112257865B (en) * | 2020-09-09 | 2023-11-03 | 中国科学院信息工程研究所 | Belief propagation method based on coloring optimization on GPU |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108491449A (en) | A kind of community discovery method based on neighbour's feature propagation label | |
| CN107729767B (en) | Social network data privacy protection method based on graph elements | |
| CN109766710B (en) | Differential privacy protection method of associated social network data | |
| Sun et al. | Scaling of the average receiving time on a family of weighted hierarchical networks | |
| CN113254669B (en) | Knowledge graph-based power distribution network CIM model information completion method and system | |
| CN111309979B (en) | RDF Top-k query method based on neighbor vector | |
| CN110334285B (en) | Symbolic network community discovery method based on structural balance constraint | |
| CN109064348A (en) | A method of it blocking rumour community in social networks and inhibits gossip propagation | |
| CN113190939B (en) | Large sparse complex network topology analysis and simplification method based on polygon coefficient | |
| CN110059731A (en) | A kind of node importance evaluation method for propagating number based on weighting K- rank | |
| Agarwal et al. | A near-linear constant-factor approximation for euclidean bipartite matching? | |
| CN115481682A (en) | Graph classification training method based on supervised contrast learning and structure inference | |
| CN112311608A (en) | Multilayer heterogeneous network space node characterization method | |
| Kathpal et al. | Hybrid PSO–SA algorithm for achieving partitioning optimization in various network applications | |
| CN109299849B (en) | Group demand level calculation method in social network | |
| CN116307328A (en) | Greedy solving method for travel business problem | |
| CN114417540B (en) | Space-earth integrated network multidimensional resource modeling method based on tree structure | |
| CN110825935A (en) | Community core character mining method, system, electronic equipment and readable storage medium | |
| CN109635183A (en) | A kind of community-based partner's recommended method | |
| Toda et al. | Autonomous and distributed construction of locality aware skip graph | |
| CN109410078A (en) | A kind of information propagation prediction method for the mobile social networking shared suitable for object oriented file | |
| CN115130044A (en) | Influence node identification method and system based on second-order H index | |
| Chakraborty et al. | Two algorithms for computing all spanning trees of a simple, undirected, and connected graph: once assuming a complete graph | |
| Huang et al. | A review of combinatorial optimization with graph neural networks | |
| CN106712900A (en) | Low-complexity message passing decoding algorithm based on factor graph evolution in sparse code multiple access |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180904 |