KR20190016249A - Method to derive cluster structure from similarity-based network - Google Patents

Abstract

According to one embodiment of the present invention, a method for deriving a cluster structure may comprise the steps of: modeling nodes into a full network, wherein the full network means that all nodes are completely connected to each other; measuring a degree of similarity between nodes included in the full network; and searching for a specific cluster in the full network based on the measured similarity. The method may perform more accurate similarity-based clustering.

Description

[0001] METHOD TO DERIVE CLUSTER STRUCTURE FROM SIMILARITY-BASED NETWORK [0002]

The following description relates to a technique for all nodes to search for clusters in a fully similar weighted-based full network of mutually connected nodes.

General Judge-Participant In the competition format, the judges are composed of several people and evaluate their performances to determine the ranking of the participants. Unlike other common forms of competition where two participants, such as football or basketball, make one-on-one matches to winners, judges' objectivity can often undermine public confidence in the fairness of competition. By modeling the jury-participant competition format as a weight-based bilateral network, we can identify how biased scores are scored and how biased scores affect competition and structure. The authoritative International Chopin Piano Competition Analysis of the scoring controversy at the 2015 competition shows that even if only a fraction of the biased edges exist, we can seriously distort our reasoning about the network structure. Therefore, the importance of biased detection and elimination in network reasoning is emphasized.

Korean Patent Laid-Open No. 10-2016-0064448 discloses a method for providing recommendation of an item based on an anticipated preference of a similar set. The method includes the steps of obtaining an overall completed prediction preference from an actual preference assigned to an item by all users, It is proposed to provide a method of recommending an item based on the anticipated similarity of the similar set, which recommends the highest real preference among the items which are most similar but which have lower actual preference than the anticipated preference of the item.

The present invention attempts to overcome the limitations of the conventional clustering search technique that could not find meaningful network clusters in the similarity weight based network by providing a technique of searching clusters in a total similarity weight based perfect network in which all nodes are completely connected to each other.

A method for deriving a cluster structure includes modeling nodes as a complete bilateral network, wherein the complete network means that all nodes are completely connected to each other; Measuring a degree of similarity between nodes included in the complete network; And searching for a specific cluster in the full network based on the measured similarity.

Modeling the nodes into a full inter-bilayer network may include modeling an edge to which a weight indicating a score is applied to a plurality of nodes (l and l are natural numbers) and a plurality of other nodes (r and r are natural numbers) , And the inter-quantum full network is given in the form of a l * r inter-quantum adjacency matrix.

Modeling the nodes into a full bilateral network includes weighting a single-mode projection to a node class of the full network and indicating the type of pair association strength between nodes through the weighted edge can do.

Modeling the nodes as a full inter-bilayer network may include: calculating a number of the plurality of nodes * of the single mode projection network by applying a cosine similarity based on the plurality of nodes (l, l is a natural number) And defining a similarity weighted adjacency matrix of the number of the plurality of nodes.

Measuring the similarity between the nodes included in the complete network may include calculating an average similarity between the plurality of nodes based on the entire node.

Wherein the step of searching for a specific cluster in the complete network based on the measured similarity includes performing hierarchical clustering for performing node classification by identifying a cluster of nodes based on pair similarity or relevance between the nodes can do.

The step of searching for a specific cluster in the full network based on the measured similarity may comprise the step of the hierarchical clustering being used to determine a module structure of the complete network.

A computer program stored on a storage medium for executing a method for deriving a cluster structure, the method comprising: modeling nodes into a full two-way network, wherein the complete network means that all nodes are completely connected to each other; Measuring a degree of similarity between nodes included in the complete network; And searching for a specific cluster in the full network based on the measured similarity.

An apparatus for deriving a cluster structure includes: a modeling unit for modeling nodes as a complete bilateral network, wherein the complete network means that all nodes are completely connected to each other; A measurement unit for measuring a degree of similarity between nodes included in the complete network; And a search unit searching for a specific cluster in the complete network based on the measured similarity.

The modeling unit modeling the nodes as a full inter-bilayer network may include a step of modeling an edge to which a weight indicating a score is applied to a plurality of nodes (l and l are natural numbers) and a plurality of other nodes (r and r are natural numbers) And the complete inter-quantum network can be given in the form of a l * r inter-quantum adjacency matrix.

The modeling unit may weight the single-mode projection to the node class of the full network, and may indicate the type of pairing strength between the nodes through the edge to which the weighting is applied.

Wherein the modeling unit calculates the number of the plurality of nodes of the single mode projection network by applying a cosine similarity based on the plurality of nodes (l, l is a natural number) to the plurality of other nodes, Adjacent matrices can be defined.

The measurement unit may calculate an average degree of similarity between a plurality of nodes based on the entire nodes.

The searching unit may perform hierarchical clustering for performing node classification by identifying a cluster of nodes based on pair similarity or relevance between the nodes.

The searching unit may use the hierarchical clustering to determine a module structure of the complete network.

The existing user clustering method first performs clustering of users and then compares the similarities between the derived clusters or compares the similarities among the users in the clusters. According to the present invention, first, After obtaining the similarity, clustering is performed based on the similarity, and more accurate clustering based on similarity can be performed.

The present invention overcomes the limitations in which a conventional cluster search method can not find a meaningful network cluster in a complete network based on similarity weights.

The present invention can be applied and applied to clustering according to similar characteristics of users in a content recommendation system by searching clusters in a network of similarity.

The present invention can be utilized to group users who have similar characteristics in the Collaborate Filtering method, which is one of the techniques that are mainly used in constructing a system for recommending personalized contents to users.

1 is a block diagram for explaining components of a community structure apparatus according to an embodiment.
2 is a flowchart illustrating a method of constructing a community structure according to an exemplary embodiment of the present invention.
FIG. 3 is a diagram illustrating a modeling of a plurality of nodes and a plurality of other nodes in a bilateral network according to an embodiment.
4 is a diagram illustrating an average pair similarity between nodes according to an exemplary embodiment of the present invention.
5 is a diagram illustrating clustering for multiple nodes in a competition according to one embodiment.
6 is a diagram illustrating that point z of a particular node in accordance with one embodiment is alternately removed from the data of each of the plurality of other nodes.
7 is a diagram illustrating a modified module and a module structure of a plurality of nodes' similarity networks according to an embodiment.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram for explaining components of a cluster structure device according to an embodiment. FIG. 2 is a flowchart illustrating a cluster structure method of a cluster structure device according to an embodiment.

The community structure apparatus 100 is for searching a community in a full network and may include a modeling unit 110, a measurement unit 120, and a search unit 130. The components of the community structure device 100 may control the community structure device 100 to perform the steps 210 to 230, which includes the community structure method of FIG.

In step 210, the modeling unit 110 may model the nodes into a full network. The modeling unit 110 may model the nodes as a weight-based inter-quantum full network. At this time, the complete network means that all the nodes are completely connected to each other. The modeling unit 110 calculates the weight of the edge that is given a weight given as a form of a matrix of l * r with respect to a plurality of nodes (l and l are natural numbers) and a plurality of other nodes (r and r are natural numbers) Can be modeled as a bilateral network. The modeling unit 110 may weight the single-mode projection to the node class of the full network and may indicate the type of pairing strength between the nodes through the weighted edge. The modeling unit 110 calculates the number of nodes of the single mode projection network by applying the cosine similarity based on the plurality of nodes (l, l is a natural number) to a plurality of other nodes, Can be defined.

In step 220, the measurement unit 120 may measure the degree of similarity between the nodes included in the complete network. The measuring unit 120 may calculate an average degree of similarity between a plurality of nodes based on the entire nodes.

In step 230, the search unit 130 may search for a specific cluster in the complete network based on the measured similarity. The search unit 130 may perform hierarchical clustering for performing node classification by identifying a cluster of nodes based on pair similarity or relevance between the nodes. The searcher 130 may use hierarchical clustering to determine the modular structure of the complete network.

FIG. 3 is a diagram illustrating a modeling of a plurality of nodes and a plurality of other nodes in a bilateral network according to an embodiment.

A plurality of nodes and a plurality of other nodes can be modeled by a bilateral network. On the other hand, in the following description, a method of modeling a plurality of judges (a plurality of nodes) and a plurality of participants (a plurality of other nodes) into a full bilateral network in a contest will be described as an example. Networks have proven useful in modeling and understanding many complex systems found in a variety of fields such as ecology, biology, social sciences, and networks are closely tied to data and information such as the Internet, the World Wide Web, and citation networks have. The class of complex systems that we study so that networks can be beneficially used means a competing system in which the competition between the constituent systems plays a key role in the function and evolution of the system.

Competition between a judge and a participant is a quantum with a weighted edge representing a given score in the form of a l * r adjacency matrix of l (r is a natural number) judges and r (r is a natural number) Can be modeled as an inter-network.

Equation (1)

For example, it may consist of 17 judges and 10 participants in the competition. Figure 3 (a) shows a network representation of the judges and participants in a jury contest. At this time, the edge width (weight) may indicate the score the judges give to the participant. Referring to FIG. 3 (b), it can be seen that the weight is applied to the single-mode projection to the node class of the inter-network. According to FIG. 3 (b), if a judge or participant of the network is selected, the participant's single-mode projection produces a weighted complete network, and the edge weight represents the degree of similarity between the judges according to the given score. At this time, cosine similarity is used as the similarity.

)를 다음과 같이 정의된 참가자에게 적용할 수 있다. Edge weights can be set to indicate the type of pairwise association strength between nodes. The present invention is based on the score of the judges'

) May be applied to a participant defined as follows.

Equation 2:

을 정의할 수 있다. This is a 17x17 positive (similarity) weighted adjacency matrix of a single-mode projection network

Can be defined.

4 is a diagram illustrating an average pair similarity between nodes according to an exemplary embodiment of the present invention.

을 가지는 Yundi가 전체적으로 다른 심사위원들과 가장 유사하며,

을 가지는 Entremont가 다른 심사위원들과 가장 덜 유사하다는 것을 보여준다. 이러한 결과에서 높은 유사도는 Yundi가 가장 전형적이고 평균적인 심사위원이라는 것을 나타내므로 그가 특히 흥미롭지 않다는 것을 의미한다. 대조적으로, Entremont는 잠재적으로 가장 흥미롭고 비정형적인 사례이다. Cho의 퍼포먼스에 대해 그가 낮은 점수를 부여한 점을 감안할 때, 비정형성이 얼마나 많은 영향을 받았는지 궁금하게 한다. 물론 다른 심사위원들과 다르다는 것은 그 자체로 문제가 되지 않으며, 창조적인 사업에서는 다른 사람과의 차이가 권장되어야 한다. 그러나, 비정형성이 불일치를 시사한다면, 이러한 비정형성이 야기할 수 있는 문제에 유의해야 한다. 따라서, 비정형성의 영향을 보기 위해, 데이터에서 Cho를 제거한 후 동일한 분석을 수행함에 따른 결과는 도 4(b)에 나타나 있다. Entremont는 유사도가 현저히 증가하여 현재 7위를 차지하였다. 이것은 Cho에 대한 그의 점수가 도 4(a)에서 보이는 그의 명백한 비정형성 뒤에 매우 강한 요소였음을 설득력 있게 보여준다.A method for determining an unstructured node in a competitive network will be described. Figure 4 (a) shows the average similarity between judges and other judges using the entire data. 4 (a)

Yundi is most similar to other judges overall,

Shows that Entremont with the least similarity to other judges. The high degree of similarity in these results means that Yundi is the most typical and average judge, so he is not particularly interesting. In contrast, Entremont is potentially the most interesting and atypical example. Given his low score for Cho's performance, I wonder how much influence non-regularization has had on him. Of course, being different from other judges is not a problem in itself, and in a creative business, differences from others should be encouraged. However, if the non-formation indicates a discrepancy, it should be noted that such non-formation may cause problems. Therefore, in order to see the influence of the atypicality, the result of performing the same analysis after removing Cho from the data is shown in FIG. 4 (b). In Entremont, the degree of similarity was significantly increased to 7th place. This convincingly demonstrates that his score for Cho was a very strong factor behind his apparent irregularity as seen in Fig. 4 (a).

Although FIG. 4 demonstrates to some extent the effect of a single non-characteristic score, averaging the edge weights (similarity) of each judge will result in a complete loss of information about the network structure and a better understanding of the problem, This is more so in relation to the relationship between individual judges. Accordingly, hierarchical clustering among various analysis and calculation methods for network analysis can be performed. Hierarchical clustering is most often used for node classification by identifying clusters of objects based on pair similarity or relevance between objects. Instrumental clustering can create a hierarchy of object groups, starting from a unique group of individuals, to every single group. The hierarchy thus found can be represented as a start through the diagram shown in FIG. 5, and the diagram is generated for the judges in the competition based on the cosine similarity of Equation 2 using cohesion clustering with mean combining .

Before focusing on the clusters, we first discuss the useful amount extracted from the hierarchy, level z, the given node joining the system for the first time. The interpretation of z is simple. For example, a node with z below a predetermined criterion initially joins the scheme and high degree of similarity with other nodes, whereas z above a predetermined criterion does not initially join the scheme as described above, And has a low degree of similarity with the node. Entremont's z is quite consistent with Fig. Entremont is one of the earliest participants as z = 3 when Cho is removed, while Entremont is the last participant in the schematic to z = 16 (the maximum possible of 17 judges) in the entire dataset. The schematic diagrams of both cases are shown in Figures 5 (a) and 5 (b). This can visualize almost the same information about the atypicality of each node while maintaining the hierarchical structure of the network. z is a simpler but more useful quantity to characterize node non-shaping. To confirm that other participants had a similar relationship with Entremont, the participants were removed from the data and Entremont's z was measured and the above procedure was repeated. The results are shown in Figure 6 (a). Other participants did not demonstrate a similar effect to Entremont's z, and the uncharacteristic character of Entremont's scores on Cho's performance can be seen. For example, one can speculate that Entremont was racially discriminated by the most popular guesses about the causes of Cho's low scores.

You can perform a similarity analysis using z to conveniently identify bias for a particular group of participants (non-white). Participants can be categorized as non-white and white, as follows, and can be divided into two groups by adding Cho to both categories. At this time, the race of the participants can be deduced from the sex included in the name.

- Non-White (5): Cho, Kobayashi, Liu, Lu, and Yang;

- White and Cho (6): Cho, Hamelin, Jurinic, Osokins, Shiskin, and Szymon.

If Entremont really treated the two ethnic groups differently, the score for Cho would be quite different in each group. 6 (a), 6 (b), and 6 (c). When Cho is excluded from the data for both cases, we can see that the z of Entrement decreases. This indicates that Cho may have been identified for reasons other than his race.

Hierarchical clustering is most often used to determine the modular structure of a network. This method can often be achieved by constructing a "cut" in the hierarchy of levels. The general method of determining the cut level is to maximize the so-called module structure Q as shown in Equation 3 below.

Equation (3)

는 Kronecker 델타이다. 괄호 안의 요소는 노드 쌍(간단한 그래프에서 0 또는 1) 사이의 실제 엣지 수와 노드의 각도를 기반으로 하는 임의의 기대 값 사이의 차이이다. 도3(b)의 단일 모드 프로젝션 네트워크에 대해 이 양을 일반화하며, 여기서 엣지는 가중치가 적용된다. 선행 상수

을 무시하는 수학식 3의 직접적인 일반화는 아래의 수학식 4와 같이 나타낼 수 있다.In Equation (3), m is the number of edges, C _i and C _j are modules to which node i belongs,

Is Kronecker delta. The elements in parentheses are the differences between the actual number of edges between the node pair (0 or 1 in the simple graph) and any expected value based on the angle of the node. This amount is generalized for the single mode projection network of Figure 3 (b), where the edges are weighted. Leading constant

The direct generalization of Equation (3) can be expressed as Equation (4) below.

Equation 4:

는 도 3(a)의 양자간 완전 네트워크에서 점수(엣지 가중치)를 무작위로 혼합하여 예상되는 유사도를 의미한다. 코사인 유사도(수학식 2)를 사용하면 분석적으로 계산할 수 있다. 코사인 유사도는

와

의 요소의 모든 순열(permutation)에 대한 평균과 동일하다. 한편,

의 요소를 순열화하는 것으로도 충분하다. In Equation 4,

Represents the degree of similarity expected by randomly mixing scores (edge weights) in the complete inter-net network of FIG. 3 (a). Using the cosine similarity (Equation 2), it can be calculated analytically. The cosine similarity is

Wow

Is the same as the mean for all permutations of the elements of. Meanwhile,

It is sufficient that the elements of " net "

Equation 5:

의 k번째 순열인

에 의해 가능한 순열에서

는

의 원소들의 합이다. 이 값을 수학식 4에 삽입하고, 네트워크에 대해 상기 값을 극대화하려고 할 때, 최대 Q'는 모든 심사위원을 포함하는 단일 모듈을 산출할 수 있다. 이러한 결과는 피갓수(summand)

의 성질에 기인한 것으로, 대다수의 노드 쌍에 대해 피갓수(Entremont가 관련될 때 조차도)는 양의 값을 가지므로 Q'에 대한 모든 노드(i, j)에 대해

을 갖는 것이 양의 값과 큰 값이 되기 위하여 유리하다. 다시 말해서, 모든 심사위원은 단일의 포괄적인 모듈에 속한다. 이와는 대조적으로 Q(수학식 3)가 간단한 네트워크에서 잘 동작하는 이유는 대부분의 피갓수가 음수이기 때문이다. 다시 말해서, 간단한 네트워크에서 대부분의 노드 쌍에 대해

이고,

은 항상 양의 값이다. 단일 그룹에 속한 모든 노드를 포함하는 것이 최적은 아니다. 이에 따라 가중치가 적용된 완전 네트워크의 보다 합리적인 모듈 특성을 식별하기 위하여 음수 항에 적절한 수를 산출하는 Q'를 수정해야 한다. 이를 위하여, 각 피갓수에서 공통의 양수 값을 감산하는 것으로서, q'의 평균값인

를 제안할 수 있다.out of r!

The kth permutation of

From possible permutations by

The

. When inserting this value into Equation 4 and trying to maximize the value for the network, the maximum Q 'can yield a single module that includes all the judges. These results are summarized as follows:

(Even when Entremont is involved) for a majority of the node pairs has a positive value, so that for all nodes (i, j) for Q '

It is advantageous to have a positive value and a large value. In other words, all judges belong to a single, comprehensive module. In contrast, Q (Equation 3) works well in a simple network because most of the contrived numbers are negative. In other words, for most node pairs in a simple network

ego,

Is always a positive value. It is not optimal to include all nodes belonging to a single group. Thus, in order to identify more reasonable modular characteristics of the weighted full network, Q ', which yields an appropriate number for the negative term, must be modified. To do this, subtracting a common positive value from each blooded bean, the average value of q '

. ≪ / RTI >

Equation (6)

는 데이터와의 실제 유사도의 평균이고,

는 수학식 5로부터의 쌍 무작위 기대이고,

는 쌍 무작위 기대의 평균이다. 수정된 모듈 방식을 아래의 수학식 7로 나타낼 수 있다.

Is the average of the actual similarities with the data,

Is the paired random expectation from equation (5)

Is the average of pair random expectations. The modified modularity can be expressed by Equation (7) below.

Equation (7)

메모리 및 쌍 유사도에 대한

이다. 유사도 계산에는

와

에 대해

시간이 필요하다. 평균 유사도와 계층적 군집화는

시간이 소요된다. 노드들 사이에 정의된 유사도만을 필요하므로, 유사도를 분리된 네트워크에도 적용시킬 수 있다. 두 개의 분리된 군집에 속한 두 노드 사이의 코사인 유사도는 단순이 0이 될 것이다.Equation (7) allows us to naturally avoid the simplest or non-informative case where all nodes have useful properties that disappear from both extremes of the hierarchy (i.e., all nodes are separated or form a single module). It is also noted that the algorithm runs in time with the secondary memory to demonstrate its applicability to large networks of this method. If you are clustering nodes,

For memory and pair similarity

to be. The similarity calculation

Wow

About

Time is needed. Average similarity and hierarchical clustering

It takes time. Since only the degree of similarity defined between nodes is needed, the similarity can be applied to a separate network. The cosine similarity between two nodes in two separate clusters will be simple.

and

) 도 7(b)와 비교할 때, 다른 모든 가능성(

의 상대적 차이에 대해 두 번째 최적해와 이것 사이에

)을 가려 Q''-최대해(z=14)가 강건함에서 우위를 점하는 것을 도 7(a)에서 확인할 수 있다. We now show Q '' again across the scheme shown in FIG. 7 for the two cases of 7 (a) with the entire data set and 7 (b) with Cho removed from the data. As shown in Fig. 7 (c), in the entire data including Cho, the maximum Q " occurs at a level of z = 14 and yields three modules. When Cho is removed, the maximum Q '' occurs at z = 15, yielding two modules (Fig. 7 (d)). The main difference between these two solutions is that Entremont's 1-node module is present in the entire data. However, without Entremont, the Q '' - maximum module structure is nearly identical. This points to an important issue that demonstrates the potential risks that may be caused by some biased edges, and we focus here. The difference between the two optimal solutions is much smaller (

and

) Compared with Fig. 7 (b), all other possibilities (

Between the second optimal solution and the relative difference between

(Z = 14) is dominant in the robustness by inserting the Q '' -maximum solution (z = 14), as shown in FIG. 7 (a).

만큼의 부분을 차지하는)로 인하여 네트워크 속성에 대해 상당히 다른 추론이 발생할 수도 있음을 의미한다. Q''의 변경에 대한 또 다른 예가 도 7(e) 및 7(f)에 도시되어 있다.Also, there are at least three different solutions with similar Q '''= (z = 13, 12, 11) in Figure 7 (b). At this time, a slightly different definition of the modularity or interclustering method may be used, given the small difference in Q '''between solutions. Although some of these or other comparable configurations may have been identified as optimal, this possibility does not appear at all in Figure 7 (a). A single non-characteristic edge (for all edges

Of the network attributes) may result in considerably different inferences about network attributes. Another example of a change in Q " is shown in Figs. 7 (e) and 7 (f).

The present invention presents a network model for judges-participant competition and suggests ways to understand the issue of fairness and prejudice. First, identify the most odd or exceptional scores by identifying the most different scores or judges. While the average similarity of other judges and individual judges provides a direct indicator of pair similarity, according to the present invention, the use of clustering schemes to graphically identify judges' non-formation of judges makes the judges more intuitive We will proceed to show that we can identify.

As a result of analyzing the standard definition of modularity in a form suitable for a weight - based perfect network, the risk posed by a small part of an abnormally deflected edge is remarkably apparent. From Entremont to Cho, a single edge has a great effect on network analogy. This led to the search for the next-generation community structure. Second, it seemed to be robust during the existence of virtually any comparable and reasonable solution. This is a major effect of our biased single edge, which is our main discovery.

The personalized contents recommendation system can be marketed because it is expected that efforts to provide content tailored to the user will continue in the future as technologies that companies that have already introduced or are already introducing to the users by the service providers of the contents.

The present invention provides a personalized recommendation system in various fields mainly for providing contents to users, and it is expected that users' desire for customized recommendation will increase in the future. Therefore, It can be commercialized by developing a user grouping tool of the system.

The apparatus described above may be implemented as a hardware component, a software component, and / or a combination of hardware components and software components. For example, the apparatus and components described in the embodiments may be implemented within a computer system, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA) , A programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For ease of understanding, the processing apparatus may be described as being used singly, but those skilled in the art will recognize that the processing apparatus may have a plurality of processing elements and / As shown in FIG. For example, the processing unit may comprise a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of the foregoing, and may be configured to configure the processing device to operate as desired or to process it collectively or collectively Device can be commanded. The software and / or data may be in the form of any type of machine, component, physical device, virtual equipment, computer storage media, or device As shown in FIG. The software may be distributed over a networked computer system and stored or executed in a distributed manner. The software and data may be stored on one or more computer readable recording media.

The method according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions to be recorded on the medium may be those specially designed and configured for the embodiments or may be available to those skilled in the art of computer software. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Magneto-optical media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (15)

A method for deriving a cluster structure,
Modeling the nodes into an inter-bilayer perfect network, wherein the complete network means that all nodes are completely connected to each other;
Measuring a degree of similarity between nodes included in the complete network; And
Searching for a specific cluster in the full network based on the measured similarity
/ RTI > The method according to claim 1,
Wherein modeling the nodes into a full bilateral network comprises:
(1, l is a natural number) and a plurality of other nodes (r and r are a natural number), and the inter-interfering perfect network is represented by l * r A step given in the form of a bilaterally adjacent matrix
/ RTI > 3. The method of claim 2,
Wherein modeling the nodes into a full bilateral network comprises:
Wherein a weight is applied to the single-mode projection to the node class of the full network and the type of pair association strength between the nodes through the weighted edge
/ RTI > 3. The method of claim 2,
Wherein modeling the nodes into a full bilateral network comprises:
The number of nodes of the single mode projection network by applying a cosine similarity based on the plurality of nodes (l, l is a natural number) to the plurality of other nodes, and defining a similarity weighted adjacency matrix of the number of the plurality of nodes Step
/ RTI > The method according to claim 1,
Wherein the step of measuring the degree of similarity between the nodes included in the complete network comprises:
Calculating the average similarity between the plurality of nodes based on the entire node
/ RTI > The method according to claim 1,
Wherein the step of searching for a specific cluster in the full network based on the measured similarity comprises:
Performing hierarchical clustering to perform node classification by identifying cluster of nodes based on pair similarity or relevance between the nodes
/ RTI > The method according to claim 6,
Wherein the step of searching for a specific cluster in the full network based on the measured similarity comprises:
Wherein the hierarchical clustering is used to determine a modular structure of the complete network
/ RTI > A computer program stored in a storage medium for executing a method for deriving a cluster structure,
Modeling the nodes into an inter-bilayer perfect network, wherein the complete network means that all nodes are completely connected to each other;
Measuring a degree of similarity between nodes included in the complete network; And
Searching for a specific cluster in the full network based on the measured similarity
≪ / RTI > An apparatus for deriving a community structure,
A modeling unit for modeling the nodes into a full bilateral network, said complete network means that all nodes are completely connected to each other;
A measurement unit for measuring a degree of similarity between nodes included in the complete network; And
A search unit for searching a specific cluster in the complete network based on the measured similarity;
Wherein the cluster structure is derived from a population structure. 10. The method of claim 9,
The modeling unit,
(1, l is a natural number) and a plurality of other nodes (r and r are a natural number), and the inter-interfering perfect network is represented by l * r A step given in the form of a bilaterally adjacent matrix
Wherein the apparatus is adapted to derive the community structure. 11. The method of claim 10,
The modeling unit,
A weight is applied to the single-mode projection to the node class of the full network, and the weight is applied across the edge to indicate the type of pair-
Wherein the apparatus is adapted to derive the community structure. 11. The method of claim 10,
The modeling unit,
The number of nodes of the single mode projection network by applying a cosine similarity based on the plurality of nodes (l, l is a natural number) to the plurality of other nodes, and defining a similarity weighted adjacency matrix of the number of the plurality of nodes doing
Wherein the apparatus is adapted to derive the community structure. 10. The method of claim 9,
Wherein the measuring unit comprises:
The average similarity between a plurality of nodes is calculated based on the entire nodes
Wherein the apparatus is adapted to derive the community structure. 10. The method of claim 9,
The searching unit searches,
Hierarchical clustering is performed to perform node classification by identifying clusters of nodes based on pair similarity or relevance between the nodes
Wherein the apparatus is adapted to derive the community structure. 15. The method of claim 14,
The searching unit searches,
Wherein the hierarchical clustering is used to determine a modular structure of the complete network
Wherein the apparatus is adapted to derive the community structure.

Images