CN112989199B - A Collaborative Network Link Prediction Method Based on Multi-Dimensional Proximity Attribute Network - Google Patents

Abstract

The invention provides a cooperative network link prediction method based on a multidimensional proximity attribute network, which belongs to the field of cooperative recommendation and comprises the following steps: respectively reserving multidimensional proximity features, local network features and global network features by using a self-encoder model, a joint probability model and an attribute Skip-Gram model; wherein the multidimensional proximity features include cognitive proximity features, geographic proximity features, and institutional proximity features; the loss function of the self-coding model, the loss function of the local network characteristic, the loss function of the global network characteristic and the loss function of the L2-norm are combined to serve as an overall objective function, and a random gradient descent method is adopted to optimize the overall objective function, so that the representation learning of the network nodes is realized; and carrying out cooperative network link prediction through the vector cosine similarity corresponding to the network node. According to the method, the network characteristics and the node attribute information are comprehensively considered, and the accuracy of the cooperative network link prediction is improved.

Description

A collaborative network link prediction method based on multi-dimensional proximity attribute network

Technical Field

The present invention belongs to the field of cooperative recommendation, and more specifically, relates to a cooperative network link prediction method based on a multi-dimensional neighboring attribute network.

Background Art

Collaborator recommendation is of great significance for promoting scientific research collaboration. Existing literature research mainly focuses on recommending the collaboration between all co-authors. Existing collaborator recommendation methods are mainly based on network models, content models and hybrid models. Among them, the collaborator recommendation methods based on network models incorporate local network features (e.g., common neighbors) or global network features (e.g., random walks with restarts RWR). The collaborator recommendation methods based on content models recommend authors by extracting content features (e.g., similarity based on LDA). The collaborator recommendation methods based on hybrid models combine network features and content features. The attribute network embedding model that combines network features and node attribute information shows good performance. Existing literature points out five dimensions of proximity in scientific research collaboration. However, existing collaborator recommendation methods only include social proximity, which belongs to network features, or cognitive proximity, which belongs to text features.

Patent document CN104573103B proposes a collaborator recommendation method in a heterogeneous network of scientific literature, but this method only considers the willingness of a pair of authors to cooperate with each other to recommend collaborators, and does not start from the perspective of the main author.

Summary of the invention

In view of the defects of the prior art, the purpose of the present invention is to provide a collaboration network link prediction method based on a multidimensional proximity attribute network, aiming to solve the problem that the existing collaborator recommendation method does not take the multidimensional proximity characteristics into consideration and cannot reflect the scientific research similarity and the probability of collaboration between authors after functional processing of the multidimensional proximity characteristics, resulting in a relatively low prediction accuracy of author collaboration.

To achieve the above object, the present invention provides a method for predicting cooperative network links based on a multi-dimensional proximity attribute network, comprising the following steps:

The autoencoder model, joint probability model and attribute Skip-Gram model are used to retain multidimensional proximity features, local network features and global network features respectively; among which, the multidimensional proximity features include cognitive proximity features, geographical proximity features and institutional proximity features;

The loss function of the autoencoder model, the loss function of the local network features, the loss function of the global network features, and the loss function of the L2-norm are combined as the overall objective function, and the stochastic gradient descent method is used to optimize the overall objective function to achieve representation learning of network nodes;

The cooperative network link prediction is performed through the cosine similarity of the vectors corresponding to the network nodes;

Among them, network nodes represent authors; cognitive proximity features represent the author's cognitive level in the field of scientific research; geographical proximity features represent the positional relationship of each author; institutional proximity features represent the similarity of the language of the author's location; local network features represent the probability of cooperation among authors; and global network features represent scientific research similarity through the likelihood value of the author's proximity vector.

Preferably, the cognitive proximity feature is expressed as:

Among them, Cp _i,y is the cumulative sum of cognitive vectors of papers published by author a _i in year y; y ₀ is the base year; Y is the year interval;

The geographic proximity feature is expressed as GG = (VG, EG), where VG is the set of geographic nodes; EG is the set of geographic edges;

Preferably, institutional proximity is measured using a continuous aggregation index of common language.

Preferably, the autoencoder model is:

h _i =σ ₁ (W ⁽¹⁾ x _i +b ⁽¹⁾ )

是解码器的重构；θ＝{W⁽¹⁾,b⁽¹⁾,W⁽²⁾,b⁽²⁾}是模型参数；σ₁(·)为激活函数中的tanh函数；Among them, _xi is the author's neighbor feature vector, _hi is the hidden layer representation of the encoder;

is the reconstruction of the decoder; θ = {W ⁽¹⁾ , b ⁽¹⁾ , W ⁽²⁾ , b ⁽²⁾ } are model parameters; σ ₁ (·) is the tanh function in the activation function;

The loss function of the autoencoder model is:

Where n is the total number of authors.

Preferably, the loss function of the local network feature is:

Among them, p _ij is the joint probability of author a _i and author a _j ; e _ij is the edge between author a _i and author a _j .

Preferably, the loss function of the global network feature is:

Among them, ai _+j is the node context in the generated node sequence, w is the window size; the conditional probability of p( _ai+j | _xi ) is the likelihood value of the proximity vector of node i given the context ai _+j ; G is the set of all network nodes; C is all random walk sequences; _ai represents the author.

In general, the above technical solution conceived by the present invention has the following beneficial effects compared with the prior art:

The present invention comprehensively covers the attribute characteristics of scientific research collaborators from the perspective of multidimensional proximity. By pre-training cognitive proximity, geographical proximity and institutional proximity and expressing them as low-dimensional vectors, the unique attributes of each author can be comprehensively considered without destroying the network characteristics. At the same time, the retained network characteristics include local networks and global networks. The local network can accurately reflect the willingness of each author to cooperate, while the global network can better highlight the similarity of scientific research. On this basis, the optimization is performed with the minimum value of the sum of the network characteristics and node attribute characteristics loss functions as the goal, thereby improving the accuracy of the cooperation network link prediction.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1( a ) is a schematic diagram of a distance-weighted network of a GPN provided in an embodiment of the present invention;

FIG. 1( b ) is a schematic diagram of a geographically adjacent network of a GPN provided in an embodiment of the present invention;

FIG1( c ) is a schematic diagram of the transition probability of a GPN node D provided in an embodiment of the present invention;

FIG. 2( a ) is a schematic diagram of an institutional proximity network provided by an embodiment of the present invention;

FIG2( b ) is a schematic diagram of the transition probability of an IPN node d provided in an embodiment of the present invention;

FIG3 is a schematic diagram of a scientific research cooperation network construction provided by an embodiment of the present invention;

FIG4 is a schematic diagram of a joint optimization framework provided by an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

The present invention provides a method for predicting cooperative network links based on a multi-dimensional proximity attribute network, comprising the following steps:

The autoencoder model, joint probability model and attribute Skip-Gram model are used to retain multidimensional proximity features, local network features and global network features respectively; among which, the multidimensional proximity features include cognitive proximity features, geographical proximity features and institutional proximity features;

The loss function of the autoencoder model, the loss function of the local network features, the loss function of the global network features, and the loss function of the L2-norm are combined as the overall objective function, and the stochastic gradient descent method is used to optimize the overall objective function to achieve representation learning of network nodes;

The cooperative network link prediction is performed through the cosine similarity of the vectors corresponding to the network nodes;

Among them, network nodes represent authors; cognitive proximity features represent the author's cognitive level in the field of scientific research; geographical proximity features represent the positional relationship of each author; institutional proximity features represent the similarity of the language of the author's location; local network features represent the cooperative relationship between authors; and global network features are based on the neighborhood similarity between authors.

The following is a detailed introduction. First, the representation learning method of multi-dimensional proximity involved in the present invention is introduced, and then the framework of the ARCR model (attribute-aware research recommendation model) provided by the present invention is explained.

1. The multidimensional proximity feature representation provided by the present invention includes cognitive proximity feature representation, geographic proximity feature representation and institutional proximity representation.

(1) Cognitive proximity feature representation:

Scientific papers are linked texts, which contain not only text but also citation links. Both text content information and link information are essential for measuring the similarity of scientific papers. In the present invention, P2V (paper-to-vector) is used to represent text content features and citation link features to obtain the cognitive proximity representation of scientific research subjects. Considering the dynamic changes in the cognitive basis of scientific research subjects, a time weight decay factor is used. For author a _i , according to the cognitive vectors of the papers published in each year, the cognitive proximity feature representation is obtained:

Among them, Cp _i,y is the cumulative sum of the cognitive vectors of the papers published by author a _i in year y; y ₀ is the base year; and Y is the year interval.

(2) Geographic proximity feature representation:

相关；

为城市u₁与城市v₁之间的距离。The Geographical Proximity Network (GPN) is defined as GG = (VG, EG), where VG is a set of geographic nodes; v ₁ ∈ VG, representing a city; EG is a set of geographic edges; e ₁ ∈ EG, representing the geographic connection between two cities, e ₁ = (u ₁ ,v ₁ ); where u ₁ is another city different from v ₁ ; e ₁ and the edge weight

Related;

is the distance between city u ₁ and city v ₁ .

Figure 1(a) is the distance-weighted network of GPN; Figure 1(b) is the geographic proximity network of GPN; Figure 1(c) is the transition probability of network node D; in the figure, A, B, C and D represent different cities; a biased random walk is performed on GPN, and its transition probability is proportional to the weight of the edge. The city node sequence is generated with the transition probability (as shown in Figure 1(c)), and finally the attribute Skip-Gram model is executed. Two cities with close geographical distances are more likely to appear together in the sampling results and obtain similar vector representations.

(3) Institutional proximity characteristics:

When scientific research subjects have similar cultural backgrounds, scientific research cooperation will be easier. Language is the core of culture. Jacques Melitz integrated the common mother tongue, common spoken language, common official language and language similarity to propose a continuous aggregate index of common language, which is better than the traditional dummy variable model. Therefore, this index is used to measure the institutional proximity of scientific research subjects.

相关，

是国家对之间共同语言的连续聚合指数；其中，u₂为异于v₂的另一个国家。The Institutional Proximity Network (IPN) is defined as GI = (VI, EI), where VI is the set of institutional nodes, v ₂ ∈VI represents a specific country; EI is the set of institutional edges; each institutional edge e ₂ ∈EI represents the institutional proximity between two countries, e ₂ = (u ₂ ,v ₂ ) and the weight of the institutional edge

Related,

is a continuous aggregation index of the common language between country pairs; where u ₂ is another country different from v ₂ .

Figure 2(a) is the institutional proximity network, and Figure 2(b) is the transition probability of network node d. In IPN, node sequences are generated based on random walks, and then the attribute Skip-Gram algorithm is performed on the node sequences. In the sampling results, two countries with greater institutional proximity are more likely to appear at the same time, resulting in similar representations.

2. Coauthorship network

The scientific research cooperation network reflects social proximity. The cooperation information of scientific research subjects co-authoring papers builds the scientific research cooperation network. Figure 3 is a schematic diagram of the construction of the scientific research cooperation network; when two scientific research subjects have a co-authored paper, there is a two-way edge with a weight of 1 between them; when two scientific research subjects have k co-authored papers, there is a connecting edge with a weight of k between them.

3. Problem Statement

G = (A, E, X) is an attributed scientific research cooperation network, where A = {a ₁ , a ₂ , ..., a _n } is the set of authors; each edge e _ij = (a _i , a _j ) ∈ E represents the scientific research cooperation relationship between authors a _i and a _j ; X ∈ R ^n×m represents the node attribute matrix; _xi is the proximity feature vector of author a _i ; the node attribute information is the union of cognitive, geographical and institutional proximity vectors; the goal is to represent each author a _i as a low-dimensional vector h _i by learning the mapping function f: a _i → _hi ∈ R ^d , where d ≤ n, and retain the network features and node attribute information; the network features include local network features and global network features. The local network feature indicates whether there is an edge between two authors; the global network feature indicates the high-order neighborhood similarity of the node.

4. Autoencoder based on proximity features

The proximity feature belongs to the attribute feature, and the autoencoder is used to save the node attribute information. The autoencoder model includes the following three layers: input layer, hidden layer and output layer; the representation function of the autoencoder is:

h _i =σ ₁ (W ⁽¹⁾ x _i +b ⁽¹⁾ )

是解码器的重构；θ＝{W⁽¹⁾,b⁽¹⁾,W⁽²⁾,b⁽²⁾}是模型参数；σ₁(·)为激活函数中的tanh函数；Among them, _xi is the author's neighbor feature vector, _hi∈Rd ^is the hidden layer representation of the encoder;

is the reconstruction of the decoder; θ = {W ⁽¹⁾ , b ⁽¹⁾ , W ⁽²⁾ , b ⁽²⁾ } are model parameters; σ ₁ (·) is the tanh function in the activation function;

The model parameters are learned by minimizing the following loss function.

In order to preserve the high nonlinearity in the attribute information, K hidden layers are used in the encoder;

…

表示作者a_i所需的低维隐藏表示形式。相应，在解码器中也采用了K层隐层。in,

represents the low-dimensional hidden representation required by the author a _i . Accordingly, K hidden layers are also used in the decoder.

5. Local Network Characteristics

Network features include local network features and global network features. The specifics of local network features are as follows:

Maximize the following likelihood estimate to preserve local network features: L _f = ∏e _ij>0 p _ij ;

Where p _ij is the joint probability of a _i and a _j :

Therefore, the negative likelihood can be minimized as follows:

6. Global Network Characteristics

In order to preserve the global network characteristics, the attribute-based Skip-Gram model is adopted. By providing the current node a _i and its proximity features x _i for all random walk sequences c ∈ C, the following negative log-likelihood is minimized:

Where a _i+j is the node context in the generated node sequence, w is the window size; the conditional probability of p(a _i+j | _xi ) is the likelihood value of the proximity vector of node i given the context a _i+j ; G is the set of all network nodes; C is the sequence of all random walks;

是上下文节点

的对应表示方式。但是该公式的计算成本较高，因此，使用负采样法，将p(a_i+j|x_i)替换为：Where f(·) is the function of the encoder part of the autoencoder model;

Is the context node

However, the computational cost of this formula is high, so negative sampling is used to replace p(a _i+j | _xi ) with:

为期望函数；

d_a是节点a的度数；Where σ ₂ (·) is the sigmoid function in the activation function, and |neg| is the number of negative samples;

is the expected function;

d _a is the degree of node a;

7. Joint Optimization Framework

的最终表示方式捕获了网络特征和节点属性信息。The joint optimization framework is shown in Figure 4. Since the autoencoder model, joint probability model, and attribute-aware skip-gram model share the same encoder layer, these models are closely connected.

The final representation captures both network characteristics and node attribute information.

The three objective functions are combined to obtain the overall objective function of the joint model:

The stochastic gradient descent algorithm is used to optimize the loss function in the above formula, and the two coupled components (αL _f +βL _ae +γL _reg , and L _h ) are iteratively optimized; L _reg is

8. Recommendation for scientific research cooperation

以及作者a_j的低维隐藏表示

计算h_i和h_j余弦相似度；最后，向目标作者推荐前k个同类作者作为潜在的科研合作对象。Based on the low-dimensional hidden representation of the author a _i

And the low-dimensional hidden representation of author a _j

Calculate the cosine similarity between _hi and _hj ; finally, recommend the top k similar authors to the target author as potential scientific research cooperation partners.

Example

Data source: Web of Science Core Citation Database was used to collect papers in the pharmaceutical category from 2010 to 2019 ("Biochemistry and Molecular Biology", "Medicine, Research and Experiment", "Pharmacology and Pharmacy", and "Toxicology"), and the search query was "WC = A AND PY = B AND LANGUAGE = 'English'", where A is the pharmaceutical field and B is 2010 to 2019. Papers with independent authors and non-journal papers were excluded, and 528,118 papers were finally retrieved. After author name disambiguation, 162,196 authors who had published more than 5 papers were screened. Geographic information was obtained using Google Maps API. Language information was obtained using CEPII language. The dataset was divided into two parts according to the year of publication: data before 2018 was used as a training set, and data from 2018 to 2019 was used as a test set.

Implementation: A three-layer autoencoder was built; the hidden dimensions of the first, second, and third layers were d(1)=600, d(2)=512, and d(3)=256, respectively; the dimensions of the geographic proximity vector and the institutional proximity vector were both 64, and the dimension of the cognitive proximity vector was 256. The weight of the local network feature loss function was set to α=1, the weight of the autoencoder loss function was set to β=10, and the weight of the L2-norm regularization γ was set to 10. 100 authors were randomly selected as target nodes and ARCR was run.

In summary, the present invention has the following advantages:

The present invention comprehensively covers the attribute characteristics of scientific research collaborators from the perspective of multidimensional proximity. By pre-training cognitive proximity, geographical proximity and institutional proximity and expressing them as low-dimensional vectors, the unique attributes of each author can be comprehensively considered without destroying the network characteristics. At the same time, the retained network characteristics include local networks and global networks. The local network can accurately reflect the willingness of each author to cooperate, while the global network can better highlight the similarity of scientific research. On this basis, the optimization is performed with the minimum value of the sum of the network characteristics and node attribute characteristics loss functions as the goal, thereby improving the accuracy of the cooperation network link prediction.

It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims (6)

1. A cooperative network link prediction method based on a multidimensional proximity attribute network is characterized by comprising the following steps:

respectively reserving multidimensional proximity features, local network features and global network features by using a self-encoder model, a joint probability model and an attribute Skip-Gram model; wherein the multidimensional proximity features include cognitive proximity features, geographic proximity features, and institutional proximity features;

the loss function of the self-coding model, the loss function of the local network characteristic, the loss function of the global network characteristic and the loss function of the L2-norm are combined to serve as an overall objective function, and a random gradient descent method is adopted to optimize the overall objective function, so that the representation learning of the network nodes is realized;

carrying out cooperative network link prediction through vector cosine similarity corresponding to the network node;

wherein the network node represents an author; the multi-dimensional proximity feature representation includes a cognitive proximity feature representation, a geographic proximity feature representation, and a institutional proximity feature representation; cognitive proximity features characterize the cognitive level of authors in the scientific domain; the geographic proximity features characterize the position relationship of each author; the system proximity feature represents the similarity of the language of the position where the author is located; the local network characteristics represent the cooperative relationship existing in each author; the global network features represent scientific research similarity through likelihood values of author proximity vectors;

wherein the geographic proximity network in the geographic proximity feature is defined as gg= (VG, EG), where VG is a set of geographic nodes; v ₁ E VG, represents a city; EG is a collection of geographic edges; e, e ₁ EG, represents the geographic relationship between two cities, e ₁ ＝(u ₁ ,v ₁ ) The method comprises the steps of carrying out a first treatment on the surface of the Wherein u is ₁ Is different from v ₁ Is a city of another city; e, e ₁ Weighting of edges

Correlation;

Is city u ₁ And city v ₁ A distance therebetween; performing biased random walk on a geographic proximity network, wherein the transition probability is in direct proportion to the weight of the edge, generating a city node sequence according to the transition probability, and finally executing an attribute Skip-Gram model; the more likely two cities with closer geographic distances are co-present in the sampling result, and similar vector representations are obtained; wherein the system proximity network in the system proximity feature is defined as gi= (VI, EI), VI is the system node set, v ₂ Epsilon VI represents a specific country; EI is the set of institutional edges; each system side e ₂ E EI, the system adjacency between two countries, e ₂ ＝(u ₂ ,v ₂ ) Weights with the system side->

Related (I)>

Is a continuous aggregation index of common language between country pairs; wherein u is ₂ Is different from v ₂ Is a country of another country; performing random walk on a system proximity network to generate a national node sequence, and performing an attribute Skip-Gram algorithm on the national node sequence; in the sampling result, the more likely two countries with greater system proximity are to appear simultaneously, resulting in a similar representation;

the self-encoder model, the joint probability model and the attribute aware skip-gram model share the same encoder layer;

the overall objective function is:

L＝L _h +αL _f +βL _ae +γL _reg

optimizing the whole objective function by adopting a random gradient descent algorithm, and iteratively optimizing the two coupling components alpha L _f +βL _ae +γL _reg And L _h ；L _f A loss function that is a local network feature; l (L) _h Is a global network feature; l (L) _ae A loss function that is a self-coding model; l (L) _reg Is a loss function of the L2-norm.

2. The cooperative network link prediction method according to claim 1, wherein the institutional adjacency is measured in terms of a continuous aggregation index in a common language.

3. The cooperative network link prediction method according to claim 2, wherein the self-encoder model is: h is a _i ＝σ ₁ (W ⁽¹⁾ x _i +b ⁽¹⁾ )，

Wherein x is _i For the author's neighboring feature vector, h _i Is a hidden layer representation of the encoder;

is a reconstruction of the decoder; θ= { W ⁽¹⁾ ,b ⁽¹⁾ ,W ⁽²⁾ ,b ⁽²⁾ -model parameters; sigma (sigma) ₁ (. Cndot.) is the tanh function in the activation function. />

4. A method of collaborative network link prediction according to claim 3, wherein the loss function of the self-encoder model is:

where n is the total number of authors.

5. The cooperative network link prediction method according to claim 2, wherein the loss function of the local network characteristic is:

wherein p is _ij Is author a _i And author a _j Is a joint probability of (2); e, e _ij For author a _i And author a _j The connecting edges between the two.

6. The cooperative network link prediction method according to claim 2, wherein the loss function of the global network feature is:

wherein a is _i+j For the node context in the generated node sequence, w is the window size; p (a) _i+j |x _i ) The conditional probability of (a) is context a _i+j Giving likelihood values of the node i proximity vector; g is a set of all network nodes; c is all random walk sequences; a, a _i Representing the author.

Images