Hypergraph-based scholar collaboration relationship prediction method
By combining hypergraph modeling and star-shaped unfolding algorithms with the XGBoost algorithm, the problem of ignoring high-order interaction relationships in existing technologies is solved, and more accurate prediction of scholar collaboration relationships is achieved.
Patent Information
- Application Number
- CN202211339487.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-29
- Publication Date
- 2026-05-15
- Estimated Expiration
- 2042-10-29
AI Technical Summary
Existing methods for predicting scholar collaborations neglect higher-order interactions during modeling, leading to inaccurate predictions. Furthermore, they rely heavily on data reliability and computational resources, resulting in a lack of interpretability.
Using hypergraph modeling technology and a star-shaped unfolding algorithm, the collaborative relationships of scholars are transformed into a binary relationship graph. The XGBoost algorithm is then used to capture the characteristics of scholars from multiple dimensions to predict collaborative relationships.
It effectively captures high-order interaction information, reduces noisy correlation data, and improves the accuracy and efficiency of predicting scholar collaboration relationships.
Smart Images
Figure CN115907095B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of big data mining technology, specifically relating to a method for predicting scholar collaboration relationships based on hypergraphs. Background Technology
[0002] With the development of mobile internet and web technologies, the urgent need for knowledge integration and sharing among researchers has led to increasingly frequent academic collaborations. Therefore, finding research collaborators provides impetus for innovation across different disciplines. For researchers, how to find new research collaborators and exploring the influencing factors on establishing research collaborations among scholars is an important research question.
[0003] This problem can be summarized as the link prediction problem in the network [1,2]. Existing research mainly includes five categories: methods based on node attribute similarity [3], methods based on network structure similarity [4], methods based on likelihood analysis [5], machine learning methods [6], and graph neural network methods [7]. Among them, the core idea of the method based on node attribute similarity is: "Birds of a feather flock together". If the attributes of two nodes are more similar, then the probability of a connection between the two nodes is greater. However, since some personal attributes of scholars, such as age and residential address, are often difficult to obtain or may be falsified, the feasibility and prediction accuracy of this type of method are limited by the reliability and authenticity of the original data. The method based on network structure similarity mainly starts from the network structure and evaluates whether a connection will be generated between two nodes by calculating some similarity indicators. This type of method can capture the similarity in network structure well, but it also relies heavily on a good network structure and performs poorly in real sparse networks. The method based on likelihood analysis calculates the probability of connection based on the current observation and network structure and maximizes the network likelihood. This type of method has high computational complexity and is not suitable for large-scale networks. Machine learning and graph neural network-based methods often treat link prediction as a binary classification problem, achieving good results by leveraging a large number of features and graph representation learning techniques. However, the former is limited by manually selected features, while the latter lacks good interpretability and sufficient computational resources.
[0004] The above five categories of methods all model the original scholar collaboration data as simple binary relationship graphs. This modeling approach ignores the natural high-order interaction relationship between scholars and assumes that the edges between each scholar are equal, which adds extra noise. Therefore, it fails to effectively solve the problem of predicting scholar collaboration relationships. Summary of the Invention
[0005] This invention aims to model existing scholar collaboration data from a high-level perspective, fully capture the multi-dimensional characteristics of scholars, and thus provide a method for predicting scholar collaboration relationships in scientific research collaboration networks.
[0006] The scholar collaboration prediction method proposed in this invention is based on hypergraph [8] technology, and its overall process is as follows: Figure 1 As shown, the process includes: data cleaning of scholars' published papers; modeling high-order interaction relationships among scholars through hypergraphs, and transforming the hypergraph of scholar collaboration into a scholar collaboration relationship graph based on binary relationships according to the star-shaped expansion algorithm; fully capturing the set of node features in multiple dimensions from multiple perspectives, including hypergraph structure, binary relationship graph structure, and scholar profiles, and using the XGBoost algorithm model
[10] to predict the collaborative relationships among scholars. This invention uses the high-order network structure of hypergraphs to characterize scholar collaboration relationships, and retains high-order interaction as much as possible based on the star-shaped expansion algorithm. On this basis, it captures the multi-dimensional features of scholars from multiple perspectives, and fully utilizes the advantages of the XGBoost model, which can effectively predict the scientific research collaboration relationships among scholars. The specific steps are as follows:
[0007] Step 1: Extract and clean data from information on collaborative publications by scholars (such as papers);
[0008] Step 2: Construct a scholar collaboration hypergraph based on the cleaned scholar collaboration data; if multiple scholars have a collaborative paper, then there is a hyperedge between these scholars; in the scholar collaboration hypergraph, nodes represent scholars, and hyperedges represent the set of scholars who co-authored the paper.
[0009] Step 3: Based on the scholar cooperation hypergraph constructed in Step 2, construct the corresponding star structure for each hyperedge according to the star expansion algorithm [9] to realize the transformation from the scholar cooperation hypergraph to the scholar cooperation relationship graph based on star expansion, and record the number of hyperedges to which each node belongs and the potential corresponding to the hyperedge; the core idea of the star expansion algorithm is: for each hyperedge in the hypergraph, introduce a new auxiliary node and establish a connection between the auxiliary node and all other nodes in the hyperedge, that is, take the newly added auxiliary node as the center point of each star structure;
[0010] Step 4: Based on the scholar cooperation relationship graph based on star-shaped expansion in Step 3, characterize the scholar features from multiple dimensions, divide the training set and test set, and use the XGBoost algorithm
[10] to predict the probability of scholar cooperation edge connection.
[0011] In this invention, step 1 specifically includes:
[0012] Step 1-1: For a specific subject area, extract scholar and paper data corresponding to consecutive time periods of Y years within the most recent 10 years, based on authoritative databases in that subject area. Here, Y needs to be set according to the number of publications in different years for different subject areas. For example, with the rapid development of publications in various subject areas in recent years, based on the DBLP scholar collaboration database in computer science, setting Y=3 can obtain hundreds of thousands of scholar data and millions of network-scale data. In addition, further filtering can be performed based on keywords. First, extract scholar and paper data corresponding to consecutive time periods of 3 years.
[0013] Steps 1-2: Next, the academic age of a scholar is defined as the difference between the time of the scholar's most recent academic paper publication and the time of the scholar's earliest academic paper publication within the selected time period; statistical analysis is performed on academic age and the number of collaborating scholars, and at the same time, referring to the experience of existing research, papers by scholars with an academic age of over 48 years old and papers with more than 50 collaborating scholars are removed.
[0014] In this invention, step 2 specifically includes:
[0015] Step 2-1: On the cleaned academic dataset, construct a scholar collaboration hypergraph based on whether there are collaborative relationships among scholars:
[0016] H = {v, E}, where V refers to the set of scholars and E refers to the set of scholar collaboration relationships; and records the set of profile features for each scholar, including the scholar's academic age, number of published papers, and number of collaborating scholars, where |V| = N, |E| = M, N is the number of scholars, and M is the number of collaboration relationships.
[0017] In this invention, step 3 specifically includes:
[0018] Step 3-1: For each hyperedge, introduce a new auxiliary node and construct a star-shaped structure based on the auxiliary node using the star expansion algorithm to complete the scholar collaboration binary relationship graph G. ＊ The construction is as follows:
[0019] G ＊ ={v ＊ E ＊},|V ＊ |=N+M,|E ＊ |=N;
[0020] Step 3-2: For each scholar i in the scholar cooperation relationship graph based on star topology, record its set of structural features in the original scholar cooperation hypergraph, including the number of hyperedges to which the original scholar belongs, the potential of the hyperedges to which the scholar belongs, and the cumulative sum of the potentials of the hyperedges to which the scholar belongs.
[0021] In this invention, step 4 specifically includes:
[0022] Step 4-1: For the scholar collaboration graph G based on star topography ＊ The graph-based structural feature set of each original scholar is calculated, including degree
[11] , closeness centrality
[11] , betweenness centrality
[11] , clustering coefficient
[12] , network constraint coefficient
[13] , effective size
[13] , and neighbor degree of one-hop neighbor scholars.
[0023] Step 4-2: Integrate the feature sets of each original scholar, including the scholar profile set, the structural feature set based on the hypergraph, and the structural feature set based on the binary relation graph;
[0024] Step 4-3: For any two pairs of nodes in the simple graph constructed based on binary relations, if there is an edge connecting them (positive sample), mark it as 1; otherwise (negative sample), mark it as 0. Randomly divide the edge set in the simple graph into training and test sets. Concatenate the feature vectors of the two scholars to obtain the feature vector of the corresponding edge. Use the XGBoost model to implement the binary classification task of whether the edge exists.
[0025] The innovation of this invention lies in its use of hypergraphs to model collaborative relationships among scholars. Based on a star-shaped expansion algorithm, new auxiliary nodes are introduced to transform the academic collaboration hypergraph into a star-shaped binary relationship graph of scholar collaboration. Multi-dimensional sets of scholar features are extracted from the scholar collaboration hypergraph, the scholar collaboration binary relationship graph, and scholar profiles. The XGBoost algorithm is then used to predict scholar collaboration relationships. This invention can capture high-order interaction information ignored by ordinary graphs and reduce the introduction of noisy correlation data, thus achieving effective prediction of scholar collaboration relationships. Attached Figure Description
[0026] Figure 1 This is a flowchart of the overall algorithm for predicting scholar collaboration relationships based on hypergraphs, as presented in this invention. Detailed Implementation
[0027] To make the above-mentioned objectives and innovations of this invention easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:
[0028] according to Figure 1 As shown, the algorithm proposed in this invention mainly includes four steps, as detailed below:
[0029] Step 1: Extract and clean data from information on collaborative (paper) publications by scholars;
[0030] Step 1-1: Taking DBLP scholar collaboration data from computer science as an example, and referring to the experience of existing research, first extract the scholar and paper data corresponding to the three consecutive years.
[0031] Steps 1-2: Next, drawing on the experience of existing research, we define a scholar's academic age as the difference between the date of their most recent academic paper and the date of their earliest academic paper within the selected time period. We conduct statistical analysis on academic age and the number of collaborating scholars, and, based on the experience of existing research, remove papers by scholars over 48 years of age and those with more than 50 collaborating scholars.
[0032] Step 2: Construct a scholar collaboration hypergraph based on the cleaned scholar collaboration data. If multiple scholars co-author a paper, then there exists a hyperedge connecting these scholars. In the scholar collaboration hypergraph, nodes represent scholars, and hyperedges represent the set of scholars who co-authored the paper.
[0033] Step 2-1: Construct a scholar collaboration hypergraph on the cleaned academic dataset based on whether there are collaborative relationships among scholars:
[0034] H = {v, E}, and records the set of features for each scholar's profile, including the scholar's academic age, number of published papers, and number of collaborating scholars, where |V| = N, |E| = M, N is the number of scholars, and M is the number of collaboration relationships.
[0035] Step 3: Based on the scholar cooperation hypergraph in Step 2, construct the corresponding star structure for each hyperedge using the star expansion algorithm [9], thereby realizing the transformation from the scholar cooperation hypergraph to the scholar cooperation relationship graph based on star expansion and recording the number of hyperedges to which each node belongs and the potential corresponding to the hyperedge. The core idea of the star expansion algorithm is: for each hyperedge in the hypergraph, introduce a new auxiliary node and establish a connection between the auxiliary node and all other nodes in the hyperedge, that is, take the newly added auxiliary node as the center point of each star structure.
[0036] Step 3-1: For each hyperedge, a star-shaped structure based on the auxiliary node is constructed by introducing a new auxiliary node based on the star expansion algorithm
[10] , thus completing the scholar collaboration binary relation graph G. ＊ The construction is as follows:
[0037] G ＊ ={v ＊ E ＊},|V ＊ |=N+M,|E ＊ |=N.
[0038] Step 3-2: For each scholar i in the scholar cooperation relationship graph based on star topology, record its set of structural features in the original scholar cooperation hypergraph, including the number of hyperedges to which the original scholar belongs, the potential of the hyperedges to which the scholar belongs, and the cumulative sum of the potentials of the hyperedges to which the scholar belongs.
[0039] Step 4: Based on the scholar cooperation relationship graph based on star-shaped expansion in Step 3, the characteristics of scholars are characterized from multiple dimensions, the training set and the test set are divided, and the XGBoost algorithm
[10] is used to predict the probability of scholar cooperation edge connection.
[0040] Step 4-1: For the scholar collaboration graph G based on star topography ＊ The graph-based structural feature set of each original scholar is calculated, including degree
[11] , closeness centrality
[11] , betweenness centrality
[11] , clustering coefficient
[12] , network constraint coefficient
[13] , effective size
[13] , and neighbor degree.
[0041] Step 4-2: Integrate the feature sets of each original scholar, including the scholar profile set, the structural feature set based on the hypergraph, and the structural feature set based on the binary relation graph.
[0042] Step 4-3: For any two nodes in the simple graph constructed based on binary relations, if an edge exists between them (positive sample), mark it as 1; otherwise, construct negative samples using negative sampling techniques and mark them as 0. Randomly divide the edge set in the simple graph into training, validation, and test sets. Concatenate the feature vectors of the two nodes to obtain the feature vector of the corresponding edge. Use the XGBoost model to perform a binary classification task to determine whether an edge exists.
[0043] This invention uses the API provided by DBLP to select scholar collaboration data from 2019 to 2021 with the keyword "graph embedding" as the experimental object, and evaluates it using the algorithm of this invention. Incomplete academic collaboration records with author and title information are deleted to obtain a cleaned dataset. An academic collaboration hypergraph is constructed based on the construction method described in step 2, and a corresponding star-shaped unfolded graph is constructed based on the method described in step 3. The statistical information of the academic collaboration hypergraph and the star-shaped unfolded graph is shown in Table 1 below:
[0044] Table 1. Basic Statistical Information of Academic Collaboration Hypergraph and Star-shaped Unfold
[0045] Number of nodes (academic collaboration hypergraph) Hyperedge number (academic collaboration hypergraph) Number of sides (star-shaped unfolded diagram) 4752 1637 10369 .
[0046] The experimental environment was set up as follows: Python was used as the programming language, specifically version 3.8.13. Representative comparative methods employed included: Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Gradient Boosting Decision Tree (GBDT), and Decision Tree (DT). Nodes represent scholars, and edges represent academic collaborations between scholars. Our task was to predict whether two scholars would establish future academic collaborations based on past collaborations. We used negative sampling to randomly sample approximately 5000 negative associations, then divided the dataset into training, validation, and test sets in a 6:2:2 ratio. The feature dimension for each scholar was 392.
[0047] The final results of predicting scholar collaboration relationships based on the aforementioned baseline method and the algorithm proposed in this invention are shown in Table 2. The results show that, compared with other baseline methods, the algorithm proposed in this invention performs best in almost all metrics, and can more accurately predict academic collaboration relationships between scholars. This is because the algorithm proposed in this invention compensates for the high-order interaction information lost by traditional methods, thus better capturing the similarity between two scholars, which further verifies the superiority of the algorithm proposed in this invention.
[0048] Table 2. Basic Statistical Information of Academic Collaboration Hypergraph and Star-shaped Unfold
[0049] Model Accuracy (%) Precision (%) Recall (%) AUC(%) LR 91.15 89.27 98.68 87.21 KNN 88.03 88.03 95.13 84.32 SVM 74.55 74.50 94.42 64.15 GBDT 93.75 92.19 99.09 90.96 DT 91.60 94.24 93.19 90.76 XGBoost 94.60 92.95 99.50 92.03 .
[0050] References:
[0051] [1] Linyuan Lü, and Tao Zhou. Link prediction in complex networks:Asurvey. Physica A: statistical mechanics and its applications, 390(6):1150-1170, 2011.
[0052] [2] Lei Cai, and Shuiwang Ji. A multi-scale approach for graph linkprediction. Proceedings of the AAAI Conference on Artificial Intelligence. 34(4):3308-3315, 2020.
[0053] [3]Yinuo Zhang,Subin Shen,and Zhenyu Wu.Improve link predictionaccuracy with node attribute similarities.Proceedings of the InternationalConference on Computer Engineering and Networks.Springer,Cham,376-384,2018.
[0054] [4]Glen Jeh,and Jennifer Widom.Simrank:a measure of structural-context similarity.Proceedings of the eighth ACM SIGKDD InternationalConference on Knowledge Discovery and Data Mining,538-543,2002.
[0055] [5]Liming Pan,Tao Zhou,Linyuan Lü,and Chin-Kun Hu.Predicting missinglinks and identifying spurious links via likelihood analysis.ScientificReports,6(1):1-10,2016.
[0056] [6]Nesserine Benchettara,Rushed Kanawati,and CelineRouveirol.Supervised machine learning applied to link prediction in bipartitesocial networks.Proceedings of the 2010 International Conference on Advancesin Social Networks Analysis and Mining,326-330,2010.
[0057] [7]Muhan Zhang,and Yixin Chen.Link prediction based on graph neuralnetworks.Advances in Neural Information Processing Systems,31,2018.
[0058] [8]Alain Bretto.Hypergraph theory.An introduction.MathematicalEngineering.Cham:Springer,2013.
[0059] [9]Sameer Agarwal,Kristin Branson,and Serge Belongie.Higher orderlearning with graphs.Proceedings of the 23rd International Conference onMachine Learning,2006.
[0060]
[10] Tianqi Chen,Carlos Guestrin.Xgboost:A scalable tree boostingsystem.Proceedings of the 22nd ACM SIGKDD International Conference onKnowledge Discovery and Data Mining.2016:785-794.
[0061]
[11] Linton C Freeman.Centrality in social networks conceptualclarification.Social Networks,1(3):215-239,1978.
[0062]
[12] Sara Nadiv Soffer,and Alexei Vazquez.Network clusteringcoefficient without degree-correlation biases.Physical Review E,71(5):057101,2005.
[0063]
[13] Ronald Stuart Burt.Structural Holes:The Social Structure ofCompetition.Harvard University Press,1992。
Claims
1. A method for predicting scholar collaboration relationships based on hypergraphs, characterized in that, include: Data cleaning and processing are performed on information related to scholars' published papers. This study utilizes hypergraph modeling to depict high-order interactions among scholars, and transforms the scholar collaboration hypergraph into a scholar collaboration relationship graph based on binary relationships using a star-shaped expansion algorithm. It fully captures multi-dimensional node feature sets from multiple perspectives, including hypergraph structure, binary relationship graph structure, and scholar profiles, and uses the XGBoost algorithm to predict collaboration relationships among scholars. The specific steps are as follows: Step 1: Extract and clean data from information published collaboratively by scholars; Step 2: Construct a scholar collaboration hypergraph based on the cleaned scholar collaboration data; if multiple scholars have a collaborative paper, then there is a hyperedge between these scholars; In the scholar collaboration hypergraph, nodes represent scholars, and hyperedges represent sets of scholars who have co-authored papers. Step 3: Based on the scholar collaboration hypergraph constructed in Step 2, construct the corresponding star structure for each hyperedge using the star expansion algorithm, realizing the transformation from the scholar collaboration hypergraph to the scholar collaboration relationship graph based on star expansion, and record the number of hyperedges to which each node belongs and the potential corresponding to the hyperedge; wherein, in the star expansion algorithm, for each hyperedge in the hypergraph, a new auxiliary node is introduced, and the auxiliary node is connected with all other nodes in the hyperedge, that is, the new auxiliary node is used as the center point of each star structure; Step 4: Based on the scholar cooperation relationship graph based on star topology in Step 3, characterize the scholar features from multiple dimensions, divide the training set and test set, and use the XGBoost algorithm to predict the probability of scholar cooperation edges. Step 2 specifically includes: On the cleaned academic dataset, a scholar collaboration hypergraph is constructed based on whether there are collaborative relationships among scholars: H={V,E}, where V refers to the set of scholars and E refers to the set of scholar collaboration relationships; and a set of profile features for each scholar is recorded, including the scholar's academic age, number of published papers, and number of collaborating scholars, where |V|=N, |E|=M, N is the number of scholars, and M is the number of collaboration relationships; Step 3 specifically includes: Step 3-1: For each hyperedge, introduce a new auxiliary node and construct a star-shaped structure based on the auxiliary node using the star expansion algorithm to complete the scholar collaboration binary relationship graph G. ＊ The construction is as follows: G ＊ ={V ＊ ,E ＊ }, |V ＊ |= N+ M,|E ＊ |=N; Step 3-2: For each scholar i in the scholar cooperation relationship graph based on star topology, record its set of structural features in the original scholar cooperation hypergraph, including the number of hyperedges to which the original scholar belongs, the potential of the hyperedges to which the scholar belongs, and the cumulative sum of the potentials of the hyperedges to which the scholar belongs. Step 4 specifically includes: Step 4-1: For the scholar collaboration graph G based on star topography ＊ Calculate the graph-based structural feature set for each original scholar, including degree value, proximity centrality, betweenness centrality, clustering coefficient, network constraint coefficient, effective size, and degree value of one-hop neighbor scholars; Step 4-2: Integrate the feature sets of each original scholar, including the scholar profile set, the structural feature set based on the hypergraph, and the structural feature set based on the binary relation graph; Step 4-3: For any two nodes in the simple graph constructed based on binary relations, mark them as 1 if there is an edge connecting them, and 0 otherwise. Randomly divide the edge set in the simple graph into a training set and a test set. Concatenate the feature vectors of the two nodes to obtain the feature vector of the corresponding edge. Use the XGBoost model to perform the binary classification task of determining whether the edge exists.
2. The method for predicting scholar collaboration relationships based on hypergraphs according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: For a specific subject area, extract the scholar and paper data corresponding to consecutive time periods of Y years within the most recent 10-year time span based on authoritative databases in that subject area; here, Y is set according to the number of publications in different years for different subject areas. Steps 1-2: Define a scholar's academic age as the difference between the time of their most recent academic paper and the time of their earliest academic paper within the selected time period; perform statistical analysis on academic age and the number of collaborating scholars, and remove papers by scholars with an academic age of over 48 years old and papers by scholars with more than 50 collaborating scholars.