A high development potential academic entity identification and prediction method based on heterogeneous graph transformer

Abstract

The application discloses a high development potential academic entity identification and prediction method based on a heterogeneous graph Transformer, comprising the following steps: firstly, a heterogeneous information graph is constructed, wherein a node represents a multi-type entity in an academic network, and an edge represents a relationship between entities; then, structural embedding and semantic embedding of the node are fused to generate a multi-modal initial feature vector; then, multi-hop neighbor information is iteratively aggregated through a multi-layer heterogeneous graph Transformer layer, weights are dynamically learned, and node representation is updated to realize global feature aggregation; subsequently, multi-type neighbor information of a target node is weighted and fused based on cross-layer attention to generate refined context features, and local background information aggregation is completed; finally, global node representation and local context features are spliced, and a classifier is used to output the probability that a target entity becomes a high development potential entity. The method can effectively handle the complex heterogeneity of the academic network, improve the prediction accuracy, and provide support for scientific research management and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data mining, and in particular to a high development potential academic entity identification and prediction method based on a heterogeneous graph Transformer. BACKGROUND

[0002] In the current academic research environment, efficiently and accurately identifying scholars or research achievements with high development potential is crucial for optimizing research resource allocation, promoting academic innovation, and driving scientific and technological progress. Traditional academic entity evaluation methods are often limited to static indicators or homogeneous graph structures, making it difficult to fully capture the multidimensional characteristics of academic entities and their dynamic evolution trends.

[0003] In today's rapidly changing research environment, accurately assessing and predicting the future development potential and influence of academic entities (such as scholars and research papers) has become a core requirement for research project management, talent development planning, and academic trend insight. Scholars with high development potential can lead research directions, produce breakthrough achievements, and bring synergistic benefits to research teams, significantly improving the quality and efficiency of project completion. Recognizing and predicting which papers will have a significant impact in the future and which scholars will achieve major accomplishments has important practical significance, which can effectively assist decision-makers in tracking emerging hotspots, optimizing information retrieval, and guiding the allocation and planning of research investment.

[0004] The academic community has proposed many methods to evaluate the existing influence of academic entities. For example, the h-index value is based on a scholar's past academic achievements and is used to comprehensively evaluate the overall influence of a scholar in the academic community, while the number of citations is often used to evaluate the influence of an article. However, these indicators only describe the achievements of academic entities, but cannot dynamically assess their future development potential. Especially for new academic entities (such as young scholars entering the academic community and newly published academic papers), the lack of prior knowledge such as past history information makes it extremely challenging to assess and model their future development potential. For example, in practical applications, it is a common but challenging task to match the most suitable scholars for different research needs, as scholars' professional knowledge, research interests, and academic influence are multidimensional and dynamically changing over time. At the same time, due to the short observation time of newly published papers, comprehensive evaluation of their influence or future potential is also challenging.

[0005] Traditional node importance estimation (NIE) methods mainly rely on graph topology analysis, such as PageRank, degree centrality, betweenness centrality, etc., to measure the existing importance of nodes in homogeneous networks. Although these methods have achieved some success in homogeneous networks, their core assumption is that all nodes and edges in the graph are homogeneous, which is contrary to the heterogeneous academic network structure that exists universally in reality. Academic networks usually contain multiple types of nodes and multiple types of semantic-rich relationships between entities. If these essential characteristics of academic networks are ignored, the model will be difficult to capture the complex interaction relationships and deep semantic information between diversified entities, thereby limiting its performance in accurately predicting future influence. Fortunately, the rapid development of the Internet and related technologies has made large-scale heterogeneous information networks (HINs) databases and advanced graph-based processing techniques widely available in academic activity analysis. Figure 1 Taking the Aminer database as an example, a typical academic social network graph paradigm is given, which contains five different entities, including authors, articles, institutions, journals, and concepts, as well as multiple relationship types between entities (such as cooperation, citation, membership, and writing).

[0006] Large-scale heterogeneous information networks (HINs) provide unprecedented opportunities for mining the potential value of academic entities from multi-dimensional and multi-modal data. However, the commonly used graph neural network (GNNs) methods still face many challenges when dealing with HINs: on the one hand, traditional GNNs methods often ignore the relationship heterogeneity in HINs (for example, the cooperation relationship between authors or the writing relationship between authors and papers), and these methods are not specifically designed for node importance prediction tasks, and their output is usually node embedding rather than directly interpretable potential prediction probability; on the other hand, even some GNNs designed for heterogeneous graphs often need to predefine meta-paths or simplify the parameterization mechanism, making it difficult to fully capture the diverse semantic relationships. Intuitively, a young scholar who publishes articles on hot topics in high-quality venues or collaborates with well-known experts is more likely to have greater potential to become a star scholar in the future, i.e., have higher development potential. In addition, most existing research focuses on node ranking or current importance estimation in static networks, and fails to effectively simulate the dynamic evolution trend of academic entities, especially for newly emerging academic entities, which face a serious cold start problem due to the lack of historical records. Therefore, there is an urgent need for a more advanced method to better understand these multi-faceted connections and reveal hidden patterns.

[0007] It should be noted that the information disclosed in the above background section is only for understanding the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0008] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide a high-potential academic entity recognition and prediction method based on heterogeneous graph Transformer.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] A high-potential academic entity recognition and prediction method based on heterogeneous graph Transformer, including the following steps:

[0011] S1. Heterogeneous graph construction: Model the multi-type entities and their relationships in the academic network as a heterogeneous information graph, where nodes represent different entities in the academic network and edges represent the relationships between the entities;

[0012] S2. Initial node embedding generation: Fusing the structural embedding and semantic embedding of the node to generate the initial feature vector for multimodal fusion;

[0013] S3, Global Feature Aggregation: Iteratively aggregates multi-hop neighbor information through multiple layers of heterogeneous graph Transformer layers, dynamically learns the weights of heterogeneous nodes and edges, and updates node representations;

[0014] S4, local context information aggregation: Based on cross-layer attention weighted fusion of multi-type neighbor information of the target node, refined context features are generated;

[0015] S5. Potential prediction: Splicing global node representations with local context features, and outputting the probability of the target entity becoming an entity with high development potential through the classifier.

[0016] Furthermore, step S2 specifically includes:

[0017] Use graph embedding tools to generate node structure embeddings and capture topological relationships;

[0018] Use pre-trained text models to extract text semantic embeddings;

[0019] Reduce the semantic embedding to the same dimension as the structural embedding and then concatenate them;

[0020] The nodes with missing text information are filled with zero vectors to uniformly form the multimodal initial feature vectors.

[0021] Furthermore, the global feature aggregation in step S3 specifically includes:

[0022] Perform linear projection on the input features to align the feature dimensions of the model;

[0023] The attention score is calculated through the heterogeneous mutual attention mechanism, fusing the edge type transformation matrix and the meta-relationship prior tensor;

[0024] Parallel execution of heterogeneous message passing, using edge-type specific transformation matrices to encode source node messages;

[0025] The target node representation is updated based on the attention weighted aggregated message, and the residual connection and normalization layers are superimposed.

[0026] Furthermore, in the heterogeneous mutual attention mechanism, the attention score calculation depends on the product of the source node Key vector, the target node Query vector and the edge type transformation matrix;

[0027] The multi-head attention results are concatenated and normalized by softmax;

[0028] The message passing uses a linear transformation matrix bound to the edge type to process source node features.

[0029] Furthermore, step S4 specifically includes:

[0030] Perform learnable weighting on the attention weights of each layer output to generate a unified attention map across layers;

[0031] Aggregate the first-order neighbor information of the target node based on the attention graph to generate a typed background vector;

[0032] The background vector is weighted by type importance through a category-aware linear transformer, and multi-type context features are adaptively fused.

[0033] Furthermore, the multi-type background information covers the directly associated entity types of the target node;

[0034] The type importance weight is calculated by using the softmax function on the linearly transformed background vector.

[0035] Furthermore, step S5 specifically includes:

[0036] The central node features encoded by the global feature aggregator are concatenated with the comprehensive context features from the local background information aggregator;

[0037] Input the multi-layer perceptron classifier output probability value;

[0038] The probability values ​​are converted into binary classification labels through a decision threshold.

[0039] Furthermore, it also includes data construction and label generation steps:

[0040] Screening emerging scholarly samples for entities first published within a specific time window;

[0041] Define high-potential labels based on citation growth patterns in future time windows;

[0042] Entities that satisfy a linear or quadratic growth trend are marked as positive samples, and zero-growth entities are marked as negative samples.

[0043] Furthermore, scholars without a continuous publication record within the verification time window are excluded during label generation;

[0044] The high-potential tag is determined to meet the requirement that the citation count sequence conforms to a linear or quadratic growth pattern;

[0045] Positive and negative samples are divided into training sets, validation sets, and test sets in proportion.

[0046] A computer program product includes a computer program, which, when executed by a processor, implements the high-potential academic entity recognition and prediction method based on heterogeneous graph Transformer.

[0047] The present invention has the following beneficial effects:

[0048] The present invention provides a method for identifying and predicting academic entities with high development potential based on heterogeneous graph Transformer, which overcomes the limitations of existing methods in dealing with the complex heterogeneity of academic networks, multimodal information of entities, and future potential prediction. This method establishes an innovative recognition, prediction and evaluation framework based on heterogeneous graph Transformer. This framework constructs multi-type academic entities and their complex relationships into heterogeneous information graphs. With the help of innovative graph representation learning methods, it comprehensively considers the structural characteristics, semantic information and cross-type interactions of academic entities, and can effectively model the complex relationships and information flows in multimodal heterogeneous academic networks. At the same time, by introducing innovative attention mechanisms and aggregation strategies, this method can more accurately predict the probability of academic entities becoming high-development-potential entities in the future time period. It is particularly suitable for processing large-scale academic networks and can effectively deal with the potential assessment problems faced by newly published papers or new scholars, thereby providing more insightful decision support for scientific research project management, talent selection and academic trend prediction.

[0049] Within the framework of this invention, we focus on leveraging the rich semantic information and complex structural relationships between entities in heterogeneous academic information networks to identify and predict future high-potential academic entities. Given a set of emerging academic entities (e.g., young scholars and recently published articles), the model can capture the characteristics of high-potential groups from the rich academic network information, such as whether young scholars work at high-caliber institutions or publish in high-profile journals, and thus estimate the probability of an emerging entity becoming a high-potential entity.

[0050] By experimentally evaluating the performance of the Heterogeneous Type Attention Graph Transformer (HTAGT) model proposed in this invention and other baseline models on high-potential scholar and paper identification tasks, it was found that the HTAGT model consistently outperformed all other competing baseline models in both author identification and paper identification tasks on relevant datasets. Its progress and effectiveness are reflected in leading scores on key evaluation metrics such as accuracy, F1 value, AUC-ROC, and Matthews correlation coefficient (MCC). This demonstrates that the method of this invention has indeed learned the rich local and global information in the heterogeneous academic graph network, thereby achieving good high-potential academic entity recognition results. In addition, the method proposed in this invention not only has superior prediction accuracy, but also can provide explainability of the model's decision-making process through a visualized attention mechanism, thereby providing more powerful support for actual scientific research management and talent identification.

[0051] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is the paradigm of academic collaboration network in Aminer database.

[0053] Figure 2 This is an overall flow chart of the high-potential academic entity recognition and prediction method based on heterogeneous graph Transformer of the present invention.

[0054] Figure 3 This is a schematic diagram of the author node representation obtained by reasoning in the reasoning stage of an embodiment of the present invention.

[0055] Figure 4 Schematic diagram of implementing high-potential academic entity recognition and prediction in the present invention.

[0056] Figure 5 This is a growth chart of the number of citations of high-potential academic entities according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present invention and its application.

[0058] This paper aims to overcome the limitations of existing technologies by proposing an innovative recognition and prediction method based on heterogeneous graph transformers. This method not only effectively models the complex relationships and information flows in multimodal heterogeneous academic networks, but also, through the introduction of innovative attention mechanisms and aggregation strategies, more accurately predicts the future development potential of academic entities, providing more insightful support for practical applications such as scientific research project management.

[0059] SeeFigure 2 The embodiment of the present invention provides a method for identifying and predicting academic entities with high development potential based on heterogeneous graph Transformer, comprising the following steps:

[0060] Step S1, heterogeneous graph construction: Model the multi-type entities and their relationships in the academic network as a heterogeneous information graph, where nodes represent different entities in the academic network and edges represent the relationships between entities.

[0061] See Figure 1 Node types can include scholars, articles, institutions, journals and concepts, and edge types can include collaboration, affiliation, employment, use, being used, writing, publishing, publishing, citing, being cited, etc.

[0062] Step S2: Initial node embedding generation: The structural embedding and semantic embedding of the nodes are fused to generate the initial feature vector for multimodal fusion.

[0063] In some embodiments, step S2 specifically includes: using a graph embedding tool to generate node structure embeddings to capture topological relationships; using a pre-trained text model to extract text semantic embeddings; reducing the semantic embeddings to the same dimension as the structural embeddings and then splicing them; filling zero vectors for nodes that are missing text information to uniformly form a multimodal initial feature vector.

[0064] Step S3, global feature aggregation: Iteratively aggregate multi-hop neighbor information through multiple layers of heterogeneous graph Transformer layers, dynamically learn the weights of heterogeneous nodes and edges, and update node representations.

[0065] In some embodiments, the global feature aggregation described in step S3 specifically includes: linearly projecting the input features to align the feature dimensions of the model; calculating the attention score through a heterogeneous mutual attention mechanism, fusing the edge type transformation matrix and the meta-relationship prior tensor; executing heterogeneous message passing in parallel, using an edge type specific transformation matrix to encode the source node message; updating the target node representation based on the attention weight aggregation message, and superimposing the residual connection and normalization layer.

[0066] In some embodiments, in the heterogeneous mutual attention mechanism, the attention score calculation depends on the product of the source node Key vector, the target node Query vector and the edge type transformation matrix; the multi-head attention results are concatenated and normalized by softmax; the message passing uses a linear transformation matrix bound to the edge type to process the source node features.

[0067] Step S4: Local background information aggregation: Generate refined context features by weightedly fusing multi-type neighbor information of the target node based on cross-layer attention.

[0068] In some embodiments, step S4 specifically includes: performing learnable weighting on the attention weights output by each layer to generate a unified attention map across layers; aggregating the first-order neighbor information of the target node based on the attention map to generate a typed background vector; weighting the background vector by type importance through a category-aware linear transformer to adaptively fuse multiple types of context features.

[0069] In some embodiments, the multi-type context information covers the directly associated entity types of the target node; the type importance weight is calculated by using a softmax function on the linearly transformed context vector.

[0070] Step S5, potential prediction: splicing the global node representation and the local context features, and outputting the probability of the target entity becoming an entity with high development potential through the classifier.

[0071] In some embodiments, step S5 specifically includes: splicing the central node features obtained by encoding the global feature aggregator with the comprehensive context features from the local background information aggregator; inputting the probability value output by the multi-layer perceptron classifier; and converting the probability value into a binary classification label through a decision threshold.

[0072] In some embodiments, the method further includes data construction and label generation steps: screening emerging academic samples of entities first published within a specific time window; defining high-potential labels based on the citation growth pattern in the future time window; marking entities that meet linear or quadratic growth trends as positive samples, and zero-growth entities as negative samples.

[0073] In some embodiments, scholars who have no continuous publication record within the verification time window are excluded when generating labels; the determination of high-potential labels must meet the requirement that the citation count sequence conforms to a linear or quadratic growth pattern; positive and negative samples are divided into training sets, verification sets, and test sets in proportion.

[0074] The high-potential academic entity recognition and prediction method based on heterogeneous graph Transformer proposed in the present invention overcomes the limitation that traditional homogeneous graph models are difficult to capture the complex heterogeneity of academic networks by constructing a heterogeneous graph structure that integrates multiple types of entities and semantic relationships; it innovatively designs a dual mechanism of global feature aggregation and local background information aggregation, and uses heterogeneous graph Transformer to dynamically learn node relationship weights and integrate multi-hop neighbor information, effectively solving the defect that static indicators cannot predict dynamic potential; especially for the cold start problem of emerging academic entities that lack historical data, combined with multimodal embedding and cross-layer attention weighted strategy, it realizes accurate probabilistic prediction of the future influence of scholars and papers; while maintaining the interpretability of the model, it significantly improves the accuracy of high-potential entity recognition and comprehensive evaluation indicators, providing decision support for scientific research resource allocation and talent selection with both global insight and local detail perception.

[0075] The following further describes specific embodiments of the present invention, its algorithm examples and experimental verification.

[0076] The problem domain targeted by this paper is complex academic network graph analysis, especially focusing on the identification and prediction of high-potential entities in academic social networks. The academic network can be regarded as a dynamic heterogeneous information graph, and the model is represented as , which consists of different types of nodes (such as scholars, articles, etc.) and the diverse edges connecting these nodes (such as cooperation relationship, reference relationship, etc.). Node type can be composed of mapping function Represented by, and the edge type can be represented by the mapping function Indicates that, and Represent the node type and edge type respectively, and for heterogeneous academic information networks, satisfy Taking the prediction of emerging scholars as an example, the prediction problem of an academic entity with high development potential can be expressed as:

[0077] Given a new scholar (defined as researchers who first published a paper in a specific time period) and their role in the comprehensive heterogeneous academic graph This task aims to identify researchers who are expected to achieve significant scientific influence or fame in the near future. This process can be represented by a mapping function that processes the scholar Various attributes of (such as publication history and citation relationships, etc.), and from the graph The complex relational structure derived from At a specified future time The probability of reaching a high potential state within . This can be expressed as:

[0078]

[0079] in, Representing scholars The eigenvector of Contains all relevant network structures. Similarly, for the paper entity, given a newly published paper , all articles published by young researchers within a specific period constitute a paper collection , so the future potential prediction of the article entity can be expressed as:

[0080]

[0081] It is worth noting that, unlike author nodes, articles usually contain rich semantic information such as title and abstract, so Contains semantic content and structural embedding. Other entity potential predictions are similar and will not be described here. Finally, the probability of reflects the probability that the academic entity will have high potential in the future. In order to use it for semi-supervised learning, it is classified into predefined categories (i.e., high potential or non-high potential), specifically expressed as:

[0082]

[0083] in is the decision threshold, which is used for parameter tuning on the validation set to select the key parameters of the optimal model by maximizing the binary classification performance.

[0084] Algorithm Design

[0085] The present invention provides a Heterogeneous Type-Attentive Graph Transformer (HTAGT) model for high-potential academic entity tasks and prediction framework. The core idea of ​​this algorithm is to learn graph representations of heterogeneous academic networks, and then predict the probability of academic entities becoming high-potential entities in the future. This framework is particularly good at encoding heterogeneous networks with rich semantic information. Unlike homogeneous network modeling and analysis methods, the present invention closely associates the structural information between different edges, thereby enabling the early identification of high-potential academic entities. Figure 3 A node representation learning architecture based on HTAGT is proposed. On the basis of obtaining global feature representation and local background information vector, Figure 4 We further describe a high-potential academic entity recognition and prediction framework based on a multi-layer perceptron classifier. Overall, the algorithmic framework consists of four key components and steps: initial node embedding, global feature aggregator, local context aggregator, and a multi-layer perceptron classifier. Figure 3 and Figure 4 In the figure, circles represent nodes, and the boxes next to the nodes represent the node vector representations obtained during the model processing. For example, the gray in step 1 represents the initial vector, the dark purple in step 2 represents the global feature vector, and the light purple and blue in step 3 correspond to the local aggregated feature vectors of the author and article respectively. In step 4, the global and feature vectors are spliced ​​accordingly for different tasks to complete the final classification task.

[0086] (a) Initial node embedding.

[0087] In order to provide entities in heterogeneous academic networks Generate comprehensive and information-rich representations, and systematically consider structural and semantic information in the node embedding process. First, the PyTorch-BigGraph (PBG) tool is used to generate low-dimensional structural vectors (e.g., 50 dimensions) for all types of nodes to capture the topological positions and relationship patterns of nodes in heterogeneous networks. Secondly, considering that the text information of some nodes (such as articles and concepts) contains rich semantic information, the Sentence-BERT model and the pre-trained all-MiniLM-L6-v2 model are used to construct high-dimensional (384 dimensions) text semantic information. Subsequently, these high-dimensional semantic vectors are reduced to the same dimension as the structural embedding (i.e., 50 dimensions) through principal component analysis (PCA). Finally, the structural embedding and the reduced-dimensional semantic embedding are concatenated, and the zero vector is uniformly filled to the input dimension required by the model (e.g., 128 dimensions) to form the initial feature vector for multimodal fusion. , as the input of the subsequent HTAGT layer.

[0088] (2) Global Feature Aggregator

[0089] Due to the small-world property of graphs, stacking multi-layer heterogeneous graph transformer (HGT) structures can effectively iteratively update the initial node embedding to capture multi-hop neighbor information and global network structure features. Therefore, we first use multi-layer HGT layers to iteratively aggregate the feature information of multi-hop neighbors in heterogeneous academic information networks. The update process of the global feature aggregator consists of four steps: 1) Linear projection. Indicates the l The output of the HGT layer, after a total of L By stacking the layers, we can get the embedded representation of all nodes in the graph. For nodes (Node type is ), its input feature matrix is (if ) or (if ). These input features then undergo a simple linear projection so that the feature dimensions of the model are aligned. h The projection of a head is defined as:

[0090]

[0091] in h is the number of attention heads.

[0092] 2) Heterogeneous Mutual Attention. For each heterogeneous edge, the heterogeneous mutual attention mechanism calculates the similarity between the source node Key and the target node Query, and combines it with the edge type specific transformation matrix and element-relation prior tensor , and obtain the multi-head attention score. This mechanism allows the model to dynamically learn the importance of different types of nodes and edges.

[0093]

[0094] Then, by connecting All attention heads and all source nodes Use softmax to normalize and get each target node The final attention vector of:

[0095]

[0096] in, The symbol represents the vector concatenation operation.

[0097] 3) Heterogeneous message passing. In parallel with attention calculation, the model also encodes the source node's message through an edge-type-specific transformation matrix:

[0098]

[0099] 4) Target node aggregation. Finally, combined with the calculated attention weights, the messages from neighboring nodes are aggregated to update the representation of the target node.

[0100]

[0101] In particular, to prevent overfitting, each layer includes residual connections, layer normalization, and dropout to enhance training stability and generalization. By stacking multiple layers, the representation of each node can incorporate information from its multi-hop neighbors, thereby achieving effective aggregation of global features.

[0102]

[0103] 3. Local Context Information Aggregator

[0104] This module is mainly used to extract and synthesize the refined, multi-type background information of the target node (such as scholars) from the final node embedding output by the global feature aggregator. First, the attention weight value of each layer obtained by the global feature aggregator is Weighted aggregation is performed to obtain a unified cross-layer attention map to reflect the comprehensive importance of edges at different depths.

[0105]

[0106] in, is the weight value that can be learned. Based on the final , we can aggregate the node’s neighbor information to get a more comprehensive node background information vector. For example, let its first-order neighbor set be (Abbreviated as ),in is an author type entity, and is the first-order neighbor node of the author, that is, the set {concept, institution, article, author}. So, The neighbor information aggregation can be expressed as:

[0107]

[0108] in It is a category-aware linear transformer. For each relevant node type (such as paper, concept, institution), the final embedding of the node of that type is subjected to type-based linear transformation and attention weighting to generate a local background vector of that type. , thereby achieving adaptive fusion of different types of context information.

[0109] (4) Multilayer Perceptron Classifier

[0110] The central node features encoded by the global feature aggregator Combined context features from local context aggregator Then, the concatenated feature vector is input into a multi-layer perceptron (MLP) classifier, which finally outputs the probability that the target entity will become an entity with high development potential in the future:

[0111]

[0112] Probability of generation Reflects the estimated importance of a node, which is used to represent the probability of becoming a high-potential entity. The core idea of ​​this framework is to learn academic entity representations based on directed heterogeneous complex network graphs, and then use predicted probabilities to achieve high-potential academic entity identification. The learned vector representations naturally include the academic entity's past scientific research collaboration history, multimodal information, and learned network topology and time-varying evolution characteristics. These representation vectors take into account both global structure and local details, have practical value, and can effectively support various downstream recognition or recommendation tasks.

[0113] Example verification

[0114] The well-known Aminer academic network dataset was used to validate the proposed high-potential academic entity identification algorithm. This dataset contains five entities: author, article, concept, institution, and journal, along with their six primary relationships, as shown in Figure 1. The dataset records a total of 848,000 author and 1.155 million paper entities. Overall, the dataset encompasses millions of nodes and tens of millions of edges, effectively supporting the proposed method's ability to handle large-scale, complex academic networks. Given that this invention primarily targets high-potential emerging entities, data sample screening and preparation were required before model testing. 2006 was set as the prediction cutoff year for the experiment, and the sample observation window spanned three years, from 2004 to 2006. First, articles first published within this observation window were selected as observation samples for article entities. For scholars, scholars whose first articles were published within this window were selected as observation samples for emerging scholar entities. Subsequently, additional data from 2007 to 2013 (i.e., the validation window VS) was obtained to calculate the number of citations per article and the total number of citations per scholar to construct the signature. To ensure the relevance and ongoing scholarly engagement of the sample, we also excluded scholars who published only between 2004 and 2006 but then did not publish in the following seven years (i.e., 2007–2013). This helped to exclude scholars who may have ceased academic research after their initial publication, thus ensuring that the researchers we observed remained active in academia.

[0115] Before validating the effectiveness of the model, it is necessary to construct labels for the emerging academic entities screened. First, obtain the time series of the cumulative number of citations of academic entities (articles and authors) in the 7 years following the validation time window (VS). Then perform linear and quadratic polynomial fitting on the citation number series of all entities. If its citation number series meets a specific linear or quadratic growth pattern, the entity is marked as a high-potential individual. If the total number of citations is always 0 within the validation time window (the next 7 years), it is marked as a non-high-potential individual. Taking the author entity as an example, Figure 5 A sample graph showing the growth trend of the cumulative citations of high-potential scholars over the next seven years is given. The numerical titles above the subgraphs represent the scholar IDs given in the database, while the solid lines in the subgraphs represent the cumulative citation trends fitted by a quadratic function, and the points on the dotted lines are the actual cumulative citation data. After the above screening process, 2,270 high-potential scholars, 4,287 non-high-potential scholars, 3,772 high-potential articles, and 52,499 non-high-potential articles were finally marked. For all experiments, the dataset was randomly divided into a training set (70%), a validation set (10%), and a test set (20%). This division was performed independently for both positive and negative samples to ensure a balanced evaluation on all subsets.

[0116] To validate the effectiveness of the proposed HTAGT method in the high-potential academic entity recognition task, we selected a series of benchmark models for comparison, including the Graph Attention Network (GAT), the Heterogeneous Graph Attention Network (HAN), and the Relational Graph Convolutional Network (RGCN). Five evaluation metrics were used to evaluate the final results: accuracy, precision, F1 score, AUC-ROC, and the Matthews correlation coefficient (MCC). The definitions of accuracy, precision, F1 score, and MCC are as follows:

[0117]

[0118] Among them, the calculation process of the F1 score uses the recall rate (Recall).

[0119] in (True Positives) the number of positive instances correctly predicted to be positive, (true negative), (false positive) and The definition of (false negative) is similar. MCC is a robust and balanced metric for measuring binary classification quality, particularly suitable for imbalanced datasets. AUC-ROC represents the area under the receiver operating characteristic (ROC) curve and is used to assess the classifier's overall ability to distinguish between positive and negative classes at different classification thresholds. Higher AUC-ROC values ​​indicate stronger discrimination ability.

[0120] Table 1 Entity recognition results of high-potential scholars

[0121]

[0122] Table 2 Entity recognition results of high-potential articles

[0123]

[0124] The performance of the proposed method and other baseline models on the high-potential scholar and paper identification tasks was evaluated, as shown in Tables 1 and 2. The experimental results show that the proposed Heterogeneous Type Attention Graph Transformer (HTAGT) model achieved significant performance improvements. First, the proposed HTAGT model consistently outperformed all other competing baseline models on the Aminer dataset, both for author identification and paper identification. This improvement and effectiveness are reflected in its leading scores in key evaluation metrics such as accuracy, F1 score, AUC-ROC, and Matthews correlation coefficient (MCC). For example, in the high-potential author identification task on the Aminer dataset, HTAGT achieved improvements of 0.16%, 0.26%, 0.27%, and 0.37% in accuracy, F1 score, AUC-ROC, and MCC, respectively, over the best-performing baseline model (the HAN model), outperforming other strong baseline models. Similarly, in the high-potential article identification task, HTAGT also outperformed other baseline models across all evaluation metrics. This shows that the proposed method has indeed learned the rich local and global information in the heterogeneous academic graph network, and thus achieved a high-potential academic entity recognition effect. This shows that when evaluating the future influence of graph structure data, it is necessary to comprehensively consider the overall structural information and local contextual information of the heterogeneous information network. If only the global topological structure information of the network is processed by the general model, it may lead to the loss of local background information of the network, thereby affecting the accuracy of the model. The method proposed in the present invention is not only superior in prediction accuracy, but also can provide explainability of the model decision-making process through the visualization of the attention mechanism, thereby providing more insightful support for actual scientific research management and talent identification.

[0125] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0126] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0127] An embodiment of the present invention further provides a processor, which executes a computer program and at least performs the method described above.

[0128] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory (Flash Memory), a magnetic surface memory, an optical disc or a read-only optical disc (CD-ROM); the magnetic surface memory can be a magnetic disk memory or a magnetic tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0129] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0130] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0131] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0132] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0133] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0134] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0135] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0136] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0137] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art will recognize that, without departing from the scope of the present invention, several equivalent substitutions or obvious variations can be made, and the performance or use of the same should be considered to fall within the scope of protection of the present invention.

Claims

1. A high-potential academic entity recognition and prediction method based on heterogeneous graph Transformer, characterized by: The following steps are involved: S1. Heterogeneous graph construction: Model the multi-type entities and their relationships in the academic network as a heterogeneous information graph, where nodes represent different entities in the academic network and edges represent the relationships between the entities; S2. Initial node embedding generation: Fusing the structural embedding and semantic embedding of the node to generate the initial feature vector for multimodal fusion; S3, Global Feature Aggregation: Iteratively aggregates multi-hop neighbor information through multiple layers of heterogeneous graph Transformer layers, dynamically learns the weights of heterogeneous nodes and edges, and updates node representations; The global feature aggregation in step S3 specifically includes: Perform linear projection on the input features to align the feature dimensions of the model; The attention score is calculated through the heterogeneous mutual attention mechanism, fusing the edge type transformation matrix and the meta-relationship prior tensor; Parallel execution of heterogeneous message passing, using edge-type specific transformation matrices to encode source node messages; Update the target node representation based on the attention weighted aggregated message, and superimpose the residual connection and normalization layer; In the heterogeneous mutual attention mechanism, the attention score calculation depends on the product of the source node Key vector, the target node Query vector and the edge type transformation matrix; The multi-head attention results are concatenated and normalized by softmax; The message passing uses a linear transformation matrix bound to the edge type to process source node features; S4, local context information aggregation: Based on cross-layer attention weighted fusion of multi-type neighbor information of the target node, refined context features are generated; Step S4 specifically includes: Perform learnable weighting on the attention weights of each layer output to generate a unified attention map across layers; Aggregate the first-order neighbor information of the target node based on the attention graph to generate a typed background vector; The background vector is weighted by type importance through a category-aware linear transformer, and multi-type context features are adaptively fused. Multi-type background information covers the directly related entity types of the target node; The type importance weight is calculated by the softmax function on the background vector after linear transformation; S5, Potential Prediction: Combine global node representations with local context features and output the probability of the target entity becoming a high development potential entity through the classifier; The method also includes data construction and label generation steps: Screening emerging scholarly samples for entities first published within a specific time window; Define high-potential labels based on citation growth patterns in future time windows; Entities that satisfy a linear or quadratic growth trend are marked as positive samples, and zero-growth entities are marked as negative samples.

2. The method according to claim 1, wherein Step S2 specifically includes: Use graph embedding tools to generate node structure embeddings and capture topological relationships; Use pre-trained text models to extract text semantic embeddings; Reduce the semantic embedding to the same dimension as the structural embedding and then concatenate them; The nodes with missing text information are filled with zero vectors to uniformly form the multimodal initial feature vectors.

3. The method according to claim 1 or 2, wherein: Step S5 specifically includes: The central node features encoded by the global feature aggregator are concatenated with the comprehensive context features from the local background information aggregator; Input the multi-layer perceptron classifier output probability value; The probability values ​​are converted into binary classification labels through a decision threshold.

4. The method according to claim 1 or 2, wherein: Scholars who have no continuous publication records within the verification time window are excluded when generating tags; The high-potential tag is determined to meet the requirement that the citation count sequence conforms to a linear or quadratic growth pattern; Positive and negative samples are divided into training sets, validation sets, and test sets in proportion.

Images