A scientific research project expert matching method based on dynamic academic collaboration network graph representation learning

Abstract

The present invention proposes a scientific research project expert matching method based on dynamic academic cooperation network graph representation learning, aiming to improve the efficiency and quality of expert matching in scientific research project management. By constructing a dynamic complex network graph, taking scholars as nodes and cooperation relationships as directed edges, this method comprehensively considers the scholars' past research history and cooperation experience. Using graph structure learning and Transformer architecture, feature vector learning is performed on scholar nodes to achieve accurate matching between scientific research projects and experts and scholars. This method has achieved a matching accuracy of more than 80% and a standardized discounted cumulative return of more than 89% on data sets in the fields of artificial intelligence and data mining, effectively solving the limitations of existing technologies in computing costs and the use of scholar feature information, and providing an innovative solution for recommending suitable collaborators for scientific research projects.

Description

Technical Field

[0001] The present invention relates to information retrieval and intelligent matching technology, and in particular to a scientific research project expert matching method based on dynamic academic cooperation network graph representation learning. Background Art

[0002] In scientific research project management, it is often necessary to screen suitable project leaders or collaborators for scientific and technological projects. Good scientific research cooperation helps to improve the quality of research, thereby promoting the high-quality and efficient completion of projects. Scientific research cooperation can always bring about synergy, that is, the combined expertise of a group of scholars and experts can often produce outstanding results far beyond individual capabilities. Therefore, how to select suitable scholars and experts for scientific research projects will directly affect the quality of related work. However, matching suitable experts for different projects is often very complicated, because the expertise, research interests and project experiences of different scholars are diverse and usually change over time, which makes it quite challenging to evaluate the compatibility of experts with projects.

[0003] The method of calculating the matching degree between projects and experts based on information retrieval technology is to convert scholar information and scientific research project information into multidimensional vectors through data mining technology, and then use vector similarity to measure the matching degree between the two. This method has derived a variety of different algorithms, such as information retrieval and collaborative filtering algorithms. In addition, with the continuous advancement of retrieval technology, more and more related technologies have been proposed, such as intelligent technologies such as data mining, vector space models and latent semantic retrieval. These algorithms are mainly used to improve the matching accuracy between experts and projects, thereby reducing tedious manual operations. With the in-depth study of matching problems, methods based on the optimal solution idea to calculate the similarity between project sets and expert sets have been applied to matching problems. For example, inspired by mathematical models, many scholars have introduced greedy algorithms and evolutionary algorithms, relational graphs, bipartite graphs, etc. into the expert matching process.

[0004] However, the existing intelligent matching technology for scientific research project experts has many limitations. On the one hand, existing research mainly focuses on algorithm optimization. Although there is a certain improvement in accuracy, it faces problems such as high computational volume and operating costs, making it difficult to apply in actual work. On the other hand, the past research experience and interest field information of experts and scholars have not been fully explored. Most algorithms mainly rely on the topological structure of the cooperative network to predict links, thereby achieving the function of project expert recommendation or matching. Since the characteristic information on the scholar side is not effectively utilized, the matching results are not ideal.

[0005] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0006] The main purpose of the present invention is to overcome the defects existing in the above-mentioned background technology and provide a scientific research project expert matching method based on dynamic academic collaboration network graph representation learning.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A scientific research project expert matching method based on dynamic academic cooperation network graph representation learning includes the following steps:

[0009] S1. Construct a dynamic complex network diagram representing the scientific research cooperation network, including scholar nodes and directed edges representing cooperation relationships, as well as timestamps associated with the edges; among them, the paper is the edge, its first author is the central node, and other authors are neighbor nodes. The direction of the edge points from the first author of the paper to other authors, and the timestamp indicates the time when the paper was published;

[0010] S2. Learn a feature vector for each scholar node in the network, which is used to capture the properties of the node;

[0011] S3. Initialize the node feature vector using the pre-trained language model;

[0012] S4. According to the constructed dynamic complex network graph and node relationships, the information of the first author node of each paper and its related neighbor nodes is integrated to construct the query sequence and key-value pair data for Transformer model training;

[0013] S5. Through the multi-head attention mechanism of the Transformer model, query sequences and key-value pair data are processed to learn the matching between scientific research projects and experts and scholars;

[0014] S6. After the model training is completed, the model parameters are fixed, and the representation vector of the scholar node is obtained through the reasoning process to match the most suitable experts and scholars for the scientific research project.

[0015] Furthermore, step S3 specifically includes:

[0016] Before model training, a warm-up method is used to extract information from the pre-trained language model for initialization of node feature vectors;

[0017] Align the initial vector of the node with the word embedding vector space of the pre-trained language model to improve the efficiency of model training;

[0018] Set the initial vector to a low dimension to prevent overfitting and provide a suitable starting point for model training;

[0019] The preheated low-dimensional vector is projected into a high-dimensional vector space to align the text embedding vector and ensure that the node representation is consistent with the model output vector dimension.

[0020] Furthermore, the operation of the warm-up method includes: introducing a negative sample set into the loss function, enhancing the model's ability to distinguish between positive and negative samples, improving the quality of the node initial vector, calculating the similarity between the node feature vector and the text vector in the form of logarithmic probability, and learning the initial vector by optimizing this similarity; thereby, a valuable initial vector is learned for each node.

[0021] Furthermore, the implementation formula of the preheating method is:

[0022]

[0023] in, represents the negative sample set composed of rich text edges, Represents the corresponding text vector; author node Learn valuable initial vectors .

[0024] Furthermore, in step S4, the BERT model is used to process the paper text and extract text feature vectors; the first author of each paper is determined as the central node, and other authors are determined as neighbor nodes, and a query sequence is constructed by concatenating the text feature vector of the central node and the representation of the neighbor nodes; and the representation vector of the central author node is used as the key (Key) and the vector representation of its historical project experience as the value (Value) to construct key-value pair data, which is used for training the Transformer model to learn the correlation between the central author node and its historical project experience, thereby enhancing the model's learning of the matching between scientific research projects and experts and scholars.

[0025] Furthermore, in step S4, the construction of the query sequence specifically includes:

[0026] The text associated with the central node is processed through the BERT model to generate a feature vector of the central text edge, and the feature vector is concatenated with the representation vector of the neighboring node to construct a query sequence for Transformer model training, which is used to match scientific research projects with experts and scholars. When constructing the query sequence, the text feature vector of the central node and the representation vector of the neighboring node are concatenated in sequence to integrate the information of the neighboring nodes to form a complete query sequence. For the central author node, its representation vector is concatenated with the vector representation of the historical project experience to form another query sequence for learning the detailed information of the central author node.

[0027] Furthermore, in step S5, when processing query sequences and key-value pair data through the multi-head attention mechanism of the Transformer model, asymmetric attention head input data is adopted to adapt to the expert matching task of scientific research projects.

[0028] Furthermore, in step S6, the representation vector of the scholar node obtained through the reasoning process includes the past scientific research cooperation history of the central author node and a vector structure similar to the query sequence.

[0029] A computer program product includes a computer program, which, when executed by a processor, implements the scientific research project expert matching method based on dynamic academic collaboration network graph representation learning.

[0030] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the scientific research project expert matching method based on dynamic academic collaboration network graph representation learning.

[0031] The present invention has the following beneficial effects:

[0032] The present invention proposes a scientific research project expert matching method based on dynamic academic cooperation network graph representation learning, which innovatively combines dynamic academic cooperation network graph representation learning and Transformer model to achieve efficient matching between scientific research projects and experts and scholars. By constructing a text complex network that comprehensively considers the past research history and cooperation experience of scholars, the present invention not only improves the matching accuracy, but also achieves a matching accuracy of more than 80% in two data sets built based on the DBLP database, and the standardized discounted cumulative return is more than 89%. The method of the present invention reconstructs complex network structure data into sequence data and uses the multi-head attention mechanism of Transformer to effectively capture the multi-dimensional characteristics of scholar nodes, including temporal characteristics and cooperation information, thereby simplifying the matching process and improving efficiency. In addition, the present invention solves the problem of ignoring the feature information on the scholar side in the prior art, and provides a more comprehensive and accurate matching framework by deeply mining the past research experience and interest field information of experts and scholars. Compared with the prior art, the present invention shows obvious advantages in technology, cost and efficiency, especially when dealing with dynamic networks with time domain and text heterogeneity, it can better model and improve matching accuracy. The practicality and innovation of this method bring new solutions to the field of scientific research project management and provide strong technical support for the process of recommending suitable collaborators in academic social network applications.

[0033] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1This is an overall flow chart of a scientific research project expert matching method based on dynamic academic collaboration network graph representation learning according to an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the reconstruction of the scientific research cooperation network diagram according to an embodiment of the present invention.

[0036] Figure 3 This is a schematic diagram of the author node representation obtained by reasoning in the reasoning phase of an embodiment of the present invention. DETAILED DESCRIPTION

[0037] 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 and application of the present invention.

[0038] See also Figure 1 The embodiment of the present invention provides a scientific research project expert matching method based on dynamic academic cooperation network graph representation learning, comprising the following steps:

[0039] Step S1. Construct a dynamic complex network diagram representing a scientific research cooperation network, including scholar nodes and directed edges representing cooperation relationships, and timestamps associated with the edges; wherein the paper is the edge, its first author is the central node, and other authors are neighbor nodes. The direction of the edge points from the first author of the paper to other authors, and the timestamp indicates the time when the paper was published.

[0040] Step S2. Learn a feature vector for each scholar node in the network, which is used to capture the attributes of the node, such as research interests and collaboration experience.

[0041] Step S3: Initialize the node feature vector using the pre-trained language model.

[0042] In a preferred embodiment, a preheating method is used to extract information from a pretrained language model and initialize node feature vectors; a low-dimensional vector is set to prevent overfitting and accelerate model training. Specifically, it includes: before model training, a preheating method is used to extract information from a pretrained language model for initialization of node feature vectors; the initial vector of the node is aligned with the word embedding vector space of the pretrained language model to improve the efficiency of model training; the initial vector is set to a low dimension to prevent overfitting and provide a suitable starting point for model training; the preheated low-dimensional vector is projected to a high-dimensional vector space to align the text embedding vector to ensure that the node representation is consistent with the model output vector dimension. In a further preferred embodiment, the operation of the preheating method specifically includes: introducing a negative sample set in the loss function to enhance the model's ability to distinguish between positive and negative samples, improving the quality of the node initial vector, using the logarithmic probability to calculate the similarity between the node feature vector and the text vector, and learning the initial vector by optimizing this similarity; thereby, a valuable initial vector is learned for each node.

[0043] Step S4. According to the constructed dynamic complex network graph and node relationships, the information of the first author node of each paper and its related neighbor nodes is integrated to construct query sequences and key-value pair data for model training based on the Transformer architecture.

[0044] In a preferred embodiment, the BERT model is used to process the paper text and extract text feature vectors; the first author of each paper is determined as the central node, and other authors are determined as neighbor nodes, and a query sequence is constructed by concatenating the text feature vector of the central node with the representation of the neighbor nodes; and the representation vector of the central author node is used as the key (Key) and the vector representation of its historical project experience is used as the value (Value) to construct key-value pair data, which is used for training the Transformer model to learn the correlation between the central author node and its historical project experience, thereby enhancing the model's learning of the matching between scientific research projects and experts and scholars.

[0045] In a preferred embodiment, the construction of the query sequence specifically includes: processing the text associated with the central node through the BERT model, generating a feature vector of the central text edge, and concatenating the feature vector with the representation vector of the neighboring node to construct a query sequence for Transformer model training for matching scientific research projects with experts and scholars; wherein, when constructing the query sequence, the text feature vector of the central node and the representation vector of the neighboring node are concatenated in sequence to integrate the information of the neighboring nodes to form a complete query sequence; for the central author node, its representation vector is concatenated with the vector representation of the historical project experience to form another query sequence for learning the detailed information of the central author node.

[0046] Step S5. Through the multi-head attention mechanism of the Transformer model, the query sequence and key-value pair data are processed to learn the matching between scientific research projects and experts and scholars.

[0047] In a preferred embodiment, when processing query sequences and key-value pair data through the multi-head attention mechanism of the Transformer model, asymmetric attention head input data is used to adapt to the scientific research project expert matching task.

[0048] Step S6. After the model training is completed, the model parameters are fixed, and the representation vector of the scholar node is obtained through the reasoning process. The scholar node representation vector obtained by reasoning is used to match the most suitable experts and scholars for the scientific research project.

[0049] Preferably, the representation vector of the scholar node obtained through the reasoning process includes the past scientific research cooperation history of the central author node and a vector structure similar to the query sequence.

[0050] The present invention proposes an innovative scientific research project expert matching method, which models and learns graph representations of scientific research cooperation networks through multiple dimensions to achieve matching of the most suitable experts and scholars for scientific research projects. The core of the method of the present invention is to use the Transformer model to process the dynamic academic cooperation network based on text edges, integrate the past research experience of experts and scholars into the scholar nodes, and form a multi-dimensional vector representation. These representations not only take into account the time characteristics, but also cover the scholars' past research experience and forms of cooperative interaction, thereby providing a comprehensive and accurate matching basis for scientific research projects. The technical advantage of the present invention is that it can significantly improve the matching accuracy to more than 80%, and the standardized discounted cumulative benefit exceeds 89% (further detailed later). The method solves the limitations of the prior art, such as high computational volume and operating costs, and the failure to fully utilize the characteristic information on the scholar side, resulting in unsatisfactory matching results. In addition, the present invention also solves a key pain point in the application of academic social networks, that is, when recommending suitable collaborators for specific scientific research projects, how to effectively deal with dynamic heterogeneity based on time domain and text, thereby avoiding the complexity and accuracy of modeling. Compared with the existing technology, the innovative graph structure learning and Transformer model application of the present invention show obvious advantages, especially in processing dynamic networks and text information, providing a more comprehensive and efficient solution for efficient and accurate matching of experts and scholars in scientific research project management.

[0051] The following further describes an algorithm example and experimental verification of a specific embodiment of the present invention.

[0052] The scientific research cooperation network can be regarded as a dynamic complex network graph, and the model is expressed as Specifically, a dynamic complex network graph consists of a set of nodes (denoted by ) and a set of directed edges (denoted by These edges show the connection relationship between nodes. Indicates scholars, and even Representing scholars and The direction of the edge is from the first author of the paper to other authors. At the same time, each edge has a timestamp. , indicating the time when the paper was published. In order to achieve the matching between scientific research projects and experts and scholars, we first need to Learn to get a feature vector , this vector fully captures the attributes or characteristics of the node, such as past research interests and collaboration experiences. Here, auxiliary information of the author node, such as the author's identity and title characteristics, is not considered. On the one hand, it is impractical and extremely time-consuming to collect this information. In fact, from the perspective of dynamic analysis, categorical variables related to authors (such as age and academic influence) change over time, and it is difficult to obtain an accurate time series. On the other hand, the attributes of the author node are relatively easier to handle than unstructured text edge information. It can be easily applied to downstream prediction or recommendation tasks through simple feature engineering techniques. Therefore, in the present invention, only the temporal characteristics and topological structure properties of nodes in dynamic complex networks are focused on, without considering the exogenous characteristics of the nodes themselves.

[0053] Thanks to the clever design of the multi-head self-attention layer (MHA) and the feed-forward network (FFN), the Transformer algorithm can fully capture the long-term dependencies in continuous sequence data. Given an input sequence ,in It is Transformer layers are located at position A vector representation of text bytes, and denote the vector dimension and the length of the sequence respectively. Subsequently, the multi-head attention layer will Mapping to triples :

[0054]

[0055] in, represents different attention heads, is the weight matrix that needs to be continuously learned by the model. The calculation method of the attention mechanism is:

[0056]

[0057]

[0058] Similarly, It is also the weight matrix that needs to be learned to map the concatenated output to fit the model dimension. In addition, the Transformer also adds a residual link. After the output of the MHA module is processed by the FFN layer to obtain the residual result, it is normalized. Finally, The output of the layer Calculated by the following formula:

[0059]

[0060] In the method of the present invention, the main focus is on matching the most appropriate experts and scholars for a specific scientific research project. Specifically, given a dynamic scientific research network, first learn to obtain valuable author node representation vectors so that the subsequent expert matching task can be carried out smoothly. In particular, the learned vectors not only contain the time-varying evolution characteristics of the dynamic network graph, but also contain the paper cooperation information between scholars.

[0061] In a textual-edge dynamic network, the encoding carries the document Attribute Edge Different from directly encoding plain text, this is mainly because the text on the edge is naturally associated with network structure information, which helps model complex networks. In view of the excellent performance of the Transformer architecture in text semantic extraction, this paper designs a simple and effective data reconstruction method so that graph network information can be easily incorporated into the Transformer modeling process. The core idea is to learn author representation based on directed complex network graphs, and then use similarity matching to achieve the problem of matching scientific research projects with scholars and experts.

[0062] Figure 2 The network graph reconstruction paradigm based on the problem of matching scientific research papers with scholars and experts is given. First, a paper is usually a bridge connecting multiple author nodes. Here, the nodes related to a paper are divided into two parts: the central node (first author) and neighbor nodes ( ,express Here, it is assumed that the first author is the most important role in a scientific research project. Ideally, given a description of a scientific research project, the algorithm aims to match the most suitable expert scholar for this project, that is, the first author of the paper. Based on these settings, query and key-value pair data are constructed for training the scientific research project expert matching model based on Transformer.

[0063] First, use the BERT model to process text All characters in the , all output vectors are combined to get , remember all neighbor nodes The node representation is Then, the query sequence is obtained by concatenating the center text edges and neighbor node representations:

[0064]

[0065] In particular, before model training, the initial vector of the node is first obtained. In order to train more efficiently, the warm-up technique is used to extract information from the semantically rich pre-trained language model (PLM) into the text-free node representation to achieve the purpose of initializing the node vector. The purpose of this step is to align the initial vector representation of the node with the BERT word embedding vector space. The specific implementation formula is:

[0066]

[0067] in, represents the negative sample set composed of rich text edges, Represents the corresponding text vector. This step can be used to Learn valuable initial vectors . In particular, it is set to a low-dimensional (64-dimensional) vector to prevent overfitting, because it is not expected that the warm-up operation will allow the node representation to contain the semantic information of all articles, but only provide a suitable initial vector to accelerate subsequent model training. Finally, in order to align the text embedding vector, it is projected into a high-dimensional vector space:

[0068]

[0069] in, is the weight matrix, is a 769-dimensional vector representation of the author node, which is the same dimension as the output vector of BERT. Similarly, the central author node ( ) and its adjacent edges ( , ,express The query sequence is obtained by concatenating the vector representations of the historical project experiences of

[0070]

[0071] Intuitively, Contains the text semantic information of each scientific research project and the author information associated with it. At the same time, with the help of model training, the vector will Aligned, so the information of the author node in the center of the project is learned. The learning process is completed with the help of the Transformer model, which learns the two through a multi-head attention mechanism:

[0072]

[0073] It can be found that the input data of the two attention heads are asymmetric. This is to better achieve the expert matching task of scientific research projects. This asymmetric task is also widely used in existing research. After the model is fully trained, the model parameters are fixed to infer the node representation. Figure 3 . Output of the inference phase model As an author node representation Therefore, this vector representation naturally contains the central author node The past scientific research cooperation history of the authors and the learned vector structure similar to the query sequence. In real business scenarios, other author information in the query sequence can be replaced by preferred experts, and can even be generalized to more general scenarios, such as recruitment activities conducted by project leaders to expand the project team. Therefore, it is believed that the author node representation obtained by reconstructing the above cooperation network graph has practical value and can be beneficial to various downstream recommendation or matching tasks.

[0074] Examples

[0075] A commonly used DBLP citation network dataset was selected to verify the proposed expert matching algorithm. The dataset contains 5.25 million paper data. Considering that the original dataset is too large, the paper data in the sub-fields were screened. Specifically, with the help of the classification catalog of high-quality scientific and technological journals in the computer field published by the China Computer Society, journals or conference papers with A ratings published in the fields of artificial intelligence (AI) and data mining (DM) in the past 10 years were selected to construct two sub-datasets, which were recorded as CCF-AI and CCF-DM.

[0076] Here, taking the CCF-AI dataset as an example, the detailed steps of screening qualified data from the DBLP database data are introduced. First, the titles and years of all articles published between 2013 and 2023 were captured from the official website of the publication using crawler technology. Subsequently, the article title information captured from the official website was matched with the entries in the DBLP database, and the successfully matched data was added to the candidate set. After that, the publication names and venue information to which these articles belong were obtained, and these publication information was used as anchors to determine whether other entries in the non-candidate set should be updated into the candidate set. For example, articles published in the "2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)" should be considered in the research of this invention. This can be easily achieved with the help of the above method. Finally, the original candidate set and the newly added entries are merged to obtain the final CCF-AI dataset.

[0077] After the above cross-validation process, the final CCF-AI dataset contains about 100,000 authors and 57,000 high-level articles. The information of these records is relatively complete, including information such as article, location, author, and year. Similarly, the CCF-DM dataset can also be obtained. The basic information of the two datasets is shown in Table 1.

[0078] Table 1 Basic information of the constructed dataset

[0079]

[0080] In the application scenario of the present invention, the task goal is to match the most suitable scholar expert for the project in the expert and scholar database given the title or project introduction of a scientific research project (the title or abstract of the paper can be used here). The expert database contains scholars' past scientific research project experience. To this end, the data set is divided into a training set and a test set with 2022 as the dividing time node. Specifically, articles published in 2022 are regarded as scientific research projects that need to be matched, and the first author corresponding to the article is regarded as the most suitable scholar expert for the article. The ultimate goal of the matching algorithm is to train the Transformer model to correctly match the two and achieve higher matching accuracy. For the data from the nine years before 2021, a training data set can be obtained through a similar data construction method for model training. In addition, part of the data from 2021 was selected for model verification to obtain the optimal model parameters.

[0081] Experimental Results

[0082] In order to verify the effect of the proposed graph representation learning in expert-scholar matching, the graph convolutional neural network GraphSAGE and the BERT model were also selected for comparison. Based on the two methods, the maximum and average pooling were used to obtain the final scholar node representation (respectively denoted as BERT-Mean, BERT-Max, GraphSAGE-Mean and GraphSAGE-Max), which were used to complete the task of matching projects with experts and scholars. In order to evaluate and compare the effectiveness of different methods, three evaluation indicators were used to evaluate the final results, namely Precision (PREC), Mean Reciprocal Rank (MRR) and Normalized Discounted Cumulative Gain (NDCG). The definitions of the three indicators are as follows:

[0083]

[0084] Where TP (True Positives) represents the number of correctly matched positive instances, and FN (False Negatives) is the number of actual positive instances that were incorrectly matched. Indicates the ranking position of the author whose matching result is consistent with the actual result, and the relevance value ( ) describes the correlation score between the position prediction result and the true result, and reflects the correlation score in an ideal order. Here, the average indicator value of the entire test set is calculated. MRR focuses on whether the appropriate experts found for the project are ranked in a more prominent position, that is, emphasizing "sequentiality", while NDCG not only considers the order, but also further considers the importance of relative position.

[0085] Table 2 Effect of expert matching in scientific research projects

[0086]

[0087] In actual tests, the method proposed in the present invention achieved relatively ideal results, as shown in Table 2. From the results in the table, the following conclusions can be drawn. On the one hand, the method TEDMatch proposed in the present invention has achieved results far higher than other benchmark models in two different fields. Taking the CCF-AI dataset of panel A as an example, TEDMatch achieved excellent performance of 0.8702, 0.9016 and 0.9232 in the three indicators of PREC, MRR and NDCG, respectively, while the second-ranked GraghSAGE-MAX method only achieved index values ​​of 0.3311, 0.4393 and 0.5537 in the three indicators, which shows that the method proposed in the present invention has indeed learned key information that helps expert matching after comprehensively considering the scholars' past research interests and cooperation history, thereby achieving better performance. On the other hand, compared with directly pooling the text vectors obtained by the BERT model (BERT-Mean and BERT-Max models), using the GraphSAGE graph neural network model for matching can achieve better results, which shows that GraphSAGE has indeed learned more information about the graph network structure, thereby improving the effect of the matching task. This shows that when processing complex text and complex network graphs, it is necessary to comprehensively consider the text content on the edges and the network topology information between the nodes. Simply aggregating the feature information of the edges into the nodes may lead to the loss of network structure information, thereby affecting the accuracy of the model.

[0088] In general, the present invention constructs a dynamic complex network diagram, takes scholars as nodes, and cooperative relationships as directed edges, taking into account the scholars' past research history and cooperation experience. Using graph structure learning and Transformer architecture, feature vector learning is performed on scholar nodes to achieve accurate matching between scientific research projects and experts and scholars. This method has achieved a matching accuracy of more than 80% and a standardized discounted cumulative return of more than 89% on data sets in the fields of artificial intelligence and data mining, effectively solving the limitations of existing technologies in computing costs and the use of scholar feature information, and providing an innovative solution for recommending suitable collaborators for scientific research projects.

[0089] 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.

[0090] 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.

[0091] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0092] 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 ferromagnetic random access memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a 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.

[0093] 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 only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, 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, 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.

[0094] 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 present embodiment.

[0095] 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.

[0096] A person skilled in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, etc. Various media that can store program codes.

[0097] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function 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 can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a 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.

[0098] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0099] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0100] 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.

[0101] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A scientific research project expert matching method based on dynamic academic cooperation network graph representation learning, characterized in that: The following steps are involved: S1. Construct a dynamic complex network diagram representing the scientific research cooperation network, including scholar nodes and directed edges representing cooperation relationships, as well as timestamps associated with the edges; among them, the paper is the edge, its first author is the central node, and other authors are neighbor nodes. The direction of the edge points from the first author of the paper to other authors, and the timestamp indicates the time when the paper was published; S2. Learn a feature vector for each scholar node in the network, which is used to capture the properties of the node; S3. Initialize the node feature vector using the pre-trained language model; S4. According to the constructed dynamic complex network graph and node relationships, the information of the first author node of each paper and its related neighbor nodes is integrated to construct the query sequence and key-value pair data for Transformer model training; S5. Through the multi-head attention mechanism of the Transformer model, query sequences and key-value pair data are processed to learn the matching between scientific research projects and experts and scholars; S6. After the model training is completed, the model parameters are fixed, and the representation vector of the scholar node is obtained through the reasoning process to match the most suitable experts and scholars for the scientific research project.

2. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning as claimed in claim 1 is characterized in that: Step S3 specifically includes: Before model training, a warm-up method is used to extract information from the pre-trained language model for initialization of node feature vectors; Align the initial vector of the node with the word embedding vector space of the pre-trained language model to improve the efficiency of model training; Set the initial vector to a low dimension to prevent overfitting and provide a suitable starting point for model training; The preheated low-dimensional vector is projected into a high-dimensional vector space to align the text embedding vector and ensure that the node representation is consistent with the model output vector dimension.

3. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning as described in claim 2 is characterized in that: The operation of the warm-up method includes: introducing a negative sample set into the loss function, enhancing the model's ability to distinguish between positive and negative samples, improving the quality of the node initial vector, calculating the similarity between the node feature vector and the text vector in the form of logarithmic probability, and learning the initial vector by optimizing this similarity; thereby, a valuable initial vector is learned for each node.

4. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning as described in claim 3 is characterized in that: The implementation formula of the preheating method is: ； in, represents the negative sample set composed of rich text edges, Represents the corresponding text vector; author node Learn valuable initial vectors .

5. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning according to any one of claims 1 to 4, characterized in that: In step S4, the BERT model is used to process the paper text and extract text feature vectors; the first author of each paper is determined as the central node, and other authors are determined as neighbor nodes. The query sequence is constructed by concatenating the text feature vector of the central node and the representation of the neighbor nodes; and the representation vector of the central author node is used as the key (Key) and the vector representation of its historical project experience as the value (Value) to construct key-value pair data. The key-value pair data is used for training the Transformer model to learn the correlation between the central author node and its historical project experience, thereby enhancing the model's learning of the matching between scientific research projects and experts and scholars.

6. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning according to any one of claims 1 to 4, characterized in that: In step S4, the construction of the query sequence specifically includes: The text associated with the central node is processed through the BERT model to generate a feature vector of the central text edge, and the feature vector is concatenated with the representation vector of the neighboring node to construct a query sequence for Transformer model training, which is used to match scientific research projects with experts and scholars. When constructing the query sequence, the text feature vector of the central node and the representation vector of the neighboring node are concatenated in sequence to integrate the information of the neighboring nodes to form a complete query sequence. For the central author node, its representation vector is concatenated with the vector representation of the historical project experience to form another query sequence for learning the detailed information of the central author node.

7. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning according to any one of claims 1 to 4, characterized in that: In step S5, when processing query sequences and key-value pair data through the multi-head attention mechanism of the Transformer model, asymmetric attention head input data is used to adapt to the expert matching task of scientific research projects.

8. The scientific research project expert matching method based on dynamic academic cooperation network graph representation learning according to any one of claims 1 to 4, characterized in that: In step S6, the representation vector of the scholar node obtained through the reasoning process includes the past scientific research cooperation history of the central author node and a vector structure similar to the query sequence.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the scientific research project expert matching method based on dynamic academic cooperation network graph representation learning as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the scientific research project expert matching method based on dynamic academic cooperation network graph representation learning as described in any one of claims 1 to 8.

Images