Tissue-oriented potential scientific research team mining method based on network diffusion

Abstract

The invention discloses a project-oriented potential scientific research team mining method based on network diffusion, and the method comprises the following steps: obtaining and preprocessing academic data, and constructing an author-paper bipartite graph network; initializing a seed paper set according to a reference coupling relationship of the target subject, and constructing an initial influence signal in a paper layer of the bipartite graph network; parallelly executing two network diffusion processes of mass diffusion and heat conduction on the bipartite graph network, and respectively calculating to obtain an author influence vector and an author influence vector under a heat conduction mechanism; normalizing the two obtained author influence vectors, and performing linear weighted fusion by adjusting parameters to obtain a final mixed influence vector; sorting all authors according to the mixed influence vector, and selecting the top authors as potential scientific research team recommendation; the method is suitable for scientific research environments in different periods.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically to a method for identifying potential research teams for specific research topics based on network diffusion. Background Technology

[0002] Currently, research team discovery and expert recommendation technologies are mainly based on data-driven algorithm recommendation methods. With the maturity of academic knowledge graphs (such as SciSciNet and MAG), bipartite graph-based recommendation algorithms have gradually become mainstream. Among them, collaborative filtering is the most common method, but it mainly relies on local similarity for prediction and struggles to capture deep structural dependencies in the network. In recent years, physical principle-based network diffusion algorithms have been introduced into this field due to their good physical interpretability, with mass diffusion and heat conduction algorithms being the most representative. Although the above-mentioned physical diffusion algorithms have achieved certain results in general recommendation systems, when directly applied to the specific scenario of "discovering research teams for research projects," existing technologies still have significant defects and irreconcilable contradictions: mass diffusion algorithms suffer from severe "authority bias": mass diffusion algorithms simulate the conserved diffusion process of matter in the network, and their core mechanism is column normalization at the source end. This mechanism leads to the tendency for influential resources to cluster towards nodes with high degrees (i.e., more published papers and extensive collaborations). In academic networks, this means the algorithm tends to over-recommend established field authorities. While its accuracy is high, the results often lack novelty and diversity, failing to identify promising scholars who, despite having fewer publications, are highly relevant to the research topic. The heat conduction algorithm suffers from both "insufficient accuracy" and "cold start" problems: it simulates the balanced heat transfer process in the network, with its core mechanism being row normalization at the target end, i.e., neighborhood averaging. This mechanism effectively suppresses the dominance of high-degree nodes, distributing energy and making it easier to discover long-tail nodes (low-degree authors). However, this "decentralized" smoothing effect results in significantly lower prediction accuracy and ranking scores compared to the quality diffusion algorithm. While the recommendation lists generated using the heat conduction algorithm alone are novel, they often contain a large amount of noise, leading to low usability in practical research management. The lack of a hybrid control mechanism for the trade-off between accuracy and diversity: Existing research teams often use either-or techniques, either falling into the Matthew effect of "the rich get richer" and leading to a rigid team structure, or focusing too much on long-tail discovery, resulting in unreliable recommendation results. Summary of the Invention

[0003] Purpose of the Invention: The purpose of this invention is to provide a potential research team discovery method based on network diffusion, which is based on mass diffusion and heat conduction. It provides matrix-based formulas and implementation details adapted to the author-paper graph, and introduces a linear hybrid model. It adopts a multi-round iterative diffusion process, which allows the influence signal to propagate through multiple layers of the network, thereby capturing the deeper structural dependencies between authors and papers and solving the problems existing in the background technology.

[0004] Technical Solution: The present invention provides a method for identifying potential research teams for specific research topics based on network diffusion, comprising the following steps:

[0005] (1) Acquire academic data and preprocess it to construct an author-paper bipartite graph network;

[0006] (2) Determine the target research topic, initialize the seed paper set based on the reference coupling relationship of the target topic, and construct the initial influence signal in the paper layer of the bipartite graph network;

[0007] (3) The two network diffusion processes of mass diffusion and heat conduction are executed in parallel on the bipartite graph network, and the author influence vector under the mass diffusion mechanism and the author influence vector under the heat conduction mechanism are calculated respectively.

[0008] (4) Normalize the two author influence vectors obtained, and perform linear weighted fusion by adjusting the parameters to obtain the final mixed influence vector;

[0009] (5) Sort all authors according to the mixed influence vector and select the top-ranked authors as potential research teams for the target topic.

[0010] Furthermore, in step (2), the initialization of the seed paper set is as follows: obtain the reference set of the target topic, and select all papers in the training data that share at least one reference with the reference set to form the seed paper set.

[0011] Furthermore, in step (3), the quality diffusion process is based on the resource conservation mechanism, and through multi-step iterative propagation from the paper layer to the author layer and back to the paper layer, it gathers author nodes with high influence degrees.

[0012] Furthermore, in step (3), the heat conduction process is based on the neighborhood averaging mechanism, and spreads through multiple iterative steps from the paper layer to the author layer and back to the paper layer, diffusing influence to author nodes with low degree.

[0013] Furthermore, in step (3), both the mass diffusion process and the heat conduction process are iterated multiple times to capture the deeper structural dependencies between the authors and the paper.

[0014] Furthermore, in step (4), the adjustment parameter is used to control the weight of the mass diffusion influence and the heat conduction influence in the final mixed result, so as to achieve a balance between accuracy and diversity in the recommendation result.

[0015] Furthermore, in step (5), the recommended team candidate list includes both authors highly relevant to the target topic and potential authors relevant to the target topic but with few published results.

[0016] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention uses a hybrid model, leveraging the MD mechanism to ensure that the recommended results include core authorities in the field (high accuracy), while utilizing the smoothing effect of the HC mechanism to uncover promising scholars in their rising phase (high novelty). The introduction of an adjustment parameter α allows for flexible switching between robust recommendations (high α) and exploratory recommendations (low α) according to actual needs. Experiments show that when α=0.8, the model outperforms the single MD algorithm in accuracy while improving novelty by approximately 20% or more. This invention is applicable to research environments at different stages. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention;

[0018] Figure 2 This is a schematic diagram comparing the propagation principles of the two diffusion mechanisms, mass diffusion and heat conduction, in the author-paper bipart diagram of the present invention.

[0019] Figure 3 This invention represents the optimal hybrid model (Hybrid16) from 1985 to 2020. A comparison of the ranking score, precision, recall, diversity, and novelty between the 0.8 (presumably referring to a specific model or algorithm) and the quality diffusion model (MD16), and the optimal (presumably referring to a specific algorithm or algorithm) model. Evaluation of time stability. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0021] like Figure 1 As shown, this embodiment of the invention provides a method for identifying potential research teams for specific research topics based on network diffusion, comprising the following steps:

[0022] Step 1: Data Preprocessing and Bipartite Graph Construction: Based on the SciSciNet academic dataset, data from a specific time window (4 years of training + 1 year of testing) is selected. First, an author-paper bipartite graph network is constructed. ,in For the collection of authors, This is a collection of papers. Define the adjacency matrix. If the author I participated in writing the paper Then matrix elements Otherwise, it is 0. Simultaneously, the degree matrix is ​​calculated for subsequent normalization:

[0023] Paper Degree Matrix ,in Indicates the paper The number of authors.

[0024] Author Degree Matrix ,in Indicates the author Number of published papers.

[0025] Step 2: Task Definition and Signal Initialization: Set the target topics to be recommended (i.e., the target papers in the test set). Initializing network signals based on the reference coupling principle: retrieving target papers. References collection And find all the related ones in the training set. Papers that share at least one reference constitute a seed node set. Initialize the influence vector of the paper layer. :

[0026]

[0027] This step ensures that the starting point of the diffusion has a substantial academic connection with the target topic.

[0028] Step 3: Implementation of the Quality Diffusion Algorithm Diffusion Operator: The quality diffusion mechanism follows the principle of resource conservation and tends to mine authoritative nodes with high influence. First, the quality diffusion transition matrix is ​​constructed. This matrix combines column normalization at the paper level and column normalization at the author level to describe the two-step propagation process from paper to author to paper:

[0029]

[0030] Perform iterative diffusion: at the paper level, according to the rules Update the influence. After reaching the preset diffusion rounds. Then, the influence is mapped unidirectionally from the paper level back to the author level to obtain the author influence vector under the quality diffusion mechanism. :

[0031]

[0032] Here The operator reflects the equal distribution characteristic of resources from papers to authors in the quality diffusion algorithm.

[0033] Step 4: As Figure 2 As shown, the heat conduction algorithm's diffusion operator is implemented as follows: the heat conduction algorithm follows the neighborhood averaging principle and tends to discover novel nodes with low degrees. First, the heat conduction transfer matrix is ​​constructed. This matrix combines author-level row normalization and paper-level row normalization to describe the heat conduction process:

[0034]

[0035] Perform iterative diffusion: at the paper level, according to the rules Heat conduction is updated. Also in the... After the round ends, the influence is mapped back to the author level to obtain the author influence vector under the heat conduction mechanism. :

[0036]

[0037] Here The operator reflects the characteristic of the author node in the heat conduction algorithm to obtain the average heat of adjacent papers.

[0038] like Figure 2 As shown, taking the first two iterations of the mass diffusion process as an example... Figure 2 The numerical notation is explained as follows: In the initial state (left side), the target paper and its associated seed papers are assigned an initial resource of 1. In the first step (paper → author): resources are evenly distributed from the paper node to its connected author nodes. For example, as shown... Figure 2 The first author in the top left corner connects to three papers: two seed papers (resource 1) and one non-seed paper (resource 0). The resource this author receives from each seed paper is the total resource of that paper divided by the paper's degree. . Figure 2 The value '5 / 6' shown represents the normalized total resources the author received from all connected papers. In the second step (Author → Paper): resources are redistributed evenly back to the paper nodes from the author nodes. The author divides all resources by their degree and passes them to the connected papers. After multiple iterations (such as...), Figure 2 The middle right side shows The state of each node represents its influence score at that moment (e.g., '71 / 108'), reflecting its importance in the network structure. The heat conduction process works similarly, except that its resource propagation follows a neighborhood averaging mechanism (row normalization). Figure 2 The lower half of the text shows the numerical changes under the heat conduction mechanism.

[0039] Step 5: Linear Mixture Aggregation To balance diversity and novelty, the two resulting author influence vectors are normalized, and then an adjustment parameter is introduced. Perform linear weighted fusion:

[0040]

[0041] After a series of parameter adjustments, it was found that when When the value is around 0.8, the model can significantly improve the novelty and diversity of the recommendation results while maintaining a low ranking score.

[0042] Step 6: Candidate Ranking and Output: Based on the final mixed influence vector All author nodes are sorted in descending order. The top-N authors (N being the actual number of authors in the paper) are selected as a recommended list of potential research teams for this target topic. This list includes both core experts in the field and promising newcomers highly relevant to the topic, achieving optimized recommendations for research team composition.

[0043] experiment:

[0044] Dataset: The experimental dataset uses the paper table, paper citation table, and paper, author and institution association table from the SciSciNet data pool.

[0045] Data preprocessing: To effectively improve the accuracy and computational efficiency of data mining by research teams, and thus enable the system to better handle different disciplines and time spans, rigorous preprocessing of the original SciSciNet data is required. This mainly includes:

[0046] Bipartite Graph Network Construction and Normalization: Due to differences in the number of publications and citation habits in different fields, it is necessary to construct an author-paper bipartite graph and perform row / column normalization on the adjacency matrix so that energy diffusion can be carried out at the same scale;

[0047] Seed signal initialization: For the target topic, the reference coupling principle is adopted to select training set papers that share references with the target paper as seed nodes, and inactive authors are filtered out by time slicing, so as to make feature extraction and relevance calculation more accurate.

[0048] Experimental platform:

[0049] CPU: Intel Xeon Gold 5220R

[0050] GPU: Nvidia A100 40G *2

[0051] System: Ubuntu 20.04.5

[0052] Experimental results:

[0053] The evaluation metrics for the model mainly include accuracy and diversity. Comparing the hybrid network diffusion model of this invention with traditional mass diffusion and heat conduction algorithms reveals a significant improvement in overall performance.

[0054] As shown in Table 1 and Figure 3 As shown, to verify the model's performance under deep information propagation, experimental data with 16 diffusion rounds were selected for comparison (at which point the model has reached a convergent and stable state, denoted as MD16, HC16, and hybrid 16, respectively). Under this setting, the proposed method of using... Compared to the classic MD16 model, which also has 16 rounds of diffusion, the hybrid model with the parameter configuration reduced the ranking score, which measures recommendation accuracy, by 2.75%, while improving precision and recall by 3.84% and 2.89%, respectively.

[0055] More importantly, this invention has achieved a breakthrough in measuring the ability to discover potential talent: the novelty index is significantly improved by 23.92% compared to the classic quality diffusion model, and the diversity index is improved by 1.50%. This shows that while maintaining or even optimizing the reliability of recommendations, this invention greatly overcomes the authority bias of traditional algorithms, verifying the effectiveness of the proposed method in discovering high-potential new talents.

[0056] Table 1 Comparison of Model Accuracy and Performance

[0057] .

[0058] Table 2 provides an explanation of all mathematical symbols and their physical meanings. Table 2. Explanation of Mathematical Symbols and Their Physical Meanings .

Claims

1. A method for identifying potential research teams for specific research topics based on network diffusion, characterized in that, Includes the following steps: (1) Acquire academic data and preprocess it to construct an author-paper bipartite graph network; (2) Determine the target research topic, initialize the seed paper set according to the reference coupling relationship of the target topic, and construct the initial influence signal in the paper layer of the bipartite graph network; (3) The two network diffusion processes of mass diffusion and heat conduction are executed in parallel on the bipartite graph network, and the author influence vector under the mass diffusion mechanism and the author influence vector under the heat conduction mechanism are calculated respectively. (4) Normalize the two author influence vectors obtained, and perform linear weighted fusion by adjusting the parameters to obtain the final mixed influence vector; (5) Sort all authors according to the mixed influence vector and select the top-ranked authors as potential research teams for the target topic.

2. The method for identifying potential research teams based on network diffusion according to claim 1, characterized in that, In step (2), the initialization of the seed paper set is as follows: obtain the reference set of the target topic, and select all papers in the training data that share at least one reference with the reference set to form the seed paper set.

3. The method for identifying potential research teams based on network diffusion according to claim 1, characterized in that, In step (3), the quality diffusion process is based on the resource conservation mechanism and spreads through multiple steps from the paper layer to the author layer and back to the paper layer, gathering author nodes with high influence degrees.

4. The method for identifying potential research teams based on network diffusion according to claim 1, characterized in that, In step (3), the heat conduction process is based on the neighborhood averaging mechanism, and spreads the influence to author nodes with low degree through multi-step iterative propagation from the paper layer to the author layer and back to the paper layer.

5. The method for identifying potential research teams for a research topic based on network diffusion according to claim 1, characterized in that, In step (3), both the mass diffusion process and the heat conduction process are iterated multiple times to capture the deeper structural dependencies between the authors and the paper.

6. The method for identifying potential research teams based on network diffusion according to claim 1, characterized in that, In step (4), the adjustment parameter is used to control the weight of mass diffusion influence and heat conduction influence in the final mixture result, so as to achieve a balance between accuracy and diversity in the recommendation result.

7. The method for identifying potential research teams based on network diffusion according to claim 1, characterized in that, In step (5), the recommended team candidate list includes both authors who are highly relevant to the target topic and potential authors who are relevant to the target topic but have few published results.

Images