Teacher and bearing relationship prediction method based on science and technology data fused with new subject map
By constructing a machine learning model based on paper and research journal data, and utilizing network representation learning and deep neural network-hybrid principal component analysis algorithms, the problems of data privacy and high manual costs in constructing mentorship relationship data in existing technologies are solved, and efficient and reliable mentorship relationship prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN TALENT DIGITAL TECH CO LTD
- Filing Date
- 2026-01-21
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies for constructing mentorship relationships suffer from data privacy and security concerns, high labor costs, and a lack of objective and scalable data systems, resulting in limited reference value for mentorship relationship data and making it difficult to construct on a large scale.
We employ a machine learning model based on data from academic papers and research journals. By combining network representation learning algorithms and deep neural network-hybrid principal component analysis (PCA) algorithms with pooling layer technology, we construct a new subject graph recognizer to predict mentorship relationships. We then use long short-term memory neural networks (LSTM) for deep optimization to establish a reliable machine learning prediction model.
It enables accurate prediction of mentorship relationships in large-scale data, reduces labor costs, improves the objectivity and reliability of data, and can identify mentorship relationships in massive amounts of academic paper data.
Smart Images
Figure CN122022787A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of information technology, specifically relating to a method for predicting mentorship relationships in a new subject map based on the fusion of scientific and technological data. Background Technology
[0002] Academic networks can form based on different types of relationships, such as those between colleagues, friends, and mentors. These relationships typically reflect different interpersonal interactions. For example, in a mentor-mentee relationship, a doctoral student's research topic is usually determined by their mentor. In a friendship, a person's daily schedule might be decided by their friend. These interactions control the dynamics and complexity of social networks. To better model interactions based on network science, a specific network is abstracted as a graph consisting of nodes and edges, where nodes represent entities and edges represent different relationships. Therefore, we can use graph theory methods and machine learning techniques to model the relationships between nodes and edges from both local and global perspectives.
[0003] In traditional interpersonal relationship construction, manual data entry and active self-reporting of mentorship relationships are the main methods. However, this method is not only difficult to collect due to subjective factors such as data privacy and security concerns, and high labor costs, but also lacks a complete, objective, and large-scale system for analysis, resulting in limited reference value for the mentorship relationship data system. If large-scale mentorship relationship construction is required (on a scale of millions), it would not only require a large number of people to conduct detailed investigations of interpersonal relationships, resulting in a huge workload, but would also be subject to the influence of subjective factors. Therefore, we need to propose a mentorship relationship prediction method that integrates science and technology data with a new disciplinary map to solve the above-mentioned problems. Summary of the Invention
[0004] The purpose of this invention is to provide a method for predicting mentorship relationships by integrating scientific and technological data into a new disciplinary map. Based on data from papers or scientific journals, a machine learning model is constructed to accurately predict mentorship relationships between talents, thereby solving the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data includes the following steps:
[0007] S1. Retrieve science and technology papers from the past 30 years from existing think tanks, aggregate the retrieved papers by name, and compile the papers corresponding to each talent node.
[0008] S2. Construct the collected papers, format the paper data, build a list of collaborators based on each paper, and build a paper data collector.
[0009] S3. Extract human-centric information from the paper data, establish edge links between collaborators, and construct a paper data representation.
[0010] S4. Represent the edges and nodes based on the network representation learning algorithm to construct a network representation learner while preserving the subject attributes of the paper.
[0011] S5. Based on the deep neural network-hybrid principal component analysis (PCA) algorithm and pooling layer technology, a mentorship relationship identifier is built. At the same time, attribute network information is added, and a new subject graph identifier is built based on the subject. Different subjects are classified and optimized. The mentorship relationship identifier is used to form a dataset of advisor-advisee pairs.
[0012] S6. Further divide the dataset, using one part for training and the other for testing. Optimize the model using the AdaBoost method to finally form the trained model.
[0013] S7. Compare the predicted results and optimize the model based on the actual data. Use the Long Short-Term Memory Neural Network (LSTM) for deep optimization to generate a complete set of mentorship relationship datasets.
[0014] Preferably, in step S5, the use of the deep neural network-hybrid principal component analysis algorithm has two purposes: first, to denoise the data, and second, to reduce the dimensionality for visualization, reducing the original dataset to an n-dimensional dataset with the minimum projection distance.
[0015] Preferably, when building the mentorship relationship recognizer, a deep neural network-hybrid principal component analysis algorithm is first used to reduce the dimensionality of the data to find the core data domain. The core data domain is one of the following: research institution, academic career age, and name ID. Since some disciplines contain tens of thousands of scholars, it is difficult to ensure the computation time and memory usage when controlling multiple vectors. Therefore, a pooling layer is added to compress the input elements. First, each adjacent vector is reduced to 1000 dimensions and the average value is calculated to reduce the vectors as input to the encoder accordingly. The entire dataset is first organized based on all disciplines, and then the new discipline map recognizer is used to segment the dataset according to discipline classification to divide the professional dataset.
[0016] Preferably, in step S2, the collaborator list is constructed based on the author's name and institution, and a unique UUID identifier is created for each author.
[0017] Preferably, in step S3, the specific steps for constructing the paper data representation are as follows:
[0018] S3.1, the edge is represented as:
[0019] Paper1---->{Person1,Person2....}, Paper2---->{Person3,Person4...}, Paper1 is a paper, and Person1 and Person2 are collaborators associated with the paper;
[0020] S3.2, Dataset Calculation --- By Year: { - }, thus obtaining the dataset of the years for which you want to perform statistics.
[0021] Preferably, in step S4, the edge represents the relationship between a person and a paper, and the node represents a paper. Preserving the subject attributes of the paper is to prepare for further optimization training based on subject classification.
[0022] Preferably, in step S1, the think tank refers to an independent research institution or organization composed of a group of experts and scholars who specialize in researching and providing public policy advice. Think tanks typically express their views and provide advice by publishing research reports, organizing seminars and forums, and participating in policy formulation and consultation. Think tanks cover fields such as science and technology, economics, international relations, energy, and the environment. In the field of science and technology, think tanks typically focus on issues related to science and technology development trends, technological innovation, and science and technology policies, providing scientific and professional advice and guidance for government and corporate decision-making.
[0023] Preferably, when retrieving paper data from the think tank, to ensure data completeness and accuracy of cooperation trends, data from the current year is removed, and the starting year is denoted as Y. start The end year is Y end .
[0024] Preferably, in step S6, when the dataset optimizes the model, the following steps are specifically included:
[0025] S6.1, Divide the training dataset into Train1;
[0026] S6.2, Divide the test dataset into Test1;
[0027] S6.3, Model training based on Train1;
[0028] S6.4, based on Test1, test the accuracy of the model and perform optimization.
[0029] Preferably, in step S7, the process of using a Long Short-Term Memory (LSTM) neural network for deep optimization is as follows:
[0030] S7.1 Data Preprocessing: Preprocess the input data, including data cleaning, standardization, and splitting the training set and test set;
[0031] S7.2 Model Structure Design: Design a suitable LSTM model structure based on the specific task requirements and data characteristics; the LSTM model structure can be further optimized by increasing the number of LSTM layers, adjusting the size of the LSTM layers, and adding other types of layers;
[0032] S7.3 Parameter tuning: Train the model using the training set and optimize model performance by adjusting different hyperparameters;
[0033] S7.4 Regularization: In order to reduce overfitting, regularization techniques are introduced into the model to help the model generalize better to unseen data;
[0034] S7.5 Gradient clipping: To avoid the gradient explosion problem, the gradient is clipped during training to limit its maximum value;
[0035] S7.6 Sequence Length Processing: For long sequence data, it is necessary to consider truncating or padding the sequence so that the model can process it better. Truncating is to select the sequence length to retain according to the task requirements, while padding is to use specific padding symbols to expand the sequence to a fixed length.
[0036] S7.7 Batch Normalization: The batch normalization layer accelerates the training process and improves model performance and robustness.
[0037] S7.8 Model Evaluation and Adjustment: Evaluate the optimized model using a test set, and adjust and improve it based on the evaluation results until the expected performance is achieved.
[0038] The present invention proposes a method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data. Compared with existing technologies, this method has the following advantages:
[0039] 1. This invention analyzes and statistically processes data from papers and patents across multiple fields, constructs a network representation learner using a network representation learning algorithm, and then builds a mentorship relationship identifier based on deep neural network-hybrid principal component analysis (PCA) algorithm and pooling layer technology. Attribute network information is also incorporated, and a new subject graph identifier is built based on disciplines. Different disciplines are classified and optimized. The mentorship relationship identifier forms an advisor-advisee dataset. Combining the mentorship relationship identifier and the new subject graph identifier, a reliable machine learning prediction model is established to predict mentor-student relationships. The invention statistically analyzes and models the collaborative network of talents in papers and patents at different times, thereby determining the mentorship relationship of the talent, such as who their mentor is and which students they have mentored. This allows for better identification of mentorship relationships in massive amounts of paper data. Attached Figure Description
[0040] Figure 1 This is a flowchart of the present invention;
[0041] Figure 2 This is a flowchart of the present invention. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] This invention provides, for example Figure 1-2 The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as shown, includes the following steps:
[0044] S1. Retrieve science and technology papers from the past 30 years from existing think tanks, aggregate the retrieved papers by name, and compile the papers corresponding to each talent node.
[0045] The term "think tank" refers to an independent research institution or organization composed of a group of experts and scholars who specialize in researching and providing public policy advice. Think tanks typically express their views and provide advice by publishing research reports, organizing seminars and forums, and participating in policy formulation and consultation. Think tanks cover fields such as science and technology, economics, international relations, energy, and the environment. In the field of science and technology, think tanks usually focus on issues related to science and technology development trends, technological innovation, and science and technology policies, providing scientific and professional advice and guidance for government and corporate decision-making.
[0046] When retrieving paper data from the think tank, to ensure data completeness and accuracy of cooperation trends, data from the current year is removed, and the starting year is denoted as Y. start The end year is Y end .
[0047] S2. Construct the collected papers, format the paper data, build a list of collaborators based on each paper, and construct a paper data collector; the collaborator list is constructed based on the author's name and institution, and a unique UUID identifier is constructed for each author.
[0048] S3. Extract human-centric information from the paper data, establish edge links between collaborators, and construct a paper data representation.
[0049] The specific steps for constructing a paper data representation are as follows:
[0050] S3.1, the edge is represented as:
[0051] Paper1---->{Person1,Person2....}, Paper2---->{Person3,Person4...}, Paper1 is a paper, and Person1 and Person2 are collaborators associated with the paper;
[0052] S3.2, Dataset Calculation --- By Year: { - }, thus obtaining the dataset of the years for which you want to perform statistics.
[0053] S4. Represent the edges and nodes based on the network representation learning algorithm to construct a network representation learner, while retaining the subject attributes of the paper; the edges represent the relationship between people and papers, and the nodes represent papers. Retaining the subject attributes of the papers is to prepare for further optimization training based on subject classification.
[0054] S5. Based on the deep neural network-hybrid principal component analysis (PCA) algorithm and pooling layer technology, a mentorship relationship identifier is built. At the same time, attribute network information is added, and a new subject map identifier is built based on disciplines to classify and optimize different disciplines. The mentorship relationship identifier is used to form an advisor-advisee pair dataset. The deep neural network-hybrid principal component analysis (PCA) algorithm has two purposes: first, to denoise the data, and second, to reduce the dimensionality for visualization, reducing the original dataset to an n-dimensional dataset with the minimum projection distance.
[0055] When building the mentorship relationship identifier, we first used a deep neural network-hybrid principal component analysis (PCA) algorithm to reduce the dimensionality of the data and identify the core data domain. The core data domain is one of the following: research institution, academic career age, and name ID. Since some disciplines contain tens of thousands of scholars, it is difficult to ensure the computation time and memory usage when controlling multiple vectors. Therefore, we added a pooling layer to compress the input features. First, we reduced each adjacent vector to 1000 dimensions and calculated the average value to reduce the vectors as input to the encoder accordingly. We first organized the entire dataset based on all disciplines, and then based on the new discipline map identifier, we segmented it according to discipline classification to divide the professional dataset.
[0056] S6. Further divide the dataset, using one part for training and the other for testing. Optimize the model using the AdaBoost method to finally form the trained model.
[0057] When optimizing the model using the dataset, the specific steps include the following:
[0058] S6.1, Divide the training dataset into Train1;
[0059] S6.2, Divide the test dataset into Test1;
[0060] S6.3, Model training based on Train1;
[0061] S6.4, based on Test1, test the accuracy of the model and perform optimization.
[0062] S7. Compare the predicted results and optimize the model based on the actual data. Use the Long Short-Term Memory Neural Network (LSTM) for deep optimization to generate a complete set of mentorship relationship datasets.
[0063] The process of deep optimization using a Long Short-Term Memory (LSTM) neural network for cabinets is as follows:
[0064] S7.1 Data Preprocessing: Preprocess the input data, including data cleaning, standardization, and splitting the training set and test set;
[0065] S7.2 Model Structure Design: Design a suitable LSTM model structure based on the specific task requirements and data characteristics; the LSTM model structure can be further optimized by increasing the number of LSTM layers, adjusting the size of the LSTM layers (the dimension of the hidden state), and adding other types of layers (such as fully connected layers, Dropout layers, etc.);
[0066] S7.3 Parameter Tuning: The model is trained using the training set, and its performance is optimized by adjusting different hyperparameters. Common hyperparameters include learning rate, batch size, and number of training iterations. Techniques such as cross-validation can be used to select the optimal combination of hyperparameters.
[0067] S7.4 Regularization: To reduce overfitting, regularization techniques are introduced into the model to help it generalize better to unseen data; such as L1 and L2 regularization, Dropout, etc. These techniques can help the model generalize better to unseen data.
[0068] S7.5 Gradient clipping: To avoid the gradient explosion problem, gradients are clipped during training to limit their maximum value.
[0069] S7.6 Sequence Length Processing: For long sequence data, it is necessary to consider truncating or padding the sequence so that the model can process it better. Truncating is to select the sequence length to retain according to the task requirements, while padding is to use specific padding symbols to expand the sequence to a fixed length.
[0070] S7.7 Batch Normalization: The batch normalization layer accelerates the training process and improves model performance and robustness.
[0071] S7.8 Model Evaluation and Adjustment: Evaluate the optimized model using a test set, and adjust and improve it based on the evaluation results until the expected performance is achieved.
[0072] In summary, assuming the following relationships exist: a) mentors typically publish more papers than students; b) mentors generally have longer academic careers and accumulated advantages in academic performance and resources; c) mentor-student collaboration networks exhibit significantly different performance compared to other general collaboration networks. When determining mentorship relationships, we analyze and statistically study paper and patent data across multiple fields, construct a network representation learner using network representation learning algorithms, and then build a mentorship relationship identifier based on deep neural networks, hybrid principal component analysis (PCA), and pooling layer techniques. Simultaneously, we incorporate attribute network information and construct a new subject graph identifier based on disciplines, classifying and optimizing different disciplines. Using the mentorship relationship identifier, we form an advisor-advisee dataset. Combining the mentorship relationship identifier and the new subject graph identifier, we establish a reliable machine learning prediction model to predict mentor-student relationships. We statistically analyze and model the collaboration network of talents in papers and patents at different times, thereby determining the talent's mentorship relationship, such as who their mentor is and which students they have supervised. This allows for better identification of mentorship relationships within massive amounts of paper data.
[0073] It should be noted that the Deep Neural Network-Hybrid Principal Component Analysis (PCA) algorithm identifies the most prominent aspects of the data and uses these aspects to represent the original data. Specifically, if our dataset is n-dimensional with m data points, we reduce the dimension of these m data points from n-dimensional to n'-dimensional, using this n'-dimensional dataset to represent the original dataset as closely as possible. It is known that reducing data from n-dimensional to n'-dimensional will inevitably result in some loss, but this can be controlled to minimize the loss. The distance from the sample points to this hyperplane should be sufficiently short, or in other words, the projections of the sample points onto this hyperplane should be as separate as possible.
[0074] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, characterized in that: Includes the following steps: S1. Retrieve science and technology papers from the past 30 years from existing think tanks, aggregate the retrieved papers by name, and compile the papers corresponding to each talent node. S2. Construct the collected papers, format the paper data, build a list of collaborators based on each paper, and build a paper data collector. S3. Extract human-centric information from the paper data, establish edge links between collaborators, and construct a paper data representation. S4. Represent the edges and nodes based on the network representation learning algorithm to construct a network representation learner while preserving the subject attributes of the paper. S5. Based on the deep neural network-hybrid principal component analysis (PCA) algorithm and pooling layer technology, a mentorship relationship identifier is built. At the same time, attribute network information is added, and a new subject graph identifier is built based on the subject. Different subjects are classified and optimized. The mentorship relationship identifier is used to form a dataset of advisor-advisee pairs. S6. Further divide the dataset, using one part for training and the other for testing. Optimize the model using the AdaBoost method to finally form the trained model. S7. Compare the predicted results and optimize the model based on the actual data. Use the Long Short-Term Memory Neural Network (LSTM) for deep optimization to generate a complete set of mentorship relationship datasets.
2. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 1, is characterized in that: In step S5, the deep neural network-hybrid principal component analysis algorithm serves two purposes: first, to denoise the data, and second, to reduce the dimensionality for visualization, reducing the original dataset to an n-dimensional dataset with the minimum projection distance.
3. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 2, is characterized in that: When building the mentorship relationship identifier, we first used a deep neural network-hybrid principal component analysis algorithm to reduce the dimensionality of the data and identify the core data domain. The core data domain is one of the following: research institution, academic career age, and name ID. Since some disciplines contain tens of thousands of scholars, it is difficult to ensure the computation time and memory usage when controlling multiple vectors. Therefore, we added a pooling layer to compress the input features. First, we reduced each adjacent vector to 1000 dimensions and calculated the average value to reduce the vectors as input to the encoder accordingly. We first organized the entire dataset based on all disciplines, and then based on the new discipline map identifier, we segmented it according to discipline classification to divide the professional dataset.
4. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 1, is characterized in that: In step S2, the collaborator list is built based on the author's name and institution, and a unique UUID identifier is constructed for each author.
5. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 1, is characterized in that: In step S3, the specific steps for constructing the paper data representation are as follows: S3.1, the edge is represented as: Paper1---->{Person1,Person2....}, Paper2---->{Person3,Person4...}, Paper1 is a paper, and Person1 and Person2 are collaborators associated with the paper; S3.2, Dataset Calculation --- By Year: { - }, thus obtaining the dataset of the years for which you want to perform statistics.
6. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 5, is characterized in that: In step S4, the edge represents the relationship between a person and a paper, and the node represents a paper. Preserving the subject attributes of the paper is to prepare for further optimization training based on subject classification.
7. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 6, is characterized in that: In step S1, the think tank refers to an independent research institution or organization composed of a group of experts and scholars who specialize in researching and providing public policy advice. Think tanks typically express their views and provide advice by publishing research reports, organizing seminars and forums, and participating in policy formulation and consultation. Think tanks cover fields such as science and technology, economics, international relations, energy, and the environment. In the field of science and technology, think tanks typically focus on issues related to science and technology development trends, technological innovation, and science and technology policies, providing scientific and professional advice and guidance for government and corporate decision-making.
8. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 7, is characterized in that: When retrieving paper data from the think tank, to ensure data completeness and accuracy of cooperation trends, data from the current year is removed, and the starting year is denoted as Y. start The end year is Y end .
9. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 8, is characterized in that: In step S6, when the dataset is used to optimize the model, the following steps are specifically included: S6.1, Divide the training dataset into Train1; S6.2, Divide the test dataset into Test1; S6.3, Model training based on Train1; S6.4, based on Test1, test the accuracy of the model and perform optimization.
10. The method for predicting mentorship relationships in a new subject map based on the fusion of science and technology data, as described in claim 9, is characterized in that: The process of deep optimization using a Long Short-Term Memory (LSTM) neural network for cabinets is as follows: S7.1 Data Preprocessing: Preprocess the input data, including data cleaning, standardization, and splitting the training set and test set; S7.2 Model Structure Design: Design a suitable LSTM model structure based on the specific task requirements and data characteristics; the LSTM model structure can be further optimized by increasing the number of LSTM layers, adjusting the size of the LSTM layers, and adding other types of layers; S7.3 Parameter tuning: Train the model using the training set and optimize model performance by adjusting different hyperparameters; S7.4 Regularization: In order to reduce overfitting, regularization techniques are introduced into the model to help the model generalize better to unseen data; S7.5 Gradient clipping: To avoid the gradient explosion problem, the gradient is clipped during training to limit its maximum value; S7.6 Sequence Length Processing: For long sequence data, it is necessary to consider truncating or padding the sequence so that the model can process it better. Truncating is to select the sequence length to retain according to the task requirements, while padding is to use specific padding symbols to expand the sequence to a fixed length. S7.7 Batch Normalization: The batch normalization layer accelerates the training process and improves model performance and robustness. S7.8 Model Evaluation and Adjustment: Evaluate the optimized model using a test set, and adjust and improve it based on the evaluation results until the expected performance is achieved.