Method and system for promoting industry-talent integration and scientific research results transformation based on big data analysis

Abstract

The present invention provides a method and system for promoting the integration of industry and talents and the transformation of scientific research results based on big data analysis, by collecting multi-source heterogeneous data related to scientific research projects, including talent skill files, scientific research results records, industry demand reports and emerging technology trends; using a pre-trained deep learning model to pre-process the collected data, extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained through historical data to identify potential talent flow patterns and non-linear relationships between scientific research results transformation; integrating the cleaned key features into a learning framework, outputting talent matching suggestions and scientific research results transformation possibility assessments to assist results transformation decisions, by identifying the precise match between talent skills and scientific research project requirements, enterprises can more effectively recruit suitable talents, while scientific researchers can find projects that best suit their expertise, thereby improving work efficiency and the quality of scientific research results.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for promoting industry-talent integration and scientific research achievement transformation based on big data analysis. Background Art

[0002] In the process of promoting the integration of industry and talent and the transformation of scientific research results, existing machine learning and data analysis models face significant limitations, which are mainly reflected in their difficulty in accurately depicting and predicting complex and highly dynamic system behaviors. The talent flow pattern and the transformation of scientific research results are often affected by a variety of factors, including personal career preferences, industry trends, technological innovation, policy orientation and even global economic fluctuations. These factors are intertwined to form a nonlinear relationship network. Traditional statistical models and machine learning algorithms may perform poorly in the face of such complex scenarios due to oversimplified assumptions or failure to effectively capture the interactions between high-dimensional features. In addition, when dealing with emerging technology fields, the generalization ability and prediction accuracy of the model are often greatly reduced due to the lack of historical data as training samples, especially in a rapidly changing market environment, the real-time adaptability and update speed of the model become a major challenge. Therefore, developing advanced algorithms that can handle large-scale unstructured data and have strong interpretability and adaptive learning capabilities will be the key to improving the efficiency of scientific research results transformation and accurately matching talent needs. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for promoting the integration of industry and talent and the transformation of scientific research results based on big data analysis, which can at least solve the problem that some knowledge graph construction processes are relatively simple, and only establish corresponding relationships based on artificial archiving of historical data without considering variable and complex factors.

[0004] In a first aspect, an embodiment of the present invention provides a method for promoting industry-talent integration and scientific research achievement transformation based on big data analysis, including:

[0005] Collect multi-source heterogeneous data related to scientific research projects, including talent skill profiles, scientific research results records, industry demand reports and emerging technology trends;

[0006] Preprocess the collected data using a pre-trained deep learning model to extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained using historical data to identify potential talent flow patterns and non-linear relationships in the transformation of scientific research results;

[0007] The cleaned key features are integrated into the learning framework to output talent matching suggestions and scientific research results transformation possibility assessment to assist in results transformation decision-making.

[0008] Furthermore, the collected data is preprocessed using a pre-trained deep learning model to extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained with historical data to identify potential talent flow patterns and non-linear relationships between scientific research results transformation, including:

[0009] Apply autoencoders or variational autoencoders to reduce the dimension of talent skill files and scientific research results records to capture potential nonlinear features;

[0010] Use graph neural networks (GNNs) to analyze the connection between talents and scientific research results, build talent-project network graphs, and identify complex network structure characteristics;

[0011] Through reinforcement learning strategies, we simulate the talent decision-making path and scientific research results diffusion process under different conditions, and generate simulated data to enhance model training.

[0012] Furthermore, the steps of applying autoencoders or variational autoencoders to reduce the dimension of talent skill profiles and scientific research achievement records to capture potential nonlinear features include:

[0013] Initialize the weight parameters of the autoencoder, set the number of neurons in the encoding layer and the decoding layer, and build the model architecture;

[0014] The talent skill files and scientific research results records are converted into vector representations as the input of the autoencoder, and forward and backward propagation are performed to minimize the reconstruction error;

[0015] The hyperparameters are adjusted, and the hyperparameters include at least a learning rate and a regularization coefficient, and the model parameters are optimized through iterative training until convergence or a predetermined stopping criterion is reached.

[0016] Furthermore, the steps of using graph neural network (GNN) to analyze the connection between talents and scientific research results, constructing a talent-project network graph, and identifying complex network structure characteristics include:

[0017] Define the feature representation of nodes and edges, map talents and scientific research results to nodes in the graph, and represent relationships such as talent participation in projects and citations between projects as edges;

[0018] Using graph convolutional networks (GCN) or graph attention networks (GAT) to learn node representations from their neighbors to capture local and global graph structure information;

[0019] During the training phase, contrastive learning or graph anomaly detection techniques are used to identify abnormal patterns and key nodes in the flow of talent and transformation of scientific research results to reveal nonlinear relationships.

[0020] Furthermore, through the reinforcement learning strategy, the talent decision-making path and scientific research results diffusion process under different conditions are simulated, and the simulation data is generated to enhance the model training steps, including:

[0021] Design state space and action space to define the state of talent transfer between different projects and the actions of scientific research results transformation at different stages;

[0022] Use Q-learning or Deep Q-Networks (DQN) algorithms to learn the optimal talent decision-making path and scientific research results transformation strategy through reward mechanisms;

[0023] Implement environmental simulation to simulate market dynamics and policy changes to test and optimize the robustness and adaptability of strategies and enhance the model's ability to cope with the complexity of the real world.

[0024] Furthermore, the Q-learning or Deep Q-Networks (DQN) algorithm is used to learn the optimal talent decision-making path and scientific research results transformation strategy through the reward mechanism, including:

[0025] Initialize Q-table or Q-network, set initial Q value or weight, and build the basic framework of the algorithm;

[0026] Implementing exploration and exploitation strategies, such as ε-greedy, which balances the algorithm’s exploration of unknown states and exploitation of known best actions;

[0027] Update the Q-table or Q-network parameters and continuously optimize the decision-making strategy based on the observed immediate rewards and future expected rewards through the Bellman equation or back-propagation algorithm.

[0028] Furthermore, the steps of updating the Q-table or Q-network parameters and continuously optimizing the decision-making strategy based on the observed immediate rewards and future expected rewards through the Bellman equation or back-propagation algorithm include:

[0029] Under the federated learning framework, a global policy network πθ and a value network Vφ are designed, as well as multiple local policy and value networks, which are distributed among different data holders to protect data privacy and reduce data transmission costs.

[0030] Using meta-learning technology, a well-initialized "meta-strategy" and "meta-value" network is pre-trained. The network can quickly adapt to new scientific research results transformation scenarios and talent flow patterns based on a small amount of sample data, accelerating the learning process.

[0031] In each training cycle, each data holder independently trains the local policy and value network using its local dataset, and then aggregates the updates through Federated Averaging or other federated learning algorithms to synchronize the global network parameters.

[0032] Introducing the Actor-Critic architecture, in which the Actor network is responsible for selecting action a based on the current state s, and the Critic network evaluates the value of the action. The two iteratively optimize each other, and the Actor network adjusts its strategy based on the feedback of the Critic network.

[0033] In the Actor-Critic architecture, the policy parameter θ is updated using the Proximal Policy Optimization (PPO) or Trust Region Policy Optimization (TRPO) algorithm, while the value parameter φ is updated using the Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC) algorithm to handle continuous action spaces and non-stationary environments.

[0034] Adopt experience replay buffer to store experience trajectory τ, combine priority experience replay and multi-step learning technology to improve data utilization and learning efficiency;

[0035] To achieve multi-agent collaborative learning, each agent represents a different stakeholder. They interact through the Actor-Critic architecture in a virtual environment and learn collaborative strategies to optimize talent allocation and the efficiency of transformation of scientific research results.

[0036] In a second aspect, an embodiment of the present invention provides a system for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis, including:

[0037] The collection module is used to collect multi-source heterogeneous data related to scientific research projects, including talent skill files, scientific research results records, industry demand reports and emerging technology trends;

[0038] A training module, for preprocessing the collected data using a pre-trained deep learning model, extracting key features and performing data cleaning, wherein the pre-trained deep learning model is trained using historical data to identify potential talent flow patterns and non-linear relationships in the transformation of scientific research results;

[0039] Integration module, which is used to integrate the cleaned key features into the learning framework, output talent matching suggestions and scientific research results transformation possibility assessment to assist in decision-making on results transformation;

[0040] The training module is specifically used to apply autoencoders or variational autoencoders to reduce the dimension of talent skill files and scientific research results records to capture potential nonlinear characteristics; use graph neural networks (GNNs) to analyze the connection between talents and scientific research results, construct a talent-project network diagram, and identify complex network structure characteristics; through reinforcement learning strategies, simulate the talent decision-making path and scientific research results diffusion process under different conditions, and generate simulated data to enhance model training.

[0041] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory and a processor, characterized in that a computer program is stored in the memory, and the processor is configured to run the computer program to execute the method for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis as described in the first aspect.

[0042] In a fourth aspect, an embodiment of the present invention provides a readable storage medium, in which a computer program is stored. The computer program includes a program code for controlling a process to execute a process. The process includes the method for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis as described in the first aspect.

[0043] The embodiment of the present invention collects multi-source heterogeneous data related to scientific research projects, including talent skill files, scientific research results records, industry demand reports and emerging technology trends; uses a pre-trained deep learning model to pre-process the collected data, extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained through historical data to identify potential talent flow patterns and non-linear relationships between scientific research results transformation; integrates the cleaned key features into a learning framework, outputs talent matching suggestions and scientific research results transformation possibility assessments to assist results transformation decisions, and by identifying the precise match between talent skills and scientific research project requirements, enterprises can more effectively recruit suitable talents, while scientific researchers can find projects that best suit their expertise, thereby improving work efficiency and the quality of scientific research results.

[0044] The beneficial effects of the embodiments of the present invention can refer to the technical effects corresponding to the technical features in the specific implementation methods, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of a method for promoting industry-talent integration and scientific research achievement transformation based on big data analysis provided in an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of a system module for promoting industry-talent integration and scientific research achievement transformation based on big data analysis provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0049] First, if Figure 1 As shown, the embodiment of the present invention provides a method for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis, including:

[0050] Step S1, collecting multi-source heterogeneous data related to scientific research projects, including but not limited to talent skill files, scientific research results records, industry demand reports and emerging technology trends;

[0051] First, data is collected from multiple sources, including but not limited to academic paper databases (such as Web of Science, Scopus), patent databases (such as WIPO, USPTO), professional social networks (such as LinkedIn), recruitment websites (such as Indeed, Glassdoor), industry reports (such as annual reports released by Deloitte and PricewaterhouseCoopers), and government-issued science and technology policies and funding information. Data types cover structured data (such as tables, database records) and unstructured data (such as text, images).

[0052] Example: Suppose we are focusing on the field of biotechnology and have collected all relevant papers published in the past ten years, biotechnology patents applied for worldwide, resumes and work experience of professionals in this field, industry trend analysis reports, and information on policy support and funding investment by various governments for the biotechnology industry.

[0053] Step S2, preprocessing the collected data using a pre-trained deep learning model, extracting key features and performing data cleaning, wherein the pre-trained deep learning model is trained using historical data to identify potential talent flow patterns and non-linear relationships between scientific research achievement transformations;

[0054] Use pre-trained deep learning models, such as BERT (Bidirectional Encoder Representations from Transformers) or its variants, to extract features from text data. The model is pre-trained on a large-scale corpus to capture the deep semantics of the language. Then, the data is cleaned to remove duplicate, irrelevant or low-quality records, fill in missing values, standardize numerical data, and encode categorical data.

[0055] Example: Use the pre-trained BERT model to process the collected academic paper abstracts, patent descriptions, and industry reports to extract key concepts and entities (such as authors, institutions, and technical terms). Then, further refine the features through natural language processing techniques such as named entity recognition (NER) and keyword extraction. Finally, through the data cleaning process, remove redundant and inaccurate information, such as deleting co-authors who have no actual contribution, or correcting misspelled terms.

[0056] In step S3, the cleaned key features are integrated into the learning framework, and talent matching suggestions and evaluation of the possibility of scientific research results transformation are output to assist in decision-making on results transformation.

[0057] Build an integrated learning framework, which can be a combination of rule-based systems, decision trees, random forests, neural networks, or deep learning models. Use the cleaned features as input and train the model to predict the degree of match between talent and project, as well as the possibility of scientific research results being transformed into actual products or services. The training goal of the model is to maximize the prediction accuracy and the generalization ability of the model.

[0058] Example: Use a deep neural network with multiple hidden layers to capture the complex relationship between talent skills, scientific research results, industry needs, and policy environment. Through supervised learning, the network is trained to predict the probability of successful transformation of given talents and scientific research results in a specific future time. The output of the model can be a probability score. The higher the score, the higher the match between the talent and the project, or the greater the possibility of scientific research results being transformed into products.

[0059] By identifying the precise match between talent skills and scientific research project requirements, the embodiments of the present invention can help enterprises recruit suitable talents more effectively, and scientific researchers can find projects that best suit their expertise, thereby improving work efficiency and the quality of scientific research results. Secondly, accurately evaluating the commercialization potential of scientific research results will help investors and enterprises make more informed investment decisions, accelerate the transformation of scientific and technological achievements to the market, shorten product development cycles, and improve market competitiveness. At the same time, government departments can formulate science and technology policies and funding allocation plans more accurately based on data analysis to ensure the effective use of public resources, while promoting the close integration of scientific research and industry, and promoting the implementation of the national science and technology innovation strategy. Finally, by identifying the potential connections between different fields, interdisciplinary research and cooperation can be promoted, innovative thinking can be stimulated, breakthrough scientific research results can be produced, and the overall progress of science and technology can be promoted.

[0060] In another embodiment provided by the present invention, a pre-trained deep learning model is used to pre-process the collected data, extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained with historical data to identify potential talent flow patterns and non-linear relationships between scientific research results transformation, including:

[0061] Apply autoencoders or variational autoencoders to reduce the dimension of talent skill files and scientific research results records to capture potential nonlinear features;

[0062] An autoencoder (AE) is an unsupervised learning model used to learn efficient encodings of data. A variational autoencoder (VAE) is a variant of an autoencoder that not only learns encodings but also learns the underlying distribution of data. In our scenario, autoencoders or variational autoencoders are used to process high-dimensional data in talent skill profiles and scientific research results records to capture nonlinear features and convert them into low-dimensional representations.

[0063] Example: Consider that talent skill profiles contain a lot of descriptive text, such as research interests, abstracts of published papers, patent descriptions, etc. Applying variational autoencoders (VAE), we can convert these texts into compact vector representations that can capture the intrinsic correlations between skills and potential skill combination patterns.

[0064] Use graph neural networks (GNNs) to analyze the connection between talents and scientific research results, build talent-project network graphs, and identify complex network structure characteristics;

[0065] Graph Neural Networks (GNNs) are a type of neural network model specifically designed to process graph-structured data. They can learn the representation of nodes (i.e., talents and scientific research projects) and capture the complex relationships between nodes. In this step, we construct a talent-project network graph, where nodes represent talents or scientific research projects, and edges represent the relationship between talents and scientific research projects.

[0066] Example: In a talent-project network diagram, a talent node may be connected to multiple project nodes, indicating that the talent has participated in multiple scientific research projects. Through GNN, we can analyze the strength of the relationship between talents and projects, identify which talents are influential in specific fields, and which projects are most likely to attract specific types of talents. This helps to discover talent flow patterns and key nodes for the transformation of scientific research results.

[0067] Through reinforcement learning strategies, we simulate the talent decision-making path and scientific research results diffusion process under different conditions, and generate simulated data to enhance model training.

[0068] Reinforcement learning (RL) is a machine learning method that enables an intelligent agent to learn through trial and error in an environment to maximize long-term rewards. In our scenario, reinforcement learning can be used to simulate the career decision-making process of talents and the diffusion process of scientific research results, thereby generating additional training data to help the model better understand the dynamic characteristics of talent flow and scientific research results transformation.

[0069] Example: We can set up a reinforcement learning environment where the "agent" represents talent and the "environment" consists of scientific research projects, industry needs, and technological trends. The agent receives rewards (such as career advancement, project success rate, etc.) through a series of decisions (such as choosing which project to join, whether to commercialize scientific research results, etc.). By simulating multiple possible decision paths, we not only enhance the training data set of the model, but also gain insight into how talent and scientific research results interact under different conditions, thereby guiding the actual decision-making process.

[0070] The embodiments of the present invention can identify and compress nonlinear relationships in data through autoencoders and variational autoencoders, which is crucial for understanding the complex relationship between talent skills and scientific research results; secondly, GNN can reveal the hidden network structure between talents and projects, which helps to identify key talents and projects, as well as important paths that affect the transformation of scientific research results. Finally, the generation of simulated data under the reinforcement learning strategy can enhance the model's adaptability to dynamic changes, making it more robust and able to predict the conversion rate of scientific research results under different conditions. It can more deeply understand the complex mechanism of talent flow patterns and scientific research results transformation, thereby providing strong data-driven decision support for promoting the integration of industry and talents and the transformation of scientific research results.

[0071] In another embodiment provided by the present invention, the steps of applying an autoencoder or a variational autoencoder to reduce the dimension of the talent skill files and scientific research achievement records to capture potential nonlinear features include:

[0072] Initialize the weight parameters of the autoencoder, set the number of neurons in the encoding layer and the decoding layer, and build the model architecture;

[0073] When building an autoencoder model, you first need to determine the architecture of the model, including the number of neurons in the input layer, hidden layer (encoding layer and decoding layer), and output layer. Suppose we are dealing with talent skill files and scientific research results records, and these data are converted into numerical vectors. We decided to use a three-layer autoencoder, in which the number of neurons in the encoding layer is less than that in the input layer and output layer, in order to compress the data and extract key features.

[0074] Example: Imagine that each talent profile we are processing contains 100 features (such as various skill scores, education background codes, past project participation, etc.), and the scientific research record contains 200 features (such as technical indicators, citation counts, patent category codes, etc.). We decided to set the number of neurons in the encoding layer to 32, which means that we will try to compress the original 100- and 200-dimensional data into a 32-dimensional feature space. When the model is initialized, we randomly assign each weight a small value, close to zero but not completely zero, to break the symmetry and ensure that each neuron can independently learn different aspects of the data in the early stages of training.

[0075] The talent skill files and scientific research results records are converted into vector representations as the input of the autoencoder, and forward and backward propagation are performed to minimize the reconstruction error;

[0076] Once the model is initialized, the next step is to convert the talent skill profile and scientific research results records into vector representations and input these vectors into the autoencoder. The encoder part of the autoencoder compresses the data into a low-dimensional representation, while the decoder part attempts to reconstruct the original data based on this low-dimensional representation. By comparing the difference between the reconstructed data and the original data, we can calculate the reconstruction error and use the back-propagation algorithm to update the model weights to reduce this error.

[0077] Example: In the specific implementation, we first convert the talent skill files and scientific research results records into numerical vectors through preprocessing, and then input these vectors into the autoencoder in batches. The encoder part of the autoencoder compresses the 100- or 200-dimensional vector into a 32-dimensional feature vector, while the decoder part tries to convert this 32-dimensional feature vector back to a 100- or 200-dimensional vector. We use the mean square error (MSE) as the loss function, which measures the average squared difference between the reconstructed vector and the original vector. Through multiple iterations, the autoencoder gradually adjusts its weights to minimize the reconstruction error, thereby learning the key features in the talent skills and scientific research results data.

[0078] Hyperparameters such as learning rate and regularization coefficient are adjusted to optimize model parameters through iterative training until convergence or a predetermined stopping criterion is reached.

[0079] When training an autoencoder, there are several hyperparameters that need to be adjusted, including the learning rate, regularization coefficient, training epochs, and batch size. The learning rate determines how fast the weights are updated, the regularization coefficient is used to prevent the model from overfitting, the training epoch determines the number of complete cycles of model training, and the batch size controls the amount of data based on which each weight update is based.

[0080] Example: When training the autoencoder, we chose a relatively small learning rate (e.g. 0.001) to ensure that the weight updates are stable and avoid oscillations around the optimal point of the loss function. We also set the L2 regularization coefficient (e.g. 0.0001) to inhibit the weights from being too large and prevent the model from overfitting the training data. We decided to train the model for 100 epochs to ensure that the model has enough time to learn the patterns in the data. For each epoch of training, we use a batch size of 256, which means that at each weight update, we consider the data of 256 archives or records. As training progresses, we monitor the reconstruction error on the validation set. If we find that the error is no longer significantly reduced, or starts to increase, this may mean that the model has learned enough information and training can be stopped at this time to avoid overfitting.

[0081] Through the above steps, the autoencoder model can effectively extract key features from talent skill files and scientific research results records, while reducing the dimension of the data, which is crucial for subsequent analysis and decision support. The feature vector after dimensionality reduction not only retains the main information of the original data, but also reduces the computational complexity, making further data mining and model training more efficient.

[0082] In another embodiment provided by the present invention, the steps of using a graph neural network (GNN) to analyze the connection between talents and scientific research results, constructing a talent-project network graph, and identifying complex network structure features include:

[0083] Define the feature representation of nodes and edges, map talents and scientific research results to nodes in the graph, and represent relationships such as talent participation in projects and citations between projects as edges;

[0084] When constructing a talent-project network graph, we first need to define the feature representation of nodes and edges. Nodes can be talents or scientific research projects, while edges represent the participation relationship between talents and projects, as well as the citation relationship between projects. Node features can include the talent's skills, experience, educational background, types of projects they have participated in, as well as the project's field, scale, achievements, number of citations, etc.

[0085] Example: Suppose we have a talent node A, whose features may include: having a Ph.D. in computer science, having published many papers in the field of artificial intelligence, and being the chief engineer in many software development projects. Another node may be a scientific research project B, whose features may include: belonging to the field of machine learning, being funded by a national foundation, and having more than 50 citation records. The features of the edge can be the role of the talent in the project (such as the main researcher, technical support), the direct citation relationship between projects, or the indirect influence.

[0086] Leveraging neighborhood aggregation operations, such as Graph Convolutional Networks (GCNs) or Graph Attention Networks (GATs), node representations are learned from the nodes’ neighbors to capture local and global graph structure information.

[0087] Graph Neural Networks (GNNs) learn node representations through neighborhood aggregation operations. Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) are two commonly used GNN models. GCNs update node representations by aggregating the direct neighbor information of nodes, while GATs give different weights to different neighbors through the attention mechanism, thereby aggregating information more finely.

[0088] Example: In our network diagram, talent node A may be connected to multiple project nodes, indicating that he is involved in these projects. When applying GCN, the new representation of node A will take into account the characteristics of all project nodes connected to it, as well as the characteristics of node A itself. When using GAT, the new representation of node A will be based on the weighted sum of the characteristics of each project node. These weights reflect the importance of the project to node A, which may be calculated based on factors such as the relevance of the project to node A and the influence of the project.

[0089] During the training phase, contrastive learning or graph anomaly detection techniques are used to identify abnormal patterns and key nodes in the flow of talent and transformation of scientific research results to reveal nonlinear relationships.

[0090] When training GNN models, we can use contrastive learning or graph anomaly detection techniques to identify key nodes and abnormal patterns in talent flow and scientific research results transformation. Contrastive learning learns the relationship between nodes by comparing the representations of similar and dissimilar nodes, while graph anomaly detection looks for nodes that behave differently from most nodes, which may represent important turning points or abnormal behaviors.

[0091] Example: Through contrastive learning, we can train the model to identify talent nodes that are similar in skills, experience, and project types, as well as project nodes that are similar in fields, influence, and citation patterns. If a node (talent or project) is found to deviate significantly from the similarity of other nodes, this may indicate that the node has a unique behavior pattern, such as a talent suddenly turning to a completely new research field, or a project suddenly gaining unprecedented attention. Such nodes may be key nodes in talent mobility or scientific research results transformation, and it is worth further studying the reasons and impacts behind them.

[0092] In an embodiment of the present invention, GNN can use graph structure information to learn the complex relationship between talents and projects, including direct and indirect connections, which is very important for understanding the dynamics of the scientific research ecosystem; through comparative learning and graph anomaly detection, we can discover abnormal behaviors and important turning points in the scientific research ecosystem. This information is crucial for predicting talent flow trends and the potential for transformation of scientific research results. Based on the node representation learned by GNN, we can recommend the most matching scientific research projects for talents, or recommend the most suitable talents for projects, and provide decision makers with a basis for scientific research resource allocation and policy formulation. Through these steps, GNN can have a deeper understanding of the interaction between talents and projects in the scientific research ecosystem, and provide a powerful data analysis tool for promoting the integration of industry and talents and the transformation of scientific research results.

[0093] In another embodiment provided by the present invention, the steps of simulating the talent decision-making path and the diffusion process of scientific research results under different conditions by using a reinforcement learning strategy, and generating simulation data to enhance the training of the model include:

[0094] Design state space and action space to define the state of talent transfer between different projects and the actions of scientific research results transformation at different stages;

[0095] In the framework of reinforcement learning (RL), the first thing that needs to be clarified is the state space and action space. For the talent decision-making path and the process of scientific research results diffusion, the specific design of these two concepts is as follows:

[0096] State space: includes factors such as the project the talent is in, the stage of the project, the talent's skill level, existing scientific research results, market conditions, policy environment, etc. For example, a state may be described as "Talent X is in the middle of project Y, the project belongs to the field of AI, the talent has deep learning expertise, the market demand for AI is strong, and the government has relevant support policies."

[0097] Action space: Defines the possible actions that talents and scientific research results can take. For talents, actions may include choosing to join new projects, improving personal skills, publishing research results, etc.; for scientific research results, it may involve applying for patents, commercializing, cooperating with other projects, etc. For example, "Talent X decided to move from project Y to project Z focusing on autonomous driving" or "Scientific research result A is applied to industrial production."

[0098] Using Q-learning or Deep Q-Networks (DQN) algorithms, the optimal talent decision path is learned through a reward mechanism. Q-learning: It is a tabular learning algorithm that attempts to learn a Q-table containing the expected long-term rewards after performing a specific action in a given state. Talents or results make decisions in different simulated states, and the Q-table is continuously updated through reward feedback to eventually find the optimal strategy.

[0099] Deep Q-Networks (DQN): is an extension of Q-learning, especially suitable for situations where the state space is very large, such as complex scenarios such as talent decision-making paths and scientific research results diffusion. DQN uses deep neural networks to approximate the Q function, allowing the processing of high-dimensional state spaces, and stabilizes the learning process through experience replay and target network techniques.

[0100] Implement environmental simulation to simulate market dynamics and policy changes to test and optimize the robustness and adaptability of strategies and enhance the model's ability to cope with the complexity of the real world.

[0101] In order to test and optimize strategies, it is necessary to create a simulation environment that can reflect the dynamics of the real world. This environment should be able to simulate market changes, policy adjustments, technological advances and other factors, so as to observe the performance of talent decision-making paths and scientific research results diffusion under different conditions:

[0102] Market dynamics: including demand fluctuations, changes in competitor strategies, economic cycles, etc.

[0103] Policy changes: such as changes in government policies on investment in specific areas, tax incentives, R&D subsidies, etc.

[0104] Technological progress: The emergence of new technologies may change the demand for talent and the application direction of scientific research results.

[0105] Through repeated iterations and learning in a simulated environment, the reinforcement learning algorithm can gradually converge to an optimal strategy that can adapt to a variety of environmental changes. This not only enhances the robustness of the model, but also improves its ability to cope with the complexity of the real world.

[0106] Through reinforcement learning strategies, especially after designing appropriate state and action spaces and adopting algorithms such as Q-learning or DQN, we can effectively simulate talent decision paths and the diffusion process of scientific research results. Combined with environmental simulation, this method can not only help us understand how talents and scientific research results evolve in a dynamic environment, but also provide decision makers with data-based insights to optimize talent allocation and scientific research results transformation strategies. This method has great potential in both theoretical research and practical applications.

[0107] In another embodiment provided by the present invention, the steps of using Q-learning or Deep Q-Networks (DQN) algorithm to learn the optimal talent decision path and scientific research achievement transformation strategy through a reward mechanism include:

[0108] Initialize Q-table or Q-network, set initial Q value or weight, and build the basic framework of the algorithm;

[0109] Q-table: In Q-learning, a table is initialized to store the Q value of each state-action pair. The initial Q value is usually set to 0 or a small positive number to ensure that the algorithm is open to all actions at the beginning.

[0110] Q-network: In DQN, a deep neural network is used instead of a Q-table. The input of the network is the state, and the output is the Q value estimate for each possible action. The weights of the network need to be randomly initialized to establish the basic framework of the algorithm.

[0111] Execute exploration and exploitation strategies, such as the ε-greedy equilibrium algorithm's exploration of unknown states and exploitation of known best actions;

[0112] ε-greedy strategy: This is a common method to balance exploration and exploitation. In each decision step, the algorithm randomly selects an action for exploration with probability ε, and selects the best action currently known for exploitation with probability 1-ε. As learning proceeds, ε can be gradually reduced, reducing the proportion of exploration and increasing the proportion of exploiting learned knowledge.

[0113] Update the Q-table or Q-network parameters and continuously optimize the decision-making strategy based on the observed immediate rewards and future expected rewards through the Bellman equation or back-propagation algorithm.

[0114] Q-learning update: Use the Bellman equation to update the values ​​in the Q-table. When a state s is observed, an action a is taken, an immediate reward r is obtained, and the next state s' is entered, the new Q value calculation formula is:

[0115]

[0116] Where α is the learning rate and γ is the discount factor, which represents the degree of discount of future rewards.

[0117] DQN Update: DQN uses the back-propagation algorithm to update the weights of the Q-network. It stabilizes the learning process by building a target network whose weights are periodically updated to the weights of the current network. The loss function is calculated as the mean square error between the Q value predicted by the current network and the target Q value predicted by the target network. The target Q value is calculated similarly:

[0118]

[0119] The network weights are then updated by minimizing the difference between the predicted Q-values ​​and the target Q-values.

[0120] The entire learning process requires continuous execution of actions in the environment, collection of feedback, and updating of the Q-table or Q-network until a stable strategy is reached, that is, the decision path for the transformation of talent and scientific research results reaches an optimal or near-optimal state. In practical applications, advanced techniques such as multi-step returns and priority experience replay may also need to be considered to improve learning efficiency and strategy quality.

[0121] In an embodiment of the present invention, the steps of updating the Q-table or Q-network parameters and continuously optimizing the decision strategy based on the observed immediate rewards and future expected rewards through the Bellman equation or the back propagation algorithm include:

[0122] Under the federated learning framework, a global policy network πθ and value network Vφ are designed, as well as multiple local policy and value networks, which are distributed among different data holders (such as enterprises, research institutions or individuals) to protect data privacy and reduce data transmission costs;

[0123] Using meta-learning technology, a well-initialized "meta-strategy" and "meta-value" network is pre-trained. The network can quickly adapt to new scientific research results transformation scenarios and talent flow patterns based on a small amount of sample data, accelerating the learning process.

[0124] In each training cycle, each data holder independently trains the local policy and value network using its local dataset, and then aggregates the updates through Federated Averaging or other federated learning algorithms to synchronize the global network parameters.

[0125] Introducing the Actor-Critic architecture, in which the Actor network is responsible for selecting action a based on the current state s, and the Critic network evaluates the value of the action. The two iteratively optimize each other, and the Actor network adjusts its strategy based on the feedback of the Critic network.

[0126] In the Actor-Critic architecture, the policy parameter θ is updated using the Proximal Policy Optimization (PPO) or Trust Region Policy Optimization (TRPO) algorithm, while the value parameter φ is updated using the Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC) algorithm to handle continuous action spaces and non-stationary environments.

[0127] Adopt experience replay buffer to store experience trajectory τ, combine priority experience replay and multi-step learning technology to improve data utilization and learning efficiency;

[0128] To achieve multi-agent collaborative learning, each agent represents a different stakeholder (such as scientists, entrepreneurs, and investors). They interact through the Actor-Critic architecture in a virtual environment and learn collaborative strategies to optimize talent allocation and the efficiency of transforming scientific research results.

[0129] In an embodiment of the present invention,

[0130] Global and local networks in a federated learning framework

[0131] Federated learning framework: Design a global policy network πθ and a value network Vφ to aggregate local learning results from different data holders (such as enterprises, research institutions or individuals). Each data holder has its own local policy and value network to process and learn its private data set, which can protect data privacy while reducing the cost and delay of intermediate transmission of large-scale data sets.

[0132] Meta-learning techniques accelerate the learning process

[0133] Meta-learning: Pre-train a "meta-strategy" and "meta-value" network that has been trained on a large number of diverse scenarios and can quickly adapt to new environments and converge quickly even on a small amount of sample data. This enables the model to more quickly understand and optimize talent decision paths and scientific research results transformation strategies, especially in the face of emerging technology fields or market changes.

[0134] Federated Averaging and Other Federated Learning Algorithms

[0135] Federated Averaging: After each training cycle, each data holder uploads its local strategy and value network weight update to the central server. The central server calculates the average of these updates and sends the updated global network parameters to all participants to achieve synchronous update of the global network.

[0136] Strategy and value optimization under the Actor-Critic framework

[0137] Actor-Critic architecture: The Actor network is responsible for selecting an action a based on the current state s, while the Critic network evaluates the value of the action. The Actor network adjusts its strategy based on the feedback from the Critic network, and the two collaborate through iterative optimization to find the optimal strategy.

[0138] Using advanced reinforcement learning algorithms

[0139] Policy Optimization Algorithm: Use the Proximal Policy Optimization (PPO) or Trust Region Policy Optimization (TRPO) algorithms to update the policy parameters θ, which can ensure the stability and security of policy updates.

[0140] Value update algorithm: The value parameter φ is updated using the Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC) algorithm. These algorithms are good at handling continuous action spaces and non-stationary environments, improving the generalization ability and adaptability of the strategy.

[0141] Experience replay and multi-step learning

[0142] Experience replay buffer: Stores the experience trajectory τ, combines priority experience replay and multi-step learning technology to ensure that the algorithm can learn from historical experience and improve data utilization and learning efficiency.

[0143] Multi-agent collaborative learning

[0144] Multi-agent system: In a virtual environment, multiple agents represent different stakeholders (such as scientists, entrepreneurs, and investors). They interact through the Actor-Critic architecture and learn collaborative strategies to optimize talent allocation and the conversion efficiency of scientific research results, thereby optimizing the overall system.

[0145] Through the above steps, this method can not only use federated learning and meta-learning technology to protect data privacy and reduce data transmission costs, but also achieve efficient, robust and adaptable optimization of talent decision-making paths and scientific research results transformation strategies through advanced reinforcement learning algorithms and multi-agent collaborative learning. This not only improves the quality of decision-making, but also promotes the rapid transformation of scientific research results to the market, and achieves in-depth optimization of the integration of production and talent.

[0146] According to another aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0147] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for promoting industry-talent integration and scientific research achievement transformation based on big data analysis, characterized in that: include: Collect multi-source heterogeneous data related to scientific research projects, including talent skill profiles, scientific research results records, industry demand reports and emerging technology trends; Preprocess the collected data using a pre-trained deep learning model to extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained using historical data to identify potential talent flow patterns and non-linear relationships in the transformation of scientific research results; Integrate the cleaned key features into the learning framework, output talent matching suggestions and scientific research results transformation possibility assessment to assist in decision-making on results transformation; the pre-trained deep learning model is used to pre-process the collected data, extract key features and perform data cleaning, wherein the pre-trained deep learning model is trained with historical data to identify potential talent flow patterns and non-linear relationships between scientific research results transformation, including: Apply autoencoders or variational autoencoders to reduce the dimension of talent skill files and scientific research results records to capture potential nonlinear features; Use graph neural networks (GNNs) to analyze the connection between talents and scientific research results, build talent-project network graphs, and identify complex network structure characteristics; Through reinforcement learning strategies, we simulate the talent decision-making paths and scientific research results diffusion process under different conditions, and generate simulated data to enhance model training; Among them, the steps of simulating the talent decision-making path and scientific research results diffusion process under different conditions through reinforcement learning strategies, and generating simulated data to enhance the training of the model include: Design state space and action space to define the state of talent transfer between different projects and the actions of scientific research results transformation at different stages; Use Q-learning or Deep Q-Networks (DQN) algorithms to learn the optimal talent decision-making path and scientific research results transformation strategy through reward mechanisms; Implement environmental simulation to simulate market dynamics and policy changes to test and optimize the robustness and adaptability of strategies and enhance the model's ability to cope with the complexity of the real world; Among them, the steps of using Q-learning or Deep Q-Networks (DQN) algorithm to learn the optimal talent decision-making path and scientific research results transformation strategy through the reward mechanism include: Initialize Q-table or Q-network, set initial Q value or weight, and build the basic framework of the algorithm; Execute exploration and exploitation strategies, such as the ε-greedy equilibrium algorithm's exploration of unknown states and exploitation of known best actions; Update the Q-table or Q-network parameters, and continuously optimize the decision-making strategy based on the observed immediate rewards and future expected rewards through the Bellman equation or back-propagation algorithm; Among them, the steps of updating the Q-table or Q-network parameters and continuously optimizing the decision-making strategy based on the observed immediate rewards and future expected rewards through the Bellman equation or back propagation algorithm include: Under the federated learning framework, a global policy network πθ and a value network Vφ are designed, as well as multiple local policy and value networks, which are distributed among different data holders to protect data privacy and reduce data transmission costs. Using meta-learning technology, a well-initialized "meta-strategy" and "meta-value" network is pre-trained. The network can quickly adapt to new scientific research results transformation scenarios and talent flow patterns based on a small amount of sample data, accelerating the learning process; In each training cycle, each data holder independently trains the local policy and value network using its local dataset, and then aggregates the updates through Federated Averaging or other federated learning algorithms to synchronize the global network parameters. Introducing the Actor-Critic architecture, in which the Actor network is responsible for selecting action a based on the current state s, and the Critic network evaluates the value of the action. The two iteratively optimize each other, and the Actor network adjusts its strategy based on the feedback of the Critic network. In the Actor-Critic architecture, the policy parameter θ is updated using the Proximal Policy Optimization (PPO) or Trust Region Policy Optimization (TRPO) algorithm, while the value parameter φ is updated using the Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC) algorithm to handle continuous action spaces and non-stationary environments. Adopt experience replay buffer to store experience trajectory τ, combine priority experience replay and multi-step learning technology to improve data utilization and learning efficiency; To achieve multi-agent collaborative learning, each agent represents a different stakeholder. They interact through the Actor-Critic architecture in a virtual environment and learn collaborative strategies to optimize talent allocation and the efficiency of transformation of scientific research results.

2. The method for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis as claimed in claim 1, wherein the step of applying an autoencoder or a variational autoencoder to reduce the dimension of talent skill files and scientific research results records to capture potential nonlinear features comprises: Initialize the weight parameters of the autoencoder, set the number of neurons in the encoding layer and the decoding layer, and build the model architecture; The talent skill files and scientific research results records are converted into vector representations as the input of the autoencoder, and forward and backward propagation are performed to minimize the reconstruction error; The hyperparameters are adjusted, and the hyperparameters include at least a learning rate and a regularization coefficient, and the model parameters are optimized through iterative training until convergence or a predetermined stopping criterion is reached.

3. The method for promoting the integration of industry and talents and the transformation of scientific research results based on big data analysis as claimed in claim 2, using a graph neural network (GNN) to analyze the connection between talents and scientific research results, construct a talent-project network diagram, and identify the complex network structure characteristics, including: Define the feature representation of nodes and edges, map talents and scientific research results to nodes in the graph, and represent relationships such as talent participation in projects and citations between projects as edges; Using graph convolutional networks (GCN) or graph attention networks (GAT) to learn node representations from their neighbors to capture local and global graph structure information; During the training phase, contrastive learning or graph anomaly detection techniques are used to identify abnormal patterns and key nodes in the flow of talent and transformation of scientific research results to reveal nonlinear relationships.

4. A system for promoting the integration of industry and talents and the transformation of scientific research results based on big data analysis, which is used to implement the method for promoting the integration of industry and talents and the transformation of scientific research results based on big data analysis as described in any one of claims 1-3, characterized in that: include: The collection module is used to collect multi-source heterogeneous data related to scientific research projects, including talent skill files, scientific research results records, industry demand reports and emerging technology trends; A training module, for preprocessing the collected data using a pre-trained deep learning model, extracting key features and performing data cleaning, wherein the pre-trained deep learning model is trained using historical data to identify potential talent flow patterns and non-linear relationships in the transformation of scientific research results; Integration module, which is used to integrate the cleaned key features into the learning framework, output talent matching suggestions and scientific research results transformation possibility assessment to assist in decision-making on results transformation; The training module is specifically used to apply autoencoders or variational autoencoders to reduce the dimension of talent skill files and scientific research results records to capture potential nonlinear features; use graph neural networks (GNNs) to analyze the connection between talents and scientific research results, construct a talent-project network diagram, and identify complex network structure features; Through reinforcement learning strategies, we simulate the talent decision-making path and scientific research results diffusion process under different conditions, and generate simulated data to enhance model training.

5. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to run the computer program to execute the method for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis as described in any one of claims 1 to 3.

6. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes a program code for controlling a process to execute a process, and the process includes a method for promoting the integration of production and talents and the transformation of scientific research results based on big data analysis according to any one of claims 1 to 3.

Images