Technology transfer opportunity prediction method and system based on science-technology knowledge hysteresis

Abstract

The invention discloses a technology transfer opportunity prediction method and system based on science-technology knowledge hysteresis. The method comprises the following steps: uniformly mapping scientific papers and patent documents to subdivision technology tags by utilizing a large language model based on fine tuning of a standard classification system; respectively constructing a time sequence scientific knowledge network and a technical knowledge network; calculating a knowledge lag value between each pair of technologies by using a time-lag cross-correlation algorithm; constructing a double-flow graph attention network model, respectively learning a historical transfer structure and a scientific lag attribute through a structure flow and an attribute flow, and fusing features by using a gating mechanism; and finally, carrying out multi-step prediction by adopting a sliding window strategy. According to the method, the problems of cross-modal data semantic alignment, scientific pilot quantization and time sequence attribute fusion are effectively solved, and the prospective accurate prediction of the technology transfer opportunity is realized.

Description

Technical Field

[0001] This invention belongs to the field of data mining and artificial intelligence application technology, specifically relating to a method and system for predicting technology transfer opportunities by analyzing the time lag relationship between scientific and technological knowledge using natural language processing and graph neural network technology. Background Technology

[0002] Against the backdrop of increasingly complex global technological innovation and intensified industrial competition, technology transfer has become a crucial bridge connecting science and industry. Accurately predicting the timing and direction of technology transfer is vital for corporate R&D decisions. Existing prediction methods are mostly based on patent data mining, network link prediction, or graph neural networks, but they have limitations in multi-source knowledge fusion and temporal lag modeling. While related research has attempted different approaches, shortcomings remain. For example: The paper "A transfer learning based patent transferprediction model: Discussion about the condition of China and US", published on January 1, 2025, provides a patent transfer prediction model based on transfer learning, but it is essentially a static model and cannot achieve real-time or forward-looking prediction. The paper "Unveiling intrinsic interactions of science and technology in artificial intelligence using a network portrait divergence approach", published on February 10, 2025, presents a study that uses a network portrait divergence approach. However, when constructing a science-technology knowledge network, it mainly relies on keywords in the title and abstract, failing to fully utilize the deep semantic information of the whole text and resulting in an incomplete representation. The paper "Neural Common Neighbor with Completion for Link Prediction", published on February 2, 2023, provides advanced link prediction models such as Neural Common Neighbor (NCN). Although it performs well on general graph data, it has not been verified and optimized for the characteristics of knowledge networks in the field of technology transfer. Chinese patent application CN120707341A, published on September 26, 2025, discloses a technology transfer supply chain management method using a dynamic matching algorithm. The proposed dynamic matching algorithm focuses on path optimization and management, but it is still insufficient in handling the dynamic temporal correlation of the transformation of scientific knowledge into technological applications.

[0003] In summary, existing technologies have the following three main shortcomings: First, the predicted technologies are too narrow. Research in the field of technology transfer mainly focuses on predicting the intrinsic value of a single technology, whether a single technology will be transferred, and when such a transfer will occur. This focus makes it difficult for technology transfer prediction to fully capture the increasingly diverse and complex market demands, and it also ignores the huge potential brought about by the synergistic evolution of different technology fields.

[0004] Secondly, in the process of constructing knowledge networks, keyword extraction often contains a large amount of non-technical noise, which still requires time-consuming manual screening. In addition, networks built based on the co-occurrence of knowledge entities tend to suffer from high sparsity, thereby affecting the robustness and accuracy of technology opportunity prediction.

[0005] Third, existing link prediction algorithms based on edge attributes fail to effectively utilize the external attributes between unconnected node pairs. How to achieve end-to-end native fusion of such external attributes and network temporal structure features in complex time scenarios remains an unsolved problem. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for predicting technology transfer opportunities based on the lag of scientific and technological knowledge. By quantifying the time lag relationship between scientific and technological knowledge and combining a dual-stream attention mechanism with a multi-step prediction strategy, it achieves high-precision forward-looking prediction of technology transfer opportunities.

[0007] According to one aspect of the present invention, a method and system for predicting technology transfer opportunities based on the lag of scientific and technological knowledge are provided, comprising: Using a large language model fine-tuned based on a standard classification system, the text content of scientific papers and patent documents is mapped to a unified set of subdivided technical tags. Based on the mapping results, a time-series science knowledge network and a time-series technology knowledge network are constructed, with subdivision technologies as nodes and co-occurrence relationships as edges. For each pair of subdivided technologies, the edge weight time series sequence of the technology in the scientific knowledge network and the technical knowledge network is extracted, and the time offset corresponding to the correlation peak between the two sequences is calculated using the time delay cross-correlation algorithm, which is used as the knowledge lag value of the technology pair. A dual-flow graph attention network model containing structural flow and attribute flow is constructed. The structural flow takes an explicit co-occurrence network constructed from historical patent transfer records as input, and the attribute flow takes a implicit attribute network constructed from knowledge lag values ​​as input. The features of the dual-flow outputs are integrated through a gating fusion mechanism to generate a comprehensive node embedding representation. A sliding window-based training strategy is adopted to predict the probability of technology transfer between subdivided technology nodes after one or more time steps in the future, using the generated node embedding representation.

[0008] As a further technical solution, the standard classification system is the ACM computational classification system; the fine-tuning process of the large language model is as follows: using the title and abstract of the scientific paper as input text, and using one or more subdivided technical tags corresponding to them in the ACM computational classification system as supervision tags, fine-tuning training is performed.

[0009] As a further technical solution, both the time-series scientific knowledge network and the time-series technical knowledge network are undirected weighted networks, where the edge weights represent the frequency with which two sub-technical tags appear together in the same scientific paper or the same patent document within a specific time slice.

[0010] As a further technical solution, the calculation of knowledge lag value using the time-delay cross-correlation algorithm includes: For each pair of sub-technologies, a sliding time window is set, and the cross-correlation function between their scientific knowledge sequence and technological knowledge sequence is calculated within the window; The time difference between the scientific knowledge sequence and the technical knowledge sequence when the cross-correlation function reaches its maximum value is determined as the knowledge lag value at the center of the window. Repeat the above calculation for all time windows to form a time series of knowledge lag values, and construct a dynamic implicit attribute network with nodes representing subdivision techniques and edge weights representing knowledge lag values.

[0011] As a further technical solution, the working method of the dual-flow graph attention network model includes: The structure flow learns the topological features of the explicit co-occurrence network through a graph attention network and outputs a structure-aware node representation. The attribute stream is based on the structure-aware node representation and runs a graph attention network on the latent attribute network to learn the latent association features represented by knowledge lag, and outputs attribute-enhanced node representations. The gated fusion mechanism uses a learnable gating function to dynamically weight and sum the structure-aware node representation and the attribute-enhanced node representation to generate the final fused node representation.

[0012] As a further technical solution, the sliding window-based training strategy is as follows: a training sample set is constructed, wherein the input of each sample is the network state and node representation of multiple consecutive historical time slices, and the output is the network connection state of the k-th future time slice; the model is trained by minimizing the loss function between the predicted connection state and the true state.

[0013] According to one aspect of the present invention, a technology transfer opportunity prediction system based on the lag of scientific and technological knowledge is provided, comprising: The multi-label classification module is used to map the text content of scientific papers and patent documents to a unified set of subdivided technical labels using a large language model finely tuned based on a standard classification system. The network construction module is used to construct a time-series science knowledge network and a time-series technology knowledge network with subdivision technology as nodes, respectively, based on the mapping results; The lag measurement module is used to extract the edge weight time series sequence of each pair of sub-technologies in the scientific knowledge network and the technical knowledge network, and to calculate the optimal time offset between the sequences using a sliding window and time-delay cross-correlation algorithm as the knowledge lag value and to construct the implicit attribute network. The prediction model module is used to construct and run a dual-flow graph attention network model that includes structural flow and attribute flow; wherein the structural flow takes an explicit co-occurrence network constructed from historical patent transfer records as input, and the attribute flow takes the implicit attribute network as input, and integrates the dual-flow features through a gating fusion mechanism to generate node embedding representations; A multi-step prediction module is used to predict the probability of future technology transfer using the node embedding representation, employing a sliding window-based strategy.

[0014] As a further technical solution, the gating fusion mechanism in the prediction model module is configured to generate a scalar gating value between 0 and 1 to dynamically adjust the dependence ratio on structural flow features and attribute flow features.

[0015] According to one aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge.

[0016] According to one aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge.

[0017] This invention systematically solves the challenges faced by existing technologies by constructing a complete technical closed loop, from cross-modal knowledge alignment to temporal lag relation quantification and dual-stream feature fusion prediction. Its beneficial effects are specifically reflected in: 1. This invention achieves precise semantic alignment across modal data, laying a high-quality data foundation for subsequent analysis. Addressing the problem of significant differences and difficulty in integrating terminology between scientific literature and patent documents in existing technologies, this invention utilizes a finely tuned large language model based on a standard classification system (such as the ACM Computational Classification System) to uniformly map both to the same set of subdivided technical tags. This technical feature directly overcomes the semantic gap between multi-source heterogeneous data, not only greatly expanding the information sources of the model but also ensuring the accuracy and consistency of the input data, thus improving the reliability of subsequent analysis from the source.

[0018] 2. This invention achieves precise quantification of the lag in scientific and technological knowledge, providing a crucial basis for forward-looking predictions. Addressing the lack of forward-looking capabilities in existing prediction methods, this invention constructs a time-series scientific knowledge network and a technological knowledge network, and uses a time-lag cross-correlation algorithm to calculate the knowledge lag value between each pair of technologies. This technique, for the first time, systematically quantifies the time difference between scientific discovery and technological application, effectively identifying potential technological connections in scientific research that are active but not yet commercialized on a large scale, thus injecting leading external attribute signals into the prediction model.

[0019] 3. A dual-stream prediction architecture integrating explicit structure and implicit attributes is proposed, achieving end-to-end deep feature fusion. Addressing the challenge of existing graph neural networks simultaneously processing network topology and external temporal attributes, this invention designs a dual-stream graph attention network model incorporating structural and attribute flows, along with a gating fusion mechanism. This architecture innovatively processes and dynamically fuses an explicit co-occurrence network based on historical transfer records and an implicit attribute network based on knowledge lag values ​​in parallel. This enables the model to learn mature market cooperation patterns while capturing potential science-driven trends, significantly improving the comprehensiveness and accuracy of predictions.

[0020] 4. Significantly enhances the robustness and predictive ability of the model in data-sparse scenarios. Addressing the prediction failure problem caused by the sparsity of patent data in emerging industries or interdisciplinary fields, the overall technical solution of this invention provides powerful external knowledge enhancement to sparse patent transfer networks by introducing scientific knowledge networks and their derived knowledge lag attributes. This enables the model to perform effective reasoning and prediction based on its solid scientific connections, even when facing technology nodes lacking historical transfer records.

[0021] 5. Supports long-term, multi-step technology transfer opportunity prediction, enhancing the strategic value of decision-making. This invention employs a sliding window-based multi-step prediction strategy, enabling the model to directly learn the long-term evolution patterns of technology networks and achieve continuous prediction of technology transfer probabilities at multiple future time steps. This not only overcomes the limitations of traditional single-step prediction but also provides time-series-based decision support for enterprises' long-term technology layout and strategic planning.

[0022] In summary, this invention systematically addresses the key shortcomings of existing technologies in data fusion, forward-looking identification, model architecture, and prediction dimensions through a series of interconnected and synergistic technical means, including cross-modal semantic alignment, temporal lag relation quantification, dual-stream feature fusion, and multi-step forward prediction. It achieves technological progress from passive analysis relying on historical data to proactive forward-looking discovery using scientific-technological correlations. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating the technology transfer opportunity prediction method based on the lag of scientific and technological knowledge provided in this embodiment of the invention; Figure 2 A schematic diagram of a technology transfer opportunity prediction system based on the lag of scientific and technological knowledge provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that the terms used in this invention specification are explained as follows: Scientific papers: These are documents published in academic journals or conferences that explain scientific research findings, theories, or experimental results, such as papers indexed by academic databases like Web of Science (WoS).

[0026] Patent documents: These typically refer to publicly disclosed patent application documents or authorized patent texts, including the title, abstract, specification, claims, etc., such as documents indexed by patent databases like Incopat.

[0027] Scientific knowledge network: In the context of this invention, it refers to a network constructed based on the content of scientific papers (such as titles and abstracts) that reflects the co-occurrence relationships of technical concepts in scientific research.

[0028] Technical knowledge network: In the context of this invention, it refers to a network constructed based on the content of patent documents (such as titles and abstracts) that reflects the co-occurrence relationships of technical concepts in technology research and development and application.

[0029] This invention addresses the shortcomings of existing technologies by first constructing a set of subdivided technology labels based on the ACM computational classification system. Then, it fine-tunes an open-source large language model using a corpus of academic papers based on this system, resulting in a fine-tuned large language model. Next, this fine-tuned model performs multi-label classification inference on retrieved scientific papers and patent documents, mapping them to corresponding subdivided technology labels. Subsequently, a temporal scientific knowledge network and a temporal technological knowledge network are constructed based on the mapping results. A time-delay cross-correlation algorithm is used to characterize the degree of lag between the scientific knowledge sequence and the technological knowledge sequence in each pair of subdivided technologies. Finally, the calculated knowledge lag values ​​are integrated into a temporal attribute and incorporated into a technology transfer opportunity network constructed from historical patent transfer records. A dual-flow graph attention network model incorporating structural and attribute flows is designed to fuse structural and attribute embeddings, and combined with a sliding window-based multi-step prediction strategy to achieve multi-step prediction of technology transfer opportunities.

[0030] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. 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. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0032] This invention discloses a method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge, used to identify commercially viable technologies for enterprises. The method specifically includes: S1: Using a large language model fine-tuned based on a standard classification system, retrieved scientific papers and patent documents are uniformly mapped to a unified set of subdivided technical tags.

[0033] S2: Based on the mapping results, construct a time-series science knowledge network and a time-series technology knowledge network, with subdivision technology as nodes and co-occurrence relationships as edges, respectively.

[0034] S3: For each pair of subdivided technologies, extract their edge weight time series sequences in the scientific knowledge network and technical knowledge network, and use the time delay cross-correlation algorithm to calculate the time offset corresponding to the correlation peak between the two sequences, which is used as the knowledge lag value of the technology pair.

[0035] S4: Construct a dual-flow graph attention network model that includes structural flow and attribute flow; wherein, the structural flow takes an explicit co-occurrence network constructed from historical patent transfer records as input, and the attribute flow takes a implicit attribute network constructed from the knowledge lag values ​​as input, and integrates the features of the dual-flow output through a gating fusion mechanism to generate a comprehensive node embedding representation.

[0036] S5: Employ a sliding window-based training strategy and utilize the generated node embedding representations to predict the probability of technology transfer between subdivided technology nodes after one or more future time steps.

[0037] As a preferred embodiment, such as Figure 1 As shown, the embodiments of the present invention are described in five aspects: fine-tuning of large language models and multi-label classification, construction of scientific and technological knowledge networks, estimation of the lag of scientific and technological knowledge, design of prediction models based on dual-stream attention mechanisms, and multi-step prediction of technology transfer opportunities.

[0038] In this invention, the multi-label classification process using a large language model fine-tuned based on a standard classification system includes constructing a detailed technology list and using the fine-tuned large language model to perform technology recognition on the text.

[0039] To address the limitations of traditional keyword-based methods, which introduce noise and ambiguity, this invention employs the ACM Computational Classification System (ACM CCS) to construct a detailed list of technologies. The construction process begins by identifying top-level ACM CCS categories relevant to specific industries. Subsequently, the tree structure under these categories is systematically traversed, collecting all leaf nodes without further subdivision, thus forming a comprehensive list of detailed technologies.

[0040] Supervised fine-tuning of the Mistral-7B model was used for subsequent multi-label classification tasks to achieve automatic mapping from text to sub-segments. A training dataset was constructed by collecting papers labeled with two or more sub-segments from an established list. During training, the titles and abstracts of these papers served as input to the model, while their corresponding sub-segments were used as output labels. Therefore, the final fine-tuned large language model was able to accurately identify multiple relevant sub-segments contained in a given text.

[0041] In the scientific and technological knowledge network construction step of this invention, a time-series scientific knowledge network and a time-series technological knowledge network are constructed based on paper data and patent data from a specific industry, respectively. The writing styles and terminology used in papers and patents differ significantly, making it challenging to identify the semantic relationships between them. To address this issue, this invention utilizes the powerful semantic embedding and contextual understanding capabilities of a finely tuned large language model to identify the semantic relationships between papers and patents.

[0042] First, we acquire WoS-indexed papers and IncoPat-indexed patents for specific industries, and then input their titles and abstracts into a finely tuned Mistral-7B model for multi-label classification inference. This process assigns multiple segmentation techniques to each text, enabling the analysis of patents and papers within a unified dimensional framework.

[0043] Subsequently, scientific knowledge networks and technical knowledge networks were constructed based on the sub-technologies corresponding to the papers and patents. Both networks are undirected, weighted, and temporal. In the networks, nodes represent sub-technologies. Edges represent the existence of text classified under the two sub-technologies, and the weight w of the edge represents the total number of texts classified in this way. The temporal network is divided into time slices according to the time granularity ∆t, forming a total of n slices from... arrive Time slices.

[0044] In this embodiment of the invention, a time-delay cross-correlation algorithm is used to quantify the lag in scientific and technological knowledge measurement. The scientific knowledge network and the technological knowledge network share the same set of nodes and are temporally aligned. For any pair of sub-technologies, their corresponding edge weights in each time slice can be extracted from both networks. This process generates two independent time series for each pair of sub-technologies: a scientific co-occurrence sequence and a technological co-occurrence sequence. These two sequences reflect the degree of technological convergence at the scientific or technological level over time. The trends of these two sequences can reveal the lead-lag relationship.

[0045] The time-lag cross-correlation algorithm determines the lag value that maximizes the correlation between two time series, thus statistically explaining the lead-lag relationship. Assume the scientific co-occurrence sequence is: The co-occurrence sequence of the technologies is as follows: The time lag cross-correlation between scientific knowledge and technological knowledge is calculated using the following formula: , in and They are time series and The average value is given by k, where t represents the lag value. For a given lag value t, a correlation peak at k=0 indicates that there is no significant lead-lag relationship between the two series. A peak at k>0 indicates that the time series... A leading trend, with a peak at k<0, indicates a time series... Leading.

[0046] Since the scientific and technological knowledge lag between two sub-technologies is not static, a sliding time window of size m is designed to calculate the scientific and technological knowledge lag. This calculation considers the two sequences over m time units preceding a given time point. Ultimately, this process generates a time series representing the scientific and technological knowledge lag for each pair of sub-technologies, from... Begin, and represent it as: .

[0047] Changes in the lag of scientific and technological knowledge should not exhibit drastic fluctuations between adjacent time windows, nor should they remain almost unchanged over long periods. To determine the optimal time window size m, this invention calculates the average sample entropy of the scientific and technological knowledge lag time series for all node pairs. The median of these average sample entropies is then selected as the optimal size of m.

[0048] In this embodiment of the invention, during the construction of the technology transfer opportunity prediction model based on the dual-flow graph attention mechanism, a time network with subdivided technologies as nodes is first constructed to predict technology transfer opportunities. This network is based on historical patent transfer records and covers technologies from... arrive The time period. The connection strength between nodes is defined by two types of weights: the first is explicit weight, obtained by calculating the co-occurrence frequency of sub-technologies in transferred patents, representing direct historical associations; the second is implicit weight, composed of the calculated scientific and technological knowledge lags between any two sub-technologies, revealing potential asynchronous relationships in the knowledge flow. To effectively integrate these two types of information, this invention proposes a technology transfer opportunity prediction model based on a dual-flow graph attention mechanism.

[0049] The core idea of ​​this method is to learn the evolutionary patterns of technology transfer opportunity networks, which are defined by their topology and implicit weights. To this end, parallel "structure flow" and "attribute flow" are designed to learn these two aspects respectively. Subsequently, these two types of information are temporally integrated by an interpretable gated fusion block to generate node representations with high predictive power.

[0050] The entire methodology consists of four core components: structural attention block, attribute attention block, temporal attention block, and gating fusion block.

[0051] The structural attention block aims to capture network structure information for each sub-technology in each snapshot, defined by direct historical associations. This module takes as input an explicit co-occurrence network constructed from the co-occurrence frequency of patent transfer opportunities, where explicit weights represent connection strength. The module consists of multiple stacked graph attention networks, employing a self-attention mechanism to assign different importance weights to each node's neighbors over time.

[0052] Specifically, for any given technical node, the model calculates attention coefficients based on its feature similarity with neighboring nodes, thereby selectively aggregating neighborhood information. Through this multi-layered stacking, nodes can capture higher-order neighborhood structural features. After processing by the structural attention block, each node generates an intermediate representation at each time slice t: .

[0053] The core task of the attribute attention block is to model implicit weight information defined by scientific and technological knowledge lags, thereby revealing potential and forward-looking knowledge flow trends in the network. In parallel with the structured flow, this module also employs a multi-layered stacked graph attention network as its core architecture. Its unique feature is that it receives an implicit attribute network defined by knowledge lag values ​​as input and uses the output from the structured attention block. This serves as the initial node feature. The rationale for this design is that the influence of lagging scientific and technological knowledge does not arise out of thin air, but rather acts on nodes that already possess a specific historical structural identity. This module learns how to leverage knowledge lag information to enhance and refine node representations based on their existing structural identity. The final output of its attribute attention block is... .

[0054] To capture the temporal evolution characteristics of the technology network, intermediate representation sequences generated from each snapshot are input into the temporal processing module. Specifically, representation sequences from the structure stream... and sequences from attribute streams The data is fed into two separate temporal attention blocks. Each temporal attention block employs a self-attention mechanism, enabling each node to focus on the key historical moments that have the greatest impact on its current state in the time dimension.

[0055] Obtaining a time-modeled structural representation and attribute representation Subsequently, a gating fusion mechanism was designed to integrate these two information flows. The final node representation... It is calculated using the following formula: .

[0056] The gating value varies between 0 and 1 and is learned by the neural network by concatenating two representations. This gating fusion mechanism not only effectively integrates information but also endows the model with excellent interpretability. When the gating value is close to 1, it indicates that the model relies more on structural information defined by historical co-occurrence for prediction. Conversely, when the value is close to 0, it indicates that the model focuses more on potential future associations revealed by the lag of scientific and technological knowledge.

[0057] In this embodiment of the invention, the multi-step prediction stage of technology transfer opportunities employs a sliding window-based multi-step prediction strategy, a technique originating from the field of time series forecasting. By constructing training samples spanning multiple steps of the entire historical data, this strategy enables the model to directly learn actual long-term evolutionary patterns, rather than performing simple extrapolation.

[0058] Taking three-step prediction as an example, the time pairings corresponding to the input and output time slices of the training samples are as follows: At the same time, using from Embedded in time slices Perform link prediction on the test set.

[0059] To enable the proposed model to predict technology transfer opportunities at any future time step, this invention designs a unified k-step prediction loss function. The goal of this function is to utilize the baseline time. Learned node embeddings To predict target time slices The network structure.

[0060] Specifically, for a given prediction step k, a set of training samples is constructed. , where (u, v) represents a pair of segmentation techniques, and y∈{0, 1} is the true label, indicating the difference between them at the target time. Does a link exist? The model's predicted score is based on the baseline time. The inner product of the node embeddings is given as follows: .

[0061] Ultimately, the general k-step prediction loss function is defined as the average binary cross-entropy over the entire training set. As shown below: .

[0062] In this embodiment of the invention, to verify the predictive performance of the proposed model, five metrics—Accuracy, Precision, Recall, F1-Score, and AUC—are used to evaluate the multi-step prediction results. These metrics are calculated as follows: .

[0063] The implementation of the various embodiments of the present invention is based on programmed processing through a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a technology transfer opportunity prediction system based on the lag of scientific and technological knowledge. This system is used to execute the technology transfer opportunity prediction method based on the lag of scientific and technological knowledge in the above method embodiments.

[0064] See Figure 2 The system includes: a multi-label classification module, used to map the text content of scientific papers and patent documents to a unified set of subdivided technology labels using a large language model fine-tuned based on a standard classification system; a network construction module, used to construct a time-series scientific knowledge network and a time-series technical knowledge network with subdivided technologies as nodes based on the mapping results; a lag calculation module, used to extract the edge weight time-series sequence of each pair of subdivided technologies in the scientific knowledge network and technical knowledge network, and use a sliding window and time-lag cross-correlation algorithm to calculate the optimal time offset between the sequences as the knowledge lag value and construct a latent attribute network; a prediction model module, used to construct and run a dual-flow graph attention network model containing structural flow and attribute flow; wherein the structural flow takes an explicit co-occurrence network constructed from historical patent transfer records as input, and the attribute flow takes the latent attribute network as input, and integrates the dual-flow features through a gating fusion mechanism to generate node embedding representations; and a multi-step prediction module, used to predict the probability of future technology transfers using the node embedding representations with a sliding window-based strategy.

[0065] The technology transfer opportunity prediction system based on the lag of scientific and technological knowledge provided in this embodiment of the invention adopts... Figure 2 This model integrates several modules with large language models and temporal graph neural network technology, proposing a novel dual-stream attention mechanism and knowledge lag quantification algorithm. This model effectively integrates leading scientific information, improving the accuracy of predicting emerging technology opportunities in the absence of historical transfer records, and contributing to enterprises' efforts to seize technological advantages and improve their industrial chain layout.

[0066] It should be noted that the system embodiments provided by this invention, in addition to implementing the methods in the above method embodiments, are also used to implement the methods in other method embodiments provided by this invention. The difference lies only in setting corresponding functional modules, and their principles are basically the same as those of the above system embodiments provided by this invention. As long as those skilled in the art, based on the above system embodiments and referring to the specific technical solutions in other method embodiments, obtain corresponding technical means and technical solutions composed of these technical means by combining technical features, and improve the modules in the above system embodiments while ensuring the practicality of the technical solutions, they can obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments. For example: Based on the above system embodiments, as a preferred embodiment, the technology transfer opportunity prediction system based on the lag of scientific and technological knowledge provided in this embodiment of the invention, wherein the multi-label classification module is further configured to execute the following instructions: Using professional data acquisition tools, we extensively retrieved raw data such as titles and abstracts of scientific papers and abstracts of patent specifications from databases such as Web of Science and Incopat. A standardized list of subdivided technical labels was constructed using the ACM computational classification system as the reference space for semantic mapping; We construct a training set of papers with explicit classification labels, and fine-tune the open-source large language model to enable it to have domain-specific multi-label classification capabilities. The large language model, after being fine-tuned, inputs the papers and patent texts to be analyzed and performs inference, outputting one or more subdivided technical tags corresponding to each document, thereby eliminating the semantic gap between heterogeneous data.

[0067] As a preferred embodiment, the technology transfer opportunity prediction system based on the lag of scientific and technological knowledge provided in this embodiment of the invention, wherein the network construction module is further configured to execute the following instructions: Using the mapped segmentation techniques as nodes, the co-occurrence frequency in scientific literature and patent documents is statistically analyzed according to time slices; Construct a scientific knowledge network with temporal characteristics, where the edge weights represent the co-occurrence strength of scientific literature within a specific time window; Construct a technology knowledge network with temporal characteristics, where the edge weights represent the co-occurrence intensity of patent documents within a specific time window; Generate a two-layer network sequence that evolves over time, providing a structured data foundation for subsequent time series analysis.

[0068] As a preferred embodiment, the technology transfer opportunity prediction system based on the lag of scientific and technological knowledge provided in this embodiment of the invention further includes a lag calculation module configured to execute the following instructions: For any pair of nodes in the network, extract their weight sequence that changes over time in the scientific knowledge network and the technical knowledge network; The time series is extracted using a sliding window, and the correlation coefficient between the two series is calculated by applying a time-delay cross-correlation algorithm to determine the optimal time offset that makes the correlation reach its peak. If the time offset shows that the scientific sequence leads the technological sequence, then the offset is defined as the knowledge lag value. The calculated knowledge lag values ​​are standardized and used as edge weights for potential associations to construct a latent attribute network containing implicit connections.

[0069] As a preferred embodiment, the technology transfer opportunity prediction system based on the lag of scientific and technological knowledge provided in this embodiment of the invention, wherein the prediction model module is further configured to execute the following instructions: A dual-flow graph attention network architecture is constructed, in which the structure flow takes an explicit co-occurrence network based on historical patent transfer records as input, and the attribute flow takes an implicit attribute network based on knowledge lag values ​​as input. In structural flow, graph attention networks are used to capture mature technological cooperation patterns and topological characteristics in the current market. In the attribute stream, based on the node features output by the structure stream, a graph attention network is used to perform message passing on the latent attribute network to capture potential technology-related features driven by scientific discovery. The design incorporates a learnable gated fusion unit that dynamically calculates adaptive weights based on the contextual information of the features. This weighted summation of structural flow features and attribute flow features generates the final node embedding vector.

[0070] The method in this embodiment of the invention is implemented using an electronic device; therefore, it is necessary to introduce the relevant electronic device. For this purpose, embodiments of the present invention provide an electronic device, such as... Figure 3 As shown, the electronic device includes: at least one processor, a communication interface, at least one memory, and a communication bus, wherein the at least one processor, the communication interface, and the at least one memory communicate with each other via the communication bus. The at least one processor invokes logical instructions stored in the at least one memory to execute all or part of the steps of the methods provided in the foregoing method embodiments.

[0071] Furthermore, when the logical instructions in at least one of the aforementioned memories are implemented as software functional units and sold or used as independent products, they are stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, is embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (a personal computer, server, or network device) to execute all or part of the steps of the methods described in the various method embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks—various media for storing program code.

[0072] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] These computer-usable program codes may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] In summary, this invention, within a time-series science-technology knowledge network, applies time-lag cross-correlation analysis to quantify the lag between any pair of sub-technologies at different time points. It employs a dual-stream attention mechanism incorporating both structural and attribute flows, combined with a time-aware graph convolution strategy, to perform multi-step technology transfer opportunity prediction. The key points of this invention are: 1. It has achieved a leap from static single-technology prediction to dynamic cross-technology forward-looking prediction, accurately capturing the potential for technological synergistic evolution.

[0076] This invention first constructs semantically unified science-technology knowledge network pairs by fine-tuning a large language model. Based on this, a time-delay cross-correlation algorithm is used to quantify the time difference (knowledge lag value) between science and technology in each pair. This design enables the model to directly identify potential technology combinations that are active in scientific research but whose technological applications are still in their infancy, thereby fundamentally shifting the predictive perspective from single technologies to the forward-looking discovery of opportunities for cross-domain technology integration.

[0077] 2. Overcame the challenges of semantic gap and network sparsity in multi-source data, laying a high-quality data foundation for accurate prediction.

[0078] To address the noise and sparsity issues inherent in traditional keyword methods, this invention introduces a large language model fine-tuning technique based on authoritative classification systems (such as ACM CCS) at the forefront. This step achieves precise and automated semantic alignment of scientific papers and patent documents at the subdivided technical level. The resulting high-quality, dense temporal knowledge network not only forms the basis for subsequent lag analysis but also directly improves the signal-to-noise ratio of the entire model's input information, ensuring the robustness and accuracy of all subsequent analyses.

[0079] 3. A dual-stream prediction architecture with native fusion of spatiotemporal attributes is proposed, which solves the core bottleneck of the difficulty in end-to-end fusion of external leading signals and historical structures.

[0080] To address the challenge of fusing temporal and attribute data, this invention designs an organically integrated dual-flow graph attention network model. This model learns the potential correlations revealed by historical transfer inertia and scientific lag through parallel structural and attribute flows, respectively, and adaptively fuses them using a learnable gating mechanism. This architecture ensures that scientific leading signals (attribute flows) can deeply interact with the current market structure (structural flows) within the same representation space, achieving true end-to-end native fusion of external attributes and network temporal structure. This significantly improves the model's predictive power and interpretability, especially for emerging and sparse technologies.

[0081] In summary, the various beneficial effects of this invention are closely related and progressively enhance each other: high-quality data alignment is a prerequisite for accurately quantifying knowledge lag; precise lag measurement provides key leading signal input for the two-stream model; and finally, the fusion decision-making capability of the two-stream model enables forward-looking prediction.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method and system for predicting technology transfer opportunities based on the lag of scientific and technological knowledge, characterized in that, include: Using a large language model fine-tuned based on a standard classification system, the text content of scientific papers and patent documents is mapped to a unified set of subdivided technical tags. Based on the mapping results, a time-series science knowledge network and a time-series technology knowledge network are constructed, with subdivision technologies as nodes and co-occurrence relationships as edges. For each pair of subdivided technologies, the edge weight time series sequence of the technology in the scientific knowledge network and the technical knowledge network is extracted, and the time offset corresponding to the correlation peak between the two sequences is calculated using the time delay cross-correlation algorithm, which is used as the knowledge lag value of the technology pair. A dual-flow graph attention network model containing structural flow and attribute flow is constructed. The structural flow takes an explicit co-occurrence network constructed from historical patent transfer records as input, and the attribute flow takes a implicit attribute network constructed from knowledge lag values ​​as input. The features of the dual-flow outputs are integrated through a gating fusion mechanism to generate a comprehensive node embedding representation. A sliding window-based training strategy is adopted to predict the probability of technology transfer between subdivided technology nodes after one or more time steps in the future, using the generated node embedding representation.

2. The method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge according to claim 1, characterized in that, The standard classification system is the ACM computational classification system; the fine-tuning process of the large language model is as follows: using the title and abstract of the scientific paper as input text, and using one or more subdivided technical labels corresponding to them in the ACM computational classification system as supervision labels, fine-tuning training is performed.

3. A method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge according to claim 1 or 2, characterized in that, Both the time-series scientific knowledge network and the time-series technical knowledge network are undirected weighted networks, where the edge weights represent the frequency with which two sub-technical tags appear together in the same scientific paper or the same patent document within a specific time slice.

4. The method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge according to claim 1, characterized in that, The calculation of knowledge lag values ​​using the time-delay cross-correlation algorithm includes: For each pair of sub-technologies, a sliding time window is set, and the cross-correlation function between their scientific knowledge sequence and technological knowledge sequence is calculated within the window; The time difference between the scientific knowledge sequence and the technical knowledge sequence when the cross-correlation function reaches its maximum value is determined as the knowledge lag value at the center of the window. Repeat the above calculation for all time windows to form a time series of knowledge lag values, and construct a dynamic implicit attribute network with nodes representing subdivision techniques and edge weights representing knowledge lag values.

5. The method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge according to claim 1, characterized in that, The dual-flow graph attention network model works as follows: The structure flow learns the topological features of the explicit co-occurrence network through a graph attention network and outputs a structure-aware node representation. The attribute stream is based on the structure-aware node representation and runs a graph attention network on the latent attribute network to learn the latent association features represented by knowledge lag, and outputs attribute-enhanced node representations. The gated fusion mechanism uses a learnable gating function to dynamically weight and sum the structure-aware node representation and the attribute-enhanced node representation to generate the final fused node representation.

6. The method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge according to claim 1, characterized in that, The sliding window-based training strategy is as follows: a training sample set is constructed, where the input of each sample is the network state and node representation of multiple consecutive historical time slices, and the output is the network connection state of the k-th future time slice; the model is trained by minimizing the loss function between the predicted connection state and the true state.

7. A technology transfer opportunity prediction system based on the lag of scientific and technological knowledge, characterized in that, include: The multi-label classification module is used to map the text content of scientific papers and patent documents to a unified set of subdivided technical labels using a large language model finely tuned based on a standard classification system. The network construction module is used to construct a time-series science knowledge network and a time-series technology knowledge network with subdivision technology as nodes, respectively, based on the mapping results; The lag measurement module is used to extract the edge weight time series sequence of each pair of sub-technologies in the scientific knowledge network and the technical knowledge network, and to calculate the optimal time offset between the sequences using a sliding window and time-delay cross-correlation algorithm as the knowledge lag value and to construct the implicit attribute network. The prediction model module is used to construct and run a dual-flow graph attention network model that includes structural flow and attribute flow; wherein the structural flow takes an explicit co-occurrence network constructed from historical patent transfer records as input, and the attribute flow takes the implicit attribute network as input, and integrates the dual-flow features through a gating fusion mechanism to generate node embedding representations; A multi-step prediction module is used to predict the probability of future technology transfer using the node embedding representation, employing a sliding window-based strategy.

8. A technology transfer opportunity prediction system based on the lag of scientific and technological knowledge according to claim 7, characterized in that, The gating fusion mechanism in the prediction model module is configured to generate a scalar gating value between 0 and 1 to dynamically adjust the dependence ratio on structural flow features and attribute flow features.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of a method for predicting technology transfer opportunities based on the lag of scientific and technological knowledge as described in any one of claims 1 to 6.

Images