Scientific and technological trend prediction method and device based on machine learning, equipment and medium

Abstract

The invention relates to a science and technology trend prediction method and device based on machine learning, equipment and a medium. According to the method, a unified database is constructed by integrating microscopic invention data, mesoscopic technology cluster data and macroscopic technology normal form data, multi-level features are extracted and fused from the unified database to generate a cross-level feature matrix, and meanwhile, a microscopic dynamic network, a mesoscopic dynamic network and a macroscopic dynamic network are constructed based on the features and association relationships and are subjected to cross-level connection; forming a multi-level dynamic network representing hierarchical interaction; further establishing an influence conduction mechanism for describing cross-level diffusion, phase change and feedback processes; and finally, comprehensively applying the feature matrix, the dynamic network and the conduction mechanism to drive a pre-training machine learning model for prediction. Systematic modeling of science and technology ecological multi-level nonlinear evolution is achieved, the cross-level conduction and amplification effect of external impact can be simulated more comprehensively and accurately, and complex emerging phenomena such as technology path locking and normal form phase change can be effectively predicted.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and in particular relates to a method, device, equipment and medium for predicting technology trends based on machine learning. Background Technology

[0002] In the current field of science and technology management and innovation research, accurately predicting technological development trends is crucial for formulating national science and technology strategies and corporate R&D plans. Existing methods for predicting technological trends mainly rely on the analysis of data at a single level. For example, one type of method focuses on micro-level invention activities, tracing the evolution of specific technologies or identifying emerging hotspots by analyzing citation networks, keyword co-occurrence, or text content in patents or academic papers. Another type of method focuses on macro-level indicators, such as R&D funding, policy text analysis, or technology maturity curves, attempting to infer the possibility of paradigm shifts from changes in the overall environment. These methods have achieved certain results at their respective levels, but they generally share a fundamental limitation: they artificially fragment the complex technological ecosystem, failing to construct a unified model that can characterize the interaction between micro-level inventions, meso-level technological fields, and macro-level technological paradigms.

[0003] Due to the lack of explicit modeling of cross-level influence transmission, existing methods struggle to explain and predict how a disruptive invention at the micro level can lead to the locking of a specific technological path, or how a macro-level science and technology policy can ultimately drive the transformation of the entire technological paradigm by influencing the development of intermediate technology clusters. They typically simplify inter-level interactions to linear or static relationships, failing to capture the inherent nonlinear dynamics, feedback loops, and the resulting "emergent" phenomena, such as the phase transitions that occur in the technological ecosystem at critical points.

[0004] Therefore, existing technologies often fall short when dealing with policy simulation problems involving multi-level nonlinear interactions, such as "how the competitiveness of a specific technology cluster will evolve if investment in basic research increases significantly." Their predictions are significantly insufficient in terms of comprehensiveness and explanatory power, making it difficult to meet the growing demand for quantitative and mechanistic analysis tools for forward-looking science and technology decision-making. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, equipment, and medium for predicting technology trends based on machine learning to address the aforementioned technical problems.

[0006] Firstly, this application provides a technology trend prediction method based on machine learning, including:

[0007] S1. Collect invention data at the micro level, technology cluster data at the meso level, and technology paradigm data at the macro level, and integrate the invention data, technology cluster data, and technology paradigm data to generate a unified database. Among them, invention data includes the text content, citation relationships, and timestamps of patents and papers; technology cluster data is obtained by clustering invention data; and technology paradigm data includes policy documents, R&D investment indicators, and technology paradigm labels.

[0008] S2. Extract features from the data in the unified database to obtain micro-level features, meso-level features and macro-level features, and fuse the micro-level features, meso-level features and macro-level features based on cross-level association to generate a cross-level feature matrix;

[0009] S3. Based on micro-level features and referencing relationships, construct a micro-dynamic network; based on meso-level features and technology cluster associations, construct a meso-dynamic network; based on macro-level features and policy associations, construct a macro-dynamic network; connect the micro-dynamic network, meso-dynamic network, and macro-dynamic network across levels to generate a multi-level dynamic network.

[0010] S4. Based on a multi-level dynamic network, establish a cross-level influence transmission mechanism; among which, the cross-level influence transmission mechanism includes the diffusion process of science and technology from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro.

[0011] S5. Based on cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms, it uses a pre-trained machine learning model to predict technology trends and outputs predictions of technology cluster activity and the probability of technology paradigm shifts.

[0012] Secondly, this application also provides a technology trend prediction device based on machine learning, used to implement the method described in the first aspect, the device comprising:

[0013] The multi-source data fusion module is used to collect invention data at the micro level, technology cluster data at the meso level, and technology paradigm data at the macro level, and integrates the invention data, technology cluster data, and technology paradigm data to generate a unified database. Among them, invention data includes the text content, citation relationships, and timestamps of patents and papers; technology cluster data is obtained by clustering invention data; and technology paradigm data includes policy documents, R&D investment indicators, and technology paradigm labels.

[0014] The cross-level feature fusion module is used to extract features from data in a unified database to obtain micro-level features, meso-level features, and macro-level features. Based on cross-level association, the micro-level features, meso-level features, and macro-level features are fused to generate a cross-level feature matrix.

[0015] The multi-level network construction module is used to construct micro-dynamic networks based on micro-level features and reference relationships; construct meso-dynamic networks based on meso-level features and technology cluster associations; construct macro-dynamic networks based on macro-level features and policy associations; and perform cross-level connections between micro-dynamic networks, meso-dynamic networks, and macro-dynamic networks to generate multi-level dynamic networks.

[0016] The influence transmission modeling module is used to establish cross-level influence transmission mechanisms based on multi-level dynamic networks. These mechanisms include the diffusion process of science and technology from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro.

[0017] The technology trend prediction module is used to predict technology trends based on cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms through pre-trained machine learning models, and outputs predictions of technology cluster activity and the probability of technology paradigm shifts.

[0018] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a technology trend prediction method based on machine learning as described in the first aspect.

[0019] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a technology trend prediction method based on machine learning as described in the first aspect.

[0020] The aforementioned machine learning-based method, apparatus, equipment, and medium for predicting technology trends integrate micro-level invention data, meso-level technology cluster data, and macro-level technology paradigm data to form a unified database. Features from each level are then extracted and fused to construct a cross-level feature matrix. Simultaneously, dynamic networks at the micro, meso, and macro levels are constructed based on features and correlations, and these networks are connected across levels to form a multi-level dynamic network characterizing inter-level interactions. Furthermore, a mechanism describing the transmission of cross-level influences (including diffusion, phase transition, and feedback processes) is established. Finally, the cross-level feature matrix, multi-level dynamic network, and cross-level influence transmission mechanism are comprehensively utilized to drive a pre-trained machine learning model for prediction. This series of coherent steps achieves a systematic modeling of the multi-level, nonlinear evolution of technology trends, enabling a more comprehensive and accurate simulation of the cross-level transmission and amplification effects of external shocks in the technology ecosystem, and predicting the technological effects of complex emergent phenomena such as technology path locking and technology paradigm phase transitions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a technology trend prediction method based on machine learning provided by this invention;

[0023] Figure 2 This is a flowchart illustrating the process of establishing a cross-level influence transmission mechanism in one optional embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of a technology trend prediction device based on machine learning provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0026] refer to Figure 1 The document presents a flowchart illustrating a technology trend prediction method based on machine learning, as provided in this application. The method includes the following steps:

[0027] S1. Collect invention data at the micro level, technology cluster data at the meso level, and technology paradigm data at the macro level, and integrate the invention data, technology cluster data, and technology paradigm data to generate a unified database. Among them, invention data includes the text content, citation relationships, and timestamps of patents and papers; technology cluster data is obtained by clustering invention data; and technology paradigm data includes policy documents, R&D investment indicators, and technology paradigm labels.

[0028] Specifically, for invention data at the micro level, the core includes two types of core information carriers: patents and papers. Patent data covers the full text of patent application documents, citation relationships between patents, and timestamps; paper data includes the title, abstract, keywords, core paragraphs of the main text, list of references, and publication date of academic papers. The above data can be obtained through official patent databases and academic literature databases via API interfaces or batch export functions, ensuring the consistency of data format during the acquisition process.

[0029] Mesoscopic-level technology cluster data is generated through clustering of microscopic invention data. The choice of clustering method needs to be combined with the characteristic attributes of the invention data. Considering that the invention data has both textual semantic features and citation relationship features, a hybrid clustering algorithm that integrates semantic similarity and topological association is advisable. In specific implementation, the text content of patents and papers is first transformed (e.g., the text is transformed into a fixed-dimensional semantic vector using the BERT model, with the vector dimension determined according to the data volume and semantic complexity to ensure effective differentiation of different technical topics). Then, an undirected graph is constructed for citation relationships (nodes are individual inventions, and edge weights are the citation strength between two inventions, which can be calculated by the number of co-citations and the citation time decay coefficient). Subsequently, the invention data is clustered based on spectral clustering or hierarchical clustering algorithms. In the spectral clustering process, the Laplacian matrix is ​​first calculated to reduce the dimensionality of the data, and the dimensionality-reduced feature vectors are used as clustering inputs. Hierarchical clustering then gradually merges similar inventions by calculating the distance between classes (using the Ward method or the average link method). Each cluster formed in the end is a technology cluster. The technology cluster data needs to record the cluster identifier, the unique ID of all inventions in the cluster, the core technical theme of the cluster (obtained by averaging the text semantic vectors in the cluster and matching them with a preset technical dictionary), and the cluster formation time (the timestamp of the earliest invention in the cluster).

[0030] The macro-level technology paradigm data includes three core types of information: policy documents covering national and local science and technology development plans, industry support policies, and technical standard documents, obtaining the full text of the documents, the issuing entity, the release date, and keywords related to the technology fields involved in the policies; R&D investment indicators including time-series data on R&D expenditure, number of R&D personnel, and number of R&D institutions at the national, industry, and enterprise levels, with data recorded at fixed time periods to ensure temporal continuity; and technology paradigm labels based on domain expert knowledge or historical technology evolution patterns, used to identify the dominant technology paradigms in different periods (such as "information technology paradigm," "artificial intelligence paradigm," etc.), with labels associated with specific time intervals and verified through multi-expert consensus to ensure accuracy.

[0031] The data integration phase is implemented in three steps: The first step is data cleaning, which addresses missing values ​​(for missing text data, interpolation with text from the same domain is used to supplement it; for missing numerical data, time-series interpolation or mean filling is used), outliers (identified using the 3σ principle or box plot method, with outliers determined by domain knowledge to determine whether to retain or correct them), and duplicate data (removed based on key fields such as invention unique ID, policy document number, and input indicator timestamps). The second step is data format unification, which converts text data into standard string format, structured data (such as citation relationships and R&D investment values) into two-dimensional table structure, and timestamps are uniformly converted to UTC time format to ensure cross-level data integrity. The time dimension can be aligned; the third step is to build a unified database, adopting a distributed database architecture (such as HBase or PostgreSQL), and designing a database partitioning strategy according to "hierarchy-data type-time interval". Among them, micro-level invention data is stored by timestamp partitioning, meso-level technology cluster data is associated with the partition index of the corresponding invention data, and macro-level technology paradigm data is partitioned by technology paradigm label. At the same time, cross-level data association indexes are established (such as the mapping relationship between technology cluster and its corresponding technology paradigm, and the mapping relationship between invention and its corresponding technology cluster). The final unified database needs to support efficient multi-dimensional queries and data calls to meet the requirements of subsequent steps for data access speed and association queries.

[0032] S2. Extract features from the data in the unified database to obtain micro-level features, meso-level features, and macro-level features. Then, based on cross-level association, fuse the micro-level features, meso-level features, and macro-level features to generate a cross-level feature matrix.

[0033] Specifically, after the unified database is built, the feature extraction stage begins. In this stage, feature extraction schemes are designed to suit the data characteristics of the three levels: micro, meso, and macro. This ensures that the features at each level can not only fully depict the technological evolution of the corresponding level, but also provide compatible feature dimensions and data formats for subsequent cross-level fusion.

[0034] For micro-level feature extraction, targeting the patent and paper text content in the invention data, the text is first preprocessed using a word segmentation tool to remove meaningless stop words and special symbols, retaining key information such as technical terms, method names, and core parameters. Then, a Bidirectional Encoder Representation Model (BERT) is used for text feature embedding. A BERT-based-uncased pre-trained model is selected and fine-tuned based on the patent and paper text datasets. For example, during fine-tuning, the batch size is set to 2, the learning rate to 2e-5, and the training epochs to 5. The final output is a text feature vector with a dimension of 768, which effectively captures semantic relationships and technical topic information in the text. For citation relationships, a citation strength quantification model is used to construct features. The citation strength calculation formula is: ,in This indicates the strength of the reference of invention i to invention j. This represents the number of references (a positive integer, extracted directly from the reference relationship data). The time decay factor is calculated as follows: , The current time for feature extraction. For the time when the reference occurred, The decay coefficient (calibrated based on the decay law of historical citations on technological influence, determined through cross-validation) is used. Simultaneously, the degree characteristics (including in-degree and out-degree) and clustering coefficients of the citation network are constructed. The in-degree represents the total number of times the invention is cited by other inventions, the out-degree represents the total number of times the invention cites other inventions, and the clustering coefficient reflects the tightness of connection between inventions directly related to the invention in the citation network. The calculation formula is... , The number of actual edges between inventions directly connected to invention i. For the degree of invention i, the clustering coefficient ranges from [0,1], with a value closer to 1 indicating a denser local citation network for that invention. For timestamps, a sliding window method is used to extract time-series features. For example, setting the window size to 6 months, the number of inventions produced and the citation growth rate within each window are calculated, and these are converted into a fixed-dimensional time feature vector. This vector is then concatenated with text features and citation features to form a micro-level feature vector with a dimension of 1024. (Among them, the text features are 768-dimensional, the citation features are 200-dimensional, and the time features are 56-dimensional. The citation and time features are merged into 256-dimensional features through the feature dimension compression algorithm PCA, and finally concatenated with the text features to form 1024-dimensional features).

[0035] For mesoscopic-level feature extraction, since technology clusters are generated by clustering invention data, it is necessary to construct aggregation features of technology clusters based on microscopic-level features. First, the scale features of the technology clusters are calculated, including the number of inventions within the cluster. (Directly count the total number of patents and papers contained within the cluster), and the time span of inventions within the cluster. ( For the latest timestamp invented within the cluster, (The first time stamp is the earliest invention within the cluster). Next, the technological association characteristics within the cluster are calculated, using mean cosine similarity to represent the technological homogeneity of inventions within the cluster. The calculation formula is: , and Let be the micro-level feature vectors of invention i and invention j within the cluster, respectively. The cosine similarity function is used. The value range is [0,1], with higher values ​​indicating a more concentrated focus on technology directions within a cluster. Simultaneously, the inter-cluster association characteristics are calculated, using the Jaccard coefficient to represent the association strength between technology clusters c1 and c2, calculated as follows: , and The core technical terms sets contained in technology clusters c1 and c2 are respectively (the top 50 terms of text features in each technology cluster are extracted using the TF-IDF algorithm); finally, the growth characteristics of the technology clusters are calculated, including the average monthly growth rate of inventions within the cluster. ( and The numbers of inventions within the cluster at times t1 and t2 are respectively the monthly average growth rate of citations within the cluster. After standardizing the size characteristics, technology association characteristics, and growth characteristics (using Z-score standardization, calculated as follows): , The characteristic mean, (The feature standard deviation) is concatenated to form a mesoscopic-level feature vector with dimension 512. .

[0036] For macro-level feature extraction, the Latent Dirichlet Allocation (LDA) model is used to extract topic features for policy documents. The number of topics can be set to 10-20 (the optimal number is determined by the perplexity index), with 1000 iterations, an alpha parameter of 0.1, and a beta parameter of 0.01. The output is the probability distribution vector of each policy document across 15 topics, serving as the policy topic features. For R&D investment indicators, including national, industry, and enterprise-level R&D expenditure, number of R&D personnel, and R&D equipment procurement, Min-Max normalization is used to map them to the [0,1] interval. The normalization calculation formula is as follows: , This is the minimum value of the indicator. To calculate the maximum value of this indicator, we also calculate derived indicators such as the annual growth rate of R&D investment and the proportion of investment by each entity, forming an R&D investment feature vector. For technology paradigm labels, we use label embedding technology (such as Word2Vec) to convert discrete labels into continuous vectors. The training window size is 5, and the vector dimension is 128, ensuring that the vectors corresponding to semantically similar technology paradigm labels have high similarity. After concatenating the policy theme features (15 dimensions), R&D investment features (32 dimensions), and technology paradigm label embedding vectors (128 dimensions), we perform feature fusion through a multilayer perceptron (MLP). The MLP contains one hidden layer (128 neurons, ReLU activation function) and outputs a macro-level feature vector with a dimension of 256. .

[0037] After feature extraction at each level, feature fusion is performed based on cross-level association to generate a cross-level feature matrix. A cross-level attention mechanism is used to calculate the weights of features at each level, first constructing the hierarchical association matrix. , elements This represents the correlation strength between micro-level features and macro-level features, calculated... and The Pearson correlation coefficient was obtained, and similarly... This indicates the strength of the correlation between mesoscopic-level features and macroscopic-level features. This represents the correlation strength between microscopic and mesoscopic level features; subsequently, attention weights and microscopic level feature weights are calculated. Mesoscopic level feature weights Macro-level feature weights ,in for The average value of the correlation strength at the meso-micro level. This represents the average strength of the mesoscopic hierarchical associations. This represents the average value of the macroscopic hierarchical correlation strength. It is an exponential function; finally, the cross-level feature matrix The calculation formula is ,in , , These are projection matrices representing microscopic, mesoscopic, and macroscopic features (dimensions of 1024×2048, 512×2048, and 256×2048, respectively, which are randomly initialized and trained and optimized together with the model). , , The transposes of the feature vectors at each level are used to generate the final cross-level feature matrix. The dimension is ( The number of samples refers to the number of time slices contained in the unified database, with each time slice corresponding to the full-level data of a single time node.

[0038] S3. Based on micro-level features and reference relationships, construct a micro-dynamic network; based on meso-level features and technology cluster associations, construct a meso-dynamic network; based on macro-level features and policy associations, construct a macro-dynamic network; and perform cross-level connections between the micro-dynamic network, meso-dynamic network, and macro-dynamic network to generate a multi-level dynamic network.

[0039] Specifically, when constructing dynamic networks at each level, the spatial association and temporal dynamics of data are characterized by combining the characteristics and relationships of the corresponding levels and using graph structure modeling. Cross-level connections establish mapping rules for nodes between levels to ensure the synergy and integrity of multi-level networks, providing a structural foundation for the establishment of subsequent cross-level influence transmission mechanisms.

[0040] For the construction of micro-dynamic networks, the definitions of network nodes and edges are first clarified. Each invention (patent or paper) at the micro-level is used as a network node. The node feature vector is composed of the micro-level features extracted in step S2, including textual features of the patent / paper (such as keyword TF-IDF weights, topic distribution output by the topic model), citation features (such as citation count, citation breadth), and temporal features (such as the time-coded value corresponding to the publication timestamp). The construction of network edges is based on the citation relationships between inventions. If invention A cites invention B, a directed edge is established between the two nodes. The edge weight needs to be calculated by combining the strength of the citation relationship and the time decay effect, using the following formula: ,in The citation strength is set (value 1-5, depending on the citation location, e.g., 5 for abstract citations and 1 for bibliography citations). The publication date of invention A. The publication date of invention B. The time decay coefficient (ranging from 0.05 to 0.1, determined through fitting historical data to ensure higher weights for recent citations) is used. Simultaneously, to capture the dynamic changes in the micro-network, the network is updated according to fixed time windows (e.g., quarterly): new invention nodes published within that time window are added, newly generated citation edges are supplemented, and the weights of existing edges are adjusted according to the time decay formula. For nodes exceeding a preset time threshold (e.g., 10 years) and without new citations, their node weights are reduced to 0.1 of the initial value (retaining the nodes to maintain network integrity, but weakening their impact on current predictions), forming a micro-dynamic network that evolves over time.

[0041] The construction of the mesoscopic dynamic network uses technology clusters as nodes, and the node feature vectors are the mesoscopic-level features extracted in step S2, including the scale features of the technology cluster (such as the number of inventions within the cluster and the invention growth rate), technical features (such as the set of core keywords within the cluster and the technology maturity index), and association features (such as the frequency of cooperation with other technology clusters). Technology cluster associations serve as the basis for constructing network edges, specifically determined by calculating the feature similarity and cooperation strength between technology clusters. First, cosine similarity is used to calculate the similarity of the feature vectors of two technology clusters:

[0042]

[0043] in, , These are the feature vectors of the two technology clusters; then, the number of times inventions within the two technology clusters are jointly cited (cooperation strength) is counted. The weight of an edge is calculated by combining similarity and cooperation strength, using the following formula: ,in These are weighting coefficients (adjusted according to the characteristics of the technical field). This represents the maximum cooperation strength among all technology cluster pairs. The update cycle of the mesoscopic dynamic network is consistent with that of the microscopic dynamic network (quarterly). The updates include: merging or splitting technology clusters (when the number of inventions within a technology cluster increases by more than 50% or a significant technological branch appears, it is re-divided through K-means clustering), updating the feature vectors of technology clusters, and adjusting the weights of edges between technology clusters (based on the newly added cooperation strength and feature changes), ensuring that the mesoscopic network can reflect the evolution and changes in the associations of technology clusters.

[0044] The macro-dynamic network uses macro-level technological paradigms as nodes, and the node feature vectors are the macro-level features extracted in step S2, including policy characteristics (such as the number of supporting policies and policy strength index), input characteristics (such as total R&D investment and investment growth rate), and output characteristics (such as the average activity level and economic contribution of technology clusters under the paradigm). Policy associations serve as the foundation for constructing network edges, determined by analyzing the association statements of different technological paradigms in policy documents: text co-occurrence analysis is used to statistically analyze keywords of two technological paradigms (…). The number of times they co-occur in the same policy document (co-occurrence frequency) The formula for calculating the policy correlation strength is as follows: (This is a formula for calculating the policy correlation strength by combining the policy implementation time and the scope of its impact.) ,in The total number of policy documents, This represents the policy impact coefficient (e.g., 1.0 for national policies, 0.7 for provincial / ministerial policies, and 0.4 for municipal policies). When When the threshold is greater than a preset threshold (e.g., 0.2), an undirected edge is established between the nodes of the two technology paradigms, and the weight of the edge is equal to the policy correlation strength. The update cycle of the macro dynamic network is annual (due to the low update frequency of macro policies and R&D investment data). The update content includes: adding or eliminating technology paradigms (for example, when the R&D investment corresponding to a certain technology paradigm declines by more than 30% for two consecutive years, it is judged as an eliminated paradigm, the node is retained but marked as "decline"), updating the feature vector of the technology paradigm (based on annual policy and investment data), recalculating the policy correlation strength and adjusting the edge weights to ensure that the macro network can reflect the long-term changes in the macro technology environment.

[0045] Cross-level connections require establishing mapping relationships between micro, meso, and macro network nodes to achieve the integration of multi-level networks. Specific mapping rules include: micro-meso mapping, mapping each micro-invention node to its corresponding meso-technology cluster node (determined by the clustering results in step S1; if an invention belongs to multiple technology clusters, mapping is performed according to the cluster with the highest membership degree). Directed edges are established between corresponding micro-nodes and meso-nodes, with the edge weight equal to the invention's contribution to the technology cluster (the contribution is calculated as the proportion of the invention's citations to the total citations within the cluster, using the formula...). ,in Cited counts for microscopic inventions (Total citation count of all inventions within the technology cluster); Mesoscopic-macroscopic mapping maps each mesoscopic technology cluster node to its corresponding macroscopic technology paradigm node (determined by matching the core keywords of the technology cluster with the definition of the technology paradigm; the matching degree is calculated using the Jaccard coefficient, formula: ...). ,in This is a collection of core keywords for a technology cluster. Given a set of keywords representing technological paradigms, the technological paradigm with the highest matching degree is selected as the mapping target. Directed edges are established between the corresponding mesoscopic and macroscopic nodes. The weight of each edge equals the contribution of the technological cluster to the technological paradigm (the contribution is calculated as the proportion of the technological cluster's R&D investment to the total R&D investment of the paradigm, using the formula...). ,in For the research and development investment of technology clusters, (This represents the total R&D investment across all technology clusters within the technological paradigm). After constructing cross-level edges, the nodes and edges of the micro, meso, and macro networks are integrated to generate an adjacency matrix for a multi-level dynamic network. The matrix dimension is the total number of nodes (number of micro nodes + number of meso nodes + number of macro nodes) × total number of nodes, and the matrix elements are the weights of the edges between corresponding nodes (weights are 0 when there is no connection). Simultaneously, the network's temporal attributes (update time, time window) are recorded to provide structured data for subsequent model input.

[0046] S4. Based on a multi-level dynamic network, establish a cross-level influence transmission mechanism; among which, the cross-level influence transmission mechanism includes the diffusion process of science and technology from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro.

[0047] Specifically, after completing the construction of the multi-level dynamic network, a cross-level influence transmission mechanism is established through quantitative modeling and logical mapping. This mechanism fully depicts the bidirectional interaction between science and technology at the micro, meso, and macro levels. Specifically, it includes the diffusion process from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro. Each process must be implemented through a computable mathematical model and an executable algorithm flow to ensure the quantification and traceability of the transmission effect.

[0048] The core of the diffusion process from the micro to the mesoscopic level is to describe how inventions at the micro level penetrate into mesoscopic technology clusters through technological linkages, triggering changes in the characteristics of these clusters. First, it's necessary to define the diffusion triggering condition: when the core technological features of a micro-invention (such as keyword weights in patent texts or the citation growth rate of papers) exceed a preset threshold, the invention is deemed to have diffusion potential. Here, the TF-IDF algorithm is used to calculate the weights of keywords in the invention text; the formula is as follows: ,in Indicates the first The first invention The weight of each keyword, For the first The first invention The frequency of each keyword To unify the total amount of invention data in the database, For including the first The number of inventions for each keyword. For example, when When the value of the core keyword in a technology cluster exceeds 1.2 times the average weight determined by K-means clustering, the diffusion process is triggered. Next, the correlation between the invention and each technology cluster at the mesoscopic level is calculated using cosine similarity, with the formula: ,in The feature vector for micro-inventions (composed of keyword weights, citation strength, etc.). This represents the central feature vector of a mesoscopic technology cluster (composed of the average of all invention feature vectors within that cluster). The top three technology clusters with the highest similarity can be selected as diffusion targets, and then the feature vectors of these clusters can be updated. ,in The weighting coefficient (dynamically adjusted based on the number of inventions within the technology cluster; the more inventions, the better). The larger the value, the more stable the characteristics of the technology cluster. Simultaneously, the timestamp of the diffusion and its intensity (determined by the cosine similarity value) are recorded to provide data support for subsequent dynamic network updates.

[0049] The mesoscopic-to-macroscopic phase transition process aims to capture how the collective changes of mesoscopic technology clusters break through critical states, triggering a shift in macroscopic technological paradigms. First, a technology cluster aggregation index is constructed to quantify the degree of synergistic development of technology clusters at the mesoscopic level. The formula is as follows: ,in The total number of mesoscopic-level technology clusters, For the first The technology cluster and the first Cosine similarity of feature vectors of a technology cluster. For example, when When the growth rate exceeds 5% for three consecutive time windows, the phase transition monitoring stage begins. At this point, a mutation detection algorithm (such as the PELT algorithm) is introduced to analyze the changing trends of macro-level technological paradigm characteristics. Taking R&D investment as an example, the second derivative of the R&D investment growth rate is calculated. (in The amount invested in research and development, (for time), when When the absolute value of the data exceeds a preset threshold (obtained through training with historical technology paradigm shift data) and persists for two consecutive time units, it is determined that the aggregation of mesoscopic technology clusters has triggered a phase transition in the macroscopic technology paradigm. Simultaneously, association rule mining algorithms (such as the Apriori algorithm) are used to find association rules between mesoscopic technology clusters and macroscopic policy documents and R&D investment indicators before and after the phase transition. The minimum support is set to 0.3, and the minimum confidence is set to 0.7. This allows for the selection of technology cluster combinations that contribute the most to the phase transition, thus clarifying the phase transition path from mesoscopic to macroscopic.

[0050] The feedback process from macro to micro levels mainly reflects the guiding role of macro policies and R&D investment on micro-level invention activities. First, semantic analysis is performed on macro-level policy documents, and a BERT pre-trained model is used to extract policy keywords and their semantic weights, generating policy feature vectors. The R&D investment indicators are categorized by technology field, and the R&D investment ratio for each technology field is calculated to generate an R&D investment feature vector. Subsequently, and By fusing the data, we obtain the macroscopic feedback feature vector. ,in The policy impact weights are adjusted based on the strength of policy implementation. Next, the matching degree between the macro-level feedback feature vector and the micro-level invention feature vector is calculated using the following formula:

[0051]

[0052] For example when When the value is greater than 0.6, the micro-invention is determined to be affected by macro-feedback. The node weights of the invention in the micro-dynamic network are then adjusted accordingly. ,in The original node weight is determined by factors such as the number of citations and publication time. This is the feedback impact coefficient (adjusted according to the growth rate of R&D investment). Through the above process, quantitative feedback of macro-level factors on micro-level invention activities is achieved, improving the cross-level impact transmission mechanism.

[0053] S5. Based on cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms, it uses a pre-trained machine learning model to predict technology trends and outputs predictions of technology cluster activity and the probability of technology paradigm shifts.

[0054] Specifically, this step uses the outputs of cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms as joint inputs to achieve quantitative prediction of science and technology trends through a pre-trained machine learning model. The core is to build a model architecture that adapts to multi-level data structures and dynamic transmission relationships to ensure the accuracy and interpretability of the prediction results. At the same time, it clarifies the quantitative form and verification standards of the prediction output to meet the needs of science and technology decision-making for quantitative analysis tools.

[0055] First, model selection and pre-training were performed. Considering the graph structure characteristics of multi-level dynamic networks and the high-dimensionality of cross-level feature matrices, a fusion model of Graph Neural Networks (GNN) and Long Short-Term Memory Networks (LSTM) (GNN-LSTM) was chosen as the basic architecture: the GNN part is used to extract the topological features of nodes and edges in the multi-level dynamic network and the inter-level correlation information, while the LSTM part is used to capture the time-series change trends of data at each level. The fusion of the two can simultaneously handle spatial correlation and temporal dynamic features. The model pre-training process used historical data from a unified database as the training set, with a time span of the past 10 years of scientific and technological data (ensuring coverage of at least one complete technological paradigm shift cycle). The data was divided into training, validation, and test sets in a 7:2:1 ratio. The inputs during the pre-training phase include the historical cross-level feature matrix (divided into multiple time steps according to time windows), the adjacency matrix of the historical multi-level dynamic network (updated with each time step), and the output parameters of the historical cross-level influence transmission mechanism (such as diffusion intensity, phase transition index, and feedback weights). The training objective is to minimize the error between the predicted and actual values ​​(such as the activity level of historical technology clusters and the occurrence of historical technology paradigm phase transitions). A hybrid loss function is used, and the formula is as follows: ,in Mean squared error (used to calculate the predicted value of technology cluster activity) Compared with the true value (error) Binary cross-entropy (used to calculate the predicted probability of a phase transition in a technological paradigm) Compared with the true value The error, This indicates that a phase transition has occurred. (This indicates that no phase transition occurred). and The weights are determined by searching the validation set error grid. The pre-training process uses the Adam optimizer, with the initial learning rate set to 0.001. When the validation set error does not decrease for 5 consecutive epochs, the learning rate is reduced to 0.5. Pre-training stops when the learning rate is below 1e-5 or the preset number of training epochs is reached, and the pre-trained model parameters are saved.

[0056] Next, input processing is performed for the prediction phase. The cross-level feature matrix at the current time step is standardized. For each feature dimension in the matrix... According to the formula Normalization is performed, where For the original value of the feature, This represents the mean of this feature dimension across the historical training set. This represents the standard deviation of this feature dimension across the historical training set, preventing interference from features of different magnitudes on the model's predictions. For the current multi-level dynamic network, its adjacency matrix is ​​constructed. ( Represents a node With nodes There is a connection. (indicating no correlation), and introduces a self-loop term ( To preserve the node's own characteristics, real-time output parameters of the cross-level influence transmission mechanism (such as the diffusion intensity from the micro to the meso, the aggregation degree from the meso to the macro, and the feedback weight from the macro to the micro) are embedded as additional features into the node feature vectors of the corresponding levels (e.g., diffusion intensity as an additional dimension for micro nodes and aggregation degree as an additional dimension for meso nodes). The processed cross-level feature matrix, adjacency matrix, and additional features are organized into a model input sequence according to time steps. The time step length of the input sequence can be set to 12 (corresponding to 12 monthly data, determined according to the update frequency of scientific and technological data) to ensure that the model can capture recent dynamic trends.

[0057] Then, prediction calculations and result output are performed. The GNN part first extracts features from a multi-level dynamic network through graph convolutional layers. The formula for calculating layer graph convolution is: ,in For the first The node feature matrix of the layer ( (where is the identity matrix) is an adjacency matrix with self-loops. For the first The trainable weight matrix of the layer, The activation function is ReLU. After three layers of graph convolution, the node feature matrices of each layer are flattened and concatenated with the cross-layer feature matrices to obtain the fused feature vector. The LSTM part inputs the fused feature vector step by step. The calculation formula of the LSTM unit includes: forget gate. Input gate Cell state update Cell state Output gate Hidden state ,in For the first The fused feature vectors at each time step For the first The hidden state of the time step. This is the weight matrix. For bias terms, This is element-wise multiplication. After passing through a fully connected layer, the output of the LSTM is divided into two branches: one branch outputs the activity level of the technology cluster through the Sigmoid activation function (the value ranges from 0 to 1, and the larger the value, the higher the activity level of the technology cluster in the future, usually with a prediction period of 6 months); the other branch outputs the probability of a technology paradigm shift through the Softmax activation function (the value ranges from 0 to 1, for example, when the probability value exceeds 0.7, it is determined that there is a risk of a technology paradigm shift in the next 12 months).

[0058] The aforementioned machine learning-based technology trend prediction method integrates micro-level invention data, meso-level technology cluster data, and macro-level technology paradigm data to form a unified database. It then extracts and fuses features from each level to construct a cross-level feature matrix. Simultaneously, based on features and correlations, it constructs dynamic networks at the micro, meso, and macro levels respectively and connects them across levels, forming a multi-level dynamic network that characterizes interactions between levels. Furthermore, it establishes a mechanism to describe the transmission of cross-level influences (including diffusion, phase transition, and feedback processes). Finally, it comprehensively utilizes the aforementioned cross-level feature matrix, multi-level dynamic network, and cross-level influence transmission mechanism to drive a pre-trained machine learning model for prediction. This series of coherent steps collectively achieves a systematic modeling of the multi-level, nonlinear evolution of technology trends, enabling a more comprehensive and accurate simulation of the cross-level transmission and amplification effects of external shocks in the technology ecosystem, and predicting the technological effects of complex emergent phenomena such as technology path locking and technology paradigm phase transitions.

[0059] In one optional embodiment, a micro-dynamic network is constructed based on micro-level features and referencing relationships; a meso-dynamic network is constructed based on meso-level features and technology cluster associations; a macro-dynamic network is constructed based on macro-level features and policy associations; and cross-level connections are made between the micro-dynamic network, the meso-dynamic network, and the macro-dynamic network to generate a multi-level dynamic network, including the following steps:

[0060] S11. Based on the micro-level features and citation relationships, construct a micro-citation network; based on the meso-level technology cluster associations, construct a meso-network; based on the macro-level technology paradigm associations, construct a macro-network. In the micro-citation network, nodes are patents or papers, and edges represent citation relationships. In the meso-network, nodes are technology clusters, and edge weights are calculated by aggregating the citation counts of their respective micro-entities. In the macro-network, nodes are technology paradigms or policy entities, and edge weights are calculated based on the co-occurrence frequency of policy documents or the relevance of R&D investment.

[0061] Specifically, the nodes of the micro-citation network directly map to the core entities at the micro-level, namely patents or papers. Each node needs to be bound to corresponding micro-level features (such as text features and timestamp features), and the edges of the network are generated entirely based on citation relationships. If patent A cites patent B, paper C cites paper D, or patent E cites paper F, then an undirected edge is established between the corresponding nodes (the citation relationship has bidirectional association attributes, which can reflect the inheritance and connection of technology).

[0062] The nodes of the mesoscopic network are technology clusters (generated by clustering micro-invention data). The weight calculation of the edges relies on the association data of micro-entities. Specifically, it is achieved by aggregating the number of citations of micro-entities between technology clusters: For any two technology clusters T1 and T2 in the mesoscopic network, the total number of times all patents / papers in T1 cite patents / papers in T2, and the total number of times all patents / papers in T2 cite patents / papers in T1 are counted. The average of the two is taken as the weight of the edge between T1 and T2, thereby quantifying the association strength between technology clusters.

[0063] The nodes of the macro network encompass both technological paradigms and policy entities. The calculation of edge weights needs to consider the characteristics of macro-level data: If two macro nodes are policy entities, the weights are determined based on the co-occurrence frequency of policy documents. The number of times two types of policy documents appear simultaneously in the same public report or planning document is counted, and then divided by the average of the total number of times each type of policy document appears, to obtain the co-occurrence frequency, which is used as the edge weight. If two macro nodes are technological paradigms and policy entities, or both are technological paradigms, the weights are calculated based on the correlation between R&D investment. The time series data of R&D investment in the technological fields corresponding to the two types of macro nodes are obtained, and the Pearson correlation coefficient between the two is calculated. The absolute value of the correlation coefficient is used as the edge weight, thereby reflecting the closeness of the relationship between macro entities.

[0064] S12. Based on the micro-reference network and the meso-network, establish micro-to-meso-connections and generate membership edges; wherein, the membership edges connect each micro-entity to its respective technology cluster node.

[0065] Specifically, the clustering results of each technology cluster in the mesoscopic network are retrieved first. This data records the list of micro-entities (patents / papers) contained in each technology cluster. The list is generated based on the similarity between the characteristics of the micro-entities and the characteristics of the technology cluster center (the similarity must meet the threshold requirements set during clustering).

[0066] Based on this list, unidirectional membership edges are established between nodes in the micro-reference network and nodes in the meso-reference network: each micro-entity node (patent / paper) is only connected to the node of its unique technology cluster, and does not establish connections with nodes of other technology clusters, thus ensuring the uniqueness of the affiliation relationship. No additional weights are required for the membership edges (the affiliation relationship is deterministic, not probabilistic); the existence of the edges only needs to reflect the core logic that "the micro-entity belongs to a certain technology cluster," laying the foundation for subsequent cross-level data transmission.

[0067] S13. Based on mesoscopic and macroscopic networks, establish connections from mesoscopic to macroscopic levels to generate influence edges. The influence edges connect each technology cluster node to the relevant macroscopic node, and the weight of the influence edges is calculated based on policy support.

[0068] Specifically, the policy entities in the macro network are first analyzed to extract the technical fields and technical directions explicitly supported in the policy documents. These are then matched with the core technical characteristics (such as technical keywords and technical function descriptions) of each technology cluster in the meso network to screen out macro nodes (technical paradigms or policy entities) that have a supporting relationship with a certain technology cluster.

[0069] The weights of the influence edges are then calculated. Policy support is determined by a combination of three dimensions: the frequency of mentions of the core features of the corresponding technology cluster in policy documents, the scale of funding for policy implementation, and the geographical scope of policy coverage. The weighted average of these three dimensions (frequency normalized to the 0-1 range, funding normalized to the industry average, and geographical scope normalized to the percentage of provinces covered) is then calculated. The result represents the weight of the influence edge between the technology cluster node and its corresponding macro-level node. A larger weight indicates a stronger influence of the macro-level entity on the technology cluster. Each technology cluster node must establish influence edges with all macro-level nodes with which it has a supporting relationship to ensure complete coverage of the influence path from the meso-to-macro level.

[0070] S14. Based on the macro network and the micro reference network, establish a macro-to-micro feedback connection and generate feedback edges; wherein, the feedback edges connect macro nodes to micro nodes, and the weight values ​​of the feedback edges are calculated based on the historical impact of policy intensity on the invention rate.

[0071] Specifically, the first step is to collect historical data, such as policy data from the past 5-10 years (policy texts, implementation time, and quantitative values ​​of policy intensity, such as policy level and funding) and micro-invention data (the number of patent / paper applications / publications within the corresponding time period, i.e., the invention rate).

[0072] Statistical analysis is used to establish the correlation between policy intensity and invention rate. For a given macro-level node (policy entity or technological paradigm), the micro-level entities affected by it are selected, and the change in invention rate in that domain before and after policy implementation is calculated. The ratio of this change to the quantified policy intensity value is used as the historical impact coefficient. The weight of the feedback edge is determined based on this historical impact coefficient. If a micro-level entity belongs to a domain affected by the macro-level node and its historical impact coefficient is positive (the policy promotes an increase in the invention rate), a feedback edge is established between the macro-level node and that micro-level entity node, with the edge weight equal to the historical impact coefficient (normalized to the 0-1 range). If the historical impact coefficient is negative or has no significant impact, no feedback edge is established. Each macro-level node needs to establish feedback edges with all micro-level entities affected by it, forming a closed-loop feedback path from macro to micro.

[0073] S15. Combine the micro-reference network, meso-network, macro-network, membership edge, influence edge, and feedback edge to generate a multi-level dynamic network.

[0074] Specifically, the combination process ensures that the network at each level and the cross-level edges form an organic whole and can reflect the "dynamic" attribute. First, the micro-reference network, meso-network, and macro-network are spatially divided according to the level (e.g., the micro-network is at the bottom level, the meso-network is in the middle level, and the macro-network is at the top level), while keeping the original structure of the nodes and edges within each level unchanged.

[0075] Subsequently, membership edges, influence edges, and feedback edges are embedded between levels according to their corresponding connection relationships. Membership edges vertically connect bottom-level micro-nodes with intermediate-level meso-nodes, influence edges vertically connect intermediate-level meso-nodes with top-level macro-nodes, and feedback edges vertically connect top-level macro-nodes with bottom-level micro-nodes, forming cross-level connection paths. Simultaneously, a time dimension marker needs to be added to the entire combined network to record the timestamps of each edge's generation (e.g., the time of reference relationship occurrence, policy implementation time). The network structure can then be updated (adding nodes / edges, adjusting edge weights) via time window sliding (e.g., by quarter, semi-annual) to achieve a "dynamic" characteristic. The resulting multi-level dynamic network encompasses both the technical connections within each level and the cross-level influence and feedback paths, comprehensively depicting the multi-level interactive relationships of the technology ecosystem.

[0076] refer to Figure 2 In one optional embodiment, a cross-level influence transmission mechanism is established based on a multi-level dynamic network, including the following steps:

[0077] S21. Based on a multi-level dynamic network, a diffusion mechanism from the microscopic to the mesoscopic level is constructed by describing the change in the activity of a technology cluster over time through a coupled diffusion equation; the expression of the coupled diffusion equation is:

[0078]

[0079] in, Representing technology clusters In time activity level Representing technology clusters The neighborhood group, Representing technology clusters and The influence weights between them This indicates the number of new inventions input from external sources. For diffusion intensity parameters, For input rate parameters, This is the attenuation coefficient parameter.

[0080] Specifically, this step aims to characterize the penetration process of new inventions at the micro level into meso-level technology clusters through a quantitative model. The core is to capture the dynamic changes in the activity of technology clusters using a coupled diffusion equation, ensuring that the diffusion process matches the correlations between technology clusters in a multi-level dynamic network. The mechanism's function is to connect micro-level invention activities with the development of meso-level technology clusters, transforming the input of new inventions at the micro level into signals of changes in the activity of meso-level technology clusters. In its implementation, the neighbor set of technology clusters is determined based on a multi-level dynamic network. This set consists of network clusters of technologies Other technology clusters that are directly related (e.g., similar technical features, shared micro-invention resources) constitute the group, ensuring that neighbor relationships accurately reflect the mutual influence between technology clusters. In the equation... As a core output metric, its initial value can be set based on historical activity data of technology clusters in a multi-level dynamic network (such as the average activity level over the past 3 time units); influencing weights Extracted from the edge weights of a multi-level dynamic network, and updated in real time with the dynamic changes in the network topology, to reflect the time dependence of the correlation strength between technology clusters; external input: number of new inventions. These are directly added from the unified database and belong to the technology cluster. Patent and paper data are generated statistically by time window (e.g., monthly). Parameters , , Calibrated based on historical data, (Diffusion intensity parameter) controls the influence of neighboring technology clusters on the target technology cluster; the closer the correlation between technology clusters, the greater the impact. The larger the value; (Input rate parameter) reflects the efficiency of new inventions in contributing to activity level; the faster the innovation and iteration in the technological field, the higher the efficiency. The larger the value; (Decrease coefficient parameter) describes the rate at which activity naturally decays over time; the higher the technological maturity, the better. The larger the value, the more accurate the calculation. This coupled diffusion equation allows for the quantitative calculation of the change in the activity of technology clusters within each time unit, enabling precise transmission of the influence of microscopic inventions to mesoscopic technology clusters.

[0081] S22. Based on a multi-level dynamic network, the phase transition probability of the technological paradigm is described by the sigmoid function, and a phase transition mechanism from mesoscopic to macroscopic is constructed; the expression for the phase transition probability is:

[0082]

[0083] in, Indicates time The technological paradigm of phase transition probability, It is the sigmoid function. It is a technology cluster The weight, It is the impact coefficient of macroeconomic policies. Indicates time The policy intensity indicator.

[0084] Specifically, the core objective of this step is to combine the activity levels of multiple technology clusters at the mesoscopic level with macroeconomic policy factors. By using the sigmoid function, multi-dimensional inputs are transformed into technological paradigm shift probabilities within the 0-1 range, achieving a quantitative mapping from mesoscopic to macroscopic impacts. The phase shift mechanism determines whether the collective activity of mesoscopic technology clusters has reached a critical state that triggers a macroeconomic technological paradigm shift. During implementation, the sigmoid function... The selection of must satisfy smoothness and boundary constraints, and its standard form is as follows: This ensures the output phase transition probability. It consistently remains between 0 (no risk of phase transition) and 1 (extremely high probability of phase transition), meeting the basic requirements of probability statistics. (Technology cluster weight) The strength of the association between technology clusters and macro-level technological paradigms in a multi-level dynamic network is determined by the strength of the association (e.g., a high degree of overlap between the core technologies of the technology cluster and the core directions of the paradigm). The larger the value, the better, and the sum of the weights of all technology clusters must be normalized to 1 to ensure a reasonable distribution of the contributions of each technology cluster; Macroeconomic policy impact coefficient This is used to adjust the degree of influence of policy factors in phase transitions; the stronger the policy's guiding role in technological development (such as national-level special policies), the better. The larger the value, the stronger the policy intensity indicator. Extracted from macro-level node data in a multi-level dynamic network, and transformed into quantifiable values ​​in the 0-1 range through semantic analysis of policy documents (such as policy objective clarity and resource commitments), this phase transition probability formula is dynamically updated over time. The core logic of this formula lies in: firstly, through... The combined active contributions of the computational mesoscopic technology cluster, superimposed It reflects the driving role of macro policies, and finally maps the comprehensive input to phase transition probability through the sigmoid function, effectively characterizing the nonlinear process of the collective change at the mesoscopic level to the macroscopic paradigm transition.

[0085] S23. Based on a multi-level dynamic network, the impact of policies on the invention rate is described through a reaction-diffusion equation, constructing a feedback mechanism from macro to micro levels; the expression of the reaction-diffusion equation is:

[0086]

[0087] in, Indicates time Micro-invention rate For the parameter of policy impact intensity, This is the attenuation rate parameter.

[0088] Specifically, this step aims to establish a reverse influence channel from macro policies to micro-level invention activities. It uses a reaction-diffusion equation to quantitatively describe how macro policies drive changes in micro-level invention rates, thus improving the two-way transmission of cross-level influence. The core function of the feedback mechanism is to translate macro-level policy guidance into quantitative changes in micro-level invention activities, ensuring that macro-control directly impacts the source of technological innovation. In the reaction-diffusion equation, the micro-level invention rate... This refers to the number of newly added patents and papers at the micro-level per unit of time. Its initial value is set based on historical invention rate data in a unified database to ensure consistency with the baseline of actual invention activities; policy impact intensity parameter. This measure assesses the impact of a unit of policy intensity on the invention rate. It is calibrated through correlation analysis of historical policy implementation effects and changes in the invention rate; the more significant the increase in the invention rate after policy implementation, the better. The larger the value, the higher the attenuation rate parameter. This describes the natural rate of decline in the invention rate without policy support, reflecting the inertia of technological innovation. The higher the technological barrier to innovation, the better. The larger the value, the more significant the change. The physical meaning of this equation is that changes in the micro-invention rate are positively driven by macro-policies. Item) and natural decay ( The invention rate tends to rise when the policy-driven effect outweighs the natural decay effect, and falls when the policy-driven effect outweighs the natural decay effect, thus accurately reflecting the dynamic feedback effect of macro policies on micro invention activities.

[0089] S24. Based on a multi-level dynamic network, a feedback loop mechanism is constructed by weighting the sum of the policy intensity index of the previous time step and the activity of the technology cluster of the previous time step to generate the policy intensity index of the current time step; the expression of the feedback loop mechanism is:

[0090]

[0091] in, This is the feedback coefficient.

[0092] Specifically, the purpose of this step is to construct a closed-loop feedback mechanism that influences across different levels, enabling adjustments to the intensity of macroeconomic policies to be dynamically optimized based on the activity of historical technology clusters, and ensuring that the entire transmission mechanism is adaptive and dynamically balanced. This represents the quantitative value of policy intensity in the macro network for the previous time unit. This represents the activity data of each technology cluster in the mesoscopic network in the previous time unit. Feedback coefficient. This is used to control the impact of historical technology cluster activity on the strength of current policies, and is set according to the sensitivity requirements of policy adjustments: if rapid policy response to changes in technology clusters is required, Take the larger value; if policy stability is required, Take the smaller value. The core logic of this feedback loop formula is that the current policy strength not only inherits the policy basis from the previous moment ( (Item), and will also be based on the activity performance of the technology cluster in the previous moment ( Adjustments will be made to the items. If the technology cluster as a whole was active in the previous moment ( If the value is large, the current policy intensity will be increased accordingly to further strengthen support for technological innovation; if the activity level of the technology cluster is low, the policy intensity adjustment will be reduced or remain stable, forming a closed loop of "policy promotes the activity of technology clusters → the activity of technology clusters feeds back into policy adjustments".

[0093] S25, integrating diffusion mechanism, phase transition mechanism, feedback mechanism and feedback loop mechanism, to obtain the cross-level influence transmission mechanism.

[0094] Specifically, this step integrates the aforementioned four sub-mechanisms to form a cross-level influence transmission system covering the complete path from "micro to meso to macro to micro," ensuring smooth data flow and tight logical connections between the mechanisms. During the integration process, the input-output relationships of each mechanism are first clarified: the output of the diffusion mechanism (S21)... (Technology cluster activity) serves as the core input to the phase transition mechanism (S22). (Item), input to the phase transition mechanism The (policy strength indicator) originates from the output of the feedback loop mechanism (S24) and the input of the feedback mechanism (S23). Similarly, the output of the feedback loop mechanism and the input of the feedback loop mechanism also depend on the feedback loop mechanism. This data originates from the output of the diffusion mechanism at the previous moment, forming a complete data flow chain. The integrated cross-level influence transmission mechanism can simultaneously characterize the diffusion of micro-level new inventions to the meso-level, the phase transition from meso-level activity to the macro-level, the feedback from macro-level policies to the micro-level, and the closed-loop adjustment between policies and technology clusters. It comprehensively reflects the nonlinear interaction between various levels of the science and technology system, providing scientific and coherent mechanism support for subsequent science and technology trend prediction.

[0095] In one optional embodiment, based on cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms, a pre-trained machine learning model is used to predict technology trends, outputting predictions of technology cluster activity and probabilities of technology paradigm shifts. This includes the following steps:

[0096] S31. Based on the micro-level features in the micro-dynamic network and cross-level feature matrix, micro-node embeddings are generated through graph convolutional networks; the operation of the graph convolutional network is represented as follows:

[0097]

[0098] in, Indicates time The micro-node embedding matrix, For the normalized microadjacency matrix, For time Micro-level characteristics, This is a trainable weight matrix.

[0099] Specifically, the micro-dynamic network provides the connection relationships between nodes, based on which an adjacency matrix is ​​constructed and normalized to obtain... Normalization can eliminate the interference of node degree differences on feature extraction, and is usually used. The way ( The degree matrix of the nodes. (as an identity matrix), ensuring a more balanced feature weight for nodes of different degrees; It extracts micro-attributes, including textual features and citation features of patents / papers, directly from the cross-level feature matrix; The trainable weight matrix for the graph convolutional layer is initialized using the Xavier initialization method and subsequently updated in reverse with the prediction error during model training.

[0100] Graph convolution operations are performed through... A linear fusion of topology and node features is achieved, followed by the introduction of nonlinearity through the ReLU activation function, resulting in the final output. (time (The micro-node embedding matrix). The choice of the ReLU function can effectively alleviate the gradient vanishing problem and ensure the feature transfer efficiency of deep networks; Each row vector in the model corresponds to an embedded representation of a micro-node (patent or paper), which includes the node's own attribute information and integrates the association features of its neighboring nodes in the micro-dynamic network, providing high-quality micro-foundation data for subsequent meso-level feature aggregation.

[0101] S32. Aggregate the micro-node embeddings to generate meso-level embeddings; wherein, the aggregation process is achieved by calculating the average embedding of the micro-nodes to which each technology cluster belongs.

[0102] Specifically, this step aims to transform the micro-level node embeddings into meso-level technology cluster feature representations. The core is to achieve cross-scale feature integration through "attribution mapping-average aggregation". Before implementation, it is necessary to clarify the attribution relationship between micro-nodes and meso-level technology clusters. This relationship originates from the logic in the previous data processing that "technology clusters are obtained by clustering invention data". A one-to-one mapping table is established to ensure that each micro-node (patent / paper) can be accurately associated with its technology cluster.

[0103] The aggregation process employs an average aggregation strategy, specifically as follows: For each technology cluster, all micro-nodes belonging to that cluster in the mapping table are selected, and these nodes are extracted... The mesoscopic embedding of the technology cluster is obtained by calculating the arithmetic mean of each dimension of the embedding vector. The core reason for choosing average aggregation is that it can evenly reflect the feature contributions of all micro-nodes within the cluster, avoiding the excessive influence of a single node (such as a highly cited patent) on the cluster features. It also has low computational complexity, strong interpretability, and facilitates subsequent model processing and result verification. During implementation, attention should be paid to data integrity verification. If the number of micro-nodes under a certain technology cluster is too small (e.g., less than 3), the average of the historical micro-node embeddings of that technology cluster can be added for smoothing to ensure the stability and effectiveness of the mesoscopic embedding.

[0104] S33. Based on the mesoscopic layer embedding and the mesoscopic layer features in the cross-layer feature matrix, the mesoscopic layer state vector is generated by processing through a long short-term memory network.

[0105] Specifically, this step uses a Long Short-Term Memory (LSTM) network to capture the temporal dynamics of mesoscopic features, generating a mesoscopic layer state vector containing temporal dimension information. The core advantage of LSTM lies in its ability to effectively handle long-term dependencies in sequential data, aligning with the dynamic characteristics of technological trends evolving over time. Mesoscopic layer embeddings reflect the spatial correlation characteristics of technology clusters, while mesoscopic layer features in the cross-level feature matrix (such as the historical activity of technology clusters and the proportion of core technologies) reflect their attribute characteristics. Combining these two as input to the LSTM enables the initial fusion of three-dimensional features: space, attribute, and time.

[0106] In implementation, the mesoscopic layer embeddings and mesoscopic layer features are first concatenated along the feature dimension to form the mesoscopic layer input sequence (the sequence length is set according to the time window, such as feature data from the past 6 time units). This sequence is then input into an LSTM network. The LSTM selectively retains key historical features and updates current feature information through the synergistic action of the forget gate, input gate, and output gate: the forget gate controls the discarding of irrelevant information from historical states, the input gate determines the retention ratio of current input features, and the output gate generates the mesoscopic layer state vector at the current time step. This state vector not only integrates the spatial and attribute features of the mesoscopic layer, but also contains trend information on the changes of technology cluster features over time (such as the rate of activity growth and changes in technology direction), laying the foundation for subsequent correlation analysis with macroscopic layer features.

[0107] S34. Based on the macroscopic hierarchical features in the cross-level feature matrix, a long short-term memory network is used for processing to generate macroscopic layer state vectors.

[0108] Specifically, this step follows the same logic as S33, focusing on the time-dynamic extraction of macro-level features and generating macro-level state vectors using LSTM. Macro-level features originate from cross-level feature matrices, including quantitative features of policy documents, trends in R&D investment, and technological paradigm maturity. These features possess significant time-series attributes (such as adjustments in policy intensity over time and annual fluctuations in R&D investment), and their long-term patterns of change are captured using LSTM.

[0109] In implementation, the macro-level features are first constructed sequentially by organizing macro-level feature data from multiple past time units in chronological order (the sequence length remains consistent with S33 to ensure time synchronization for subsequent feature associations), forming the macro-level input sequence. This sequence is then input into an independent LSTM network (without sharing LSTM parameters with S33 to avoid interference between macro- and meso-level features). The LSTM's gating mechanism filters out short-term noise in the macro-level features (such as abnormal fluctuations in R&D investment within a specific time unit), while retaining core trend information (such as the long-term growth direction of R&D investment and the duration of policy impact). Finally, the macro-level state vector is output. This vector can accurately reflect the temporal dynamic characteristics of macroeconomic factors (such as policy guidance and the overall R&D environment), providing a basis for measuring the strength of the correlation between macroeconomic and mesoeconomic factors.

[0110] S35. Based on the mesoscopic layer state vector and the macroscopic layer state vector, the attention weight is calculated using an attention mechanism; the formula for calculating the attention weight is:

[0111]

[0112] in, This represents the mesoscopic layer state vector corresponding to the i-th technology cluster in the mesoscopic layer. The macro-level state vector corresponding to the j-th technological paradigm or policy entity in the macro-level. Attention weights between them.

[0113] Specifically, this step quantifies the strength of the correlation between mesoscopic technology clusters and macroscopic factors through an attention mechanism. The core is to highlight macroscopic-mesoscopic correlation pairs that have a more significant impact on technological trends through weight allocation. In implementation, the computational object of the attention mechanism is first defined: the set of mesoscopic layer state vectors. ( (Number of technology clusters) and set of macroscopic layer state vectors ( This refers to the quantity of macro-level entities, such as policies and technological paradigms.

[0114] Attention weight The calculation follows the "dot product-softmax" logic, first calculating the... Mesoscopic state vector of a technology cluster With the Macro-level state vector of a macro entity The larger the dot product, the higher the feature similarity and the stronger the correlation between the two. Then, for each mesoscopic layer state vector, the exponent of its dot product with all macroscopic layer state vectors is taken (to avoid the influence of negative values), and then divided by the sum of the exponents of the dot products of all macroscopic layer state vectors. The weights are normalized by the softmax function to ensure that the sum of all attention weights corresponding to each mesoscopic layer state vector is 1. The physical meaning is "macroscopic entity". technology cluster The higher the weight of the macro entity, the more significant its influence on the trend evolution of the corresponding technology cluster, providing a priority basis for subsequent feature fusion.

[0115] S36. Based on the attention weight, the state vectors of the mesoscopic layer and the macroscopic layer are concatenated to obtain the fused feature vector; the fused feature vector is input into the fully connected layer for nonlinear transformation to generate the initial prediction vector; the initial prediction vector is constrained and optimized based on the cross-level influence transmission mechanism to output the technology cluster activity prediction and the technology paradigm phase transition probability.

[0116] Specifically, this step is the core of the prediction result generation process, achieving the transformation from features to prediction results through "feature fusion - nonlinear transformation - constraint optimization". First, feature fusion is performed: for each mesoscopic layer state vector... This is compared with all macroscopic layer state vectors. By attention weight A weighted summation is performed to obtain the macroscopic correlation features of the technology cluster. Subsequently, the mesoscopic state vector is concatenated with the corresponding macroscopic correlation features to form a fused feature vector. This vector contains both the temporal dynamic features of the mesoscopic technology cluster itself and integrates the weighted influence of macroscopic factors, achieving deep fusion of cross-level features.

[0117] Next, the fused feature vector is input into the fully connected layer. The fully connected layer maps the fused features to the prediction dimension (e.g., technology cluster activity corresponds to a 1D output, and technology paradigm phase transition probability corresponds to a 1D output) through a trainable weight matrix. Simultaneously, a nonlinear transformation is performed using a sigmoid or tanh activation function to generate an initial prediction vector. The initial prediction vector may deviate from the actual laws of technological development; therefore, a cross-level influence transmission mechanism is introduced for constraint optimization. Specifically, the initial prediction result is checked for consistency with the diffusion mechanism (e.g., the diffusion law of technology cluster activity) and the phase transition mechanism (e.g., the reasonable range of phase transition probability) within the cross-level influence transmission mechanism. If the initially predicted technology cluster activity exceeds the reasonable range derived from the diffusion equation, or the phase transition probability does not conform to the phase transition law characterized by the sigmoid function, the weights of the fully connected layer are fine-tuned through gradient backpropagation until the prediction result satisfies the constraints of the transmission mechanism.

[0118] The final output prediction results consist of two parts: the prediction of the activity of technology clusters (within the range of 0-1, the larger the value, the more active the innovation activities of the technology cluster will be in the future) and the probability of technological paradigm shift (within the range of 0-1, the closer the value is to 1, the higher the possibility of technological paradigm change in the future). Both have been verified for the rationality of cross-level influence transmission mechanism, and can provide quantitative and institutional decision-making basis for science and technology strategy formulation and R&D planning.

[0119] In one alternative embodiment, the total loss function used when training the machine learning model consists of a regression loss function, a classification loss function, and a dynamic constraint loss function;

[0120] The expression for the regression loss function is:

[0121]

[0122] in, For regression loss function, It is a prediction of the activity level of technology clusters. It reflects the actual activity level of the technology cluster. It is the number of training samples;

[0123] The expression for the classification loss function is:

[0124]

[0125] in, For classification loss function, It is the probability of a real technological paradigm phase transition. It is the predicted probability of a technological paradigm phase transition;

[0126] The expression for the dynamic constraint loss function is:

[0127]

[0128] in, The loss function is the dynamic constraint function.

[0129] The expression for the total loss function is:

[0130]

[0131] in, For the total loss function, These are the parameters used to balance the loss term.

[0132] Specifically, the core function of the regression loss function is to quantify the deviation between the model's predicted activity of technology clusters and the actual activity of technology clusters, focusing on solving the error optimization problem of predicting continuous value activity of technology clusters. In the specific calculation, the regression loss function adopts the mean squared error form, and its expression is: .in, As the output value of the regression loss function, its magnitude directly reflects the overall deviation between the predicted activity and the actual activity. The smaller the value, the higher the accuracy of the model in predicting continuous values. Representative model in time For the The predicted activity output by each technology cluster is calculated by the model based on cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms. This is the time extracted from the unified database. No. The true activity level of a technology cluster is calculated by comprehensively analyzing data such as the number of new patents, the frequency of paper publications, and the rate of citation growth for that technology cluster within the corresponding time window, ensuring that the true value is objective and traceable. The total number of training samples covers activity data pairs from different time windows and technology clusters. By averaging these pairs, the excessive influence of single-sample errors on the overall loss calculation is avoided, allowing the loss value to better reflect the model's predictive performance on the global data. The design logic of this regression loss function is to amplify the loss contribution of larger deviations through squaring, prompting the model to prioritize optimizing predictions that deviate significantly from the true values, thereby improving the overall accuracy of technology cluster activity prediction.

[0133] Secondly, a classification loss function is constructed to measure the difference between the model's predicted probability of a technology paradigm shift and the actual probability of a phase shift, thus adapting to the classification task requirement of predicting the probability value of a technology paradigm shift within the 0-1 interval. Its expression is set as follows: ,in The smaller the loss value, the closer the model's prediction of the phase transition probability is to the true situation; It is time The true probability of a technological paradigm shift, if in time... A technological paradigm shift actually occurred. The value is 1 if no phase transition occurs, and 0 if no phase transition occurs. Alternatively, the value can be between 0 and 1 based on the actual intensity of the phase transition (such as the percentage of affected technology clusters) to ensure that the true value accurately describes the phase transition state. Is the model in time The predicted probability of a technological paradigm shift is generated by the model through analysis and calculation of macroscopic-level features and the aggregation state of mesoscopic technology clusters, with a value constrained between 0 and 1. The classification loss function adopts the cross-entropy loss form. Its core advantage lies in the fact that when the model's predicted probability deviates significantly from the true probability (e.g., the true value is 1 while the predicted value is close to 0), it generates a larger loss value, thereby enabling stronger gradient updates to the model parameters. This guides the model to quickly adjust its prediction logic, improving the prediction accuracy for this crucial event of a technological paradigm shift.

[0134] Subsequently, a dynamic constraint loss function is constructed. This loss function is crucial for ensuring that the model's predictions conform to the inherent evolutionary laws of the technological system. Its role is to constrain the rate of change in the activity of the predicted technology clusters. It must be consistent with the coupling diffusion equations characterizing the micro-to-meta-scale diffusion process in the cross-level influence transmission mechanism, avoiding predictions that are superficially accurate but do not conform to the actual logic of technological evolution obtained solely through data fitting. Its expression is: ,in The deviation is calculated using the L2 norm squared value to represent the dynamic constraint loss function value, ensuring that the deviation value is non-negative and can amplify the effect of a large deviation. It is the model's prediction of the first A cluster of technologies in time The rate of change in activity, through the predicted activity within a continuous time window. Numerical differentiation (e.g., using the central difference method) yields the following; the part within parentheses on the right side of the formula is the core expression of the coupling-diffusion equation in the cross-level influence transmission mechanism, representing the theoretically calculated value of the rate of change of technology cluster activity, where... For diffusion intensity parameters, For technology cluster and Influence weight, For input rate parameters, For external input of the number of new inventions, These parameters, representing attenuation coefficients, are consistent with the calibrated parameters in the cross-level influence transmission mechanism, ensuring that the theoretical values ​​accurately reflect the true evolution of technology cluster activity. This loss function forces the model's predicted rate of change in activity to converge with the theoretical evolution, improving the physical rationality and interpretability of the prediction results.

[0135] Finally, a total loss function is constructed, which merges the three loss functions mentioned above to achieve comprehensive optimization of the model's prediction accuracy, classification accuracy, and dynamic rationality. The expression for the total loss function is as follows: ,in The final output value of the total loss function is the core optimization objective for updating model parameters; This is a loss term balancing parameter, and its value is adjusted according to the specific needs of model training. When it is crucial to ensure that the prediction results conform to the dynamic laws, it can be... The value should be set between 0.5 and 1.0; when the focus is more on improving prediction accuracy and classification accuracy, the value can be... The value is adjusted to be between 0.1 and 0.3, typically by performing a grid search on the validation set (e.g., in the test). Determine the optimal value to ensure that the three loss terms can work together, so that the model can make accurate predictions while possessing the inherent logic that conforms to the evolution of the science and technology system, thus providing reliable model support for the prediction of science and technology trends.

[0136] The aforementioned machine learning-based technology trend prediction method integrates micro-level invention data, meso-level technology cluster data, and macro-level technology paradigm data to form a unified database. It then extracts and fuses features from each level to construct a cross-level feature matrix. Simultaneously, based on features and correlations, it constructs dynamic networks at the micro, meso, and macro levels respectively and connects them across levels, forming a multi-level dynamic network that characterizes interactions between levels. Furthermore, it establishes a mechanism to describe the transmission of cross-level influences (including diffusion, phase transition, and feedback processes). Finally, it comprehensively utilizes the aforementioned cross-level feature matrix, multi-level dynamic network, and cross-level influence transmission mechanism to drive a pre-trained machine learning model for prediction. This series of coherent steps collectively achieves a systematic modeling of the multi-level, nonlinear evolution of technology trends, enabling a more comprehensive and accurate simulation of the cross-level transmission and amplification effects of external shocks in the technology ecosystem, and predicting the technological effects of complex emergent phenomena such as technology path locking and technology paradigm phase transitions.

[0137] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0138] Based on the same inventive concept, this application also provides an apparatus for implementing the aforementioned machine learning-based technology trend prediction method. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations of one or more machine learning-based technology trend prediction apparatus embodiments provided below can be found in the limitations of the machine learning-based technology trend prediction method described above, and will not be repeated here.

[0139] In one exemplary embodiment, such as Figure 3 As shown, a technology trend prediction device 30 based on machine learning is provided to implement the methods in the above-described method embodiments. The device includes:

[0140] The multi-source data fusion module 31 is used to collect invention data at the micro level, technology cluster data at the meso level, and technology paradigm data at the macro level, and integrate the invention data, technology cluster data, and technology paradigm data to generate a unified database. Among them, the invention data includes the text content, citation relationships, and timestamps of patents and papers; the technology cluster data is obtained by clustering the invention data; and the technology paradigm data includes policy documents, R&D investment indicators, and technology paradigm labels.

[0141] The cross-level feature fusion module 32 is used to extract features from data in a unified database to obtain micro-level features, meso-level features and macro-level features, and to fuse the micro-level features, meso-level features and macro-level features based on cross-level association to generate a cross-level feature matrix.

[0142] The multi-level network construction module 33 is used to construct a micro-dynamic network based on micro-level features and reference relationships; construct a meso-dynamic network based on meso-level features and technology cluster associations; construct a macro-dynamic network based on macro-level features and policy associations; and perform cross-level connections between the micro-dynamic network, meso-dynamic network, and macro-dynamic network to generate a multi-level dynamic network.

[0143] The influence transmission modeling module 34 is used to establish a cross-level influence transmission mechanism based on a multi-level dynamic network. The cross-level influence transmission mechanism includes the diffusion process of science and technology from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro.

[0144] The technology trend prediction module 35 is used to predict technology trends based on cross-level feature matrices, multi-level dynamic networks, and cross-level influence transmission mechanisms through a pre-trained machine learning model, and outputs predictions of technology cluster activity and the probability of technology paradigm shifts.

[0145] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0146] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0147] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0148] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A technology trend prediction method based on machine learning, characterized in that, The method includes: S1. Collect invention data at the micro level, technology cluster data at the meso level, and technology paradigm data at the macro level, and integrate the invention data, technology cluster data, and technology paradigm data to generate a unified database; wherein, the invention data includes the text content, citation relationships, and timestamps of patents and papers, the technology cluster data is obtained by clustering the invention data, and the technology paradigm data includes policy documents, R&D investment indicators, and technology paradigm tags; S2. Extract features from the data in the unified database to obtain micro-level features, meso-level features and macro-level features, and fuse the micro-level features, meso-level features and macro-level features based on cross-level association to generate a cross-level feature matrix; S3. Based on the micro-level features and the reference relationships, construct a micro-dynamic network; based on the meso-level features and technology cluster associations, construct a meso-dynamic network; based on the macro-level features and policy associations, construct a macro-dynamic network; perform cross-level connections between the micro-dynamic network, the meso-dynamic network, and the macro-dynamic network to generate a multi-level dynamic network. S4. Based on the multi-level dynamic network, establish a cross-level influence transmission mechanism; wherein, the cross-level influence transmission mechanism includes the diffusion process of science and technology from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro. S5. Based on the cross-level feature matrix, the multi-level dynamic network, and the cross-level influence transmission mechanism, a pre-trained machine learning model is used to predict technology trends and output the prediction of technology cluster activity and the probability of technology paradigm shift.

2. The method according to claim 1, characterized in that, The system constructs a micro-dynamic network based on the micro-level features and the reference relationships; a meso-dynamic network based on the meso-level features and technology cluster associations; and a macro-dynamic network based on the macro-level features and policy associations. Cross-level connections are made between the microscopic dynamic network, the mesoscopic dynamic network, and the macroscopic dynamic network to generate a multi-level dynamic network, including: S11. Based on the micro-level features and citation relationships in the micro-level, construct a micro-citation network; based on the technology cluster associations in the meso-level, construct a meso-network; based on the technology paradigm associations in the macro-level, construct a macro-network; wherein, the nodes of the micro-citation network are patents or papers, and the edges of the micro-citation network are the citation relationships; the nodes of the meso-network are technology clusters, and the edge weights of the meso-network are obtained by aggregating the citation counts of their respective micro-entities; the nodes of the macro-network are technology paradigms or policy entities, and the edge weights of the macro-network are calculated based on the co-occurrence frequency of policy documents or the relevance of R&D investment; S12. Based on the micro-reference network and the meso-network, establish micro-to-meso-connections and generate membership edges; wherein, the membership edges connect each micro-entity to its respective technology cluster node; S13. Based on the mesoscopic network and the macroscopic network, establish a connection from mesoscopic to macroscopic to generate influence edges; wherein, the influence edges connect each technology cluster node to the relevant macroscopic node, and the weight value of the influence edges is calculated based on policy support. S14. Based on the macro network and the micro reference network, establish a macro-to-micro feedback connection and generate feedback edges; wherein, the feedback edges connect macro nodes to micro nodes, and the weight values ​​of the feedback edges are calculated based on the historical impact of policy intensity on the invention rate; S15. The micro-reference network, the meso-network, the macro-network, the membership edge, the influence edge, and the feedback edge are combined to generate the multi-level dynamic network.

3. The method according to claim 1, characterized in that, The establishment of a cross-level influence transmission mechanism based on the multi-level dynamic network includes: S21. Based on the aforementioned multi-level dynamic network, a diffusion mechanism from the microscopic to the mesoscopic level is constructed by describing the change in the activity of the technology cluster over time using a coupled diffusion equation; the expression of the coupled diffusion equation is: in, Representing technology clusters In time activity level Representing technology clusters The neighborhood group, Representing technology clusters and The influence weights between them This indicates the number of new inventions input from external sources. For diffusion intensity parameters, For input rate parameters, This refers to the attenuation coefficient parameter; S22. Based on the aforementioned multi-level dynamic network, a phase transition mechanism from mesoscopic to macroscopic is constructed by describing the phase transition probability of the technological paradigm using the sigmoid function; the expression for the phase transition probability is: in, Indicates time The technological paradigm of phase transition probability, It is the sigmoid function. It is a technology cluster The weight, It is the impact coefficient of macroeconomic policies. Indicates time Policy intensity indicators; S23. Based on the aforementioned multi-level dynamic network, the impact of policies on the invention rate is described through a reaction-diffusion equation, constructing a feedback mechanism from macro to micro levels; the expression of the reaction-diffusion equation is: in, Indicates time Micro-invention rate For the parameter of policy impact intensity, This is the attenuation rate parameter; S24. Based on the multi-level dynamic network, the current policy intensity index is generated by weighting the sum of the policy intensity index of the previous time step and the activity level of the technology cluster of the previous time step, thus constructing a feedback loop mechanism; the expression of the feedback loop mechanism is: in, For feedback coefficients; S25. Integrate the diffusion mechanism, the phase transition mechanism, the feedback mechanism, and the feedback loop mechanism to obtain the cross-level influence transmission mechanism.

4. The method according to any one of claims 1 to 3, characterized in that, The method of predicting technology trends using a pre-trained machine learning model based on the cross-level feature matrix, the multi-level dynamic network, and the cross-level influence transmission mechanism, outputting a prediction of technology cluster activity and a probability of technology paradigm shifts, includes: S31. Based on the micro-dynamic network and the micro-level features in the cross-level feature matrix, a graph convolutional network is used for processing to generate micro-node embeddings; the operation of the graph convolutional network is represented as follows: in, Indicates time The micro-node embedding matrix, For the normalized microadjacency matrix, For time The aforementioned micro-level features, The weight matrix is ​​trainable. S32. The micro-node embeddings are aggregated to generate meso-level embeddings; wherein, the aggregation process is achieved by calculating the average embedding of the micro-nodes to which each technology cluster belongs; S33. Based on the mesoscopic layer embedding and the mesoscopic layer features in the cross-level feature matrix, the mesoscopic layer state vector is generated by processing through a long short-term memory network. S34. Based on the macro-level features in the cross-level feature matrix, a long short-term memory network is used for processing to generate a macro-level state vector. S35. Based on the mesoscopic layer state vector and the macroscopic layer state vector, an attention mechanism is used to calculate the attention weights; the formula for calculating the attention weights is: in, This represents the mesoscopic layer state vector corresponding to the i-th technology cluster in the mesoscopic layer. The macro-level state vector corresponding to the j-th technological paradigm or policy entity in the macro-level. The attention weights between them; S36. Based on the attention weight, the mesoscopic layer state vector and the macroscopic layer state vector are concatenated to obtain a fused feature vector; the fused feature vector is input into a fully connected layer for nonlinear transformation to generate an initial prediction vector; the initial prediction vector is constrained and optimized based on the cross-level influence transmission mechanism to output the technology cluster activity prediction and the technology paradigm phase transition probability.

5. The method according to claim 3, characterized in that, The total loss function used when training the machine learning model consists of a regression loss function, a classification loss function, and a dynamic constraint loss function; The expression for the regression loss function is: in, Let the regression loss function be... It is a prediction of the activity level of technology clusters. It reflects the actual activity level of the technology cluster. It is the number of training samples; The expression for the classification loss function is: in, Let be the classification loss function. It is the probability of a real technological paradigm phase transition. It is the predicted probability of a technological paradigm phase transition; The expression for the dynamic constraint loss function is: in, Let the dynamic constraint loss function be denoted as . The expression for the total loss function is: in, Let the total loss function be... These are the parameters used to balance the loss term.

6. A technology trend prediction device based on machine learning, used to implement the method of any one of claims 1 to 5, characterized in that, The device includes: A multi-source data fusion module is used to collect invention data at the micro level, technology cluster data at the meso level, and technology paradigm data at the macro level, and integrate the invention data, technology cluster data, and technology paradigm data to generate a unified database. The invention data includes the text content, citation relationships, and timestamps of patents and papers; the technology cluster data is obtained by clustering the invention data; and the technology paradigm data includes policy documents, R&D investment indicators, and technology paradigm tags. The cross-level feature fusion module is used to extract features from the data in the unified database to obtain micro-level features, meso-level features and macro-level features, and to fuse the micro-level features, meso-level features and macro-level features based on cross-level association to generate a cross-level feature matrix. A multi-level network construction module is used to construct a micro-dynamic network based on the micro-level features and the reference relationships; construct a meso-dynamic network based on the meso-level features and technology cluster associations; construct a macro-dynamic network based on the macro-level features and policy associations; and perform cross-level connections between the micro-dynamic network, the meso-dynamic network, and the macro-dynamic network to generate a multi-level dynamic network. The influence transmission modeling module is used to establish a cross-level influence transmission mechanism based on the multi-level dynamic network; wherein, the cross-level influence transmission mechanism includes the diffusion process of science and technology from the micro to the meso, the phase transition process from the meso to the macro, and the feedback process from the macro to the micro. The technology trend prediction module is used to predict technology trends based on the cross-level feature matrix, the multi-level dynamic network, and the cross-level influence transmission mechanism, through a pre-trained machine learning model, and output the technology cluster activity prediction and the probability of technology paradigm phase transition.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images