A data processing method and related apparatus

Abstract

This application discloses a data processing method and related apparatus, comprising: for nodes i and j in a first node set, determining the normalized weights of edges with nodes i and j as endpoints based on the number of times node i is referenced by node j and the time decay factor; determining the multi-type information fusion feature vector of node i based on the fusion feature vector of node i, the aggregated feature vectors of each type of neighboring nodes of node i, and the attention weights between node i and its neighboring nodes; determining whether to add a semantic enhancement edge between nodes i and j based on the relationship between the semantic similarity between nodes i and j and a preset semantic similarity threshold; if a semantic enhancement edge is added between nodes i and j, determining the normalized weights of the semantic enhancement edge based on the semantic similarity between nodes i and j, the time-modulated semantic similarity between nodes i and j, and the structural correlation between nodes i and j.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a data processing method and related apparatus. Background Technology

[0002] Knowledge flow networks are dynamic network structures that describe the transmission, sharing, association, and transformation of different knowledge subjects (such as papers or patents) through specific channels (such as downloading and citation).

[0003] Traditional methods for analyzing knowledge flow networks include word frequency analysis, co-occurrence analysis, topic modeling, and hotspot identification using standard graph neural networks. However, these traditional methods neglect the temporal dimension of the data, failing to effectively capture the temporal changes in data flow and leading to inaccurate predictions for emerging interdisciplinary fields.

[0004] How to capture information about knowledge nodes and the time dimension between them, and improve the accuracy of predictions for emerging interdisciplinary fields, has become one of the urgent technical problems to be solved in the field of data processing technology. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a data processing method that can capture the evolution of knowledge nodes and the relationships between nodes over time, thereby improving the accuracy of predictions for emerging interdisciplinary fields.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] The first aspect of this application discloses a data processing method, including:

[0008] Obtain an initial knowledge network; the initial knowledge network includes a first set of nodes, a first set of edges, a first feature matrix, and a first set of time elements; wherein, the first set of nodes includes two types: paper nodes and patent nodes;

[0009] For node i and node j in the first node set, the normalized weight of the edge with node i and node j as endpoints is determined based on the number of times node i is referenced by node j and the time decay factor.

[0010] For node i in the first node set, the multi-type information fusion feature vector of node i is determined based on the fusion feature vector of node i, the aggregated feature vector of each type of neighboring node of node i, and the attention weight between node i and the neighboring node.

[0011] For node i and node j in the first node set, based on the relationship between the semantic similarity between node i and node j and the preset semantic similarity threshold, it is determined whether to add a semantic enhancement edge between node i and node j;

[0012] If a semantically enhanced edge is added between node i and node j, the normalized weight of the semantically enhanced edge is determined based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j.

[0013] In one alternative implementation, obtaining the initial knowledge network includes:

[0014] Obtain an initial dataset; the initial dataset includes multiple data entries; each data entry includes multiple fields; the multiple fields include keywords, abstract, publication date, and citation count; the data is literature or patent;

[0015] The initial dataset is cleaned to obtain an intermediate dataset;

[0016] For each data point in the intermediate dataset, semantic feature extraction, temporal feature extraction, and network topology feature extraction are performed on the data point to obtain the semantic feature vector, temporal feature vector, and topology feature vector corresponding to the data point.

[0017] For each data point in the intermediate dataset, the semantic feature vector, temporal feature vector, and topological feature vector corresponding to that data point are fused to obtain the fused feature vector corresponding to that data point.

[0018] Each data point in the intermediate dataset is treated as a node to construct the first node set; the first edge set is constructed based on the association between each data point in the intermediate dataset and other data points; the first feature matrix is ​​constructed based on the fusion feature vector corresponding to each data point in the intermediate dataset; and the first time meta set is constructed based on the publication time of each data point in the intermediate dataset.

[0019] The first set of nodes, the first set of edges, the first feature matrix, and the first set of time elements are used as the initial knowledge network.

[0020] In one optional implementation, determining the normalized weights of edges with nodes i and j as endpoints, based on the number of times node i is referenced by node j and a time decay factor, for nodes i and j in the first node set, includes:

[0021] The basic weight is obtained by performing a logarithmic operation on the number of times node i is referenced by node j;

[0022] Substitute the basic weight, the absolute value of the time difference between the publication time of node i and the publication time of node j, and the time decay factor into... In the process, the normalized weights of the edges with endpoints i and j are obtained; where, This represents the normalized weight of the edge with nodes i and j as endpoints; Indicates the time decay factor; These are the basic weights.

[0023] In one optional implementation, for node i in the first node set, determining the multi-type information fusion feature vector of node i based on the fusion feature vector of node i, the aggregated feature vectors of each type of neighboring nodes of node i, and the attention weights between node i and its neighboring nodes includes:

[0024] Determine all neighboring nodes of node i in the first node set, and the node type of each neighboring node of node i; the types of neighboring nodes include paper node type and patent node type;

[0025] For each neighboring node of node i, the attention weight between node i and the neighboring node is calculated based on the fusion feature vector of node i, the node type of the neighboring node, and the fusion feature vector of the neighboring node.

[0026] For multiple neighboring nodes of the paper node type corresponding to node i, the aggregated feature vector of the paper type neighbors is obtained based on the fusion feature vector of each of the multiple neighboring nodes and the attention weight between each of the multiple neighboring nodes and node i.

[0027] For multiple neighboring nodes of the patent node type corresponding to node i, the aggregated feature vector of the patent type neighbors is obtained based on the fusion feature vector of each of the multiple neighboring nodes and the attention weight between each of the multiple neighboring nodes and node i.

[0028] For node i, the fusion feature vector of node i, the aggregated feature vector of the paper type neighbors corresponding to node i, and the aggregated feature vector of the patent type neighbors corresponding to node i are weighted and fused to obtain the multi-type information fusion feature vector of node i.

[0029] In one optional implementation, determining the normalized weight of the semantically enhanced edge based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j includes:

[0030] The semantic similarity is calculated based on the semantic feature vector of node i and the semantic feature vector of node j;

[0031] Based on the semantic feature vector of node i and the temporal context information of node i, calculate the time-aware feature vector of node i; and based on the semantic feature vector of node j and the temporal context information of node j, calculate the time-aware feature vector of node j.

[0032] The time-sensing feature vector of node i and the time-sensing feature vector of node j are used to calculate the time modulation semantic similarity.

[0033] Based on the intersection of the neighbor node set of node i and the neighbor node set of node j, and the union of the neighbor node set of node i and the neighbor node set of node j, calculate the similarity between the neighbor node sets of node i and node j.

[0034] Based on the betweenness centrality of node i and the betweenness centrality of node j, calculate the centrality difference between node i and node j;

[0035] The structural association degree is obtained by weighting the similarity of the neighbor node set and the centrality dissimilarity.

[0036] Based on the normalization function and the multilayer perceptron, the first weight of the semantic similarity, the second weight of the temporal modulation semantic similarity, and the third weight of the structural association are determined.

[0037] The first weight, the second weight, the third weight, the semantic similarity, the temporal modulation semantic similarity, and the structural correlation are weighted and calculated to obtain the normalized weight of the semantically enhanced edge.

[0038] In one alternative implementation, after determining the normalized weights of the semantically enhanced edges, the method further includes:

[0039] Identify the isolated nodes in the first node set;

[0040] Based on the fused feature vector, variational autoencoder, and decoder of the isolated node, candidate connection nodes of the isolated node are generated.

[0041] Based on the preset prompt word template and large language model, the correlation degree between the isolated node and the candidate connection node is calculated.

[0042] Based on the correlation degree value and the time decay factor, the normalized weights of the edges with the isolated node and the candidate connected node as endpoints are determined.

[0043] In an alternative implementation, after determining the normalized weights of the edges with endpoints at the isolated node and the candidate connected node, the method further includes:

[0044] For each neighboring node of node i in the first node set, the weight coefficient of that neighboring node, the time influence factor between that neighboring node and node i, and the position of that neighboring node in the time frame of node i. The neighbor aggregation operation is performed on the aggregated neighbor feature vectors to obtain the th neighbor feature vector. The nodes described in the layer The aggregated neighbor feature vector; the weight coefficient of the neighbor node is determined according to the node type of the neighbor node;

[0045] For node i in the first node set, based on the... The nodes described in the layer The aggregated neighbor feature vector, preset balancing parameters, and knowledge-enhanced transformation vector are used to calculate the embedded feature vector corresponding to node i; the knowledge-enhanced transformation feature vector is based on node i. The corresponding type of learnable relation vector and the temporal relation vector of node i are generated.

[0046] In one optional implementation, after obtaining the embedded feature vector corresponding to node i, the method further includes:

[0047] For node i and node j in the first node set, the embedded feature vectors of node i and node j are substituted into the bilinear bisection function to obtain the original association score between node i and node j.

[0048] The original correlation score is adjusted based on the time decay modulation function to obtain the corrected correlation score; the expression of the time decay modulation function is: Where γ is the time decay parameter, Δt ij Let p(i,j) be the time difference between the publication time of node i and the publication time of node j, p(i,j) be the original association score between node i and node j, and p'(i,j) be the modified association score between node i and node j.

[0049] Using a breadth-first search algorithm, search for all propagation paths from node i to node j whose path length does not exceed a preset path length threshold, and calculate the propagation strength score for each propagation path.

[0050] The maximum propagation intensity score among the multiple propagation intensity scores is taken as the multi-step inference score between node i and node j, and the propagation path corresponding to the maximum propagation intensity score is taken as the optimal propagation path between node i and node j.

[0051] Determine whether the multi-step inference score is greater than a preset score threshold;

[0052] If the multi-step inference score is determined to be greater than a preset score threshold, the optimal propagation path is added to the knowledge enhancement network; the knowledge enhancement network is generated based on the initial knowledge network.

[0053] In one alternative implementation, the method further includes:

[0054] Cluster the nodes in the first node set to obtain multiple clusters, and determine the central feature vectors corresponding to the clusters;

[0055] For each node in the cluster, the embedded feature vector corresponding to the node and the central feature vector of the cluster to which the node belongs are substituted into the preset quality scoring function to obtain the cluster rationality score corresponding to the node.

[0056] For each node in the cluster, if the cluster rationality score corresponding to the node is greater than or equal to the preset rationality score, then the node is determined to belong to the corresponding cluster; if the cluster rationality score corresponding to the node is less than the preset rationality score, then the node is deleted from the corresponding cluster to obtain the target cluster.

[0057] In one alternative implementation, the method further includes:

[0058] Determine the innovation frontier index corresponding to each of the target clusters;

[0059] Select multiple innovation frontier indices whose values ​​are greater than the preset index threshold;

[0060] Multiple target clusters corresponding to multiple innovation frontier indices are treated as multiple hotspot clusters.

[0061] In one alternative implementation, the method further includes:

[0062] For each hotspot cluster, obtain the text information corresponding to the hotspot cluster, and generate a text summary of the hotspot cluster based on the text information, the target prompt word template, and the large language model.

[0063] In one alternative implementation, the method further includes:

[0064] Obtain the comparative learning similarity between the candidate document and the hot spot cluster to which the candidate document belongs;

[0065] The recommendation score for the candidate document is determined based on the comparative learning similarity, the innovation frontier index of the hot spot cluster to which the candidate document belongs, and the number of years since the candidate document was published.

[0066] A second aspect of this application provides a data processing apparatus, comprising:

[0067] An initial knowledge network acquisition module is used to acquire an initial knowledge network; the initial knowledge network includes a first set of nodes, a first set of edges, a first feature matrix, and a first set of time elements; wherein, the first set of nodes includes two types: paper nodes and patent nodes;

[0068] The first normalized weight determination module is used to determine the normalized weight of the edge with nodes i and j as endpoints for nodes i and j in the first node set, based on the number of times node i is referenced by node j and the time decay factor.

[0069] The multi-type information fusion feature vector determination module is used to determine the multi-type information fusion feature vector of node i for node i in the first node set based on the fusion feature vector of node i, the aggregated feature vector of each type of neighboring node of node i, and the attention weight between node i and the neighboring node.

[0070] The semantic enhancement edge addition module is used to determine whether to add a semantic enhancement edge between node i and node j in the first node set, based on the relationship between the semantic similarity between node i and node j and a preset semantic similarity threshold.

[0071] The second normalized weight determination module is used to determine the normalized weight of the semantic enhancement edge based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j if a semantic enhancement edge is added between node i and node j.

[0072] A third aspect of this application discloses a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0073] A fourth aspect of this application discloses an electronic device, comprising: a memory storing a computer program thereon; and a processor for executing the computer program in the memory to implement the steps of the method described in any of the first aspects.

[0074] Compared with the prior art, this application has the following beneficial effects:

[0075] The data processing method disclosed in this application improves the accuracy of the normalized weight results by considering both the number of references and the time decay factor during the calculation of edge normalization weights. During node feature construction, it integrates its own features, multi-type neighbor aggregation features, and attention weights to deeply capture the influence of node attributes and neighbor differences, thus strengthening the feature expression capabilities of nodes. In the construction of semantically enhanced edges, weights are calculated based on semantic similarity thresholds, time modulation factors, and structural correlation, ensuring semantic and temporal rationality while compensating for the sparsity of the explicit network. This results in an optimized network possessing both explicit reference relationships and implicit semantic connections, providing more comprehensive structural information and feature support for subsequent knowledge mining tasks such as community discovery and hotspot identification. Attached Figure Description

[0076] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0077] Figure 1 A flowchart illustrating a data processing method provided in an embodiment of this application;

[0078] Figure 2 This application provides a schematic diagram of the structure of an initial knowledge network as an embodiment of the present application.

[0079] Figure 3 A flowchart illustrating a method for data processing of isolated nodes provided in this application embodiment;

[0080] Figure 4 This application provides a schematic diagram of the structure of an enhanced knowledge network.

[0081] Figure 5 A flowchart of a knowledge flow prediction method provided in an embodiment of this application;

[0082] Figure 6 A schematic diagram of the structure of a target knowledge flow network provided in an embodiment of this application;

[0083] Figure 7 A flowchart illustrating a data clustering analysis method provided in this application embodiment;

[0084] Figure 8 A schematic diagram illustrating a target clustering result provided in an embodiment of this application;

[0085] Figure 9 A schematic diagram of an innovation frontier index provided for an embodiment of this application;

[0086] Figure 10 A schematic diagram of a hotspot recommendation list provided in an embodiment of this application;

[0087] Figure 11 A schematic diagram illustrating a recommendation result provided in an embodiment of this application;

[0088] Figure 12 A schematic diagram of a trend analysis report provided in an embodiment of this application;

[0089] Figure 13 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. Detailed Implementation

[0090] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0091] Figure 1 This is a flowchart illustrating a data processing method provided in an embodiment of this application. (In conjunction with...) Figure 1 As shown, the data processing method disclosed in this application includes:

[0092] S101, Obtain the initial knowledge network.

[0093] The initial knowledge network in this application includes a first set of nodes, a first set of edges, a first feature matrix, and a first set of time elements. The first set of nodes includes two types: paper nodes and patent nodes. Among them, the papers corresponding to the paper nodes include, but are not limited to, academic papers, journal articles, and conference papers.

[0094] In one alternative implementation, the steps for obtaining the initial knowledge network are as follows:

[0095] A1, Obtain the initial dataset.

[0096] The initial dataset in this application contains data from academic literature (academic papers, dissertations, journal articles, conference papers, or research reports) or patent data.

[0097] For example, the initial dataset can be obtained through a pre-built data query system. The pre-built data query system in this application supports multi-dimensional data filtering (domain classification identifier, geographical region restriction, and language type); at the same time, the query time range (15-year time window) supported by the pre-built data query system in this application matches the periodic characteristics of knowledge evolution to ensure that the complete knowledge development cycle can be captured.

[0098] After entering multi-dimensional data information to be queried in the data query field of a pre-built data query system, the system will automatically generate an initial dataset. This application does not limit the format of the initial dataset; it can be an Excel spreadsheet, a JSON file, an XML document, a database table structure (such as a MySQL table or an Oracle table in a relational database), or a data table visualization view, etc. However, this application requires that each piece of initial data in the initial dataset has multiple fields, including but not limited to keywords, abstract, publication date, number of downloads, and number of citations.

[0099] A2. Clean the initial dataset to obtain an intermediate dataset.

[0100] Understandably, the initial dataset contains a large amount of unstructured noisy data, invalid words, inconsistent terminology, multilingual mixed text, and special data with inconsistent formats. It is necessary to clean the initial dataset through methods such as text cleaning to obtain an intermediate dataset.

[0101] First, HTML tags and special symbols are removed from the initial dataset. Then, stop word removal is performed based on a domain-specific thesaurus. This thesaurus not only covers general stop words but also includes transitional words with no actual semantic meaning, such as "in summary," commonly used in academic settings, to avoid interference from invalid words in subsequent data processing. Next, terminology standardization is performed. On the one hand, a domain-specific dictionary is used to standardize synonyms such as "AI" and "artificial intelligence," and on the other hand, abbreviations such as "DOI" (Digital Object Identifier) ​​are expanded to ensure consistency in terminology across the data. Subsequently, for potentially multilingual text, language detection algorithms (such as CLD3) are used to automatically identify the language of the text (e.g., English, Chinese, Japanese), and different language texts are categorized and stored or processed according to preset rules. Finally, regular expressions are used to standardize date information in the data, ensuring the consistency of date data format. After completing the entire text cleaning process, a structurally sound and accurately accurate intermediate dataset is obtained.

[0102] After obtaining the intermediate dataset, the quality of the intermediate dataset can be checked and controlled by methods such as checking the integrity of citation relationships, evaluating text quality, and verifying the validity of time information.

[0103] A3. For each data point in the intermediate dataset, semantic feature extraction, temporal feature extraction, and network topology feature extraction are performed on that data point to obtain the corresponding semantic feature vector, temporal feature vector, and topology feature vector.

[0104] Each data point in the intermediate dataset is input into the PubMedBERT (PubMed Bidirectional Encoder Representations from Transformers) model. The PubMedBERT model encodes the textual information corresponding to each data point, resulting in a semantic feature vector for each data point. The PubMedBERT model used in this application is a pre-trained model specifically for biomedical literature, exhibiting better domain adaptability.

[0105] The feature extraction paradigm of PubMedBERT in this application is shown in Equation (1):

[0106] (1)

[0107] In formula (1) For nodes The semantic feature vector, [CLS] is the sentence beginning marker of the BERT model, and [SEP] is the separator. It is the title of the data. It is a summary of the data; one data point in the intermediate dataset corresponds to one node.

[0108] Meanwhile, this application employs a sine-cosine positional encoding scheme to capture the temporal dimension features of each data point in the intermediate dataset and generate a temporal feature vector corresponding to each data point.

[0109] The encoding function for time information in the sine-cosine position coding scheme is shown in Formula 2:

[0110] (2)

[0111] In formula (2) It refers to the time difference between nodes (i.e., the year difference between the publication time of the data corresponding to the two nodes). For dimensional indexing, Encoding vector dimension (64-dimensional); PE ( ) is an abbreviation for "Positional Encoding".

[0112] Meanwhile, this application employs the graph embedding algorithm GraphSAGE (Graph Sample and Aggregate) to generate topological feature vectors for nodes. The graph feature extraction formula is shown in formula (3):

[0113] (3)

[0114] In formula (3) Represents the topological feature vector of node i; It is a graph sampling and aggregation algorithm; Represents the basic diagram structure; This indicates that the GraphSAGE algorithm uses 2 layers, meaning it performs neighbor sampling and aggregation operations twice. This indicates that the number of neighbors sampled at each layer is 25, which is used to control the size of the neighbors selected during each aggregation.

[0115] Using the method in A3, we can obtain the semantic feature vector, temporal feature vector, and topological feature vector corresponding to each data point in the initial dataset.

[0116] A4. For each data point in the intermediate dataset, the semantic feature vector, temporal feature vector, and topological feature vector corresponding to that data point are fused to obtain the fused feature vector corresponding to that data point.

[0117] First, the Z-score normalization method is used to normalize the target feature vectors (semantic feature vector, temporal feature vector, and topological feature vector) corresponding to each data point in the intermediate dataset, resulting in standard target feature vectors. The Z-score normalization method is shown in formula (4):

[0118] (4)

[0119] In formula (4) Represents the standard target feature vector. and These are the mean and standard deviation of the target feature vector, respectively. This represents the target feature vector.

[0120] Through the above processing, we can obtain the standard semantic feature vector, standard temporal feature vector, and standard topological feature vector corresponding to each data point in the intermediate dataset.

[0121] Then, a multilayer perceptron is used to perform nonlinear feature fusion on the three standard feature vectors (standard semantic feature vector, standard temporal feature vector, and standard topological feature vector) of each data point to form the fused feature vector corresponding to that data point. The expression of the fused feature vector is shown in formula (5):

[0122] (5)

[0123] In formula (5) Let i be the fused feature vector corresponding to node i. For nodes The corresponding semantic feature vector, It is the time feature vector corresponding to node i. This represents the topological feature vector corresponding to node i, and MLP (Multilayer Perceptron) represents a multilayer perceptron.

[0124] The multilayer perceptron in this application has an input layer dimension of 960 (formed by fusing 768, 64, and 128-dimensional features), a hidden layer consisting of three fully connected networks (with dimensions of 960, 512, and 256 respectively), and uses the ReLU (Rectified Linear Unit) function and Dropout (with a probability of 0.1) as the activation mechanism. The output layer is a 256-dimensional unified knowledge representation.

[0125] The feature fusion mechanism in this application effectively preserves the multidimensional features of textual semantic information, temporal evolution features, and network topology attributes, and can accurately characterize the distribution location of nodes in the abstract knowledge space.

[0126] Furthermore, after obtaining the fusion feature vector corresponding to each data point in the intermediate dataset, the distance between any two data points, i.e., two nodes, can be measured based on formula (6).

[0127] The expression for formula (6) is:

[0128] (6)

[0129] In formula (5) It is the fused feature vector corresponding to node i. It is the fused feature vector corresponding to node j. It is the distance between node i and node j in the knowledge space; using the distance between node i and node j in the knowledge space instead of the traditional geographic spatial distance is more in line with the characteristics of academic knowledge spreading in spatial networks.

[0130] A5, treat each data point in the intermediate dataset as a node, and construct the first node set.

[0131] Each piece of data in the intermediate dataset (such as a document, a patent, etc.) is regarded as an independent node in the network, and the set of all these nodes is the first node set.

[0132] A6. Based on the relationship between each data point and other data points in the intermediate dataset, construct the first edge set.

[0133] Based on the relationships between data, such as the reference relationships between data, connections are established between corresponding nodes, that is, the edges described in this application are established, and the set of all these edges forms the first edge set.

[0134] The edges in the first edge set record the number of references between nodes, and the weight information for each edge in the first edge set is empty.

[0135] It should be noted that the first node set in this application includes two types: paper nodes and patent nodes. In practice, for these two types of nodes, there are cases where patent nodes cite paper nodes, but there are no cases where paper nodes cite patent nodes.

[0136] A7. Construct the first feature matrix based on the fused feature vector corresponding to each data point in the intermediate dataset.

[0137] The fusion feature vector corresponding to each data point contains information about the semantics, time, and structure of that data. Arranging the fusion feature vectors of all nodes in order forms the first feature matrix.

[0138] A8 constructs a first-time metaset based on the publication time of each data item in the intermediate dataset.

[0139] The publication time of each piece of data is used as the timestamp of that data, and these time information are integrated into the first time element set.

[0140] A9 uses the first set of nodes, the first set of edges, the first feature matrix, and the first set of time elements as the initial knowledge network.

[0141] The initial knowledge network in this application is denoted as D. The initial knowledge network D adopts a standardized graph data format and includes a first set of nodes denoted as V, a first set of edges E, a first feature matrix H, and a first set of time elements T.

[0142] Figure 2 This is a schematic diagram of the structure of an initial knowledge network provided in an embodiment of this application. (Combined with...) Figure 2 It can be seen that the initial knowledge network includes nodes, edges, and metadata. Among them, nodes include node identifier, type, title, summary, keywords, year, citation count, and features such as semantics, time, structure, and fused knowledge representation; edges include edge identifier, source node, target node, weight, type, and timestamp; metadata covers total number of nodes, total number of edges, time range, processing time, and quality report information such as completeness, accuracy, and coverage uniformity.

[0143] It should be noted that, in order to improve data processing efficiency, this application adopts a distributed parallel query architecture to process massive amounts of academic data. The degree of parallelism is set according to the formula (6):

[0144] (6)

[0145] In formula (6) For optimal parallelism, Total data volume For batch size, The available computing cores are specified. This application employs asynchronous I / O and connection pooling techniques to ensure efficient utilization of database connections. An adaptive request frequency control mechanism is implemented to address API (Application Programming Interface) limitations of different data sources, preventing service interruptions due to excessive requests.

[0146] After obtaining the initial knowledge network, a heterogeneous knowledge network graph G=(V,E,T,Φ) is constructed. In the heterogeneous knowledge network graph G, V is the first set of nodes, including all paper nodes and patent nodes; E is the first set of edges; T is the first set of time elements, which records the time information of each node and each edge; Φ is the node type mapping function, Φ:V→{0,1}, mapping nodes to type identifiers. Where V={v_paper∪v_patent}; v_paper is the paper node, and v_patent is the patent node.

[0147] In this application, the node type classification adopts a discrete coding scheme; wherein, paper node: type_id=0, type_embedding =[1,0]; patent node: type_id=1, type_embedding=[0,1].

[0148] Discrete encoding schemes facilitate neural network processing while maintaining explicit type information and computational efficiency. Graph initialization follows three basic principles: balanced distribution of node types, reasonable heterogeneous connections, and consistency of temporal information, laying a structural foundation for subsequent heterogeneous graph neural network processing.

[0149] S102, for node i and node j in the first node set, based on the number of times node i is referenced by node j and the time decay factor, determine the normalized weight of the edge with node i and node j as endpoints.

[0150] Understandably, the citation relationships between academic literature and patents exhibit a significant long-tail distribution, meaning that a small number of high-impact documents or patents receive a large number of citations, becoming core nodes in the field; while the vast majority of documents or patents have extremely low citation counts, occupying the long tail end of the distribution. Directly using the original citation count as the network edge weights would cause highly cited documents and patents to dominate the network, excessively amplifying the weights of their associated edges. This could mask the potential citation relationships between long-tail documents and patents, thereby distorting the true flow path of knowledge within the literature and patent network and causing the modeling of knowledge dissemination patterns and field evolution trends to deviate from reality.

[0151] To address the aforementioned issues, this application employs logarithmic transformation and time decay mechanisms to determine the normalized weights of edges in the first node set with nodes i and j as endpoints.

[0152] The logarithmic transformation and time decay mechanism specifically include the following steps:

[0153] B1 is the base weight obtained by performing a logarithmic operation on the number of times node i is referenced by node j.

[0154] Specifically, substituting the number of times node i is referenced by node j into formula (7) yields the basic weights of the edges with nodes i and j as endpoints. The expression for formula (7) is:

[0155] (7)

[0156] in, This represents the basic weight of the edge with nodes i and j as endpoints; This represents the number of times node j references node i.

[0157] Understandably, in the actual scientific research and industrial transformation chain, academic papers and patent documents bear different core values, and the difference in their positioning directly determines the unidirectional nature of the citation relationship: academic papers, as the core carrier of scientific research, focus on cutting-edge theoretical exploration, breakthroughs in basic principles, and innovation in research methods, and are the theoretical cornerstone for building a knowledge system in a field; while patent documents are engineering application-oriented, focusing more on the feasibility verification of technical solutions, practical design, and industrial implementation protection, and are the bridge connecting theoretical achievements and actual products.

[0158] This difference in positioning, when mapped onto citation relationships, manifests as a clear one-wayness: in the process of technological research and development, the technical solutions of patents often need to be based on scientific principles, experimental data, or innovative methods that have been verified in papers. For example, a patent for a certain type of new material may cite the composition design theory of relevant materials science papers, and a patent for a certain type of algorithm may draw on the core logic of papers in the field of computer science. Therefore, it is a common practice for patents to cite papers, which is essentially the reference and application of basic scientific research results by engineering practice.

[0159] In contrast, the core objective of academic papers is to present original theoretical explorations, and their value is judged based on academic innovation rather than the feasibility of technological implementation. The technical solutions disclosed in patent literature mostly pertain to application-level implementation details and typically lack the theoretical breakthroughs or academic foresight required for academic papers. Therefore, they are unlikely to serve as the core basis for a paper's argument. Consequently, in academic research settings, the citation of patents is extremely rare, almost negligible. This one-way citation relationship also reflects, from another perspective, the typical path of transforming basic research into engineering applications.

[0160] B2, by substituting the basic weights of the edges with nodes i and j as endpoints, the absolute value of the time difference between the publication time of node i and the publication time of node j, and the time decay factor into the first objective formula, we obtain the normalized weights of the edges with nodes i and j as endpoints.

[0161] The formula for the first objective is shown in formula (8):

[0162] (8)

[0163] In formula (8) This represents the normalized weight of the edge with nodes i and j as endpoints; This represents the time decay factor, used to control the decay rate (typical value 0.1). The basic weights are the edges with nodes i and j as endpoints, which are obtained by performing a logarithmic operation on the number of times node i is referenced by node j in this application. It is an exponential decay function, simulating the weakening of the influence of knowledge over time.

[0164] S103, for node i in the first node set, based on the fusion feature vector of node i, the aggregated feature vector of each type of neighboring node of node i, and the attention weight between node i and the neighboring node, determine the multi-type information fusion feature vector of node i.

[0165] Paper nodes and patent nodes have different attribute structures, citation patterns and evolutionary characteristics, so a special attention mechanism is needed to model this heterogeneity.

[0166] This step uses the HGT (Heterogeneous Graph Transformer) framework, which uses a type-aware attention mechanism to achieve information propagation and feature updates between heterogeneous nodes.

[0167] The core of the HGT mechanism is the calculation of heterogeneous attention weights. In the process of calculating attention weights, this application considers three key factors simultaneously: differences in node types, temporal distance, and semantic relevance.

[0168] In this application, the multi-head attention mechanism uses 8 attention heads and the dropout rate is set to 0.1 to balance model complexity and the risk of overfitting.

[0169] For node i and node j in the first node set, the attention weight between node i and node j can be calculated using formula (9).

[0170] The expression for formula (9) is:

[0171] (9)

[0172] In formula (9) This represents the attention weight of node i to node j, and its range is [0,1]. It is a node The fused feature vector, It is a node fused feature vectors; The learnable weights for the query (256×64); The learnable weight matrix for the key (256×64); The dimension of the attention head (64); Indicates node type bias (determined based on the type of neighboring nodes); It represents the time distance offset (calculated based on the time difference); the node types include paper nodes and patent nodes.

[0173] For node i in the first node set, based on the first edge set corresponding to the first node set, all neighboring nodes of node i and the types of neighboring nodes can be determined; among them, the types of neighboring nodes are patent nodes or paper nodes.

[0174] It is understandable that those skilled in the art can also use formula (9) to calculate the attention weight between node i and its neighboring nodes after determining each of node i's neighboring nodes.

[0175] After determining the attention weight of node i to node j, and the type of each neighbor node of node i, the aggregate feature vector of the neighbor nodes of node i can be calculated by formula (10).

[0176] The expression for formula (10) is:

[0177] (10)

[0178] In formula (10) Represents a node A specific type of neighbor set, This is the aggregated feature vector of this type of neighbor; This represents the attention weight of node i to node j; Represents a node The fused feature vector.

[0179] It is understandable that since the neighboring nodes include two types of nodes, namely paper nodes and patent nodes, the aggregated feature vector of the paper type neighbor corresponding to node i and the aggregated feature vector of the patent type neighbor corresponding to node i can be calculated according to formula (10).

[0180] Furthermore, the multi-type information fusion feature vector of node i is determined by formula (11);

[0181] (11)

[0182] In formula (11), The multi-type information fusion feature vector of node i; Let i be the self-feature transformation matrix of node i (256×256). The information transformation matrix (256×256) of the paper type neighbors of node i. The information transformation matrix (256×256) of the patent type neighbors of node i. It is the aggregated feature vector of the neighboring nodes of the paper type of node i. Let be the aggregated feature vector of the neighboring nodes of the patent type of node i.

[0183] The mechanism in this step enables each node to obtain different information updates based on the differences in the types of its neighbors. Paper nodes pay more attention to the academic influence from paper neighbors, while patent nodes pay more attention to the technical connections with patent neighbors. At the same time, cross-type connections are used to model the bidirectional flow of academic and technical knowledge.

[0184] The multi-type information fusion feature vectors of each node in the first node set are added to the first feature matrix to obtain the second feature matrix.

[0185] S104, for node i and node j in the first node set, based on the relationship between the semantic similarity between node i and node j and the preset semantic similarity threshold, determine whether to add a semantic enhancement edge between node i and node j.

[0186] The dissemination of academic knowledge relies not only on explicit citations but also on a wealth of potential connections based on topical similarity. Many highly related research papers may not have direct citations due to factors such as similar publication dates, separate author networks, or journal differences, but they should still possess semantic connections within the knowledge flow network. This step automatically discovers and supplements these hidden knowledge connections by calculating the deep semantic similarity between nodes.

[0187] For nodes i and j in the first node set, the semantic similarity between nodes i and j can be calculated based on formula (12). The expression of formula (12) is:

[0188] (12)

[0189] In formula (12) For nodes and nodes The semantic similarity between the contents (also known as semantic similarity) ranges from [0,1]. It is the semantic feature vector of node i (also known as the text semantic embedding vector (768-dimensional)). The semantic feature vector of node j.

[0190] Understandably, cosine similarity calculation can measure the angular similarity of two vectors in a high-dimensional space, is not affected by the length of the vectors, and is suitable for comparing the semantic relevance of texts of different lengths.

[0191] After obtaining the node and nodes After determining the semantic similarity between nodes, this semantic similarity can be compared with a preset semantic similarity threshold. If the nodes... and nodes If the semantic similarity between nodes i and j is greater than the semantic similarity threshold, it is inferred that nodes i and j have a potential topical association in the knowledge space. A semantic enhancement edge is automatically added between nodes i and j in the first edge set, and the base weight of this semantic enhancement edge is the node's weight. and nodes Semantic similarity between nodes; if nodes and nodes If the semantic similarity between nodes is not greater than the semantic similarity threshold, then it is inferred that nodes i and j do not have potential topical association in the knowledge space.

[0192] It should be noted that in this application, by controlling the size of the preset semantic similarity threshold, the maximum number of semantic edges that can be added to each node is limited to 20, so as to avoid the graph structure being too dense.

[0193] For example, the semantic similarity threshold is set to 0.7 in this application.

[0194] Understandably, content similarity modeling can automatically discover potential knowledge flow pathways that are not formally cited but are highly related to the topic, thus making up for the edge-missing problem in explicit citation networks.

[0195] S105, if a semantic enhancement edge is added between node i and node j, the normalized weight of the semantic enhancement edge is determined based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j.

[0196] Knowledge flow is not a static process, but rather exhibits significant evolutionary characteristics over time. On a timeline, specific research topics undergo a dynamic life cycle, from their initial emergence and rapid development to their maturation and gradual decline. Traditional semantic analysis methods often build models based on a static perspective, focusing only on textual information within a specific time slice, completely ignoring the evolutionary clues inherent in the temporal dimension. This results in an inability to effectively capture the temporal development patterns of research hotspots, and consequently, a difficulty in accurately depicting the dynamic characteristics of knowledge flow over time.

[0197] To address the aforementioned issues, this step introduces time series modeling techniques. By dynamically tracking and associating the semantic features of nodes under different time windows, a time-sensitive feature representation method is constructed. This method not only preserves the core semantic information of nodes but also integrates the temporal dependence of knowledge flow into the feature representation, providing more realistic evolutionary feature inputs for subsequent tasks predicting the direction and intensity of knowledge flow.

[0198] For node i in the first node set, after determining the semantic feature vector of node i, the time-aware feature vector of node i can be established based on formula (13). The expression of formula (13) is:

[0199] (13)

[0200] In formula (13) It is a node Time-aware feature vectors; For nodes The time context information, including the year of publication and relative time position; It is the semantic feature vector of node i; LSTM is a time series encoder that learns the dynamic change pattern of node semantic features in the time dimension, and can capture the evolution trajectory and periodic features of the research topic.

[0201] After obtaining the node After obtaining the time-aware feature vectors, the temporal modulation semantic similarity between node i and node j can be calculated based on formula (14). The expression of formula (14) is:

[0202] (14)

[0203] In formula (14) The temporal modulated semantic similarity between node i and node j reflects the true correlation between the two nodes after considering temporal evolution. It is a node Time-aware feature vectors; It is a node The time-aware feature vector.

[0204] To model time sensitivity, this application introduces the exponential decay of time difference to calculate the time influence factor, and the calculation formula is shown in formula (15):

[0205] (15)

[0206] In formula (15) For nodes and nodes The time influence factor between them, with a value range of [0,1]. It is in S102 (Time decay factor); The time decay parameter is set to 0.05. For nodes Release time and node The time difference between the release times; The temporal modulation semantic similarity between node i and node j; The term represents the exponential decay, simulating the time decay effect of knowledge dissemination.

[0207] Time Influence Factor By combining semantic relevance and temporal proximity, smaller time differences are assigned higher influence weights, consistent with the principle in knowledge dissemination that "new research is more likely to influence concurrent research." This design enables the model to distinguish between contemporaneous related research and indirect influences across eras, improving the temporal accuracy of knowledge flow modeling.

[0208] The topology of a network contains rich implicit relationship information. The structural position of a node in the network, its neighbor distribution, and its global importance all reflect its role and influence in the flow of knowledge. Nodes with similar neighbor structures often have potential connections in research topics, methodologies, or application areas, even if there is no direct citation relationship between them.

[0209] This step analyzes the topological characteristics of nodes to discover potential connections based on network location and influence. In citation and cooperative networks, the neighbor structure and global centrality information of nodes are crucial for predicting potential connections.

[0210] In this application, neighbor structural similarity is quantified using the Jaccard coefficient; the expression for the Jaccard coefficient is shown in formula (16). Formula (16) is as follows:

[0211] (16)

[0212] In formula (16) For nodes and nodes The similarity of the set of neighboring nodes, with values ​​ranging from [0,1]. and They are nodes and nodes The set of neighboring nodes; This represents the size of the intersection of two neighbor sets, i.e., the number of common neighbors; This represents the size of the union of two neighbor sets, i.e., the total number of distinct neighbors. The Jaccard coefficient measures the degree of overlap between the neighbors of two nodes; a higher value indicates that their local environments in the network are more similar.

[0213] In this application, the difference in global importance of nodes is quantified by betweenness centrality, as shown in formula (17):

[0214] (17)

[0215] In formula (17) For nodes and nodes The centrality dissimilarity, with a value range of [0,1]; and Representing nodes respectively and The betweenness centrality of a node is used to measure its bridging importance in the shortest path. The absolute difference in the centrality of the two nodes; Used to normalize the denominator, ensuring that the degree of difference is in the range [0,1].

[0216] After obtaining the node and nodes Similarity of the set of neighboring nodes and nodes and nodes Centrality Dissimilarity Then, the structural correlation degree can be determined based on formula (18). The expression of formula (18) is:

[0217] (18)

[0218] In formula (18) The structural correlation between node i and node j is represented by a weight of 0.5, which is used to balance the contributions of local similarity and global differences. This design can identify node pairs that have similar research backgrounds but different levels of influence, providing a structural basis for modeling potential academic collaborations and knowledge transfer. The meanings of the remaining letters in formula (18) are described in the previous embodiments and will not be repeated here.

[0219] Based on the foregoing description, we can obtain the semantic similarity (content similarity) between node i and node j, the temporal modulation semantic similarity (temporal influence factor), and the similarity (structural association) of the neighbor node set. These three types of similarity information capture different aspects of knowledge flow and need to be integrated through an effective fusion mechanism to form a comprehensive assessment of association strength.

[0220] This application designs an adaptive weight learning method based on an attention mechanism, which allows the model to automatically adjust the contribution weights of each dimension of the features according to the specific node.

[0221] The adaptive weights are calculated using a multilayer perceptron and softmax normalization; the specific calculation process is shown in formula (19):

[0222] (19)

[0223] in, This is the adaptive weight for the content dimension, i.e., the first weight, with a value range of [0,1]. This is the adaptive weight for the time dimension, i.e., the second weight, with a value range of [0,1]. The adaptive weights for the structural dimension, i.e., the third weights, have a value range of [0,1]; MLP represents a multilayer perceptron with a structure of [3,16,8,3] and the activation function is ReLU; The concatenated vector of the three-dimensional features is used as the input to the MLP; the meanings of the other letters in formula (19) are as described in the previous embodiment, and will not be repeated here.

[0224] After obtaining the weights for each of the three dimensions, the association strength between node i and node j can be obtained based on formula (20). The expression for formula (20) is:

[0225] (20)

[0226] In formula (20) For nodes and nodes The correlation strength between nodes i and j, with a value range of [0,1], represents an adaptive combination of content similarity (semantic similarity), temporal correlation (temporally modulated semantic similarity), and structural correlation (similarity of neighboring node sets) between nodes i and j.

[0227] It should be noted that the nodes in this application and nodes The strength of the correlation between , which serves as the normalized weight for adding semantically enhanced edges between node i and node j.

[0228] It is understandable that after obtaining the normalized weight of each edge in the first edge set, the basic weight of the corresponding edge in the first set can be updated using the normalized weight. At the same time, after adding semantic enhancement edges between node pairs (such as node i and node j) in the first node set and obtaining the weight of the semantic enhancement edges, the semantic enhancement edges can be added to the first node set, and the normalized weight of the semantic enhancement edges can also be added to the corresponding semantic enhancement edges to obtain the second edge set.

[0229] The mechanism in S105 can dynamically adjust weights based on the specific features of each node pair. For example, it increases the weight of the time dimension for node pairs with similar publication times, and increases the weight of the structural dimension for node pairs with high neighbor overlap. Adaptive weight learning avoids the subjectivity of manually setting fixed weights, thus improving the objectivity and accuracy of the fusion results.

[0230] In actual academic literature and patent networks, the existence of isolated nodes is a common and unavoidable phenomenon, and its occurrence is mostly related to the timeliness or domain characteristics of the node itself. For example, newly published academic literature and early patents in emerging technology fields have not yet accumulated enough citation relationships due to their short publication time; or interdisciplinary research results are difficult to integrate into the explicit connection system of the existing network due to the strong cross-domain nature of the fields involved.

[0231] In traditional network analysis based on explicit relationships (such as direct references and cooperative associations), these isolated nodes are often excluded from the analysis scope due to a lack of quantifiable connection information. However, in fact, they often carry the exploratory results of cutting-edge research directions or groundbreaking technical solutions, and are important carriers reflecting the innovative vitality of the field.

[0232] To address this issue, this step introduces a generative learning approach. By mining the semantic features (such as document topics and patent technology keywords), domain background (such as subject classification and technical field tags) and potential association patterns of isolated nodes, it automatically discovers and establishes potential connections that conform to domain logic. This effectively fills the connection gaps in the network structure, enhances the integrity and representativeness of the academic patent network, and ensures that cutting-edge innovative achievements are not overlooked in network analysis.

[0233] Figure 3 This is a flowchart illustrating a method for data processing of isolated nodes, provided as an embodiment of this application. (In conjunction with...) Figure 3 As shown, the method for data processing of isolated nodes disclosed in this application includes:

[0234] S301, Identify the isolated nodes in the first node set.

[0235] In the context of knowledge networks (or graph structures), an isolated node is a node that is not connected to any other node. In this application, an isolated node is a node in the first set of nodes that is not connected to any other node, meaning that the node has no corresponding edge in the first set of edges. For example, if a document (corresponding to node A) does not cite any other document and is not cited by other documents, then node A is an isolated node.

[0236] After obtaining the second edge set in this application, all isolated nodes in the first node set can be determined.

[0237] It is understood that the number of isolated nodes in the first node set can be one or more. In the following embodiments of this application, an isolated node is used as an example to introduce the specific implementation of determining the candidate connected nodes of the isolated node, the correlation degree between the isolated node and the corresponding candidate connected node, and the weight of the edge with the isolated node and the candidate connected node as endpoints.

[0238] S302, Based on the fused feature vector, variational autoencoder and decoder of the isolated node, candidate connection nodes of the isolated node are generated.

[0239] After identifying isolated nodes, the fused feature vectors of the isolated nodes are input into a variational autoencoder. The variational autoencoder learns the latent distribution of node features to obtain the distribution parameters of isolated nodes in the latent space. Then, the features are reconstructed based on the decoder to obtain candidate connected nodes of the isolated nodes.

[0240] The expression for the encoder's latent distribution parameters is shown in Equation (21), the expression for latent variable sampling is shown in Equation (22), the expression for the decoder's reconstructed features is shown in Equation (23), and the expression for candidate connection search is shown in Equation (24).

[0241] , (twenty one)

[0242] (twenty two)

[0243] (twenty three)

[0244] (twenty four)

[0245] In formulas (21) to (24), and These are nodes The mean vector and standard deviation vector in the latent space; For the encoder network, 256-dimensional features are mapped to a 128-dimensional latent space; The latent variables (128 dimensions) are sampled from a Gaussian distribution. For the decoder network, it reconstructs from the latent space to the original feature space; The reconstructed node feature vectors; This is a nearest neighbor search algorithm based on cosine similarity; the meanings of the other letters are explained in the previous embodiments and will not be repeated here.

[0246] The reconstruction process of variational autoencoders can learn the inherent laws and distribution patterns of node features. By sampling and reconstructing in the latent space, it can discover candidate connected nodes with similar features to isolated nodes.

[0247] S303, based on the preset prompt template and large language model, calculate the correlation degree between the isolated node and the candidate connection node.

[0248] It is understandable that an isolated node can have one or more candidate connected nodes. In this step, we will take one candidate connected node corresponding to an isolated node as an example to introduce the specific method for calculating the correlation degree between an isolated node and its candidate connected nodes.

[0249] Large language models possess rich domain knowledge and commonsense reasoning capabilities, enabling them to analyze potential knowledge relationships between nodes based on text content. This step utilizes the large language model to verify the semantic rationality and evaluate the association strength of candidate connections generated by the variational autoencoder; and leverages the commonsense reasoning capabilities of the large language model to generate knowledge propagation paths between isolated nodes and their corresponding candidate connection nodes.

[0250] After identifying isolated nodes and their corresponding candidate connection nodes, the core features of the isolated nodes and individual candidate connection nodes (such as the research topic of the literature, the technical solution of the patent, etc.) are transformed into structured text prompts that conform to the input format of the large language model based on the prompt word template. Then, the prompts are input into the large language model, and the model's ability to understand semantic and technical correlations is used to output the correlation degree between the isolated node and the candidate connection node, as well as the knowledge propagation path between the isolated node and the candidate connection node.

[0251] For example, the prompt template could be:

[0252] "Based on the information from the following two research nodes, analyze the possible knowledge dissemination paths between them:"

[0253] Node A: Title = {title_A}, Keywords = {keywords_A}, Abstract Excerpt = {abstract_A[:200]}

[0254] Node B: Title = {title_B}, Keywords = {keywords_B}, Abstract Excerpt = {abstract_B[:200]}

[0255] Please output: 1) Relevance score (0-1); 2) Explanation of knowledge dissemination path.

[0256] S304, Based on the correlation degree value and the time decay factor, determine the normalized weight of the edge with the isolated node and the candidate connected node as endpoints.

[0257] After obtaining the correlation value between the isolated node and a single candidate connection node, the normalized weight of the edge with the isolated ground and the candidate connection node as endpoints can be obtained based on formula (25).

[0258] The expression for formula (25) is:

[0259] (25)

[0260] In formula (25) It is an isolated node. It is an isolated node Candidate connection nodes; It is the normalized weight of the edge with nodes a and b as endpoints, and its range is [0,1]. It is the correlation degree between node a and node b, and its range is [0,1]. It is the time decay factor between node a and node b; It is the absolute value of the time difference between the publication time of node a and the publication time of node b.

[0261] It should be noted that this application only retains The method employs high-confidence connections, limiting each isolated node to a maximum of five candidate connection nodes, and randomly selecting 10% of the generated connections for expert evaluation. This combination of variational autoencoder and large language model leverages both data-driven feature learning and knowledge-driven semantic reasoning capabilities, enabling the generation of high-quality potential connections for isolated nodes and significantly improving the network's integrity and coverage.

[0262] In passing Figure 2 The method shown in the diagram determines isolated nodes in the first node set and candidate connected nodes corresponding to the isolated nodes. Then, the normalized weights of the edges with the isolated nodes and candidate connected nodes as endpoints can be calculated. The edges connecting the isolated nodes and candidate connected nodes, along with their normalized weights, are added to the second edge set to obtain the third edge combination.

[0263] In this application, the network consisting of the first set of nodes, the second feature matrix, the third set of edges, and the first set of time elements is used as the enhanced network G'.

[0264] It should be noted that the operations involved in S102-S105 and S301-S304 in the aforementioned embodiments of this application are all localized processing steps for nodes during the network construction phase, and their core function is to complete the initial construction of the network structure. After the complete network structure is generated, this application will carry out deep graph neural network (GNN) learning based on the complete network structure. In this process, relying on the hierarchical progression characteristics of the multi-layer GNN architecture, the richer connection relationships and topological information in the network will be fully utilized to achieve deeper semantic and structural feature representation learning of nodes, thereby improving the effectiveness and discriminability of node representation.

[0265] After completing the graph structure construction and semantic enhancement, a deep graph neural network needs to be designed to learn high-quality node representations. Traditional graph neural networks typically only consider topological information, ignoring temporal evolution features and differences in heterogeneous node types. This step designs a unified neighbor aggregation framework that simultaneously handles heterogeneous type differences and temporal sensitivity, achieving more accurate knowledge flow modeling.

[0266] The unified neighbor aggregation formula based on neural networks is shown in formula (26):

[0267] (26)

[0268] In formula (26) Indicates the first Layer nodes Aggregated neighbor feature vectors (256 dimensions); For the first Layer nodes Aggregated neighbor feature vectors (256 dimensions); Based on neighbor nodes The weight coefficients are determined by the node type; The time-related factor; For nodes The set of all neighboring nodes; This is a type-time-aware aggregate function.

[0269] in, .

[0270] In the unified neighbor aggregation process of the neural network in this application, the node feature update uses residual connections to ensure training stability, as shown in formula (27):

[0271] (27)

[0272] In formula (27) It is a learnable transformation matrix (512×256). For the activation function, please refer to the explanation in formula (26) for the meaning of the other letters, which will not be repeated here.

[0273] To better model the combined effects of temporal evolution and heterogeneous types in the node embedding space, this step introduces a relation transformation mechanism based on RotatE. This mechanism draws on the relation modeling ideas in knowledge graph embedding, expressing complex relational patterns between nodes through geometric transformations in the feature space, particularly evolutionary relationships in the temporal dimension and transformation relationships between heterogeneous types.

[0274] The process of constructing the time-type joint relation vector is shown in formula (28):

[0275] (28)

[0276] In formula (28) It is an enhanced relation vector. Represents a node The corresponding type of learnable relation vector (256 dimensions), paper type Patent type ; For nodes The corresponding time relationship vector; This enables element-wise multiplication operations for Hadamard, allowing for the interactive integration of type and time information.

[0277] The calculation process of the knowledge-enhanced transformation vector is shown in formula (29):

[0278] (29)

[0279] In formula (29) Enhance knowledge by transforming vectors; For the current layer node Feature representation; It is an enhanced relation vector; the Hadamard product enables the transformation vector to be adaptively adjusted according to the type of node and time characteristics.

[0280] The calculation process for feature enhancement and update is shown in formula (30):

[0281] (30)

[0282] In formula (30) This is the node feature vector after knowledge enhancement transformation (also known as the embedded feature vector corresponding to node i). To balance the parameters, the transformation intensity is controlled (typical value 0.3); the addition operation ensures residual connectivity and maintains the stability of gradient propagation.

[0283] The mechanism of relation transformation based on RotatE works by using nodes. Learnable relation vectors of corresponding types Capture the intrinsic feature patterns of different types of nodes; through nodes Relationship vectors corresponding to time The modeling approach establishes general laws governing temporal evolution; it achieves interactive modeling of type and time through Hadamard product; and it ensures the controllability of transformations and training stability through residual connections. The RotatE-based relation transformation mechanism enables node embeddings to form type-related and time-sensitive structures in geometric space, improving the accuracy of similar node retrieval and relation prediction in downstream tasks.

[0284] Furthermore, to ensure that similar nodes are close in the embedding space, maintain the ability to adapt to nodes and abnormal structures, ensure that important information is not lost, and avoid the model from overfitting the training data, after the node features have been integrated with rich semantic, temporal and structural information through multi-layer neighbor aggregation and relation transformation, this application introduces a feature reconstruction mechanism to perform aligned feature reconstruction.

[0285] The feature reconstruction loss of the variational autoencoder style is shown in Equation (31).

[0286] (31)

[0287] In formula (31) The feature reconstruction loss consists of two parts: reconstruction error and distribution alignment. The original node feature vector (256 dimensions); The encoder maps features to a latent space (128 dimensions). The decoder reconstructs the original space (256 dimensions) from the latent space. This represents the L2 reconstruction error, ensuring that information is preserved during the encoding-decoding process; KL divergence encodes the distribution. Align to prior distribution ; is the encoding vector in the latent space.

[0288] The system's total loss function integrates graph learning and reconstruction constraints. The expression for the system's total loss function is shown in formula (32):

[0289] (32)

[0290] In formula (32) It is the total loss. The learning loss of the graph neural network model (loss for tasks such as node classification and edge prediction). For feature reconstruction loss To reconstruct the loss weight coefficients (typically 0.1), weight balancing ensures that the main task is optimized while maintaining the good properties of the feature space.

[0291] The above methods, through an encoder-decoder loop, verify the quality and stability of the node representations, providing a reliable feature foundation for subsequent downstream tasks such as link prediction and cluster analysis. The introduction of reconstruction loss further enhances the model's generalization ability, enabling it to better handle dynamically changing knowledge network structures.

[0292] It is understandable that after the data in the augmented network G' is processed by formulas (26) to (32), the third feature matrix (a matrix composed of the embedding feature vectors corresponding to each node) is obtained, also known as the node embedding matrix H'.

[0293] It should be noted that each node in the node embedding matrix H' corresponds to a 256-dimensional dense vector representation, which includes semantic information, temporal information, structural information, and heterogeneous information; this 256-dimensional dense vector representation is the result after L2 norm normalization.

[0294] Figure 4 This is a schematic diagram illustrating the structure of an enhanced knowledge network provided in an embodiment of this application. (Combined with...) Figure 4 As shown, the enhanced knowledge network G' includes a node set, an edge set, and edge statistics. The node set in the enhanced knowledge network G' has the same format as the first node set D, and the features of the nodes in this node set have been updated using HGT (Heterogeneous Graph Transformer). The edge set in the enhanced knowledge network G' contains three types of edges: reference edges with updated weights, semantically enhanced edges with semantic similarity greater than 0.7, and generated edges (completed edges) obtained by combining a variational autoencoder (VAE) with a large language model (LLM). The edge statistics in the enhanced knowledge network G' cover the number of original edges (original_count = 250,000), the number of newly added semantic edges (semantic_added = 15,000), the number of newly added generated edges (generated_added = 3,000), and the total number of edges in the network (total_edges = 268,000).

[0295] After obtaining the knowledge-enhancing network G' and the node embedding matrix H', it is possible to base it on... Figure 5 The method described herein predicts potential future knowledge flow connections.

[0296] Figure 5 A flowchart illustrating the knowledge flow prediction method provided in this application's embodiments. (In conjunction with...) Figure 5 As shown, the flowchart of the knowledge flow prediction method provided in this application embodiment includes:

[0297] S501, for node i and node j in the first node set, substitute the embedded feature vector of node i and the embedded feature vector of node j into the bilinear bisector function to obtain the original association score between node i and node j.

[0298] Predicting knowledge flow essentially involves predicting edges that don't currently exist in the network but may appear in the future. Based on node embeddings learned from graph neural networks, the likelihood of nodes forming connections can be assessed by calculating the similarity of node pairs in the embedding space. Node embeddings encode rich semantic, temporal, and structural information; therefore, distances in the embedding space can reflect the strength of potential associations between nodes.

[0299] This application employs a bilinear scoring function to define the original correlation score between node pairs in the first node set. The expression for the bilinear bisection function is shown in formula (33):

[0300] (33)

[0301] In formula (33) For nodes and nodes The original correlation score between them, with a value range of (0,1); It is the embedded feature vector (256 dimensions) of node i. It is a node Embedded vector (256 dimensions); The learnable relation projection matrix, with dimensions 256×256, is used to capture complex interaction patterns between nodes. It is a bilinear operation, which can express complex similarity patterns better than the simple inner product; The bias term between node i and node j is used to model asymmetric association patterns. The Sigmoid function normalizes the score to the range (0,1).

[0302] Bilinear scoring has stronger expressive power than simple inner product calculation, enabling it to learn nonlinear interactions between node features. Bias term. The design takes into account structural biases in academic networks, such as the fact that literature from high-impact journals is more likely to be cited and research from well-known institutions is more likely to receive attention.

[0303] S502, the original correlation score is adjusted based on the time decay modulation function to obtain the corrected correlation score.

[0304] The citation and dissemination of academic knowledge exhibit significant time constraints. Generally, research primarily cites literature published before its initial publication date, rarely involving "past citations of future works." Furthermore, the probability of establishing a citation relationship decreases as the time lag increases, meaning researchers tend to cite the most recent relevant work. This step modulates the original relevance score using a time decay mechanism to ensure the prediction results conform to the temporal logic of knowledge dissemination.

[0305] The expression for the time decay modulation function is shown in formula (34):

[0306] (34)

[0307] In formula (34) The corrected correlation score between node i and node j after time modulation, with a value range of (0,1); The corrected association score between node i and node j; This is a time decay hyperparameter that controls the strength of time sensitivity (typical value 0.2). Let be the time difference between the publication time of node i and the publication time of node j; It is an exponential decay function, simulating the phenomenon that the citation probability decreases as the time difference increases.

[0308] In citation relationship modeling, when the publication time of node i is significantly later than that of node j, the probability of a true citation association between the two is usually low; based on this reason, an exponential decay function is introduced. This function adaptively adjusts the edge weights to correct the potential reference probability over different time spans.

[0309] The core value of this design lies in enabling the model's reference prediction logic to break through the single embedding similarity dependency and further incorporate the constraints of the time dimension: on the one hand, through the implicit filtering of the time window, it ensures that the predicted edges, in addition to semantic association and network structure adaptability, also meet the temporal rationality of knowledge dissemination; on the other hand, it effectively avoids prediction results that violate objective laws, such as "time travel references," thereby improving the scientificity and credibility of the model's predictions.

[0310] S503, using a breadth-first search algorithm, search for all propagation paths from node i to node j whose path length does not exceed a preset path length threshold, and calculate the propagation intensity score for each propagation path.

[0311] Direct node similarity scoring can only identify direct potential connections between two nodes, but real-world knowledge propagation often occurs through intermediary nodes and indirect paths. For example, research by node A may inspire node B, and node B's work further influences node C, thus forming a knowledge propagation chain A→B→C. To model this potential multi-hop diffusion pattern, this module introduces a multi-step probabilistic path aggregation strategy. By analyzing multi-step path patterns in the network, it discovers these indirect knowledge flow relationships, improving the ability to model complex propagation patterns.

[0312] For node i and node j in the first node set, a breadth-first search algorithm is used to search for all propagation paths from node i to node j whose path length does not exceed a preset path length threshold. For example, in this application, the preset path length threshold is set to 3.

[0313] The set of all paths from node i to node j whose length does not exceed a preset path length threshold is shown in formula (35):

[0314] (35)

[0315] In formula (35) It is a preset path length threshold. The starting point of the path is node i. The endpoint of the path is node j; This represents the set of all propagation paths from node i to node j whose path length does not exceed a preset path length threshold.

[0316] Path collection The propagation strength score of each propagation path is calculated by multiplying the edge weights, as shown in formula (36):

[0317] (36)

[0318] In formula (36) This represents the propagation intensity score for a single propagation path, with a value range of [0,1]. For the node To the node A specific transmission path; This refers to an edge in the propagation path; The weight of an edge e in the propagation path, i.e., the corrected correlation score in S502. If a propagation path consists of three edges, then the propagation strength score of that path is the product of the three weights corresponding to the three edges. Through this multiplication operation, the gradual attenuation of information during multi-step propagation is simulated.

[0319] S504, the maximum propagation intensity score among the multiple propagation intensity scores is taken as the multi-step inference score between node i and j, and the propagation path corresponding to the maximum propagation intensity score is taken as the optimal propagation path between node i and j.

[0320] The maximum propagation intensity score among multiple propagation intensity scores is taken as the multi-step inference score between nodes i and j. The calculation process is shown in formula (36):

[0321] (36)

[0322] In formula (36) This represents a path from node i to node j that is in the set of all propagation paths whose length does not exceed a preset path length threshold. The propagation strength score represents the path. The multi-step inference score between nodes i and j represents the node's inference score. Affect nodes through arbitrary paths The strongest probability; This indicates a maximum value operation, preserving the strength of the optimal propagation path.

[0323] Furthermore, the propagation path corresponding to the maximum propagation intensity score is taken as the optimal propagation path between nodes i and j.

[0324] In modeling the information propagation mechanism, this application employs a cumulative multiplication method to quantify the probability decay of each step in the propagation path, accurately simulating the characteristic of information gradually being lost along the transmission links during the diffusion process. The maximum value strategy, based on the core assumption that "the strongest path determines the propagation effectiveness," retains only the most influential path as the dominant factor in scenarios with multiple parallel propagation paths. The combination of these two approaches can not only effectively uncover overlooked indirect propagation patterns in explicit networks but also predict potential reference bridges formed through intermediary nodes, identifying potential "knowledge bridging" paths across topics and domains, providing crucial information for the early identification of cutting-edge hot topics.

[0325] Through the operations in S501 to S504, a multi-step inference score can be generated for any pair of nodes in the first node set, and the optimal propagation path corresponding to any pair of nodes can be determined.

[0326] From the core requirements of network optimization, the quantity and quality of predicted edges need to maintain a dynamic balance: including too many predicted edges will significantly increase network redundancy and noise levels, interfering with the accuracy of subsequent analyses based on the network structure; including too few predicted edges will cause the mined potential correlation information to be over-filtered, limiting the ability to explain and predict knowledge evolution trends. Therefore, during network expansion, not all optimal propagation paths determined by the model are necessary to include in the network. A scientific screening mechanism is needed to verify the prediction results to ensure the quality and reliability of the network expansion.

[0327] S505, determine whether the multi-step inference score is greater than a preset score threshold.

[0328] For the optimal propagation paths corresponding to multiple node pairs in the first node set, the optimal propagation paths of node pairs whose multi-step inference scores exceed the preset score threshold are selected as high-confidence prediction edges through a preset confidence threshold screening strategy. This process is shown in formula (37):

[0329] (37)

[0330] In formula (37) This represents the set of high-confidence predicted edges obtained through filtering. This represents the multi-step inference score between nodes i and j. This represents a preset score threshold. For example, the balanced result based on cross-validation and expert evaluation will... Set to 0.6.

[0331] If the multi-step inference score between node i and node j is greater than the preset score threshold, then the optimal propagation path between node i and node j is determined to be the high-confidence prediction edge.

[0332] S506, if it is determined that the multi-step inference score is greater than a preset score threshold, then the optimal propagation path is added to the knowledge enhancement network.

[0333] If the multi-step inference score between node i and node j is greater than the preset score threshold, then the optimal propagation path between node i and node j is determined to be the high-confidence prediction edge.

[0334] After obtaining the set of high-confidence predicted edges, batch insertion is used to add all high-confidence predicted edges to the knowledge enhancement network G', realizing the batch expansion operation of the network and obtaining the final extended knowledge flow network. This process is shown in formula (38):

[0335] (38)

[0336] In formula (38) To ultimately expand the knowledge flow network (target knowledge flow network). It is a knowledge-enhancing network; This is a set union operation; It is a high-confidence prediction edge.

[0337] The weights of the high-confidence prediction edges are determined by formula (39). Formula (39) is as follows:

[0338] (39)

[0339] Formula (39) The weights represent the high-confidence predicted edges corresponding to nodes i and j. This represents the multi-step inference score between nodes i and j. This is the predicted edge weight decay coefficient, used to reflect the uncertainty of the predicted connection relative to the known connection. For example, setting... The value is 0.8.

[0340] Figure 6 This is a schematic diagram of the structure of a target knowledge flow network provided in an embodiment of this application. (Combined with...) Figure 6 As shown, the target knowledge flow network includes a first set of nodes, a fourth set of edges, extended statistical information, and edge types. The content of the nodes in the first set is related to the enhanced knowledge network. The nodes in the same set of nodes have the same content; the fourth set of edges includes original reference edges, semantically enhanced edges, generated completion edges, and predicted future edges marked with confidence scores; the extended statistics include the number of predicted edges, average confidence, coverage, and density increase; the edge type list includes reference edges, semantic edges, generated edges, and predicted edges.

[0341] Figure 5 The method shown transforms a static historical knowledge network into a forward-looking dynamic network through predictive link generation, providing a crucial evolutionary information foundation for identifying cutting-edge hotspots and predicting trends. The target knowledge flow network G'' not only retains historical knowledge flow patterns but also incorporates predictive modeling of future development trends, significantly enhancing the system's predictive capabilities and application value.

[0342] In the context of knowledge evolution within scientific literature and patent networks, knowledge dissemination typically revolves around specific themes, forming tightly structured citation clusters and collaborative subnetworks. These clusters, as core carriers of knowledge aggregation, correspond to different research directions, technological fields, or disciplinary branches. Their structural characteristics (size, internal connection density) and association attributes (cross-domain association strength) are key indicators reflecting the activity level, technological maturity, and innovation potential of the corresponding research fields.

[0343] Therefore, after obtaining the target knowledge flow network G'' and the node embedding matrix H', clustering methods can be used to analyze the target knowledge flow network G'' and the node embedding matrix H' to obtain multiple clusters. Traditional structural clustering methods only consider topological connection information and are easily affected by network sparsity and noisy edges. This step adopts a strategy that combines embedding space clustering with graph structure constraints, which utilizes the high-quality feature representation extracted by deep learning while maintaining consistency with the original network structure.

[0344] Figure 7 This is a flowchart illustrating a data clustering analysis method provided in an embodiment of this application. (In conjunction with...) Figure 7 As shown, the data clustering analysis method disclosed in this application includes:

[0345] S701, cluster the nodes in the first node set to obtain multiple clusters, and determine the central feature vector corresponding to the clusters.

[0346] The clustering objective function used in the clustering of nodes in the first node set is shown in formula (40):

[0347] (40)

[0348] In formula (40) This is the total clustering loss function, used to balance embedding similarity and structural rationality; For nodes The 256-dimensional embedded feature vector; For nodes The central feature vector of the cluster it belongs to (256 dimensions); For nodes Cluster assignment labels, ; This is the structural constraint weight coefficient, balancing embedding clustering and structural rationality (typical value 0.5). It is a target knowledge flow network. It is the initial knowledge network; This is the network degree module, which measures the structural quality of community partitioning (cluster partitioning), with a value range of [-1, 1]. The meanings of the other letters are explained in the previous embodiments and will not be repeated here.

[0349] The expression for the network degree module is shown in formula (41):

[0350] (41)

[0351] In formula (41) For the purpose of knowledge flow network The total number of sides in; For the elements of the adjacency matrix, if the node and If there is an edge between them, the value is 1; otherwise, the value is 0. and Each node and The degree (number of connected edges); For nodes in a random network and nodes The expected probability of the connection; This is an indicator function for the same cluster; it is 1 for the same cluster and 0 for different clusters.

[0352] S702, for each node in the cluster, the embedded feature vector corresponding to the node and the central feature vector corresponding to the cluster to which the node belongs are substituted into the preset quality scoring function to obtain the cluster rationality score corresponding to the node.

[0353] The expression for the preset quality scoring function is shown in formula (42):

[0354] (42)

[0355] In formula (42) Represents a node Belonging to a cluster The clustering rationality score (quality score) ranges from [0,1]. Represents clusters The center vector (256 dimensions); Represents a node To cluster The square of the distance between the centers; Represents clusters The maximum squared distance from the embedded feature vectors of all nodes to the central feature vector; the normalization design makes the quality scores comparable, with 1 representing the best assignment and 0 representing the worst assignment.

[0356] S703, for each node in the cluster, if the cluster rationality score corresponding to the node is greater than or equal to the preset rationality score, then the node is determined to belong to the corresponding cluster; if the cluster rationality score corresponding to the node is less than the preset rationality score, then the node is deleted from the corresponding cluster to obtain the target cluster.

[0357] For example, the preset reasonableness score is set to 0.6.

[0358] For each node in a cluster, if the cluster rationality score corresponding to the node is greater than or equal to a preset 0.6, then the node is determined to belong to the corresponding cluster; if the cluster rationality score corresponding to the node is less than 0.6, then the node is deleted from the corresponding cluster to obtain the target cluster.

[0359] Furthermore, the size of the target cluster can be controlled. For example, small clusters with fewer nodes than the preset target node data (such as 10) can be deleted.

[0360] Figure 8 This is a schematic diagram illustrating a target clustering result provided in an embodiment of this application. Combined with... Figure 8 As shown, the target clustering results contain two core parts: First, detailed information about each knowledge community. Taking one cluster (cluster_id) as an example, it includes the set of nodes it belongs to (such as nodes "node_1", "node_2", etc.), the 256-dimensional center embedding vector, the cluster size (size is 150), the quality score (quality_score is 0.85), the dominant research field (such as "machine learning" and "computer vision"), and the time span (time_span is 2018-2024); Second, clustering statistics, including the total number of communities (total_communities is 28), the average size of the communities (average_size is 125), the modularity score (modularity_score is 0.72), and the number of isolated nodes (isolated_nodes is 45).

[0361] After obtaining multiple target clusters, the innovation frontier index of each target cluster can be calculated. Specifically, for each target cluster, the interdisciplinary index, novelty index, and knowledge mobility index corresponding to that target cluster are calculated. Then, by weighting the interdisciplinary index, novelty index, and knowledge mobility index, the innovation frontier index of that target cluster is obtained.

[0362] Among them, the interdisciplinary index, novelty index and knowledge mobility index are well known to those skilled in the art, and the determination process of the interdisciplinary index, novelty index and knowledge mobility index will not be described in detail in this application.

[0363] In this application, formula (42) is used to calculate the weighted interdisciplinary index, novelty index, and knowledge mobility index. The specific expression of formula (42) is as follows:

[0364] (42)

[0365] In formula (42) Innovation Frontier Index of Target Clusters It is the interdisciplinary index of the target cluster. It is the novelty index of the target cluster. It is the knowledge mobility index of the target cluster. These are the weights corresponding to the interdisciplinary index. These are the weights corresponding to the novelty index. It is the weight corresponding to the knowledge mobility index.

[0366] Optionally, in this application , and It is calculated automatically using the entropy weight method, which avoids subjective bias and ensures the objectivity of weight allocation and data-driven characteristics.

[0367] Figure 9 This is a schematic diagram of an innovation frontier index provided in an embodiment of this application. Figure 9 The document displays the values ​​of the innovation frontier index of the clusters, as well as related information on the values ​​of the interdisciplinary index, knowledge mobility index, and novelty index.

[0368] In one optional implementation, after obtaining the innovation frontier index corresponding to each of the multiple target clusters, the innovation frontier index of each target cluster can be compared with a preset index threshold. Multiple innovation frontier indices with values ​​greater than the preset index threshold are selected, and the multiple target clusters corresponding to the multiple innovation frontier indices are taken as multiple hot spot clusters.

[0369] The selection of hotspot clusters is achieved through formula (42), which is expressed as follows:

[0370] (42)

[0371] In formula (42) It is a set of hotspot clusters that have passed the screening. Indicates the first One community; Indicates community Innovation Frontier Index; The index threshold (also known as the hotspot identification threshold, with a recommended value of 0.7) is used.

[0372] It should be noted that setting the index threshold too high may miss important emerging directions; setting it too low may result in too many hot topics, thus losing its screening value. Therefore, the index threshold needs to be set reasonably based on the distribution characteristics of the innovation frontier index corresponding to multiple hot topic clusters.

[0373] In one alternative implementation, multiple hotspot clusters can be sorted in descending order of their cutting-edge innovation index.

[0374] In one alternative implementation, higher weights can be assigned to hotspot clusters with larger cutting-edge innovation indices, thus giving them greater weight in subsequent recommendations.

[0375] After identifying the hotspot clusters, it is necessary to provide users with a main overview and background explanation of the hotspot clusters so that users can quickly and accurately grasp the core information of the hotspot clusters.

[0376] In one alternative implementation, for each hotspot cluster, the text information corresponding to that hotspot cluster is obtained, and a text summary corresponding to that hotspot cluster is generated based on the text information, the target prompt word template, and the large language model.

[0377] For example, a target cue word template could be:

[0378] Based on the following collection of research literature, a core overview of this hot topic is generated:

[0379] Document titles: {titles_list}

[0380] Key terms: {key_terms}

[0381] Representative abstracts: {selected_abstracts}

[0382] Requirements: A summary of no more than 200 words, highlighting the innovative points, application value, and development potential.

[0383] This application utilizes text mining and natural language generation techniques to automatically extract core concepts and generate structured summaries from a collection of documents in hotspot clusters.

[0384] It is understandable that research hotspots exhibit lifecycle characteristics, progressing through different stages such as emergence, rapid development, maturity, and decline. After identifying multiple hotspot clusters, it is necessary to analyze these clusters to predict future development trends. To identify the development stages of hotspots and predict future trends, this application provides a historical time series and prediction model of an innovation frontier index based on hotspot clusters to determine the future development trends of these clusters.

[0385] The expression for the historical time series characteristics of the frontier innovation index of hot spot clusters is shown in formula (43):

[0386] (43)

[0387] In formula (43) Fg represents the historical time series features corresponding to the hotspot cluster g; n represents the time window length, usually 5-7 years; t represents the current time point, Fg( ) represents the time-series feature function defined for hotspot clusters.

[0388] The prediction model is a combination of a neural network model and a bidirectional long short-term memory network. The expressions for the prediction model are shown in equations (44), (45), and (46):

[0389] (44)

[0390] (45)

[0391] (46)

[0392] The meanings of the letters in formulas (44) to (46) are as follows: It is the time-series feature sequence input into the prediction model, i.e. ;CNN(·) is the operation function of a convolutional neural network layer; This indicates that the kernel size is 3. This represents the output of the convolutional neural network; BiLSTM(·) represents the operation function of the bidirectional long short-term memory network layer. This indicates that the hidden layer dimension of BiLSTM is 128; This represents the output of the BiLSTM layer; Dense(·) represents the operation function of the fully connected layer. It is the prediction output of the prediction model, which represents the estimate of the time series characteristics of the hot spot cluster at the next time point t+f (t is the current time point, f is the prediction step size, such as f=1 means predicting the next time unit). (For example, predicting the number of publications or the innovation index of the cluster in the next year.)

[0393] The loss function of the prediction model is shown in formula (47):

[0394] (47)

[0395] In formula (47) It is the total loss value of the predicted name. This represents the predicted value of future time-series features output by the prediction model; This represents the true observed values ​​that indicate the future time-series characteristics of the prediction model; The weight coefficient for the smoothness regularization term is 0.1. The smoothing regularization term prevents the prediction curve from fluctuating too drastically; MSE (Mean Squared Error) reflects the overall deviation between the predicted value and the true value by calculating the squared mean of the difference between the two.

[0396] The core of literature recommendation is calculating the relevance of candidate documents to hot topics. Candidate documents are derived from all nodes in the extended network G'', and the degree of matching between candidate documents and hot topic clusters can be evaluated through semantic similarity.

[0397] In one alternative implementation, the matching degree between candidate documents and hotspot clusters can be evaluated using formula (48). The expression for formula (48) is:

[0398] (48)

[0399] In formula (48) This represents the similarity between hotspot clusters and candidate documents learned through comparative learning. Embed the 256-dimensional feature vector at the center of the hotspot cluster; The embedded feature vector (256-dimensional) of the candidate documents; This is the set of embedded feature vectors for all candidate documents; Temperature hyperparameter controls the sharpness of the similarity distribution (typical value 0.1). It is an exponential function that converts the dot product into a positive probability.

[0400] After obtaining the similarity between the hotspot cluster and the candidate documents, the final recommendation score for the candidate documents is obtained based on formula (49). The expression of formula (49) is:

[0401] (49)

[0402] In formula (49) This indicates the final recommendation score for the candidate literature; This represents the similarity between hotspot clusters and candidate documents learned through comparative learning. The innovation frontier index represents the hotspot cluster to which the candidate literature belongs; Indicates the number of years since the candidate literature was published. This is the time decay parameter (usually set to 0.1).

[0403] When the frontier innovation index of the hot spot cluster to which the candidate literature belongs is large (e.g., greater than 0.8), the candidate literature is more likely to be recommended; when the frontier innovation index of the hot spot cluster to which the candidate literature belongs is small (e.g., less than 0.5), the candidate literature is less likely to be recommended.

[0404] In one alternative implementation, a large language model can be invoked to perform secondary evaluation and generative analysis on candidate documents, thereby ensuring recommendation quality and providing trend explanations.

[0405] The prompt word template for calling a large language model to evaluate and analyze candidate documents can be:

[0406] "Evaluate the relevance of the following literature to hot topics:"

[0407] Hot topics: {hotspot_summary}

[0408] Candidate references: {doc_title, doc_abstract, doc_keywords}

[0409] Requirements: 1) Relevance score (0-1); 2) Reasons for recommendation (within 50 words); 3) Evaluation of innovation value.

[0410] In one alternative implementation, a large language model can be invoked to perform generative analysis on hotspot clusters, thereby obtaining a description of the future development path of the hotspot clusters.

[0411] A prompt word template for using a large language model to perform generative analysis of hotspot clusters can be:

[0412] Based on the development history and predicted trends of the following popular online communities, we simulate the research and development path for the next three years:

[0413] Community Overview: {community_summary}

[0414] Historical trend: {trend_history}

[0415] Predicted metrics: {predicted_scores}

[0416] Requirements: Generate a structured description of the development path, including key milestones and turning points.

[0417] After performing generative analysis on candidate documents and hotspot clusters using a large language model, high-quality recommendations with scores greater than a preset score threshold (e.g., 0.7) can be retained; candidate documents can be re-ranked according to a comprehensive score (the product of the candidate document's final recommendation score and its large language score); and explanations of the reasons for recommendation and its innovative value can be obtained.

[0418] In one alternative implementation, a comprehensive analysis of hotspot clusters and candidate files can be performed based on a large language model to obtain a comprehensive analysis report.

[0419] For example, the comprehensive analysis report includes the ranking of innovation frontier indices, basic information on hotspot clusters, the change curve of innovation frontier indices over the next three years, the assessment of development stages, the prediction of key turning points, a list of recommended candidates with risks, reasons for recommending candidate literature, and risk warnings, etc.

[0420] A prompt word template that uses a large language model to comprehensively analyze hotspot clusters and candidate documents can be:

[0421] Based on the following analysis results, a complete report on cutting-edge hot topics will be generated:

[0422] Hotspot ranking: {hotspot_ranking}

[0423] Trend predictions: {trend_predictions}

[0424] Recommended literature: {recommendation_list}

[0425] Requirements: Generate a structured report of approximately 1000 words, including four parts: overview, trends, recommendations, and suggestions.

[0426] For example, the final generated hot topic recommendation list is as follows: Figure 10 As shown, the recommendation results are as follows: Figure 11 As shown in the trend report Figure 12 As shown.

[0427] Based on the same inventive concept, this application discloses a data processing device. Figure 13 This is a schematic diagram of the structure of a data processing apparatus provided in an embodiment of this application. (In conjunction with...) Figure 13 As shown, the data processing apparatus 130 disclosed in this application includes:

[0428] The initial knowledge network acquisition module 1301 is used to acquire an initial knowledge network; the initial knowledge network includes a first node set, a first edge set, a first feature matrix, and a first time element set; wherein, the first node set includes two types: paper nodes and patent nodes;

[0429] The first normalized weight determination module 1302 is used to determine the normalized weight of the edge with nodes i and j as endpoints for nodes i and j in the first node set, based on the number of times node i is referenced by node j and the time decay factor.

[0430] The multi-type information fusion feature vector determination module 1303 is used to determine the multi-type information fusion feature vector of node i for node i in the first node set based on the fusion feature vector of node i, the aggregated feature vector of each type of neighboring node of node i, and the attention weight between node i and the neighboring node.

[0431] The semantic enhancement edge adding module 1304 is used to determine whether to add a semantic enhancement edge between node i and node j in the first node set based on the relationship between the semantic similarity between node i and node j and a preset semantic similarity threshold.

[0432] The second normalized weight determination module 1305 is used to determine the normalized weight of the semantic enhancement edge based on the semantic similarity between the nodes i and j, the temporal modulation semantic similarity between the nodes i and j, and the structural correlation between the nodes i and j if a semantic enhancement edge is added between the nodes i and j.

[0433] In one alternative implementation, the initial knowledge network acquisition module 1301 includes:

[0434] An initial data acquisition unit is used to acquire an initial dataset; the initial dataset includes multiple data entries; each data entry includes multiple fields; the multiple fields include keywords, abstract, publication time, and citation count; the data is literature or patents;

[0435] An intermediate data acquisition unit is used to clean the initial dataset to obtain an intermediate dataset.

[0436] The feature vector acquisition unit is used to extract semantic features, temporal features, and network topology features from each piece of data in the intermediate dataset, and obtain the semantic feature vector, temporal feature vector, and topology feature vector corresponding to the data.

[0437] The feature fusion unit is used to fuse the semantic feature vector, temporal feature vector and topological feature vector corresponding to each data point in the intermediate dataset to obtain the fused feature vector corresponding to the data point.

[0438] The initial knowledge network construction first unit is used to construct the first node set by treating each data point in the intermediate dataset as a node; construct the first edge set based on the association relationship between each data point in the intermediate dataset and other data points; construct the first feature matrix based on the fusion feature vector corresponding to each data point in the intermediate dataset; and construct the first time element set based on the publication time of each data point in the intermediate dataset.

[0439] The second unit for constructing the initial knowledge network is used to take the first set of nodes, the first set of edges, the first feature matrix, and the first set of time elements as the initial knowledge network.

[0440] In one alternative implementation, the first normalized weight determination module 1302 includes:

[0441] The basic weight acquisition unit is used to perform a logarithmic operation on the number of times node i is referenced by node j to obtain the basic weight;

[0442] The first normalized weight calculation unit is used to substitute the basic weight, the absolute value of the time difference between the publication time of node i and the publication time of node j, and the time decay factor into the calculation. In the process, the normalized weights of the edges with endpoints i and j are obtained; where, This represents the normalized weight of the edge with nodes i and j as endpoints; Indicates the time decay factor; These are the basic weights.

[0443] In one alternative implementation, the multi-type information fusion feature vector determination module 1303 is used for:

[0444] The neighbor node type determination unit is used to determine all neighbor nodes of node i in the first node set, and the node type of each neighbor node of node i; the neighbor node types include paper node types and patent node types;

[0445] The attention weight determination unit is used to calculate the attention weight between node i and each neighboring node i based on the fusion feature vector of node i, the node type of the neighboring node, and the fusion feature vector of the neighboring node.

[0446] The first aggregated feature vector determination unit is used to obtain the aggregated feature vector of the paper type neighbors for the multiple neighbor nodes corresponding to the paper node type of the node i, based on the fusion feature vector of each of the multiple neighbor nodes and the attention weight between each of the multiple neighbor nodes and the node i.

[0447] The second aggregated feature vector determination unit is used to obtain the aggregated feature vector of the patent type neighbors for the multiple neighbor nodes of the patent node type corresponding to the node i, based on the fusion feature vector of each neighbor node in the multiple neighbors and the attention weight between each neighbor node and the node i.

[0448] The multi-type information fusion feature vector determination unit is used to perform weighted fusion of the fusion feature vector of node i, the aggregated feature vector of the paper type neighbor corresponding to node i, and the aggregated feature vector of the patent type neighbor corresponding to node i for node i, so as to obtain the multi-type information fusion feature vector of node i.

[0449] In one alternative implementation, the second normalized weight determination module 1305 includes:

[0450] A semantic similarity calculation unit is used to calculate the semantic similarity based on the semantic feature vector of node i and the semantic feature vector of node j;

[0451] A time-aware feature vector calculation unit is used to calculate the time-aware feature vector of node i based on the semantic feature vector of node i and the time context information of node i; and to calculate the time-aware feature vector of node j based on the semantic feature vector of node j and the time context information of node j.

[0452] The time modulation semantic similarity calculation unit is used to calculate the time modulation semantic similarity based on the time-aware feature vector of node i and the time-aware feature vector of node j;

[0453] The similarity calculation unit for neighbor node sets is used to calculate the similarity between the neighbor node sets of node i and node j based on the intersection of the neighbor node sets of node i and the neighbor node sets of node j, and the union of the neighbor node sets of node i and the neighbor nodes of node j.

[0454] A centrality dissimilarity calculation unit is used to calculate the centrality dissimilarity between node i and node j based on the betweenness centrality of node i and the betweenness centrality of node j.

[0455] The structural correlation calculation unit is used to perform a weighted calculation on the similarity of the neighbor node set and the centrality difference to obtain the structural correlation.

[0456] The basic weight calculation unit is used to determine the first weight of the semantic similarity, the second weight of the temporal modulation semantic similarity, and the third weight of the structural correlation based on the normalization function and the multilayer perceptron.

[0457] The second normalized weight determination subunit is used to perform a weighted operation on the first weight, the second weight, the third weight, the semantic similarity, the time-modulated semantic similarity, and the structural correlation to obtain the normalized weight of the semantic enhancement edge.

[0458] In one alternative implementation, the data processing device 130 further includes:

[0459] An isolated node determination module is used to determine isolated nodes in the first node set;

[0460] The candidate connection node determination module is used to generate candidate connection nodes for the isolated node based on the fused feature vector, variational autoencoder, and decoder of the isolated node.

[0461] The correlation degree value determination module is used to calculate the correlation degree value between the isolated node and the candidate connection node based on the preset prompt word template and the large language model.

[0462] The basic normalized weight determination module is used to determine the normalized weight of the edge with the isolated node and the candidate connected node as endpoints based on the correlation degree value and the time decay factor.

[0463] In one alternative implementation, the data processing device 130 further includes:

[0464] The aggregation neighbor feature vector determination unit is used to determine the weight coefficient of each neighbor node of node i in the first node set, the time influence factor between the neighbor node and node i, and the neighbor node's position in the time frame of node i. The neighbor aggregation operation is performed on the aggregated neighbor feature vectors to obtain the th neighbor feature vector. The nodes described in the layer The aggregated neighbor feature vector; the weight coefficient of the neighbor node is determined according to the node type of the neighbor node;

[0465] Embedded feature vector determination unit, used for node i in the first node set, based on the first feature vector determination unit. The nodes described in the layer The aggregated neighbor feature vector, preset balancing parameters, and knowledge-enhanced transformation vector are used to calculate the embedded feature vector corresponding to node i; the knowledge-enhanced transformation feature vector is based on node i. The corresponding type of learnable relation vector and the temporal relation vector of node i are generated.

[0466] In one alternative implementation, the data processing device 130 further includes:

[0467] The original association score determination unit is used to, for nodes i and j in the first node set, substitute the embedded feature vectors of node i and node j into the bilinear bisector function to obtain the original association score between node i and node j.

[0468] The correction correlation score determination unit is used to adjust the original correlation score based on a time decay modulation function to obtain a corrected correlation score; the expression of the time decay modulation function is: Where γ is the time decay parameter, Δt ij Let p(i,j) be the time difference between the publication time of node i and the publication time of node j, p(i,j) be the original association score between node i and node j, and p'(i,j) be the modified association score between node i and node j.

[0469] The propagation path determination unit is used to search for all propagation paths from node i to node j with a path length not exceeding a preset path length threshold using a breadth-first search algorithm, and to calculate the propagation intensity score of each propagation path.

[0470] The optimal propagation path determination unit is used to take the maximum propagation intensity score among the multiple propagation intensity scores as the multi-step inference score between node i and node j, and to take the propagation path corresponding to the maximum propagation intensity score as the optimal propagation path between node i and node j.

[0471] The first judgment unit is used to determine whether the multi-step inference score is greater than a preset score threshold.

[0472] A knowledge enhancement network acquisition unit is used to add the optimal propagation path to the knowledge enhancement network if it is determined that the multi-step inference score is greater than a preset score threshold; the knowledge enhancement network is generated based on the initial knowledge network.

[0473] In one alternative implementation, the data processing device 130 further includes:

[0474] The cluster acquisition unit is used to cluster the nodes in the first node set to obtain multiple clusters and determine the central feature vectors corresponding to the clusters.

[0475] The clustering rationality determination unit is used to, for each node in the cluster, substitute the embedded feature vector corresponding to the node and the central feature vector corresponding to the cluster to the node into a preset quality scoring function to obtain the clustering rationality score corresponding to the node.

[0476] The target cluster acquisition unit is used to determine that a node belongs to the corresponding cluster if the cluster rationality score of the node is greater than or equal to a preset rationality score; and to delete the node from the corresponding cluster if the cluster rationality score of the node is less than the preset rationality score, thereby obtaining the target cluster.

[0477] In one alternative implementation, the data processing device 130 further includes:

[0478] The innovation frontier index determination unit is used to determine the innovation frontier index corresponding to each of the target clusters;

[0479] The Innovation Frontier Index Screening Unit is used to filter out multiple Innovation Frontier Indices whose values ​​are greater than a preset index threshold.

[0480] The hotspot cluster determination unit is used to identify multiple target clusters corresponding to the multiple innovation frontier indices as multiple hotspot clusters.

[0481] In one alternative implementation, the data processing device 130 further includes:

[0482] The text summarization module is used to obtain the text information corresponding to each hotspot cluster, and generate a text summary of the hotspot cluster based on the text information, the target prompt word template and the large language model.

[0483] In one alternative implementation, the data processing device 130 further includes:

[0484] The learning similarity determination unit is used to obtain the comparative learning similarity between the candidate document and the hot spot cluster to which the candidate document belongs;

[0485] The recommendation score determination unit is used to determine the recommendation score of the candidate document based on the comparative learning similarity, the innovation frontier index of the hot spot cluster to which the candidate document belongs, and the number of years since the candidate document was published.

[0486] Based on the data processing method and apparatus provided in the foregoing embodiments, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the data processing method mentioned above.

[0487] Based on the data processing method and apparatus provided in the foregoing embodiments, this application also provides an electronic device, including:

[0488] A memory on which computer programs are stored;

[0489] A processor is configured to execute the computer program in the memory to implement some or all of the steps in the data processing method provided in the foregoing embodiments.

[0490] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated 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 embodiment solution according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0491] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data processing method, characterized in that, The method includes: Obtain an initial knowledge network; the initial knowledge network includes a first set of nodes, a first set of edges, a first feature matrix, and a first set of time elements; wherein, the first set of nodes includes two types: paper nodes and patent nodes; For node i and node j in the first node set, the normalized weight of the edge with node i and node j as endpoints is determined based on the number of times node i is referenced by node j and the time decay factor. For node i in the first node set, the multi-type information fusion feature vector of node i is determined based on the fusion feature vector of node i, the aggregated feature vector of each type of neighboring node of node i, and the attention weight between node i and the neighboring node. For node i and node j in the first node set, based on the relationship between the semantic similarity between node i and node j and the preset semantic similarity threshold, it is determined whether to add a semantic enhancement edge between node i and node j; If a semantically enhanced edge is added between node i and node j, the normalized weight of the semantically enhanced edge is determined based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j.

2. The method according to claim 1, characterized in that, The acquisition of the initial knowledge network includes: Obtain an initial dataset; the initial dataset includes multiple data entries; each data entry includes multiple fields; the multiple fields include keywords, abstract, publication date, and citation count; the data is literature or patent; The initial dataset is cleaned to obtain an intermediate dataset; For each data point in the intermediate dataset, semantic feature extraction, temporal feature extraction, and network topology feature extraction are performed on the data point to obtain the semantic feature vector, temporal feature vector, and topology feature vector corresponding to the data point. For each data point in the intermediate dataset, the semantic feature vector, temporal feature vector, and topological feature vector corresponding to that data point are fused to obtain the fused feature vector corresponding to that data point. Each data point in the intermediate dataset is treated as a node to construct the first node set; the first edge set is constructed based on the association between each data point in the intermediate dataset and other data points; the first feature matrix is ​​constructed based on the fusion feature vector corresponding to each data point in the intermediate dataset; and the first time meta set is constructed based on the publication time of each data point in the intermediate dataset. The first set of nodes, the first set of edges, the first feature matrix, and the first set of time elements are used as the initial knowledge network.

3. The method according to claim 1, characterized in that, The step of determining the normalized weights of edges with nodes i and j as endpoints, based on the number of times node i is referenced by node j and the time decay factor, for nodes i and j in the first node set, includes: The basic weight is obtained by performing a logarithmic operation on the number of times node i is referenced by node j; Substitute the basic weight, the absolute value of the time difference between the publication time of node i and the publication time of node j, and the time decay factor into... In the process, the normalized weights of the edges with endpoints i and j are obtained; where, This represents the normalized weight of the edge with nodes i and j as endpoints; Indicates the time decay factor; These are the basic weights.

4. The method according to claim 1, characterized in that, For node i in the first node set, the multi-type information fusion feature vector of node i is determined based on the fusion feature vector of node i, the aggregated feature vectors of each type of neighboring nodes of node i, and the attention weights between node i and its neighboring nodes, including: Determine all neighboring nodes of node i in the first node set, and the node type of each neighboring node of node i; the types of neighboring nodes include paper node type and patent node type; For each neighboring node of node i, the attention weight between node i and the neighboring node is calculated based on the fusion feature vector of node i, the node type of the neighboring node, and the fusion feature vector of the neighboring node. For multiple neighboring nodes of the paper node type corresponding to node i, the aggregated feature vector of the paper type neighbors is obtained based on the fusion feature vector of each of the multiple neighboring nodes and the attention weight between each of the multiple neighboring nodes and node i. For multiple neighboring nodes of the patent node type corresponding to node i, the aggregated feature vector of the patent type neighbors is obtained based on the fusion feature vector of each of the multiple neighboring nodes and the attention weight between each of the multiple neighboring nodes and node i. For node i, the fusion feature vector of node i, the aggregated feature vector of the paper type neighbors corresponding to node i, and the aggregated feature vector of the patent type neighbors corresponding to node i are weighted and fused to obtain the multi-type information fusion feature vector of node i.

5. The method according to claim 1, characterized in that, The determination of the normalized weights of the semantically enhanced edges based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j includes: The semantic similarity is calculated based on the semantic feature vector of node i and the semantic feature vector of node j; Based on the semantic feature vector of node i and the temporal context information of node i, calculate the time-aware feature vector of node i; and based on the semantic feature vector of node j and the temporal context information of node j, calculate the time-aware feature vector of node j. The time-sensing feature vector of node i and the time-sensing feature vector of node j are used to calculate the time modulation semantic similarity. Based on the intersection of the neighbor node set of node i and the neighbor node set of node j, and the union of the neighbor node set of node i and the neighbor node set of node j, calculate the similarity between the neighbor node sets of node i and node j. Based on the betweenness centrality of node i and the betweenness centrality of node j, calculate the centrality difference between node i and node j; The structural association degree is obtained by weighting the similarity of the neighbor node set and the centrality dissimilarity. Based on the normalization function and the multilayer perceptron, the first weight of the semantic similarity, the second weight of the temporal modulation semantic similarity, and the third weight of the structural association are determined. The first weight, the second weight, the third weight, the semantic similarity, the temporal modulation semantic similarity, and the structural correlation are weighted and calculated to obtain the normalized weight of the semantically enhanced edge.

6. The method according to any one of claims 1-5, characterized in that, After determining the normalized weights of the semantically enhanced edges, the method further includes: Identify the isolated nodes in the first node set; Based on the fused feature vector, variational autoencoder, and decoder of the isolated node, candidate connection nodes of the isolated node are generated. Based on the preset prompt word template and large language model, the correlation degree between the isolated node and the candidate connection node is calculated. Based on the correlation degree value and the time decay factor, the normalized weights of the edges with the isolated node and the candidate connected node as endpoints are determined.

7. The method according to claim 6, characterized in that, After determining the normalized weights of the edges with the isolated node and the candidate connected node as endpoints, the method further includes: For each neighboring node of node i in the first node set, the weight coefficient of that neighboring node, the time influence factor between that neighboring node and node i, and the position of that neighboring node in the time frame of node i. The neighbor aggregation operation is performed on the aggregated neighbor feature vectors to obtain the th neighbor feature vector. The nodes described in the layer The aggregated neighbor feature vector; the weight coefficient of the neighbor node is determined according to the node type of the neighbor node; For node i in the first node set, based on the... The nodes described in the layer The aggregated neighbor feature vector, preset balancing parameters, and knowledge-enhanced transformation vector are used to calculate the embedded feature vector corresponding to node i; the knowledge-enhanced transformation feature vector is based on node i. The corresponding type of learnable relation vector and the temporal relation vector of node i are generated.

8. The method according to claim 7, characterized in that, After obtaining the embedded feature vector corresponding to node i, the method further includes: For node i and node j in the first node set, the embedded feature vectors of node i and node j are substituted into the bilinear bisection function to obtain the original association score between node i and node j. The original correlation score is adjusted based on the time decay modulation function to obtain the corrected correlation score; the expression of the time decay modulation function is: Where γ is the time decay parameter, Δt ij Let p(i,j) be the time difference between the publication time of node i and the publication time of node j, p(i,j) be the original association score between node i and node j, and p'(i,j) be the modified association score between node i and node j. Using a breadth-first search algorithm, search for all propagation paths from node i to node j whose path length does not exceed a preset path length threshold, and calculate the propagation strength score for each propagation path. The maximum propagation intensity score among the multiple propagation intensity scores is taken as the multi-step inference score between node i and node j, and the propagation path corresponding to the maximum propagation intensity score is taken as the optimal propagation path between node i and node j. Determine whether the multi-step inference score is greater than a preset score threshold; If the multi-step inference score is determined to be greater than a preset score threshold, the optimal propagation path is added to the knowledge enhancement network; the knowledge enhancement network is generated based on the initial knowledge network.

9. The method according to claim 8, characterized in that, The method further includes: Cluster the nodes in the first node set to obtain multiple clusters, and determine the central feature vectors corresponding to the clusters; For each node in the cluster, the embedded feature vector corresponding to the node and the central feature vector of the cluster to which the node belongs are substituted into the preset quality scoring function to obtain the cluster rationality score corresponding to the node. For each node in the cluster, if the cluster rationality score corresponding to the node is greater than or equal to the preset rationality score, then the node is determined to belong to the corresponding cluster; if the cluster rationality score corresponding to the node is less than the preset rationality score, then the node is deleted from the corresponding cluster to obtain the target cluster.

10. The method according to claim 9, characterized in that, The method further includes: Determine the innovation frontier index corresponding to each of the target clusters; Select multiple innovation frontier indices whose values ​​are greater than the preset index threshold; The multiple target clusters corresponding to the multiple innovation frontier indices are used as multiple hot spot clusters.

11. The method according to claim 10, characterized in that, The method further includes: For each hotspot cluster, obtain the text information corresponding to the hotspot cluster, and generate a text summary of the hotspot cluster based on the text information, the target prompt word template, and the large language model.

12. The method according to claim 10, characterized in that, The method further includes: Obtain the comparative learning similarity between the candidate document and the hot spot cluster to which the candidate document belongs; The recommendation score for the candidate document is determined based on the comparative learning similarity, the innovation frontier index of the hot spot cluster to which the candidate document belongs, and the number of years since the candidate document was published.

13. A data processing apparatus, characterized in that, The device includes: An initial knowledge network acquisition module is used to acquire an initial knowledge network; the initial knowledge network includes a first set of nodes, a first set of edges, a first feature matrix, and a first set of time elements; wherein, the first set of nodes includes two types: paper nodes and patent nodes; The first normalized weight determination module is used to determine the normalized weight of the edge with nodes i and j as endpoints for nodes i and j in the first node set, based on the number of times node i is referenced by node j and the time decay factor. The multi-type information fusion feature vector determination module is used to determine the multi-type information fusion feature vector of node i for node i in the first node set based on the fusion feature vector of node i, the aggregated feature vector of each type of neighboring node of node i, and the attention weight between node i and the neighboring node. The semantic enhancement edge addition module is used to determine whether to add a semantic enhancement edge between node i and node j in the first node set, based on the relationship between the semantic similarity between node i and node j and a preset semantic similarity threshold. The second normalized weight determination module is used to determine the normalized weight of the semantic enhancement edge based on the semantic similarity between node i and node j, the temporal modulation semantic similarity between node i and node j, and the structural correlation between node i and node j if a semantic enhancement edge is added between node i and node j.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-12.

15. An electronic device, characterized in that, include: A memory on which computer programs are stored; A processor for executing the computer program in the memory to implement the steps of the method according to any one of claims 1-12.

Images