A technology knowledge big graph high-order link prediction method fusing technology life cycle and multi-agent role constraints

Abstract

The application discloses a technology knowledge big graph high-order link prediction method fusing technology life cycle and multi-agent role constraints, constructs a three-layer technology knowledge big graph, differentiates coding of text and structural feature dominant class nodes, and learns multi-granularity representation through hierarchical perception hypergraph convolution and adaptive fusion; determines technology life cycle phases based on entity time series, constructs differentiated time sequence attenuation kernel and asymmetric directional constraints, constructs time sequence meta-path and calculates semantic scores, and removes homogenization candidates in combination with multi-agent role complementarity matrix; constructs a three-element coupling tensor for institutional collaborative relationship and adds a conduction time delay constraint; through bidirectional cross verification of dimensionality increasing prediction and dimensionality reducing verification and three types of hierarchical consistency constraints, logical self-consistency is guaranteed, field rules are relaxed into differentiable constraints and are adaptively fused with neural network scores; based on path backtracking, a structured interpretable evidence chain is generated, and a closed-loop evolution mechanism is formed through incremental retraining.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge graph technology, specifically to a method for predicting high-order links in a large graph of scientific and technological knowledge that integrates the technology lifecycle and multi-agent role constraints. Background Technology

[0002] The scientific and technological innovation activities carried by science and technology intelligence data have unique characteristics such as significant multi-entity collaboration, asymmetric temporal evolution, cross-domain coupling, hierarchical relationships, and interdisciplinary intersection. Complex relationships, manifested in collaborative innovation, technological integration, institutional synergy, event co-evolution, and interdisciplinary intersection, are prevalent among various entities including papers, patents, policies, projects, institutions, enterprises, researchers, and technological concepts. Formalizing this multi-source heterogeneous data into a large-scale science and technology knowledge graph representing multi-entity high-order relationships with hyperedges, and accurately predicting potential high-order links within it, is of great significance for assisting in research layout decisions, judging technological evolution trends, evaluating policy effectiveness, identifying industry-academia-research collaborative partners, and discovering cutting-edge interdisciplinary collaborations.

[0003] However, existing technologies have the following limitations: First, most existing knowledge graph link prediction methods are based on binary relation modeling, which is insufficient for characterizing multi-agent high-order relations involving three or more entities. Second, some existing technologies introduce tools such as hypergraph structures, simplex complexes, and tensor decomposition, but their objects are mostly homogeneous dynamic high-order networks. They do not introduce ontology semantic constraints and differentiated modeling of entity role assignments within hyperedges, causing the evaluation to mainly rely on the strength of topological associations, making it difficult to filter out type conflict candidates and eliminate pseudo-cooperative candidates with highly homogeneous roles.

[0004] Third, existing time-series graph link prediction methods typically employ a uniform attenuation coefficient and a uniform bandwidth time-series random walk mechanism, using a homogeneous time-series weighting method for all node pair interactions. This fails to characterize the differences in the time-value of entities at different life-cycle stages and the irreversible directionality of scientific and technological evolution from basic research to application. Fourth, existing technologies mostly complete prediction and feature fusion at a single level, without cross-layer joint modeling of the hierarchical structural features of the entity layer, binary relation layer, and higher-order relation layer. This may lead to logical conflicts such as abnormally high prediction confidence of the upper layer compared to the lower layer, lack of binary relation support for higher-order hyperedges, and type roles not satisfying relation type templates.

[0005] Fifth, existing technologies lack specific constraints designed for certain types of high-order relationships in the field of science and technology intelligence. For high-order relationships with distinct cross-domain characteristics, such as institutional collaboration, there is a lack of structured modeling of the temporal transmission and content coupling patterns between policies, technological fields, and industries. Furthermore, there is a lack of meta-path systems and interdisciplinary measurement methods for cross-disciplinary scenarios. Sixth, existing high-order link prediction methods typically fail to generate structured and interpretable chains of evidence while outputting predicted probabilities, making it difficult to provide science and technology intelligence decision-makers with traceable and verifiable reasoning. The method of directly generating explanations based on large language models, without strict boundary indications, factual verification, and logical consistency checks, is prone to producing illusory statements that contradict the facts of the graph.

[0006] In summary, existing technologies suffer from numerous technical problems, including insufficient ability to model high-order relationships for scientific and technological intelligence data, overly homogeneous temporal and hierarchical modeling, a lack of dedicated constraint mechanisms for cross-domain collaboration and interdisciplinary relationships, weak interpretable reasoning capabilities, and difficulty in supporting the dynamic evolution of scientific and technological knowledge graphs. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a high-order link prediction method for science and technology knowledge graphs that integrates the technology life cycle and multi-subject role constraints. This method addresses the problems of existing knowledge graph link prediction methods, such as being based on binary relation modeling, lacking ontology and role semantic modeling capabilities for science and technology intelligence data, using homogenized processing methods for temporal and hierarchical structure modeling, lacking dedicated modeling mechanisms for cross-domain collaboration and interdisciplinary relationships, and having insufficient interpretable reasoning capabilities.

[0008] Technical Solution: The present invention provides a high-order link prediction method for a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints, comprising the following steps:

[0009] Step 1: Construct a three-layer science and technology knowledge graph and formalize it into an octet; apply hard constraints on type and role compatibility to the science and technology meta-ontology and eliminate candidates that violate these constraints.

[0010] Step 2: Text encoding is used for semantic nodes and topological encoding is used for structural feature nodes. A unified embedding space is formed through type projection. The information of binary neighborhood, hyperedge neighborhood and node itself is aggregated using hierarchical hypergraph convolution. Differentiated weighting is applied to the neighborhood aggregation through relation type attention parameters and role perception parameters. The aggregated information of entity layer, binary relation layer and higher-order relation layer is dynamically weighted and fused through hierarchical fusion weight.

[0011] Step 3: Determine the phase of the technology life cycle based on the time series of entity attributes, construct a time series decay kernel by setting differentiated decay coefficients for the nascent stage, growth stage, maturity stage and decline stage, and apply asymmetric directional constraints to the phase sequence of the meta-path instance to characterize the irreversibility of the transformation of technology evolution from basic research to application.

[0012] Step 4: Construct temporal meta-paths for collaborative innovation, technological integration, institutional synergy, event co-evolution, and interdisciplinary intersection, and calculate the semantic score of the meta-paths through an attention mechanism;

[0013] Step 5: Construct a multi-agent role complementarity matrix, calculate the complementarity score of any two entity roles within the candidate hyperedge, and eliminate role homogeneous candidates with scores below the threshold.

[0014] Step 6: Construct a three-element coupling tensor of policy, technology, and industry for institutional synergy and decompose it to alleviate sparsity. Add a policy transmission delay window constraint, requiring that the response times of policy, technology, and industry strictly increase and fall within the preset window, and eliminate candidates that fail the verification.

[0015] Step 7: Achieve cross-layer joint prediction through bidirectional cross-validation of dimensionality-up prediction and dimensionality-down verification, and perform hierarchical consistency verification using super-edge substructure support constraints, cross-layer confidence transfer constraints, and type-role consistency constraints;

[0016] Step 8: Use the hard constraints of the ontology as the first layer of filtering, relax the various domain rules into differentiable constraint losses through product T-norm, and combine them with multiple losses in a weighted manner to form a joint loss. Then, through type condition gating, adaptively fuse the neural prediction score, meta-path semantic score, rule compliance score, role complementarity score and ternary coupling score to obtain a comprehensive prediction score.

[0017] Step 9: For candidate higher-order links that exceed the threshold, extract six types of evidence elements based on path backtracking, and generate a structured and interpretable evidence chain through a triple mechanism of boundary constraints, post-fact verification and logical consistency verification.

[0018] Step 10: Feed back the verified prediction results to the knowledge graph and include them in the set to be updated. Update the technology life cycle phase of the affected entities. When the set to be updated accumulates to a threshold, trigger incremental retraining based on elastic weight consolidation to form a closed-loop evolution mechanism.

[0019] Furthermore, in step 1, the octet formal definition uses the science and technology meta-ontology as a unified semantic standard, abstracting the science and technology knowledge graph into a unified representation of entity nodes, binary edges, hyperedges and their relation types, role assignments, time information, technology life cycle phases and relation levels; the ontology hard constraints are specifically: based on the predefined type compatibility rules and role compatibility rules in the science and technology meta-ontology, the candidate higher-order links are subject to mandatory verification, and those that violate any rule are directly eliminated.

[0020] Furthermore, in step 2, relationship-type-aware attention weights and role-aware weights based on role assignment are introduced into the hierarchical hypergraph convolution to achieve differentiated aggregation of neighbors of different relationship types and roles; the hierarchical fusion weights are adaptively generated by the sparsity of the node neighborhood and the type distribution, automatically increasing the entity layer weights of long-tail entities with sparse neighborhoods, and automatically increasing the higher-order relationship layer weights of core entities with rich interactions.

[0021] Furthermore, in step 3, the temporal decay kernel function of the technology life cycle phase perception is assigned a differentiated decay coefficient according to the different stages of the entity's nascent, growth, maturity, or decline, so that entities at different evolutionary stages have different historical information decay rates; the asymmetric directional constraint determines the monotonicity of the phase sequence of entities involved in the meta-path, applies a positive constraint to sequences that conform to the unidirectional evolution law from the nascent stage to the decline stage, and applies a constraint to strictly reverse sequences that violate this law.

[0022] Furthermore, in step 4, the semantic score of the metapath is the weighted sum of the attention weight and the temporal rationality score of each metapath instance; among them, the collaborative innovation metapath requires that the active time of adjacent entities is not strictly monotonically increasing; the technology integration metapath requires that the active time intervals of participating technologies overlap; the institutional collaboration metapath requires that the node response time is strictly monotonically increasing; the event co-evolution metapath requires that the time of the preceding and following events be strictly sequential and causally related; the interdisciplinary cross-cutting metapath requires that the distance between the two disciplines in the discipline classification tree exceeds a preset threshold.

[0023] Furthermore, in step 5, the role complementarity matrix is ​​pre-calibrated by domain experts based on the mechanism of technological innovation collaboration, and is used to quantify the strength of synergy and complementarity between different roles;

[0024] Furthermore, in step 6, the ternary coupling tensor is obtained by decomposing the outer product of multiple factor vectors; the policy transmission delay constraint requires that the response times of policy, technology, and industry increase sequentially and that adjacent time intervals fall within a preset reasonable delay window.

[0025] Furthermore, in step 7, the superedge substructure support constraint is used to verify whether the candidate superedge contains a sufficient proportion of observed or high-confidence binary relations; the cross-layer confidence propagation constraint is used to prevent the prediction confidence of the higher-order relation layer from being unreasonably significantly higher than the reasonable aggregation value based on the confidence of the lower-level entity layer and the binary relation layer; the type and role consistency constraint is used to verify whether the type and role assignment of each entity in the candidate superedge meets the compatibility requirements of the predetermined relation type template; only when the joint verification of the three types of constraints passes, the candidate superedge enters the subsequent comprehensive prediction score calculation process.

[0026] Furthermore, in step 8, the fusion weights of the comprehensive prediction scores are adaptively generated by the type-conditional gating network based on the relationship type of the candidate hyperedges. For candidate hyperedges that are not institutional collaborative relationships, the weights of their ternary coupling score items are set to zero and the corresponding weights are redistributed to the remaining score items.

[0027] Furthermore, in step 9, the prompt boundary constraint is used to restrict the large language model to interpret only based on the given structured prompts, the post-fact verification is used to reverse the factual statements in the evidence chain, and the logical consistency check is used to perform internal consistency checks on the output evidence chain based on the rule base.

[0028] Furthermore, in step 10, the incremental retraining adopts an elastic weight consolidation strategy, which prevents new data from overwriting existing knowledge by applying elastic constraints to existing parameters.

[0029] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0030] (1) Significantly improves the semantic acceptability and collaborative value recognition capability of high-order link prediction. This invention uses the technology meta-ontology as a unified semantic standard to impose hard constraints on type compatibility and role compatibility, and eliminates role homogeneous candidates through a multi-subject role complementarity matrix, effectively avoiding type conflict links and pseudo-collaborative candidates without real collaborative value.

[0031] (2) Improve the accuracy of node representation and prediction accuracy in time-sensitive scenarios. This invention uses a differential encoder for semantically dominant and structurally dominant nodes, combined with hierarchical fusion weights adaptively generated by neighborhood sparsity and type distribution; constructs a phase-aware temporal decay kernel through a technology lifecycle phase mapping function and differential decay coefficient, and verifies the monotonicity of the phase sequence with asymmetric directional constraints, thereby accurately characterizing the semantic emphasis of different types of nodes and the time-sensitivity characteristics of entities at different evolution stages, and improving the accuracy of emerging technology identification and technology evolution prediction.

[0032] (3) Expand the coverage of high-order link prediction tasks and improve the prediction quality of cross-domain collaborative relationships. This invention constructs a differentiated time-series meta-path system for five types of high-order relationships; for institutional collaborative relationships, it models the content coupling rules of policies, technologies and industries through ternary coupling tensors and CP decomposition, and adds policy transmission delay constraints, which significantly improves the prediction quality in cross-domain scenarios such as policy effectiveness assessment and industry response analysis.

[0033] (4) Ensure the hierarchical logical consistency of the prediction results while taking into account both data-driven learning and domain knowledge constraints. This invention avoids logical conflicts such as the lack of support from the lower layer in the upper-level prediction and the violation of relationship type templates by the hierarchical progressive reasoning and three types of hierarchical consistency constraints; it encodes multiple types of domain rules into differentiable constraint terms and integrates them with the neural prediction scores in a weighted manner, so that the model can capture the data-driven pattern rules and follow the ontology and logical constraints during the training process, and maintain high prediction robustness under sparse data conditions.

[0034] (5) Provide a structured chain of evidence for science and technology intelligence decision-making and support the dynamic evolution of the science and technology knowledge graph. This invention extracts six types of evidence elements based on path backtracking and avoids the illusion of large language models through a triple controlled generation mechanism, outputting a traceable and verifiable structured chain of evidence; constructs a prediction-verification-update-phase evolution closed loop and adopts an elastic weight consolidation strategy to achieve incremental retraining, enabling the science and technology knowledge graph to have continuous dynamic evolution capabilities. Attached Figure Description

[0035] Figure 1 A flowchart of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] like Figure 1 As shown, this embodiment of the invention provides a high-order link prediction method for a large-scale science and technology knowledge graph that integrates the technology lifecycle and multi-agent role constraints. This embodiment is implemented in Python, with the model built on the PyTorch and PyTorch Geometric frameworks. Graph data storage uses the Neo4j graph database, text encoding employs the Qwen3-Embedding large language model, and model training uses the Adam optimizer with a learning rate decay strategy. The engineering implementation details of steps S1 to S10 are described below.

[0038] S1. Construction of a large graph of scientific and technological knowledge and filtering of ontology hard constraints A three-layered science and technology knowledge graph, comprising an entity layer, a binary relation layer, and a higher-order relation layer, is constructed based on multi-source heterogeneous science and technology intelligence data. Hard constraints are applied for filtering using the science and technology meta-ontology as a unified semantic standard. The specific implementation process is as follows: (1) Definition of the Science and Technology Meta-ontology. The defined set of entity types includes nine categories: enterprises, researchers, technological concepts, projects, patents, policies, papers, institutions, and disciplines. The defined set of relationship types covers binary relationships such as affiliation, funding, invention, citation, cooperation, and competition, as well as five higher-order relationships such as collaborative innovation, technological integration, institutional collaboration, event co-evolution, and interdisciplinary intersection. The defined type constraints include type compatibility constraints and role compatibility constraints for the entities participating in the relationship. For example, the collaborative innovation relationship requires that the superedge contain at least one researcher entity, one institution entity, and one technological concept entity; the institutional collaboration relationship requires that the superedge contain at least one policy entity, one technology-related entity, and one industry entity, and the role allocation of each entity within the superedge must fall within the set of roles allowed by the relationship type.

[0039] (2) Formalization of the science and technology knowledge graph. The science and technology knowledge graph is formalized into octet groups. ,in For a set of entity nodes, For a set of binary edges, Let be a set of hyperedges and for any satisfy ; For type mapping functions, Assign functions to the hyperedge role. For time mapping function, For the technology lifecycle phase mapping function, This is a relational hierarchy mapping function.

[0040] (3) Construction of the node-hyperedge incidence matrix. Construct the incidence matrix. If and only if hour .

[0041] (4) Ontology Hard Constraint Filtering. For any candidate higher-order link, hard constraint verification is performed based on the type compatibility and role compatibility constraints of the ontology. Candidates that violate the constraints are directly eliminated and do not enter the subsequent evaluation process. For example, for a candidate collaborative innovation hyperedge, if it does not contain any scientific researcher type entity, it is directly eliminated; for a candidate institutional collaboration hyperedge, if it does not simultaneously contain policy entities, technology-related entities, and industry entities, it is also directly eliminated. In this embodiment, the science and technology knowledge graph is stored and managed based on the Neo4j graph database, and the ontology hard constraint verification of candidate hyperedges is implemented through Cypher query statements.

[0042] S2, Semantic-focused differential projection and hierarchical awareness multi-granularity representation learning To address the semantic differences in different node types, a differential encoding strategy and hierarchical-aware hypergraph convolution are employed to achieve multi-granularity node representation learning. The specific implementation process is as follows: (1) Differentiated encoding. For semantically dominant nodes such as researchers, papers, patents, and policies, a text-driven encoder based on large language models such as Qwen3-Embedding is used to encode entity names, attribute descriptions, and contextual corpora. For structurally dominant nodes such as institutions, enterprises, projects, and technical concepts, a topology-driven encoder is used, with node degree vector, centrality indices (betweenness centrality, proximity centrality, and eigenvector centrality) and neighbor type distribution vector as the main features.

[0043] (2) Type projection. For each type of node... Set independent linear projections respectively Project the differential coding results onto a dimension of A unified embedded space, For nodes Differential coding results , The projection weights and biases are for the corresponding node types.

[0044] (3) Hierarchical-aware hypergraph convolution. The embedding of the (l+1)th layer node is updated as follows: ; in, For ELU activation function; , , , , are learnable linear transformation matrices for information about themselves, their binary neighborhood, and their hyperedge, respectively; , Hyperparameters for balancing the contributions of binary neighborhood aggregation and hyperedge neighborhood aggregation; For a node's binary neighbor set; Perceive attention weights for relation types; Assign role-aware weights based on hyperedge roles and relationship types. Attention weights are calculated using the following formula: ; in, A learnable attention vector; To share the linear transformation matrix; This is the embedding vector of the relation type to which the edge belongs.

[0045] (4) Hierarchical fusion mechanism. This is achieved through weighting... The aggregated information from the entity layer, binary relation layer, and higher-order relation layer is weighted and fused to satisfy... The weights are adaptively generated based on the sparsity of the node's neighborhood and the type distribution: ; in, , , They are nodes In terms of degree at the entity layer, binary relation layer, and higher-order relation layer. Automatically increase the size of long-tail entities with sparse neighborhoods. Automatically enlarge core entities with rich interactions In this embodiment, the unified embedding space dimension is set to... The number of hypergraph convolutional layers is set to , With a negative slope set to 0.2, the text encoder employs the Qwen3-Embedding model and compresses the output 1024-dimensional vector into a unified embedding space via linear projection. The time complexity of the entire hypergraph convolution layer is O(n log n). .

[0046] S3. Construction of temporal decay kernel function and asymmetric directional constraints for phase-aware technology lifecycle. A temporal decay kernel function and asymmetric directional constraints for phase-aware technology lifecycle are constructed to characterize the asymmetric temporal evolution of scientific and technological knowledge. The specific implementation process is as follows: (1) Lifecycle phase determination. For each entity Time series based on correlation attributes (number of published papers, number of patent applications, corporate R&D investment, etc.) Perform sliding window smoothing, slope analysis, and inflection point detection, combined with cumulative thresholding. Phase determination with first and second derivatives: ① Budding stage P1: and ; ② Growth stage P2: and ; ③ Maturity stage P3: and ; ④ Decline phase P4: and .

[0047] in For cumulative statistics, The threshold for determining zero slope is set. Different combinations of judgment attributes are used for different types of entities.

[0048] (2) Differentiated attenuation coefficients and time-series attenuation kernels. Differentiated attenuation coefficients are set for the four types of phases. satisfy This means that entities in the germination stage decay slowly and those in the decay stage decay quickly. In this embodiment, the four decay coefficients are set to 0.05, 0.15, 0.30, and 0.50 respectively. The temporal decay kernel function is constructed as follows: ; in This is to predict the difference between the target time and the historical observation time.

[0049] (3) Asymmetric directional constraints. Construction Evaluation of the phase sequence of entities involved in the metapath Monotonicity: Instances of metapaths with monotonically non-decreasing phase sequences receive a positive constraint score (+1), strictly reversed metapaths receive a negative constraint score (−1), and other cases are zero.

[0050] (4) Timing rationality score of metapath instances: ,in This represents the average time difference between adjacent nodes within the path instance. This is a key node in the metapath instance. Completeness of entity type coverage for metapath instances.

[0051] S4. Construction of a temporal meta-path system oriented towards higher-order relations and attention scoring. For five types of high-order relationships, namely collaborative innovation, technological integration, institutional synergy, event co-evolution, and interdisciplinary intersection, a temporal meta-path system is constructed, and the importance weight of each meta-path instance to the candidate superedge is calculated based on the attention mechanism.

[0052] (1) Construction of five types of meta-paths. (Collaborative innovation) consists of a sequential link of scholars, papers, technical concepts, projects, and institutions, requiring that the active time of adjacent entities is not strictly monotonically increasing; (Technology integration) consists of multiple technologies, application fields, and implementing entities, requiring that the active time periods of each technology overlap; (Institutional coordination) consists of policies, technological fields, target enterprises, and response results, requiring that the response time of each node strictly increases monotonically; (Event co-evolution) consists of a preceding event, a causal node, and a subsequent event, requiring a strict chronological order and a causal relationship; (Interdisciplinary) A symmetrical link is formed by discipline A, scholar, paper, technical concept, paper, scholar, and discipline B, requiring that the distance between the two disciplines in the discipline classification tree exceeds a threshold. Define interdisciplinary metrics.

[0053] ; in, The path distance between two disciplines on the subject classification tree; , , These are the disciplines corresponding to the two ends and the intermediate nodes of the meta-path, respectively. This is the semantic relevance function for intermediate nodes.

[0054] (2) Attention mechanism weight calculation. For candidate hyperedges Metapath instance collection The path embedding is obtained by encoding the node sequence using an LSTM or Transformer encoder. And calculate the attention weights using the following formula: ; (3) Metapath semantic score: ; A controlled depth-first search is used to enumerate meta-path instances, and a maximum path length is set. With sample number Controlling the size, the overall time complexity is In this embodiment, it is set that... , When the overall weight of all metapath instances is low under a certain type of reasoning task, a novel metapath discovery process is triggered. The large language model generates candidate metapath templates by combining relevant entities, which are then included in the metapath set after verification by domain experts.

[0055] S5. Construction of Multi-Agent Role Complementarity Matrix and Evaluation of Candidate Hyperedge Roles (1) Construction of the role complementarity matrix. The matrix elements represent the complementary strength between roles, with values ​​ranging from... It was calibrated by domain experts based on the collaborative mechanism of technological innovation; This is the set of role types under this relationship type. Based on its scale, the complementarity strength between funding roles and technology roles, R&D roles and transformation roles, policy supply roles and policy response roles, and interdisciplinary donors and recipients roles is relatively high (typically 0.7-0.9); the complementarity strength between similar or related roles is relatively low (typically 0.1-0.3). Taking collaborative innovation relationships as an example, its role type set can include academic research roles, industrial R&D roles, funding supply roles, policy support roles, and technology transfer roles, etc. In the matrix, the (academic research role, industrial R&D role) item can be marked as 0.85, the (funding supply role, academic research role) item can be marked as 0.80, and the (academic research role, academic research role) item can be marked as 0.20, reflecting the basic collaborative law that the collaborative value between similar roles is low and the complementary value between heterogeneous roles is high.

[0056] (2) Calculation of role complementarity score: ; in Let be the size of the candidate superedge. When If a character is deemed to be homogeneous, it will be removed. In this embodiment... .

[0057] S6. Construction of the ternary coupling tensor and cross-domain constraints of policy, technology, and industry. To address the institutional synergy, a tripartite coupling tensor of policy, technology, and industry is constructed, with an additional constraint on the policy transmission delay.

[0058] (1) Construction of a ternary coupled tensor. And perform CP decomposition on it to alleviate sparsity: ; in For the outer product operator, To decompose the rank, factor vector , , It is obtained by training by minimizing the Frobenius norm reconstruction error.

[0059] (2) Policy transmission delay window. (Preset) Describe the reasonable transmission time among the three parties. Require policy release. Technical response Industry response satisfy And adjacent delays all fall within The lower bound excludes spurious correlations with excessively short delays, while the upper bound excludes cases of conduction failure. In this embodiment... , It is determined by experts in the field, taking into account the statistical characteristics of the industrial policy response cycle.

[0060] (3) Coupling strength is calculated using the frequency of effective event chains: ; in, For policy events Technical events Industry events The validity of the event chain consisting of the three elements is determined by setting 1 to indicate that the ternary event chain is a valid transmission chain, and setting 0 to indicate that it is invalid. The event that occurs; The characteristic function takes the value 1 if the condition is true, and 0 otherwise; tensor elements This is the cumulative frequency of valid event chains.

[0061] (4) Tripartite Coupling Score. Policy categories are extracted from candidate hyperedges of institutional synergy relationships. Technical Field Industry type The score is .

[0062] (5) Candidate compliance verification. A time-delay window verification is performed on candidate superedges for institutional collaboration relationships. Those that fail the verification are directly eliminated. For candidate superedges that are not institutional collaboration relationships, this step does not participate in the evaluation, and the corresponding score is set to zero. The remaining weights are then normalized and redistributed by the subsequent type condition gating mechanism.

[0063] S7. Hierarchical Progressive Reasoning Mechanism and Hierarchical Consistency Constraints A hierarchical progressive reasoning mechanism enables cross-layer joint prediction of entity layer, binary relation layer and higher-order relation layer, and three types of hierarchical consistency constraints ensure the hierarchical logical consistency of the prediction results.

[0064] (1) Bidirectional progressive reasoning. This includes two directions: dimensionality-up prediction (entity layer → binary relation layer → higher-order relation layer) and dimensionality-down verification (higher-order relation layer reversely verifies the lower layer). The confidence level of the entity layer is obtained by weighting attribute integrity, neighborhood stability, and data source evidence strength: ; in, , , These represent attribute integrity, neighborhood stability, and data source evidence strength, respectively. , , These are the corresponding weighting coefficients. Confidence level of the binary relation layer. Calculation based on relation type priors, co-occurrence frequency, and triangular closure; confidence level of higher-order relations. Calculated based on neural prediction score and metapath semantic score.

[0065] (2) Hyperedge substructure support constraints. A sufficient proportion of the binary relation subsets within candidate hyperedges must be observed or have high confidence: ; in, The high confidence threshold for binary relations; This is a preset threshold. When Substructures with insufficient support are identified and removed. A verification function is introduced. .

[0066] (3) Cross-level confidence propagation constraint. It requires that the confidence level of higher-order relations should not be significantly higher than the reasonable aggregation of the confidence levels of lower-order relations. ; in, The propagation coefficient for the confidence level of the lower-level binary relation; These are the aggregation coefficients for entity-level confidence, and both control the extent to which higher-level confidence values ​​are allowed to exceed lower-level aggregation values. A verification function is introduced. The value is 1 if the inequality is satisfied, and 0 otherwise.

[0067] (4) Type and Role Consistency Constraint. Each entity's (type, role) must satisfy the relation type template compatibility requirement: ; in, The set of combinations allowed for a relation type, i.e., a predefined relation type template.

[0068] (5) Joint verification. The combined effect of the three types of constraints: Only when When candidate superedges are included in the comprehensive prediction score calculation process: S8. Calculation of Neural Symbol Integration and Comprehensive Prediction Score (1) Formalization of domain rules. Rule base These are rules for ontology type compatibility, technology lifecycle monotonicity, role complementarity, policy transmission, and hierarchical consistency. Each rule is accompanied by domain expert-calibrated confidence levels. .

[0069] (2) Layered processing of hard constraints and differentiable constraints. Violation candidates are directly eliminated as hard constraints; the remaining rules are relaxed to differentiable constraint losses through product T-norm relaxation: ; in, as a candidate Regarding the rules The degree of violation is obtained by product T-norm relaxation of the corresponding rule predicate, and takes the value [0,1].

[0070] (3) Joint loss function: ; in Predicting cross-entropy loss for binary relationships, Margin ranking loss for semantic matching. For hypergraph structure loss, This is a penalty term for violations of timing direction. In this embodiment... Determined through a validation set grid search. Neural prediction scores are obtained by pooling the outputs of hypergraph convolutions. ,in For the multilayer perceptron scoring head. Rule compliance score: ; (4) Type-based gating fusion. The fusion weights are adaptively adjusted according to the hyperedge type through the gating network: ; in, , For gating network parameters; Embedding for relation types. Enforcing super-edges for non-institutional collaborative relations. The remaining weights are then normalized and redistributed. Overall prediction score: ; , , , , These are the fusion weights for neural prediction score, meta-path semantic score, rule compliance score, role complementarity score, and triadic coupling score, respectively. When The output is then sent to the evidence chain generation step.

[0071] S9. Path Backtracking and Controlled Chain of Evidence Generation (1) Extraction of six domain-specific evidence elements. For effective prediction candidates... ,extract : ① Supporting subgraph : Binary relations, child hyperedges, and entity attribute records of candidate hyperedges; ② Critical Meta Path Instance Top-K instances selected according to S4 attention weights; ③ Technology life cycle basis : Entity phase, phase sequence monotonicity, attenuation intensity; ④ Basis for complementary roles Role allocation, strength of role complementarity, and overall complementarity score; ⑤ Policy Transmission Basis (Institutional coordination only): Coupling strength, delay verification results, and time window position; ⑥ Interdisciplinary measurement criteria (Interdisciplinary only): Distance between subject classification trees and semantic relevance of intermediate nodes.

[0072] (2) Structured prompt template. The organization provides a structured cue template for consumption by the large language model, explicitly restricting its interpretation to be based solely on the provided evidentiary elements. Key fields of the cue template include: candidate hyperedge identifiers, hyperedge relation types, participating entities and their role assignments, key summaries of supporting subgraphs, descriptions of key meta-path instances, technology lifecycle phase and attenuation strength, query results for the complementarity strength of role pairs, policy transmission basis (if applicable), and interdisciplinary measurement basis (if applicable). The template explicitly requires the large language model to generate only semantic descriptions and causal explanations of the given evidence, prohibiting the introduction of entity, event, or relation statements beyond the scope of the given evidence.

[0073] (3) Triple controlled generation mechanism. ① Boundary constraints: explicitly write constraint clauses that are only based on the given evidence, and constrain the output format and fields; ② Post-fact verification: reverse trace the factual statements in the output involving specific entities, time, and reference relationships, and remove statements that cannot find corresponding support as illusions; ③ Logical consistency verification: perform internal logical consistency checks on the output evidence chain based on the rule base, and trigger regeneration if it fails.

[0074] (4) Standardized output: ,in It is a natural language interpretation verified by a triple-controlled generation mechanism.

[0075] S10, Closed-Loop Feedback and Phase Evolution of the Technology Life Cycle Construction of the set to be updated. Valid predictions verified by the chain of evidence are included in the set to be updated. ,Right now Chain of evidence Validity verification through S9's triple controlled generation mechanism Redundancy and conflict checks. ① Redundancy check: If a candidate hyperedge completely overlaps with an existing hyperedge at the entity set and relation type levels, it will not be written repeatedly, but the confidence record will be updated instead. ② Conflict check: Candidates that conflict with existing relations in terms of type or time will be placed in a manual review queue or automatically invoked to make a decision based on auxiliary verification logic. Technology lifecycle phase reassessment. The set to be updated involves the entity set and its one-hop neighbor entity set. The phase determination module of S3 is invoked again to recalculate the cumulant and the first and second derivatives to determine the current phase. When the determination result is inconsistent with the original phase, a phase evolution event is triggered, and the attenuation coefficient and directional constraint are updated synchronously. Incremental retraining is based on elastic weight consolidation. When the cumulative set to be updated reaches a threshold... Incremental retraining is triggered at certain times, and the expanded loss function is: ; in For newly added data, For the current parameter, For parameters before retraining, For the diagonal elements of the Fisher information matrix, The strength of the elastic constraint. This strategy can mitigate catastrophic forgetting while absorbing the benefits of new data.

[0076] A closed-loop mechanism is formed. S1-S10 form a closed-loop mechanism that connects prediction, verification, updating, and phase evolution through the cycle of the set to be updated, phase re-estimation, incremental retraining, and re-prediction, enabling the science and technology knowledge graph to have dynamic evolution capabilities.

[0077] The specific parameter values ​​in the embodiments of this invention are merely examples. Those skilled in the art can adjust parameters such as the number of model layers, embedding dimension, attenuation coefficient, latency window, and threshold according to specific application scenarios. The method and system described in this invention can be applied to various science and technology intelligence decision-making scenarios: in scientific research layout decision-making scenarios, predicting collaborative innovation relationships helps identify potential industry-academia-research collaborative partners; in technology evolution trend judgment scenarios, predicting technology integration relationships reveals possible future technology combination directions; in policy effectiveness evaluation scenarios, predicting institutional synergy relationships assesses the transmission effect of specific policies on related technology fields and industries; in cutting-edge interdisciplinary discovery scenarios, predicting interdisciplinary relationships uncovers potential disciplinary integration opportunities. Any modifications and equivalent transformations made based on the technical concept of this invention should fall within the protection scope of this specification.

Claims

1. A method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints, characterized in that, Includes the following steps: Step 1: Construct a three-layer science and technology knowledge graph and formalize it into an octet; apply hard constraints on type and role compatibility to the science and technology meta-ontology and eliminate candidates that violate these constraints. Step 2: Text encoding is used for semantic nodes and topological encoding is used for structural feature nodes. A unified embedding space is formed through type projection. The information of binary neighborhood, hyperedge neighborhood and node itself is aggregated using hierarchical hypergraph convolution. Differentiated weighting is applied to the neighborhood aggregation through relation type attention parameters and role perception parameters. The aggregated information of entity layer, binary relation layer and higher-order relation layer is dynamically weighted and fused through hierarchical fusion weight. Step 3: Determine the phase of the technology life cycle based on the time series of entity attributes, construct a time series decay kernel by setting differentiated decay coefficients for the nascent stage, growth stage, maturity stage and decline stage, and apply asymmetric directional constraints to the phase sequence of the meta-path instance to characterize the irreversibility of the transformation of technology evolution from basic research to application. Step 4: Construct temporal meta-paths for collaborative innovation, technological integration, institutional synergy, event co-evolution, and interdisciplinary intersection, and calculate the semantic score of the meta-paths through an attention mechanism; Step 5: Construct a multi-agent role complementarity matrix, calculate the complementarity score of any two entity roles within the candidate hyperedge, and eliminate role homogeneous candidates with scores below the threshold. Step 6: Construct a three-element coupling tensor of policy, technology, and industry for institutional synergy and decompose it to alleviate sparsity. Add a policy transmission delay window constraint, requiring that the response times of policy, technology, and industry strictly increase and fall within the preset window, and eliminate candidates that fail the verification. Step 7: Achieve cross-layer joint prediction through bidirectional cross-validation of dimensionality-up prediction and dimensionality-down verification, and perform hierarchical consistency verification using super-edge substructure support constraints, cross-layer confidence transfer constraints, and type-role consistency constraints; Step 8: Use the hard constraints of the ontology as the first layer of filtering, relax the various domain rules into differentiable constraint losses through product T-norm, and combine them with multiple losses in a weighted manner to form a joint loss. Then, through type condition gating, adaptively fuse the neural prediction score, meta-path semantic score, rule compliance score, role complementarity score and ternary coupling score to obtain a comprehensive prediction score. Step 9: For candidate higher-order links that exceed the threshold, extract six types of evidence elements based on path backtracking, and generate a structured and interpretable evidence chain through a triple mechanism of boundary constraints, post-fact verification and logical consistency verification. Step 10: Feed back the verified prediction results to the knowledge graph and include them in the set to be updated. Update the technology life cycle phase of the affected entities. When the set to be updated reaches the threshold, trigger incremental retraining based on elastic weight consolidation to form a closed-loop evolution mechanism.

2. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 1, the octet formal definition uses the science and technology meta-ontology as a unified semantic standard, abstracting the science and technology knowledge graph into a unified representation of entity nodes, binary edges, hyperedges and their relation types, role assignments, time information, technology life cycle phases and relation levels; the ontology hard constraints are as follows: based on the predefined type compatibility rules and role compatibility rules in the science and technology meta-ontology, the candidate higher-order links are forcibly verified, and those that violate any rule are directly eliminated.

3. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, is characterized in that... In step 2, relationship-type-aware attention weights and role-aware weights based on role assignment are introduced into the hierarchical hypergraph convolution to achieve differentiated aggregation of neighbors of different relationship types and roles. The hierarchical fusion weights are adaptively generated by the sparsity of the node neighborhood and the type distribution. For long-tail entities with sparse neighborhoods, the entity layer weights are automatically increased, and for core entities with rich interactions, the higher-order relationship layer weights are automatically increased.

4. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 3, the temporal decay kernel function of the technology life cycle phase perception is assigned a differentiated decay coefficient according to the different stages of the entity's nascent, growth, maturity, or decline, so that entities at different evolutionary stages have different historical information decay rates; the asymmetric directional constraint determines the monotonicity of the phase sequence of entities involved in the meta-path, applies a positive constraint to sequences that conform to the unidirectional evolution law from the nascent stage to the decline stage, and applies a constraint to strictly reverse sequences that violate this law.

5. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 4, the meta-path semantic score is the weighted sum of the attention weight and temporal rationality score of each meta-path instance; among them, the collaborative innovation meta-path requires that the active time of adjacent entities is not strictly monotonically increasing; the technology integration meta-path requires that the active time intervals of participating technologies overlap; the institutional collaboration meta-path requires that the node response time is strictly monotonically increasing; the event co-evolution meta-path requires that the time of the preceding and following events be strictly sequential and causally related; and the interdisciplinary cross-path requires that the distance between the two disciplines in the discipline classification tree exceeds a preset threshold.

6. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 5, the role complementarity matrix is ​​pre-calibrated by domain experts based on the mechanism of technological innovation collaboration, and is used to quantify the strength of synergy and complementarity between different roles.

7. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 6, the ternary coupling tensor is obtained by decomposing the outer product of multiple factor vectors; the policy transmission delay constraint requires that the response times of policy, technology and industry increase sequentially and that adjacent time intervals fall within a preset reasonable delay window.

8. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 7, the superedge substructure support constraint is used to verify whether the candidate superedge contains a sufficient proportion of observed or high-confidence binary relations; the cross-layer confidence propagation constraint is used to prevent the prediction confidence of the higher-order relation layer from being unreasonably significantly higher than the reasonable aggregation value based on the confidence of the lower entity layer and the binary relation layer. The type and role consistency constraint is used to verify whether the type and role assignment of each entity within the candidate hyperedge conforms to the compatibility requirements of the predetermined relationship type template; only when the joint verification of the three types of constraints passes will the candidate hyperedge enter the subsequent comprehensive prediction score calculation process.

9. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 8, the fusion weights of the comprehensive prediction scores are adaptively generated by the type-conditional gating network based on the relationship type of the candidate hyperedges. For candidate hyperedges that are not institutional collaborative relationships, the weights of their ternary coupling score items are set to zero and the corresponding weights are redistributed to the remaining score items.

10. The method for predicting high-order links in a large-scale science and technology knowledge graph that integrates technology lifecycle and multi-agent role constraints as described in claim 1, characterized in that, In step 9, the cue boundary constraint is used to restrict the large language model to interpret only based on the given structured cue, the post-fact verification is used to reverse the factual statements in the evidence chain, and the logical consistency check is used to perform internal consistency of the output evidence chain based on the rule base; in step 10, the incremental retraining adopts the elastic weight consolidation strategy, which prevents new data from overwriting existing knowledge by applying elastic constraints to the existing parameters.

Images