A sci-tech information recommendation method based on a time sequence heterogeneous graph

Abstract

This invention belongs to the field of graph representation learning and scientific and technological intelligence analysis, specifically involving a scientific and technological intelligence recommendation method based on temporally heterogeneous graphs. First, temporally context-aware encoding is performed on various types of entities and their relationships in complex temporally sequenced scientific and technological intelligence, such as authors, papers, institutions, and research topics, mapping time scalars to contextual features containing absolute evolutionary cycles and relative time decay. Second, a time-guided dynamic collaborative attention mechanism is used to adaptively allocate the fusion ratio of multiple relationships based on temporal signals. Next, an evolutionary state update module based on gated recurrent units is introduced to iteratively update local instantaneous features into global long-term memory. Then, combined with a temporally smoothing alignment mechanism, joint optimization is performed by penalizing the Euclidean distance between memory states of adjacent time steps. Finally, the learned node representations are used for scientific and technological intelligence association recommendation tasks such as potential collaboration discovery, related results recommendation, and topic association analysis.

Description

Technical Field

[0001] This invention belongs to the field of graph representation learning and scientific and technological intelligence analysis, specifically involving a scientific and technological intelligence recommendation method based on temporal heterogeneous graphs, which can be used for practical discovery of potential collaborating experts, recommendation of relevant results, and analysis of associations of cutting-edge topics. Background Technology

[0002] Existing science and technology intelligence service platforms typically need to aggregate, clean, map, and intelligently recommend multi-source data from paper databases, patent databases, project databases, achievement databases, and institutional databases to support scientific research management, technology assessment, expert discovery, achievement transformation, and thematic monitoring. Science and technology intelligence data usually exhibits characteristics of multiple entities, multiple relationships, and strong temporal evolution. For example, there are various relationships among researchers, papers, patents, projects, institutions, and research topics, such as publication, application, citation, collaboration, affiliation, co-occurrence, participation, and transformation, and these relationships continue to evolve over time. Traditional recommendation methods often struggle to simultaneously consider the heterogeneity, temporal sequence, and complex relationships of science and technology intelligence objects, resulting in insufficient accuracy, timeliness, and interpretability of the platform's recommendation results.

[0003] Current mainstream graph representation learning algorithms for complex interaction scenarios can be divided into the following categories: Heterogeneous graph representation learning algorithms based on a static perspective: Dong et al. proposed metapath2vec in their 2017 paper "metapath2vec: Scalable representation learning for heterogeneous networks" published at KDD. Metapath2vec constructs node sequences by designing a random walk strategy based on metapaths and learns node embeddings using a Skip-gram model. Fu et al. proposed MAGNN in their 2020 paper "Magnn: Metapath aggregatedgraph neural network for heterogeneous graph embedding" published at the WWW conference. MAGNN introduces an internal aggregation mechanism within metapaths, encoding the features of intermediate nodes on the path before fusing the semantics of the metapath. Schlichtkrull et al. proposed Relational Graph Convolutional Network (R-GCN) in their 2018 paper "Modeling relational data with graph convolutional networks" published at ESWC. R-GCN assigns a specific transformation matrix to each edge type in the graph and processes the graph structure through a relation-specific message passing mechanism. Sun et al. proposed GTC in their 2025 paper "GNN-transformer co-contrastive learning for self-supervised heterogeneous graph representation" published in Neural Networks. GTC improves the representation capabilities of heterogeneous graphs through a co-contrastive learning framework combining GNN and Transformer. However, these methods inevitably assume that connections between nodes are singular or static, treating multiple relationships simply as independent channels and ignoring the dynamic evolution of interaction behavior over time. This limits the model's application potential in complex, multi-association, dynamic scenarios, making it unable to dynamically adapt to real-time changes in interaction relationships.

[0004] Heterogeneous graph representation learning algorithms based on dynamic perspectives: These methods attempt to capture the evolutionary patterns of graph structures over time. Pareja et al., in their 2020 AAAI paper "Evolvegcn: Evolving graph convolutional networks for dynamic graphs," proposed EvolveGCN, which uses a recurrent neural network (RNN) to evolve the weight parameters of the GCN, thus adapting to dynamic changes in the graph structure. Wang et al., in their 2020 TKDE paper "Dynamicheterogeneous information network embedding with meta-path based proximity," proposed a model called HIN based on meta-path proximity, which updates heterogeneous node representations by capturing small changes between first-order and second-order meta-path snapshots. Zhang et al., in their 2023 AAAI paper "Dynamic heterogeneous graph attention neural architecture search," proposed DHGAS, which handles dynamic heterogeneous graphs by jointly modeling heterogeneous neighbors and temporal dependencies across different snapshots. In their 2020 paper "Odeling Dynamic Heterogeneous Network for Link Prediction Using Hierarchical Attention with Temporal RNN," published in ECML PKDD, Xue et al. proposed DyHATR, which introduces a hierarchical attention mechanism combined with Transformer to capture the evolutionary patterns of different types of neighbors over continuous time. However, most of these methods focus on capturing the temporal evolution of a single type of edge, or treat edge type as a simple feature label, making it difficult to distinguish the complex temporal logic and nonlinear time decay differences between different interaction types. For example, in a science and technology intelligence service platform, the concentrated citation of a hot paper or the short-term high-frequency collaboration of a research team often has a short-term promoting effect on the association of subsequent results and expert recommendations; while changes such as the shift in research direction, long-term lack of collaboration, and adjustments in institutional layout may have a longer-term impact on the evolution of subsequent relationships. Existing methods lack an effective mechanism to align the evolutionary trajectory of multiple interactions on the time axis, thus losing key time-sensitive information useful for learning dynamic heterogeneous graph representations, resulting in existing models only achieving suboptimal performance.

[0005] The recommendation functions in existing science and technology intelligence platforms typically model heterogeneous relationships as static graph snapshots for spatial aggregation, or use only simple temporal models to handle dynamic edges of a single type. Therefore, they neglect the temporal sequence of heterogeneous interactions and the characteristics of their changing influence over time. In the actual operation of science and technology intelligence service platforms, different types of interactions often have different temporal impact patterns. For example, recent citations of papers and patents, short-term active collaborations within research teams, and rapid co-occurrence of keywords usually promote subsequent results recommendations and the discovery of potential collaborations; while changes such as shifts in research interests, long-term lack of collaboration, institutional restructuring, and declining technological popularity can have a lasting impact on subsequent interactions. These complex spatiotemporal collaborative patterns have often been overlooked in previous research, making it difficult for existing platforms to accurately distinguish the true contribution of interactions occurring at different points in time to the current state. This results in an inability to fully characterize knowledge dissemination paths, the evolution of collaborative relationships, and changes in entity states, ultimately leading to biased recommendation results. Summary of the Invention

[0006] To address the problems of existing technologies, the present invention aims to provide a science and technology intelligence recommendation method based on temporal heterogeneous graphs, applicable to science and technology intelligence service scenarios. This invention solves the suboptimal recommendation performance problem caused by neglecting the timeliness of interaction sequences and the collaborative evolution of multiple spatial relationships in complex, dynamic, and heterogeneous interaction scenarios such as science and technology intelligence service platforms by deeply injecting the temporal dimension into the heterogeneous relationship modeling process. Using the technical solution of this invention, the fusion weights of multiple relationships can be adaptively guided by the temporal context, distinguishing the short-term enhancement effects and long-term decay patterns of different interaction behaviors, learning effective node representations that include long-range evolutionary dependencies, and outputting results such as potential collaboration discovery, related achievement recommendations, patent association analysis, topic association analysis, and science and technology trend judgment.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for recommending scientific and technological information based on temporal heterogeneous graphs is proposed. First, a set of scientific and technological information objects and interaction data are acquired. The set of scientific and technological information objects includes one or more of the following: papers, patents, projects, researchers, institutions, and keywords. The interaction data consists of a set of data representing the interaction behaviors between scientific and technological information objects within a preset time range. Each interaction behavior includes one or more of the following: source object, target object, relationship type, and timestamp. First, each interaction behavior with a timestamp is encoded using temporal context awareness, uniformly mapping the time scalar to contextual features containing absolute evolutionary cycles and relative time decay. Second, a temporally guided dynamic collaborative attention mechanism is used to adaptively allocate the fusion ratio of multiple spatial relationships. Next, an evolutionary state update module based on gated recurrent units is introduced to update local instantaneous features into global long-term memory. Finally, joint optimization is performed using temporal smoothing alignment regularization to output associated recommendation results, and continuous optimization is performed until the optimal model is reached. The specific steps are as follows: Step (1): Perform temporal context-aware encoding to extract the timeliness features of interactive behaviors; 1.1) For the current interaction behavior in the interaction data between scientific and technological intelligence objects, let the timestamp of the current interaction behavior be denoted as . The timestamp is extracted by harmonic time transformation. High-dimensional absolute time vector ,in, for 3D real vector space, The dimension representing the absolute time-coded vector. Represents the absolute time-coded vector The Each component is calculated as follows: in, and The first The frequency and phase parameters of each component are adaptively learned during model training.

[0008] 1.2) For the current interaction behavior record in the interaction data, let the timestamp of the current interaction behavior record be... Let the timestamp of the last interaction between the technology intelligence object corresponding to the current interaction record be . Calculate the time span between two interactions. The relative decay is measured by an exponential time interval encoding with a learnable scaling factor, calculated as follows: in, Indicates time span The generated relative time decay encoding vector; This represents a learnable parameter that controls the rate of time decay; and Representing time span The corresponding linear transformation matrix and bias terms; express A real matrix with 1 row and 1 column, express A dimensional real vector space.

[0009] 1.3) By using a multilayer perceptron and a nonlinear activation function, absolute time evolution coding and relative time decay coding are fused to generate a comprehensive temporal context representation of the current interaction event. : Where || denotes the vector concatenation operation. Represents the activation function of a linear unit. This is a linear transformation matrix that fuses and maps the absolute time-coded vector with the relative time-decayed coded vector. In the initial stage of evolution... , each node The historical interactive memory state is initialized to , Represents a node In the historical interactive memory state at the initial time step Represents a node The static embedding vector at the initial time step, Indicates based on node The node identifier is an initial static embedding vector obtained by mapping or retrieving from a trainable node embedding matrix. In the constructed temporal heterogeneous graph, scientific and technological intelligence objects are represented as nodes, including paper nodes, patent nodes, project nodes, and researcher nodes.

[0010] Step (2): Based on a time-guided dynamic collaborative attention mechanism, multi-dimensional spatial relationships are integrated; 2.1) Given a central node (The technology intelligence object node currently awaiting update) in time Through relationships The set of connected neighbor nodes The spatial preliminary aggregate representation is extracted through graph convolution operations: in, Indicates the central node At time step relation type The initial spatial aggregation representation below; Indicates at time step By relation type With the central node The set of connected neighbor nodes; Represents the set of neighboring nodes The number of nodes in; Represents any neighbor node in the set of neighbor nodes; Representing neighboring nodes In the previous time step eigenvectors; The learnable weight matrix representing the aggregation of local spaces; This represents a non-linear activation function.

[0011] 2.2) Integrated Central Node Previous global memory state Synthetic temporal context representation and relation types The initial spatial aggregation representation below Calculate relation type For the target node Dynamic importance score The calculation is as follows: in, This represents the projection matrix corresponding to the attention mechanism; This represents the transpose of the learnable parameter vector.

[0012] Dynamic collaborative attention weights are derived by normalizing the results over all relation types using the Softmax function. : in, Indicates at time step Lower relation type For the target node Dynamic collaborative attention weights; Represents the set consisting of all relation types; Represents a set of relation types Index variable of any relation type in the table.

[0013] 2.3) Perform weighted summaries based on this dynamic weight, and generate the time. Local comprehensive representation below : in, Indicates the central node At time step The local comprehensive representation after incorporating the influence of multiple relationship types; Representing relation type The corresponding feature transformation matrix.

[0014] Step (3): Update the spatiotemporal evolution state through a gated loop mechanism; 3.1) A gated recurrent unit (GRU) is used to control the fusion ratio of historical information and current features, and the update gate is calculated. and reset door The calculation process is as follows: in, Represents a node At time step The update gate is used to characterize the proportion of historical state information retained and current feature information fused at the current moment; Represents a node At time step The reset gate is used to characterize the degree of participation of historical state information in the candidate state calculation; This refers to the node generated in step 2.3). At time step Local comprehensive representation; Represents a node In the previous time step The global memory state; and These represent the input feature weight matrices corresponding to the update gate and the reset gate, respectively; and These represent the historical state weight matrices corresponding to the update gate and the reset gate, respectively; and These represent the bias vectors corresponding to the update gate and the reset gate, respectively; This represents the Sigmoid activation function.

[0015] 3.2) Calculate the node based on the output of the reset gate. At time step candidate states : in, Represents a node At time step Candidate states; and These represent the input feature weight matrix and the historical state weight matrix corresponding to the candidate state, respectively; This represents the bias vector corresponding to the candidate state; Represents the hyperbolic tangent activation function; This represents the Hadamard product, which is the element-wise product.

[0016] 3.3) By using an update gate to perform linear interpolation between the historical state and the candidate state, the final update of the node's evolutionary memory is completed: At the last time step The output will be represented Embedded as the final node Used for downstream tasks.

[0017] Step (4): Introduce smooth alignment for joint optimization and output the prediction results; 4.1) Embed the final node obtained in the last time step of step 3. As input to the downstream task decoder, the task output prediction results are analyzed based on preset technology intelligence, and the main task loss is calculated. The prediction results include predictions of the correlations between scientific and technological intelligence objects or predictions of the state values ​​of scientific and technological intelligence objects.

[0018] Specifically, the association prediction results are used to characterize whether two scientific and technological intelligence objects have a target association relationship in a later time step, and the probability or strength of such association relationship. The target association relationship includes collaboration, publication, application, citation, participation, affiliation, co-occurrence, and transformation, etc., and can be used for identifying potential collaborating experts, recommending related papers, recommending related patents, discovering project associations, analyzing technology transfer paths, and analyzing associations of cutting-edge topics. The state value prediction results are used to characterize the continuous attribute values ​​of a single scientific and technological intelligence object in a later time step. These continuous attribute values ​​include the expected citation frequency of papers or patents, the activity level of researchers, the level of institutional research output, and the level of project attention, etc. The results can be used for judging scientific and technological trends, monitoring hot topics, and assessing the activity status of scientific and technological intelligence objects.

[0019] When performing the association prediction task, the final node embeddings corresponding to the two science and technology intelligence objects to be predicted and the relationship features of the target relationship type are input into the association prediction decoder, and the probability value of the target association relationship between the two science and technology intelligence objects is output. When performing the state value prediction task, the final node embedding corresponding to the target science and technology intelligence object is input into the regression prediction layer, and the continuous attribute prediction value of the target science and technology intelligence object in the next time step is output.

[0020] 4.2) To further mitigate the abnormal fluctuations caused by the sparsity of dynamic heterogeneous graph interaction data, a temporal smoothing alignment regularization term is calculated by penalizing the Euclidean distance between the memory states of the same node in adjacent time steps. : 4.3) Calculate the overall objective function using an end-to-end joint optimization strategy. Train and optimize the model parameters until the model converges to its optimal state. This represents the weight coefficient of the temporal smoothing alignment regularization term, which is used to control the influence of the temporal smoothing alignment regularization term on the overall objective function.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention enables unified modeling of multi-source scientific and technological intelligence objects, including papers, patents, projects, researchers, institutions, and topics, and deeply couples temporal evolution patterns with spatial multi-association semantics. By using temporal signals to adaptively guide the fusion weights of various relationships, it accurately distinguishes the short-term enhancement effects and long-term decay patterns of different interactive behaviors on the time axis. Combined with a temporal smoothing alignment mechanism that forces features to transition smoothly between adjacent time steps, it learns robust and coherent entity representations. This effectively supports platform business functions such as potential collaborative expert discovery, achievement recommendation, patent association analysis, cutting-edge topic tracking, institutional assessment, and scientific and technological trend monitoring, while also considering generalization capabilities for other dynamic and heterogeneous interaction scenarios. This invention considers the nonlinear decay of interaction timeliness and the dynamic transfer patterns of multiple types of relationships in scientific and technological intelligence scenarios. Through temporal collaborative modeling, it learns coherent and robust node representations and improves the effectiveness of scientific and technological intelligence association recommendation. Attached Figure Description

[0022] Figure 1 This is the basic framework of the method of this invention. Detailed Implementation

[0023] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0024] The basic process of the technology information recommendation method based on temporal heterogeneous graphs of the present invention is as follows: Figure 1 As shown. First, acquire the set of scientific and technological intelligence objects and interaction data; wherein, the set of scientific and technological intelligence objects includes one or more of papers, patents, projects, researchers, institutions and keywords; interaction data refers to the data set consisting of the interaction behaviors between the scientific and technological intelligence objects within a preset time range, and each interaction behavior includes at least the source object, target object, relationship type and timestamp. Perform the following 4 steps: (1) Temporal context-aware encoding to extract the timeliness features of interaction behaviors in the interaction data; (2) Based on the temporal guidance dynamic collaborative attention mechanism, integrate multi-dimensional spatial relationships; (3) Update the cross-temporal evolution state through the gating loop mechanism; (4) Introduce smooth alignment for joint optimization and output the prediction results.

[0025] The first step is to use temporal context-aware encoding to extract the timeliness features of interactive behaviors.

[0026] 1) For interactive behaviors in the interactive data, the timestamp is: Timestamps are extracted using harmonic time transformation. High-dimensional absolute time encoding vector This is used to characterize the evolutionary position information of the current associated event record in the absolute time dimension.

[0027] 2) Regarding the relative time decay between actions, given the current interaction time... Time since last interaction Calculate the time span An exponential time interval encoding with a learnable scaling factor is introduced to measure relative decay. This is used to characterize the time decay effect of historical related events on current related events.

[0028] 3) By using a multilayer perceptron and a nonlinear activation function, absolute time evolution coding and relative time decay coding are fused to generate a comprehensive temporal context representation of the current interaction event. .

[0029] The second step involves integrating multi-dimensional spatial relationships through a time-guided dynamic collaborative attention mechanism.

[0030] 1) Given a central node At time step Below by relation type The central node is extracted from the set of connected neighboring nodes using graph convolution operations. In relation types The initial spatial aggregation representation below .

[0031] 2) Integrated Central Node Global memory state at the previous time step The first step generates a comprehensive timing context representation. and relation types The initial spatial aggregation representation below Relationship types are calculated through the feedforward attention layer. For the target node Dynamic importance score .

[0032] 3) Based on the calculated dynamic weights, the central node... The feature representations under different relational channels are weighted and summarized to generate a local comprehensive representation. .

[0033] The third step is to update the spatiotemporal evolution state through a gating loop mechanism.

[0034] 1) A gated loop unit is used to control the fusion ratio of historical information and current features; given a node Global memory state at the previous time step and the local synthesis representation generated in the second step compute nodes At time step Update Gate and reset door .

[0035] 2) Calculate the node based on the output of the reset gate. At time step candidate states .

[0036] 3) By updating the door Regarding the historical state With candidate state Perform linear interpolation to complete the node. At time step global memory state The final update. At the last time step. , will node Global memory state Embedded as the final node This is used for subsequent downstream tasks.

[0037] The fourth step involves introducing smooth alignment for joint optimization and outputting the prediction results.

[0038] 1) Embed the final node obtained in step 3 As input to the downstream task decoder, the system outputs prediction results and calculates the main task loss based on the preset science and technology intelligence analysis task. The prediction results include correlation prediction results between science and technology intelligence objects or state value prediction results for science and technology intelligence objects. When performing a correlation prediction task, the final node embeddings corresponding to the two science and technology intelligence objects to be predicted, along with the relationship features of the target relationship type, are input to the correlation prediction decoder, which outputs the probability value of a target correlation between the two science and technology intelligence objects. When performing a state value prediction task, the final node embedding corresponding to the target science and technology intelligence object is input to the regression prediction layer, which outputs the continuous attribute prediction value of the target science and technology intelligence object at the next time step.

[0039] 2) To mitigate the abnormal fluctuations caused by the sparsity of dynamic graph interaction data, a temporal smoothing alignment regularization term is calculated by penalizing the Euclidean distance between the memory states of the same node in adjacent time steps. .

[0040] 3) Calculate the overall objective function using an end-to-end joint optimization strategy. The model parameters are trained and optimized based on the overall objective function until the model converges. The overall objective function is composed of a weighted sum of the main task loss and the temporal smoothing alignment regularization term, with weight coefficients... Used to control the influence of the temporal smoothing alignment regularization term on the overall objective function.

[0041] Based on the solution of this invention, the experimental analysis is as follows: This method is constructed and tested using dynamic heterogeneous network datasets from real-world scenarios such as publicly available scientific and technological intelligence, and compared with current recommendation and prediction techniques to evaluate the effectiveness of this method.

[0042] (1) Introduction to the dataset OGBN-MAG was used as the primary scenario validation dataset to simulate a multi-entity relationship network of "author-paper-institution-research topic" in a science and technology intelligence service platform. In practical applications, this relationship network can be further expanded to a multi-entity science and technology intelligence network of "researcher-paper-patent-project-institution-research topic". Simultaneously, generalization performance was compared and tested on the COVID and tgbl-wiki datasets. COVID corresponds to the scenario of topic-based emergency scientific research knowledge association, while tgbl-wiki corresponds to the scenario of open knowledge collaborative maintenance. Detailed statistical information for the datasets is shown in Table 1.

[0043] The detailed information of the dataset is shown in Table 1: Table 1. Statistical information of the dataset

[0044] (2) Comparison of experimental results between this method and other mainstream methods This method was compared with state-of-the-art baseline models, including LSTM, Transformer, GCN, GAT, HGT, DySAT, and HTGNN, on dynamic association prediction tasks (using AUC and AP as evaluation metrics) and node regression tasks (using MAE and RMSE as evaluation metrics). The link prediction comparison experiments on the OGBN-MAG and tgbl-wiki datasets are shown in Table 2. Table 2. Results of the comparative experiment on association prediction

[0045] Table 3 shows the comparative experiment on state value prediction in COVID: Table 3 Comparison of State Value Prediction Experiment Results

[0046] As can be seen from the results in Tables 2 and 3, the proposed method achieves better results in most cases, indicating that the method proposed in this invention has good effectiveness, stability and application value in complex interactive scenarios.

Claims

1. A method for science and technology intelligence recommendation based on a time-series heterogeneous graph, characterized in that, First, a set of scientific and technological intelligence objects and interaction data are acquired. The set of scientific and technological intelligence objects includes one or more of the following: papers, patents, projects, researchers, institutions, and keywords. The interaction data consists of a set of data on the interaction behaviors between scientific and technological intelligence objects within a preset time range. Each interaction behavior includes at least a source object, a target object, a relationship type, and a timestamp. First, each interaction behavior with a timestamp is encoded using temporal context awareness, uniformly mapping the time scalar to contextual features containing absolute evolutionary cycles and relative time decay. Second, a temporally guided dynamic collaborative attention mechanism is used to adaptively allocate the fusion ratio of multiple spatial relationships. Next, an evolutionary state update module based on gated recurrent units is introduced to update local instantaneous features into global long-term memory. Finally, joint optimization is performed by combining temporal smoothing alignment regularization to output associated recommendation results, and the model is continuously optimized to the optimal model. 2.The technology intelligence recommendation method based on a time-ordered heterogeneous graph according to claim 1, characterized in that, The specific steps are as follows: Step (1): Perform temporal context-aware encoding to extract the timeliness features of interactive behaviors; 1.1) for a current interaction behavior in the interaction data between the objects of scientific information, let the time stamp of the current interaction behavior be , extract the high-dimensional absolute time vector of the time stamp by harmonic time transformation, where, is a -dimensional real vector space, is the representation dimension of the absolute time encoding vector, represents the th component of the absolute time encoding vector ; 1.2) For the current interaction behavior record in the interaction data, let the timestamp of the current interaction behavior record be , let the timestamp of the last interaction on the sci-tech information object corresponding to the current interaction behavior record be , calculate the time span between the two interactions , and introduce an exponential time interval coding with a learnable scaling factor to measure the relative decay; 1.3) fuse the absolute time evolution encoding and the relative time decay encoding by a one-layer multi-layer perceptron with a non-linear activation function to generate a comprehensive temporal context representation of the current interaction event ; Step (2): Based on a time-guided dynamic collaborative attention mechanism, multi-dimensional spatial relationships are integrated; 2.1) Given a center node i.e. the sci-tech intelligence object node whose representation is to be updated, at time Through relations Connected neighbor node set , a spatial preliminary aggregated representation is extracted through graph convolution operation; 2.2) integrated center node previous time global memory state , integrated temporal context representation and spatial preliminary aggregation representation under relationship type , calculate relationship type dynamic importance score of target node ;​ The dynamic collaborative attention weight is normalized on all the set of relation types by using a Softmax function ; 2.3) Weighted aggregation according to the dynamic weights, generating a temporal local summary representation ; Step (3): Update the spatiotemporal evolution state through a gated loop mechanism; 3.1) The proportion of fusion of historical information and current features is controlled by the gating recurrent unit GRU to calculate the update gate and the reset gate ; 3.2) Compute the nodes that incorporate the outputs of the reset gates At time step of the candidate states ; 3.3) Final update of the node evolution memory by linearly interpolating the history states and the candidate states through the update gates : At the last time step The output representation As final node embeddings For downstream tasks; Step (4): Introduce smooth alignment for joint optimization and output the prediction results; 4.1) embedding the final node obtained at the last time step in step 3 As the input of the downstream task decoder, the prediction result is output according to the preset science and technology information analysis task, and the main task loss is calculated ; wherein the prediction result includes an association prediction result between science and technology information objects or a state value prediction result of the science and technology information objects; 4.2) To further mitigate the abnormal fluctuations caused by the sparsity of dynamic heterogeneous graph interaction data, a temporal smoothing alignment regularization term is calculated by penalizing the Euclidean distance between the memory states of the same node at adjacent time steps ; 4.3) calculate the overall objective function by using end-to-end joint optimization strategy training the optimized model parameters until the model converges to the optimal, represents a weight coefficient of the time series smooth alignment regular term, used to control the influence strength of the time series smooth alignment regular term on the overall objective function.

3. The method for recommending scientific and technological information based on temporal heterogeneous graphs according to claim 2, characterized in that: In step 1.1), denotes the absolute time encoding vector of the first component, which is calculated as follows: wherein, and are the first component frequency and phase parameters learned adaptively during model training. In step 1.2), an exponential time interval encoding with a learnable scaling factor is introduced to measure the relative decay, calculated as follows: wherein, denotes a relative time-decay encoding vector generated by a time span ; denotes a learnable parameter controlling the rate of time decay; and denotes a linear transformation matrix and a bias term corresponding to the time span ; denotes a real matrix of size 1 x C, denotes a real vector space of dimensionality D. In step 1.3), the integrated temporal context representation of the current interaction event The formula for calculating the value of the parameter is as follows: wherein || denotes a vector concatenation operation, denotes a linear unit activation function, is a linear transformation matrix for fusing and mapping the absolute time encoding vector and the relative time decay encoding vector; in the initial stage of evolution , the historical interaction memory state of each node is initialized as , denotes the historical interaction memory state of the node at the initial time step, denotes the static embedding vector of the node at the initial time step, denotes the initial static embedding vector obtained by mapping or querying from the trainable node embedding matrix according to the node identifier of the node ; wherein the sci-tech information objects are respectively taken as nodes in the constructed time-heterogeneous graph.

4. The method for recommending scientific and technological information based on temporal heterogeneous graphs according to claim 2, characterized in that: In step 2.1), the preliminary spatial aggregation is represented as follows: in, Indicates the central node At time step relation type The initial spatial aggregation representation below; Indicates at time step By relation type With the central node The set of connected neighbor nodes; Represents the set of neighboring nodes The number of nodes in; Represents any neighbor node in the set of neighbor nodes; Representing neighboring nodes In the previous time step eigenvectors; The learnable weight matrix representing the aggregation of local spaces; Represents a nonlinear activation function; In step 2.2), the relation type For the target node Dynamic importance score The calculation is as follows: in, This represents the projection matrix corresponding to the attention mechanism; Represents the transpose of a learnable parameter vector; Dynamic collaborative attention weights The calculation is as follows: in, Indicates at time step Lower relation type For the target node Dynamic collaborative attention weights; Represents the set consisting of all relation types; Represents a set of relation types Index variable of any relation type in; In step 2.3), time Local comprehensive representation below The calculation is as follows: in, Indicates the central node At time step The local comprehensive representation after incorporating the influence of multiple relationship types; Representing relation type The corresponding feature transformation matrix.

5. The method for recommending scientific and technological information based on temporal heterogeneous graphs according to claim 2, characterized in that: In step 3.1), update the door. and reset door The calculation process is as follows: in, Represents a node At time step The update gate is used to characterize the proportion of historical state information retained and current feature information fused at the current moment; Represents a node At time step The reset gate is used to characterize the degree of participation of historical state information in the candidate state calculation; This refers to the node generated in step 2.3). At time step Local comprehensive representation; Represents a node In the previous time step The global memory state; and These represent the input feature weight matrices corresponding to the update gate and the reset gate, respectively; and These represent the historical state weight matrices corresponding to the update gate and the reset gate, respectively; and These represent the bias vectors corresponding to the update gate and the reset gate, respectively; This represents the Sigmoid activation function; In step 3.2), the node At time step candidate states The calculation is as follows: in, Represents a node At time step Candidate states; and These represent the input feature weight matrix and the historical state weight matrix corresponding to the candidate state, respectively; This represents the bias vector corresponding to the candidate state; Represents the hyperbolic tangent activation function; This represents the Hadamard product, which is an element-wise product. In step 3.3), the final update of the node evolution memory is calculated as follows: 。 6. The method for recommending scientific and technological information based on temporal heterogeneous graphs according to claim 2, characterized in that: In step 4.1), the association prediction result is used to characterize whether there is a target association relationship between two scientific and technological intelligence objects in the next time step and the association probability or association strength of the target association relationship. The target association relationship includes cooperation, publication, application, citation, participation, affiliation, co-occurrence, and transformation. The state value prediction result is used to characterize the continuous attribute value of a single scientific and technological intelligence object in the next time step. The continuous attribute value includes the expected citation popularity of the paper or patent, the activity level of researchers, the research output level of the institution, and the attention of the project. When performing the association prediction task, the final node embeddings corresponding to the two science and technology intelligence objects to be predicted and the relationship features of the target relationship type are input into the association prediction decoder, and the probability value of the target association relationship between the two science and technology intelligence objects is output. When performing the state value prediction task, the final node embedding corresponding to the target science and technology intelligence object is input into the regression prediction layer, and the continuous attribute prediction value of the target science and technology intelligence object in the next time step is output.

7. The method for recommending scientific and technological information based on temporal heterogeneous graphs according to claim 2, characterized in that: In step 4.2), the time-series smoothing alignment regularization term The calculation is as follows: 。

Images