Knowledge recommendation method fusing knowledge graph and graph convolutional neural network

Abstract

The application discloses a kind of knowledge recommendation methods of fusing knowledge graph and graph convolutional neural network, it is related to knowledge graph technical field.The method first constructs application context, knowledge unit and the interaction matrix of both;After using LightGCN, the collaborative features of application context and knowledge unit are extracted, and the feature vector of the application context to be recommended is obtained by similarity matching and KNN near neighbor weighting, the semantic features of knowledge unit are extracted using TransD, KGAT and hierarchical aggregation, finally the collaborative features and semantic features of knowledge unit are fused, the matching degree of application context collaborative features and knowledge unit comprehensive features is calculated and Top-N recommendation result is generated.The recommendation method of the application, recommendation no longer only depends on single user portrait or keyword, and the matching degree of final recommendation result is higher.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically to a knowledge recommendation method that integrates knowledge graphs and graph convolutional neural networks. Background Technology

[0002] The nuclear fuel manufacturing process involves a wide range of knowledge, including process tasks, equipment, parameters, equipment malfunctions, technical specifications, process manuals, maintenance records, and operating procedures. This knowledge is typically from diverse sources, comes in various formats, is dense with technical terminology, and is often scattered across document systems, equipment ledgers, production management systems, and maintenance records, easily creating knowledge silos. When faced with tasks such as process design, parameter setting, equipment operation and maintenance, or troubleshooting, frontline engineers often need to manually search, filter, and understand relevant knowledge from a large amount of data, resulting in low retrieval efficiency, difficulty in discovering knowledge connections, and insufficient knowledge reuse.

[0003] Existing knowledge management systems mostly employ keyword retrieval, category directory retrieval, or recommendation methods based on content similarity. These methods typically rely on explicit keyword matching, making it difficult to understand the deep semantic relationships between process tasks, equipment entities, parameter constraints, and reference documents. When the task description entered by the user does not completely match the expression in the document, retrieval omissions or recommendation biases are prone to occur. Although traditional collaborative filtering models can learn user preferences by utilizing historical browsing, bookmarking, and downloading behaviors, in industrial knowledge services, sparse interaction data and the frequent appearance of new users and new tasks lead to significant cold start problems. Furthermore, collaborative filtering models cannot fully express the structured semantics within knowledge units composed of process tasks, equipment, parameters, and documents.

[0004] Knowledge graphs can represent complex industrial knowledge in an entity-relationship-entity format, providing a foundation for the structured organization of knowledge related to nuclear fuel manufacturing processes and equipment. Graph convolutional neural networks can perform neighborhood aggregation and higher-order association learning on interaction graphs or knowledge graphs, providing deep feature representations for recommendation systems. However, simply using knowledge graph embedding easily overlooks the dynamic preferences formed by users in different task contexts, while simply using graph convolutional collaborative filtering struggles to handle the semantic structure of knowledge units and cold starts in new application contexts. Therefore, a knowledge recommendation method is needed that unifies the modeling of knowledge graph semantic features, graph convolutional collaborative features, task context features, and user background features. Summary of the Invention

[0005] To address at least one of the aforementioned problems, this invention proposes a knowledge recommendation method that integrates knowledge graphs and graph convolutional neural networks.

[0006] The technical solution of this invention is: a knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks, comprising the following steps: S1. Based on historical data, construct knowledge units and application context objects for process equipment, and construct an interaction matrix by combining the interaction between knowledge units and application context objects; the application context objects include user contexts and task contexts. S2. Based on the interaction matrix and the LightGCN model based on the layer attention mechanism, update the application context objects and knowledge units and calculate the weight of each layer. Perform weighted fusion of the output of each layer to obtain the application context collaborative feature vector and the knowledge unit collaborative feature vector. S3. For the correspondence between the context object to be recommended and the application context object, extract the application context collaborative feature vector, construct a knowledge graph embedding task based on TransD to reduce the computational burden caused by high-dimensional sparse representation, and obtain the initial semantic representation of entities and relations. Based on KGAT, add the local structural relations and high-order neighborhood semantics of the knowledge graph to the knowledge item representation. Based on the hierarchical feature representation learning architecture, perform weighted aggregation on multiple entities to obtain the semantic feature vector of the knowledge unit. S4. The knowledge unit collaborative feature vector and the knowledge unit semantic feature vector are fused using balanced hyperparameters to obtain the knowledge unit comprehensive feature vector. The inner product of the application context collaborative feature vector and the knowledge unit comprehensive feature vector is calculated, and the inner product is converted into a matching degree using the sigmoid nonlinear activation function. The knowledge units corresponding to the comprehensive feature vectors of the knowledge units with the highest matching degree (TOP-N) are selected as recommended knowledge.

[0007] The beneficial effects of this invention are as follows: 1. By unifying user background and current project tasks into a unified application scenario, recommendations no longer rely solely on a single user profile or keyword.

[0008] 2. LightGCN is used to learn the high-order collaborative relationship between application contexts and knowledge units, and KNN nearest neighbor weighting is used to construct features for new application contexts to alleviate the cold start of application contexts.

[0009] 3. Use TransD, KGAT and threshold filtering to extract semantics from the knowledge graph to make up for the lack of features caused by sparse interaction and lack of interaction records for new knowledge units.

[0010] 4. By fusing collaborative features with semantic features, Top-N knowledge unit recommendations are generated, improving the matching degree and interpretability of recommendation results with specific task semantics. Attached Figure Description

[0011] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a two-part diagram of the application scenario object-knowledge unit in an embodiment of the present invention; Figure 3This is an embedded diagram of the knowledge graph of nuclear fuel manufacturing process equipment based on the TransD model of this invention. Detailed Implementation

[0012] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0013] like Figure 1 As shown, a knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks includes the following steps: S1. Based on historical data, construct knowledge units and application context objects, and construct an interaction matrix by combining the interaction between knowledge units and application context objects; the application context objects include user contexts and task contexts. In historical data, knowledge units refer to the specific classification of knowledge components, including: manufacturing process equipment, manufacturing tasks, equipment failures, process equipment reference knowledge, process parameters, projects, organizational units, and professional fields.

[0014] In this embodiment, the nuclear fuel manufacturing process and equipment typically include a six-roll reversible cold rolling mill, a continuous push-boat type hydrogen-protected high-temperature sintering furnace, and a vacuum degreasing-high-temperature sintering integrated furnace. Manufacturing tasks typically include powder pressing, green billet sintering, pellet grinding and testing, and end plug welding. Equipment malfunctions typically include abnormal temperatures and inability to start. Reference knowledge for the process equipment typically includes equipment operation manuals, technical specifications, industry standards, maintenance procedures, and fault diagnosis guidelines. Process parameters typically include temperature, pressure, voltage, push-boat speed, hydrogen flow rate, and atmosphere dew point, such as a sintering temperature of 1740–1760°C. Projects typically include uranium powder preparation line modification projects and uranium assembly production projects. Organizational units typically include manufacturers, departments, workshops, production lines, and employees. Employees may include process engineers, equipment engineers, quality controllers, and equipment maintenance personnel. Professional fields typically include mechanical engineering, chemical engineering, electrical engineering, nuclear engineering, quality management, and computer science.

[0015] After obtaining the aforementioned historical data, a knowledge graph based on triples is constructed. This method is existing technology, so its specific operation will not be described in detail here.

[0016] The application context objects in historical data consist of two parts: user context and task context. The user context is mainly used to describe the basic attributes and ability characteristics of the people to be recommended, and it is used to characterize the differences between different users in terms of professional ability, business responsibilities and knowledge permissions. The task context is used to describe the task background corresponding to the knowledge application process.

[0017] In actual production processes, user contexts may include information such as user employee ID, department, position, major, education, skills, length of service, and system usage time. This type of information can be represented in different formats. For example, a user employee ID consists of a series of strings; different departments can be numbered. Specific examples are shown in Table 1. Of course, those skilled in the art will understand that the method shown in Table 1 is only one example, and can be modified according to actual circumstances.

[0018] Table 1 User Context Information Task contexts describe the specific task background corresponding to knowledge application, and can be further divided into process task contexts and maintenance task contexts. Process task contexts include the process task ID and process task description, while maintenance task contexts include equipment number, fault location, fault cause, and fault phenomenon, used to describe the knowledge application requirements in equipment operation, maintenance, and fault handling scenarios. All task contexts are text-based data.

[0019] For ease of description, a set of application context objects is established: , In the formula, Context represents the collection of application context objects; C m Let m represent the m-th application context object; n(m) represents the total number of application context objects; This represents the set of user scenarios for the m-th application scenario object; This represents the set of task contexts for the m-th application context object.

[0020] and , In the formula, context process Represents the context of a technological task; maintee This indicates the maintenance and inspection task scenario.

[0021] Since a knowledge unit is a set of knowledge items organized for a task, rather than a single isolated node, a knowledge unit set is established: In the formula, Knowledge represents the set of knowledge units, and K n Let n represent the nth knowledge unit; n(n) represents the total number of knowledge units.

[0022] For knowledge unit K n Its construction method is as follows: A knowledge unit is a set of knowledge items organized for a task, not a single isolated node. The knowledge unit set for nuclear fuel manufacturing process equipment is defined as follows: in Let n represent the nth knowledge unit, and n(n) represent the total number of knowledge units. Based on the knowledge graph task representation, knowledge units are defined for both process tasks and maintenance tasks.

[0023] For a process task, the knowledge unit consists of the process task, nuclear fuel manufacturing process equipment, process parameters, and process equipment reference knowledge, as defined in the following formula: in and For a single knowledge item, and It can contain multiple knowledge items, and the process task can be uniformly expressed as shown in the following formula: in, Representing knowledge units The Middle One knowledge item, This indicates the total number of knowledge items in a knowledge unit.

[0024] For maintenance tasks, the knowledge unit consists of maintenance tasks, nuclear fuel manufacturing process equipment, equipment failures, and process equipment reference knowledge, as defined in the following formula: At this time only It can be multiple knowledge items; the knowledge unit of the maintenance and inspection task can be uniformly expressed as: In the formula, Representing knowledge units The Middle One knowledge item, This indicates the total number of knowledge items in a knowledge unit.

[0025] When building the interaction matrix, a bipartite graph structure is constructed based on historical data, consisting of the knowledge unit set and the application context object set, such as... Figure 2 As shown.

[0026] In actual historical data, the interaction patterns of application context objects with a particular knowledge unit vary. For example, important knowledge might be saved or downloaded, less important knowledge might only be browsed, and completely unrelated knowledge might not be interacted with at all. Therefore, based on the actual situation, various interaction types are scored, as shown in Table 2. Those skilled in the art will understand that Table 2 is only one scoring method, and the interaction values ​​can be modified according to the actual situation.

[0027] Table 2 Interaction values ​​for different interaction types For the same application scenario, multiple interactions between the user and knowledge units are represented by the mean value to indicate their overall preference: In the formula, W mn C represents m For K n Overall preferences; f represents the total number of knowledge items in the knowledge unit; f represents the knowledge unit K. n The sequence number of the internal knowledge item, i.e., the f-th knowledge item.

[0028] The final interaction matrix is ​​shown below: In the formula, Interaction represents the interaction matrix.

[0029] S2. Based on the interaction matrix and the LightGCN model based on the layer attention mechanism, update the application context objects and knowledge units and calculate the weight of each layer. Perform weighted fusion of the output of each layer to obtain the application context collaborative feature vector and the knowledge unit collaborative feature vector. In this step, the LightGCN model based on layer attention introduces layer attention at the output of each layer of the LightGCN model, which can automatically learn the weights of different layers in the LightGCN model.

[0030] Specifically, in the initial layer of the model (i.e., layer 0), the application context and knowledge unit are each initialized with a dimension of size [missing value]. The embedding vector, denoted as and . Representing the application context The vector in the initial layer, and This represents a knowledge unit. The initial vectors are used as trainable underlying parameters during subsequent end-to-end model training, and are continuously optimized and updated through backpropagation of the loss function.

[0031] Subsequently, the interaction matrix is ​​incorporated, and LightGCN is used for multi-layer neighborhood propagation on the application context-knowledge unit bipartite graph. Application context object nodes update their representations by aggregating information from adjacent knowledge units, while knowledge unit nodes update their representations by aggregating information from adjacent application contexts. Specifically, application context nodes learn collaborative preferences for knowledge usage among user-task combinations by aggregating features of knowledge units they have historically interacted with; knowledge unit nodes learn which user backgrounds and task contexts are typically applicable to a knowledge unit by aggregating features of application contexts they have interacted with. Through multi-layer propagation, higher-order collaborative relationships such as "similar application contexts have accessed similar knowledge units" can also be discovered.

[0032] In this process, the aggregation update formulas for application context object nodes and knowledge unit nodes are as follows: , In the formula, Representing the application context In the The feature vector of the layer, and This represents a knowledge unit. Node at the The feature vector of the layer. Representing the application context The sum of interaction values ​​(such as browsing 0.4, downloading 1.0, etc.) between the knowledge units connected to it. Representing knowledge units The sum of interaction values ​​with each associated application context. The denominator is symmetrically normalized to attenuate the absolute dominance of high-frequency active nodes with excessively high degrees during propagation, thus maintaining the stability of the node feature space distribution in the network. After multiple layers of propagation, the model can uncover higher-order collaborative relationships beyond direct interactions, such as the potential pattern of "similar application contexts accessing similar knowledge units."

[0033] Multi-layer graph convolutions produce neighborhood representations of different orders. Lower-order representations emphasize direct interactions, while higher-order representations can incorporate collaborative information over longer distances but may also introduce noise. Therefore, this embodiment does not employ conventional average fusion. Instead, after calculating the weights of each layer through a layer attention mechanism, it performs weighted fusion of the multi-layer collaborative features of the application context and knowledge units separately to form the final high-order collaborative feature vector, as shown in the following equation: In the formula, Application scenario C mThe weights of the l-th layer features in the final collaborative representation This represents the feature vector of its l-th layer. Represents knowledge unit K n The weights of the l-th layer features in the final collaborative representation W represents its l-th layer feature vector. atte Let L be the learnable layer attention transformation weight vector, and L be the number of graph convolution layers. e Cm( syn) e Kn (syn) represents the application context collaborative feature vector and the knowledge unit collaborative feature vector, respectively.

[0034] This process allows the model to retain direct interaction information, while utilizing higher-order collaborative information and reducing the interference of irrelevant distant neighbor nodes on the recommendation results.

[0035] For LightGCN, the training data comes from the historical operation logs of the enterprise production management system. The logs record the browsing, liking, collection, and downloading behaviors of application scenarios on knowledge items or knowledge units, and assign interaction values ​​of 0, 0.4, 0.6, 0.8, and 1.0 to construct the application scenario-knowledge unit bipartite graph and interaction matrix.

[0036] During training, the initial embeddings of application contexts and knowledge units are randomly generated from a standard normal distribution and used as trainable parameters. LightGCN performs multi-layer neighborhood propagation according to interaction edges and behavior weights, and then fuses the outputs of different layers through layer attention. The overall model uses BPR loss, which makes the prediction score of positively interacted samples higher than that of non-interactive samples generated by negative sampling, and updates the parameters through backpropagation and Adam optimizer until the loss converges.

[0037] However, in some cases, if the context object to be recommended and the application context object do not correspond, the similarity between the context object to be recommended and the application context object is calculated. When the similarity is greater than the first threshold, the context object to be recommended is regarded as not corresponding to the application context object and step S3 is executed. When the similarity is less than the first threshold, the top k application context objects with the highest similarity to the context object to be recommended are obtained through the KNN model. The similarity between the context object to be recommended and the top k application context objects is combined with the application context collaborative feature vectors of the top k application context objects to perform weighted aggregation, and a comprehensive feature vector of the context object to be recommended is obtained. The operation of S3 is then executed using this comprehensive feature vector.

[0038] Specifically, the similarity calculation formula is as follows: In the formula, Sim(C new C m () indicates similarity; This represents the i-th new application scenario; This represents the i-th application scenario in the interaction matrix; This represents the degree of matching between the current context and the historical context on the i-th basic attribute; This represents the range of values ​​for the i-th attribute in a data type attribute, and is used to normalize numerical differences.

[0039] If a certain historical application scenario exists in the historical interaction record to make it full Then determine the current input application context. Historical application context If they are from the same application context, then determine the current input application context. For the new application scenario, an improved KNN model is used to find the top-k historical application scenarios similar to the new application scenario, forming a nearest neighbor set, as shown in the following formula: In the formula, A set of semantic vectors representing all historical application scenarios in the system. The hyperparameter represents the ranking of a historical application context's similarity score within the entire set. The size of the reference neighborhood was controlled.

[0040] After locking down the nearest neighbor set, the similarity between the new application scenario and each nearest neighbor scenario is used as a weight to perform weighted aggregation of the collaborative feature vectors learned by these historical application scenarios in LightGCN, resulting in a comprehensive feature vector for the new application scenario, as shown in the following formula: In the formula, This represents the comprehensive feature vector of the new application scenario obtained through weighted aggregation; Indicates a new application scenario; Let j represent the j-th nearest neighbor historical application scenario, that is, the historical application scenario with the top k similarity to the new application scenario; This represents the collaborative feature vector that the j-th nearest neighbor historical application scenario has learned in LightGCN.

[0041] S3. For the correspondence between the context object to be recommended and the application context object, extract the application context collaborative feature vector, construct a knowledge graph embedding task based on TransD to reduce the computational burden caused by high-dimensional sparse representation, and obtain the initial semantic representation of entities and relations. Based on KGAT, add the local structural relations and high-order neighborhood semantics of the knowledge graph to the knowledge item representation. Based on the hierarchical feature representation learning architecture, perform weighted aggregation on multiple entities to obtain the semantic feature vector of the knowledge unit. For a context object to be recommended, it can typically be matched with application context objects in historical data. In this case, the contextual collaborative feature vector of the application context object that matches the context object to be recommended is extracted and used as the feature vector of the context object to be recommended.

[0042] When the context object to be recommended cannot be matched with the application context object in the historical data, the feature vector of the context object to be recommended is obtained by referring to the method in S2.

[0043] The knowledge graph embedding task is then constructed using the contextual collaborative feature vectors and the TransD model, as detailed below: TransD models are widely used in knowledge graph embedding tasks, employing methods such as... Figure 3 The TransD model shown introduces independent projection vectors for entities and relations, and dynamically constructs a mapping matrix based on these two. In the TransD model, for any triple (head entity, relation, tail entity) in the nuclear fuel manufacturing process equipment knowledge graph, let the semantic representation vectors of the head entity and tail entity in the entity space be respectively... The semantic representation vector in the relation space is The projection vectors corresponding to the triplet are respectively and Then construct head entity mapping matrices for the head entity and tail entity respectively. Tail entity mapping matrix (in This is used to transform entity features into a relation space. The corresponding mathematical expression of the mapping matrix is ​​shown below: In the above formula, Represents the identity matrix.

[0044] After the mapping matrix is ​​constructed, the head entity and the tail entity are mapped to the target relation space respectively, and the calculation method is shown in the following formula: , In the formula, e h e t e represents the semantic representation vector of the head entity h and the tail entity t in the entity space; r M represents the semantic representation vector of relation r in the relation space; rh M rt These represent the dynamic mapping matrices used to map the head and tail entities to the relation space r, respectively; e h⊥ e t⊥These represent the mapping vectors of the head entity and the tail entity in the semantic space of relation r, respectively.

[0045] In the relation space, TransD measures the semantic reasonableness of triples through a scoring function, as shown in the following equation: In the formula, f r (e h ,e t ) represents the score of the triple (h,r,t) under relation r; ||·||2 represents the L2 norm, and the square represents the square of the Euclidean distance; the negative sign indicates that the smaller the distance, the higher the score.

[0046] The model training employs an interval-based ranking loss function to ensure that real triples outperform fake triples in scoring. The optimization objective is shown in the following equation: In the formula, B represents the set of valid positive triplets in the knowledge graph; B′ represents the set of negative triplets constructed by randomly replacing the head or tail entity in the same concept class; (h,r,t) represents the positive triplet; (h′,r,t′) represents the corresponding negative triplet; L represents the loss function; γ is the boundary value that controls the score difference between positive and negative examples, usually taken as 1; express Operation. When the difference between the scores of positive and negative examples exceeds the set boundary value. When the time comes, the model stops updating and outputs the corresponding values. , and .

[0047] TransD obtains the initial semantic embeddings of entities and relations. In order to further explore the potential structural relationships in the knowledge graph, the Knowledge Graph Attention Network (KGAT) mechanism is adopted to further aggregate the attention of the neighborhood semantics of knowledge items.

[0048] First, the correlation score between the target node and its neighbors is calculated using the following formula: In the formula, and Represents the learnable weight parameter matrix, and tanh is the hyperbolic tangent activation function. These represent the initial feature vectors of the target node and its neighboring nodes, respectively. This represents the feature vector of the corresponding relationship. After obtaining the relevance score, the Softmax function needs to be used to analyze all local neighbors of the target node. Weight normalization is performed within the formula shown below: In the formula, s(h,r,t) represents the normalized attention weights corresponding to relation r and neighbor node t within the neighbor set of target node h; s(h,r,t) represents the relevance score; exp(·) represents the exponential function; N(h) represents the set of all local neighbors of target node h.

[0049] Not all neighbors in the graph have equal reference value, so a relevance score threshold is set to retain only those with a relevance score of not less than [a certain threshold]. The neighboring nodes constitute a high-value neighbor subset. The attention weights are then renormalized. Based on the optimized attention weights, the target node aggregates the features of high-value neighbors to obtain a comprehensive neighborhood feature vector, as shown in the following equation: In the formula, : The neighborhood comprehensive feature vector obtained by weighted summation of all high-value neighbors of target node h; N*(h) represents the high-value neighborhood subset obtained by retaining neighbors s(h,r,t)≥δ from N(h).

[0050] To determine the relevance score threshold, this invention employs univariate sensitivity analysis to determine the threshold δ, rather than directly specifying it based on experience. Keeping other parameters constant, δ is tested at 0.1 intervals within the range of 0 to 1, and HR@10 and NDCG@10 are compared. The results show that the model performs best when δ=0.4; therefore, this study uses 0.4. The decision logic is: when δ<0.4, more low-relevance neighbors are retained, easily introducing noise; when δ>0.4, the filtering is too strict, pruning useful information.

[0051] After neighborhood aggregation is completed, a bidirectional interactive aggregator is used to deeply fuse the target node's own features with the comprehensive features of its neighborhood. The calculation formula is shown below: In the formula, and These are all network parameters that need to be optimized. This represents the element-wise multiplication operation of a vector.

[0052] After multiple layers of propagation, to avoid dimensionality expansion caused by direct concatenation, a hierarchical attention mechanism is used to fuse the outputs of different layers, resulting in the final semantic feature vector of the target node, as shown below: In the formula, These are the learnable hierarchical attention weight vectors in the model. The level traversal variable is introduced to solve the normalized denominator. L* is the level of KGAT.

[0053] A knowledge unit consists of multiple knowledge items, each playing a different role within the unit. Some knowledge items also contain multiple knowledge entities. The overall semantic feature vector of a knowledge unit can be expressed using the semantic feature vectors of each knowledge item within it. A hierarchical feature representation learning architecture is employed, combining intra-group attention-weighted aggregation, inter-group feature concatenation, and multilayer perceptron fusion. For an entity set with multiple entities, the semantic feature vectors of entities of the same type are first aggregated using attention-weighted aggregation to represent the overall semantic features of that type of entity. Then, inter-group feature fusion based on entity types is performed to obtain the knowledge unit. semantic feature vector .

[0054] Intra-group feature aggregation: Taking the set of process parameters in a process task as an example, using process task and equipment features as global context information, for the first group feature aggregation... Each process parameter node has its attention weight. The calculation formula is as follows: In the formula, This indicates a splicing operation. The feature transformation weight matrix, This represents a learnable attention parameter vector. After obtaining the attention weights, the process parameter terms are linearly weighted and aggregated to obtain the overall semantic features of this type of entity set, as shown in the following formula: Similarly, the same attention mechanism is used to calculate the overall semantic feature vector of the process equipment reference knowledge set, denoted as... .

[0055] Inter-group feature fusion: Taking industrial tasks as an example, the overall semantics of the process task, process equipment, and process parameters, along with the overall semantics of the process equipment reference knowledge, are concatenated in a fixed order to form a joint feature vector of knowledge units, as shown in the formula below: In the formula, This represents a vector concatenation operation, the concatenated vector... Dimensional expansion .

[0056] Finally, the joint vector is input into a multilayer perceptron for nonlinear feature crossing and dimensionality compression, outputting a global semantic feature vector for the process task knowledge unit, as shown in the following formula: in, and These are the feature transformation weight matrices for the first and second fully connected layers, respectively. and is the corresponding bias vector, and ReLU is a non-linear activation function.

[0057] S4. Using balanced hyperparameters, the collaborative feature vector and semantic feature vector of knowledge units are fused to obtain a comprehensive feature vector for knowledge units. The inner product of the application context collaborative feature vector and the comprehensive feature vector of knowledge units is calculated, and the inner product is transformed into a matching degree using a sigmoid non-linear activation function. Knowledge units corresponding to the comprehensive feature vectors of knowledge units with higher matching degrees are selected as recommended knowledge. The collaborative feature vector of a knowledge unit reflects its behavioral preferences in historical application contexts, while the semantic feature vector reflects its professional structural meaning within the knowledge graph. To utilize both simultaneously, this invention introduces a balancing hyperparameter to linearly balance and fuse the collaborative and semantic feature vectors of knowledge units, resulting in a comprehensive feature vector for the knowledge unit. The calculation formula is shown below: In the formula, e* Kn Represents knowledge unit K n The comprehensive feature vector; e Kn (sem) represents the semantic feature vector of a knowledge unit; e Kn (syn) represents the collaborative feature vector of knowledge units; β is a balancing hyperparameter, usually restricted to [0,1]; 1-β is the weight of the collaborative features. The larger β is, the more the model relies on the semantics of the knowledge graph; the smaller β is, the more the model relies on historical interaction and collaborative signals.

[0058] When the balance parameter is large, the model relies more on high-order semantic rules in the knowledge graph; when the balance parameter is small, the model focuses more on collaborative filtering signals in historical interactions.

[0059] After obtaining the comprehensive feature vector of a knowledge unit, the system maps the feature vector of the current application context and the comprehensive feature vector of the candidate knowledge unit into the same latent space, and calculates the matching degree between the two through the inner product. The larger the inner product, the more likely the knowledge unit is to meet the requirements of the current task. The calculation formula is as follows: In the formula, CK Indicates the current application scenario C new With candidate knowledge unit K n The original matching score; e CnewThis represents the comprehensive feature vector of the current new application scenario. The superscript T indicates transpose, and the multiplication of two vectors represents the inner product.

[0060] Subsequently, the inner product value is converted into a probabilistic matching degree using the sigmoid nonlinear activation function, as shown in the following formula: Md represents the matching degree between the current application context and a candidate knowledge unit after Sigmoid transformation.

[0061] Based on the above steps, the system can generate comparable matching scores for different candidate knowledge units, sort the candidate knowledge units in descending order, and form a Top-N recommendation ranking.

[0062] To facilitate a better understanding of the methods of the embodiments of the present invention by those skilled in the art, specific test examples are given below.

[0063] Test Example 1 The experimental dataset consists of historical operation logs from the enterprise production management system and a constructed knowledge graph of nuclear fuel manufacturing process equipment. The "application context-knowledge unit" pairs in the historical interaction records serve as implicit positive feedback samples, and negative samples are generated through negative sampling. The statistical results are shown in Table 3.

[0064] Table 3 Initial Experimental Data Statistics In the experiment, 10% of the application scenarios were randomly selected as new application scenarios to test the expressive ability of the KNN module in the cold start scenario; the remaining historical application scenarios were divided into training and test sets in an 8:2 ratio for model training and recommendation effect evaluation.

[0065] The training parameters of the model in this embodiment are shown in Table 4. BPR-MF, NGCF, LightGCN and KGAT are compared. The experiments use the same training rounds, optimizer, learning rate, regularization coefficient and embedding dimension to ensure consistency in the comparison between different recommendation models.

[0066] Table 4 Model Parameter Table The final test results are shown in Table 5.

[0067] Table 5 Comparison of Evaluation Results for Each Model The meanings of each indicator in Table 5 are as follows: HR@5, HR@10, HR@20: Top 5, Top 10, and Top 20 hit rates, used to measure whether the target knowledge unit that has actually been interacted with appears in the top N recommendation results given by the model. The higher the value, the better.

[0068] NDCG@5, NDCG@10, NDCG@20: The top 5, top 10, and top 20 normalized loss cumulative gains. It not only determines whether a hit has occurred, but also considers the position of the target knowledge unit in the recommendation list. The higher the ranking, the greater the contribution, and the higher the value, the better.

[0069] Response time (s): The time required for the model to complete one recommendation calculation and generate a recommendation list, measured in seconds. A lower value usually indicates higher online response efficiency.

[0070] As shown in Table 5, the model in this embodiment achieves optimal results on metrics such as HR@5, NDCG@5, HR@10, NDCG@10, HR@20, and NDCG@20. Compared to methods that rely solely on collaborative filtering or knowledge graph semantic aggregation, the model in this embodiment improves the hit rate and ranking quality of the recommendation list in terms of real knowledge requirements through the fusion of "LightGCN collaborative features + KNN novel contextual representation + TransD / KGAT semantic features".

[0071] Test Example 2 A process engineer with employee number 2156 was selected as the current user. This user belongs to the Process Department, specializes in Materials Science and Engineering, and has core skills in process design. The current input task scenario is "high-temperature sintering process," and the system needs to proactively recommend relevant process equipment, process parameters, and reference knowledge around this task. The recommendation results of the method in this embodiment are shown in Table 6.

[0072] Table 6 Recommended Results of the Method in this Embodiment To further illustrate the recommendation advantages of the method in this embodiment, it is compared with the "application context pre-filtering" strategy. The application context pre-filtering strategy only makes recommendations within the set of knowledge units that have interacted with similar historical contexts retrieved by KNN, and the results are shown in Table 7.

[0073] Table 7 Recommendation Results of Application Context Pre-filtering Strategies The recommendation results that rely solely on LightGCN collaborative features without considering the semantic features of knowledge units are used as a comparison. The final results are shown in Table 8.

[0074] Table 8. Recommendation results relying solely on LightGCN collaborative features. As shown in Table 6, under the task scenario of "high-temperature sintering process", the recommended results are concentrated on the "green sintering" process task, and further related to equipment such as continuous push-boat type hydrogen-protected high-temperature sintering furnace, vacuum degreasing-high-temperature sintering integrated furnace, and walking beam high-temperature hydrogen sintering furnace. At the same time, key process parameters such as sintering temperature, push-boat speed, hydrogen flow rate, atmosphere dew point, and isothermal time are given, as well as relevant operation and maintenance manuals, technical specifications and process control procedures, indicating that the recommended results have clear engineering applicability.

[0075] Comparing Tables 6 and 7, it can be seen that while the context-based pre-filtering strategy can recommend some highly relevant knowledge units, its candidate range is limited by the historical interaction set, easily overlooking knowledge units with fewer historical interactions but higher semantic relevance in the knowledge graph. The model in this paper calculates the comprehensive matching degree for the entire set of knowledge units, thus leveraging entity relationships in the knowledge graph to supplement potential relevant knowledge, making the recommendation results more complete. Comparing Tables 6 and 8, it can be seen that when using only collaborative features, the model is more susceptible to the influence of historical interaction frequency, potentially prioritizing knowledge units with high interaction frequency but insufficient semantic matching with the current task. By integrating semantic features from the knowledge graph, the system can more accurately understand the professional connections between "high-temperature sintering process" and sintering furnace equipment, sintering parameters, and reference documents, thereby improving the task adaptability and interpretability of the recommendation results.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks, characterized in that, Includes the following steps: S1. Based on historical data, construct knowledge units and application context objects for process equipment, and construct an interaction matrix by combining the interaction between knowledge units and application context objects; the application context objects include user contexts and task contexts. S2. Based on the interaction matrix and the LightGCN model based on the layer attention mechanism, update the application context objects and knowledge units and calculate the weight of each layer. Perform weighted fusion of the output of each layer to obtain the application context collaborative feature vector and the knowledge unit collaborative feature vector. S3. For the correspondence between the context object to be recommended and the application context object, extract the application context collaborative feature vector, construct a knowledge graph embedding task based on TransD to reduce the computational burden caused by high-dimensional sparse representation, and obtain the initial semantic representation of entities and relations. Based on KGAT, add the local structural relations and high-order neighborhood semantics of the knowledge graph to the knowledge item representation. Based on the hierarchical feature representation learning architecture, perform weighted aggregation on multiple entities to obtain the semantic feature vector of the knowledge unit. S4. The knowledge unit collaborative feature vector and the knowledge unit semantic feature vector are fused using balanced hyperparameters to obtain the knowledge unit comprehensive feature vector. The inner product of the application context collaborative feature vector and the knowledge unit comprehensive feature vector is calculated, and the inner product is converted into a matching degree using the sigmoid nonlinear activation function. The knowledge units corresponding to the comprehensive feature vectors of the knowledge units with the highest matching degree (TOP-N) are selected as recommended knowledge.

2. The knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks according to claim 1, characterized in that, The knowledge units include manufacturing process equipment, manufacturing tasks, equipment failures, process equipment reference knowledge, process parameters, projects, organizational units, and professional fields; the user context is used to describe the basic attributes and ability characteristics of the personnel to be recommended; the task context is used to describe the task background corresponding to the knowledge application process.

3. The knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks according to claim 1, characterized in that, In S2, different interaction situations are scored to obtain interaction values, and an interaction matrix is ​​constructed based on the interaction values. The interaction situations include no interaction, browsing, favorite, and download in order of interaction degree. The higher the interaction degree, the higher the interaction value.

4. The knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks according to claim 1, characterized in that, S2 includes the following steps: S21. Transform the knowledge unit and application context object into a knowledge unit embedding vector and an application context embedding vector, respectively; S22. Establish a LightGCN model based on layer attention mechanism. Embed the knowledge unit embedding vector and the application context embedding vector into the initial layer of the LightGCN model and perform multi-layer neighborhood propagation. The application context embedding vector is updated by aggregating adjacent knowledge units, and the knowledge unit embedding vector is updated by aggregating adjacent application contexts. The updated knowledge unit embedding vector and the updated application context embedding vector are used to calculate the weight of each layer through the layer attention mechanism. S23. Based on the weights of each layer, the updated application context embedding vector and the updated knowledge unit embedding vector output by each layer are weighted and fused to obtain the application context collaborative feature vector and the knowledge unit collaborative feature vector.

5. The knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks according to claim 1, characterized in that, S2 further includes the following steps: If the context object to be recommended and the application context object do not correspond, calculate the similarity between the context object to be recommended and the application context object. When the similarity is greater than the first threshold, the context object to be recommended is regarded as not corresponding to the application context object and the steps of S3 are executed. When the similarity is less than the first threshold, the top k application context objects with the highest similarity to the context object to be recommended are obtained through the KNN model. The similarity between the context object to be recommended and the top k application context objects is combined with the application context collaborative feature vectors of the top k application context objects to perform weighted aggregation to obtain the comprehensive feature vector of the context object to be recommended. The operation of S3 is then executed using this comprehensive feature vector.

6. The knowledge recommendation method integrating knowledge graphs and graph convolutional neural networks according to claim 1, characterized in that, In S3, the weighted aggregation includes the following steps: First, the semantic feature vectors of entities of the same type are weighted by attention to obtain a first aggregate vector. Then, the aggregate vectors of different entities are concatenated and nonlinear feature crossing and dimensional compression are performed using a multilayer perceptron to obtain the semantic feature vector of the knowledge unit.

Images