Manufacturing service recommendation method based on knowledge graph and graph convolutional network

Abstract

The invention provides a manufacturing service recommendation method based on a knowledge graph and a graph convolutional network, which fully excavates information contained in manufacturing services and tasks in a manufacturing service-task information knowledge graph, and realizes feature representation of the manufacturing services and the tasks through the graph convolutional network. In the process of aggregating manufacturing service neighbor information and task neighbor information, neighbors are selected according to degree centrality of nodes, neighbor information is selectively aggregated according to relation scores and self-attention scores of manufacturing service neighbor nodes, noise is reduced, manufacturing service representation is generated, and task neighbor information is aggregated by using a self-attention mechanism. The problem that the recommendation result has limitation due to the fact that modeling representation at the other end is insufficient is effectively solved, and the accuracy of the recommendation result is improved.

Description

[Technical field]

[0001] The present invention relates to the field of service recommendation technology, and in particular to a manufacturing service recommendation method based on knowledge graph and graph convolutional network. [Background technology]

[0002] Service-oriented manufacturing is a new manufacturing model that has emerged in the transformation of traditional manufacturing to service-oriented manufacturing. This model relies on the manufacturing service platform. The manufacturing service providers on the platform will encapsulate their own manufacturing resources and capabilities as manufacturing services and upload them to the platform. The platform reviews each manufacturing service and includes it in the service resource pool. Users with manufacturing needs (task demanders) use the manufacturing service platform to publish tasks. After the platform responds to the task, it will push the manufacturing services selected from the resource pool to the task demander for selection. However, with the deepening of the transformation of the manufacturing industry and the expansion of the manufacturing scale, the number of manufacturing service providers on the manufacturing service platform has increased sharply, and a large number of manufacturing services are stored in the service resource pool. In this case, the platform cannot select appropriate manufacturing services for push in a timely manner, and the task demand cannot select the messy manufacturing service information, resulting in information overload.

[0003] Although the research and application of traditional recommendation algorithms in manufacturing service platforms have achieved certain results, the problems of cold start and insufficient information utilization still affect the accuracy of recommendation results.

[0004] Although incorporating knowledge graphs as auxiliary information into recommendation methods can achieve better recommendation performance, a large amount of information irrelevant to the recommendation will also introduce noise problems. At the same time, most recommendation methods based on knowledge graphs have the problem of insufficient modeling representation. Specifically, when capturing knowledge associations on the graph, knowledge selection and aggregation are not distinguished, which introduces some noise. Secondly, most recommendation methods only aggregate information on manufacturing services in the knowledge graph to enrich the vector representation of the manufacturing service side, and rarely consider the task information in the knowledge graph, resulting in insufficient modeling representation of the task information, which leads to certain limitations in the recommendation results. [Summary of the invention]

[0005] In order to solve the above shortcomings, the present invention proposes a manufacturing service recommendation method based on knowledge graph and graph convolutional network. In the manufacturing service-task information knowledge graph, the information contained in the manufacturing service and task is fully mined, and the characteristic representation of the manufacturing service and task is realized through the graph convolutional network. In the process of aggregating the manufacturing service neighbor information and the task neighbor information, the neighbors are selected according to the degree centrality of the node instead of random sampling. The neighbor information is selectively aggregated according to the relationship score and self-attention score of the manufacturing service neighbor node, the noise is reduced, and the manufacturing service representation is generated. The self-attention mechanism is used to aggregate the task neighbor information and generate the task representation, which effectively solves the problem that the recommendation results are limited due to insufficient modeling representation on the other end.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A manufacturing service recommendation method based on knowledge graph and graph convolutional network, characterized by comprising:

[0008] Step S1: construct a manufacturing service-task information knowledge graph G through manufacturing service history data and task data, and perform data processing and information reading on the manufacturing service-task information knowledge graph G;

[0009] Step S2: Calculate the degree centrality DC of each node in G, sample a fixed number of neighbors according to the calculated results, and construct the adjacency matrix of entities and relationships.

[0010] Step S3: Input the data obtained in steps S1 and S2 into the manufacturing service neighborhood information representation unit, and calculate the manufacturing service relationship score through preference propagation And normalize it, according to the normalized score Get the weighted neighbor vector Then, the self-attention mechanism is used on the processed neighbor vectors to obtain the linear combination of the manufacturing service neighborhood representation vectors.

[0011] Step S4: Input the data obtained in step S1 and step S2 into the task neighborhood information representation unit, and obtain a weighted task neighborhood representation vector according to the self-attention mechanism.

[0012] Step S5: Input the linear combination of the manufacturing service neighborhood representation vector obtained in step S3 and step S4, the task neighbor representation vector, and the manufacturing service and task initial vectors into the graph convolutional neural network model, and use the summation and aggregation method to extract features to obtain the final representation vector of the manufacturing service and task.

[0013] Step S6: Apply the inner product to the final representation vector of the manufacturing service and task obtained in step S5 to predict the interaction probability, and make recommendations based on the interaction probability.

[0014] Optionally, the manufacturing service-task information knowledge graph G in step S1 includes manufacturing service information and task information; the manufacturing service information includes manufacturing service capabilities, historical tasks, historical task objects, and the industry neighborhood to which they belong; the task information includes the manufacturing service capabilities required by the task, the task object, the industry neighborhood of the task, etc. The read information includes the number of manufacturing services z, the number of tasks, the number of entities and relations r contained in G, and the training set, validation set, and test set divided in proportion.

[0015] Manufacturing service relationship score in step S3 Calculate using the inner product function g:R d ×R d →R calculates the score, the formula is as follows:

[0016]

[0017] In the formula, z0 represents the initial vector of manufacturing service, r0 represents the initial vector of relationship;

[0018] Normalized score The calculation is as follows:

[0019]

[0020] Where N(z) is the neighbor set of manufacturing service z after degree centrality processing;

[0021] Neighbor vector after normalized score calculation The calculation is as follows:

[0022]

[0023] where e z The initial vector of neighbors for manufacturing service z is collectively referred to;

[0024] The linear combination calculation formula of the weighted manufacturing service neighborhood representation vector is as follows:

[0025]

[0026] in is the self-attention score and d is the vector dimension.

[0027] The linear combination of the weighted task neighborhood representation vectors in step S4 is calculated as follows:

[0028]

[0029] in is the self-attention score of the neighbor vector of task t, e tis the initial neighbor vector of task t, d is the vector dimension; N(t) is the neighbor set of the target node of task t after degree centrality processing.

[0030] Furthermore, the calculation method of the final representation vector of the manufacturing service and task obtained by summing and aggregating in step S5 is as follows:

[0031] agg z =σ(W Z ·(Z0+u k s(z) )+b z )

[0032] agg t =σ(W t ·(Z t +u k s(t) )+b t )

[0033] where agg z 、agg t are the final representation vectors of manufacturing services and tasks, σ is a nonlinear function such as ReLU, and W Z , W t and b z 、b t They are the transformation weights of manufacturing service aggregation and task aggregation and the deviations of manufacturing service aggregation and task aggregation, u k s(z) is the k-layer neighborhood combination vector of manufacturing service z, u k s(t) is the k-layer neighborhood combination vector of task t.

[0034] Optionally, step S6 recommends the top K manufacturing services according to the interaction probability, namely, Top-K recommendation.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes a manufacturing service recommendation method based on a knowledge graph and a graph convolutional network. When selecting a neighborhood, this method divides the importance of each neighbor in the neighborhood by a degree centrality index that measures the importance of the node, and selects according to the degree centrality, thereby reducing noise. At the same time, a task neighborhood information representation unit is added to the existing service-side neighborhood information representation unit, and the self-attention mechanism is combined to selectively and biasedly aggregate the manufacturing service neighbor information and the task neighbor information, and the feature representation of the manufacturing service and the task is realized through the graph convolutional network. The present invention can not only fully learn the structural information and semantic information of the knowledge graph, but also realize a two-end modeling representation, thereby improving the accuracy of the recommendation results.

Brief Description of the Drawings

[0036] Figure 1 It is a schematic flow chart of the main steps of the method of the present invention.

[0037] Figure 2 It is a schematic diagram of a knowledge graph of an embodiment of the method of the present invention.

[0038] Figure 3 Schematic diagram of neighbor sampling of degree centrality in an embodiment of the method of the present invention. [Specific implementation method]

[0039] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with embodiments and drawings. It should be understood that the exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined in any way as long as they do not conflict with each other.

[0040] See also Figure 1-3 The present invention provides a manufacturing service recommendation method based on knowledge graph and graph convolutional network, which is characterized by comprising:

[0041] Step S1: construct a manufacturing service-task information knowledge graph G through manufacturing service history data and historical task data; and read data information from the manufacturing service-task information knowledge graph G;

[0042] In an exemplary embodiment, as shown in the attached Figure 2 As shown, the manufacturing service-task information knowledge graph G in step 1 includes manufacturing service information and task information; the manufacturing service information includes manufacturing service skills, historical service task status, historical tasks, historical task objects and the industry neighborhood to which they belong; the task information includes the manufacturing service skills required for the task, the task objects and the task industry neighborhood.

[0043] Optionally, when a manufacturing service and a task have the same information, they can point to the same neighborhood entity in the manufacturing service-task information knowledge graph G.

[0044] In one embodiment, as shown in Table 1, the data information reading includes processing and reading the number of manufacturing services, the number of tasks, the number of triples, the number of neighborhoods, the number of skill types, the number of relationship types, and the relationship information between each data in the knowledge graph, dividing the data information according to the ratio set according to actual needs, and using them for training sets, verification sets, and test sets for model training, verification, and testing, respectively.

[0045] Table 1 Information reading data

[0046]

[0047] Step S2: Calculate the degree centrality DC of each node in the manufacturing service-task information knowledge graph G, and perform a set number of neighbor sampling according to the calculated result;

[0048] In an exemplary embodiment, the DC calculation formula in the prior art may be used:

[0049]

[0050] In the formula, K e is the number of edges connected to node e or the degree of node e, N is the total number of nodes including node e and its connected nodes, and N-1 is the number of edges connecting node e to all other nodes.

[0051] Through the sampled neighbor relationship information, the adjacency matrix of entities and relationships is constructed, the node and relationship vectors are initialized, and an initialized low-dimensional vector is allocated to each entity and relationship according to actual needs.

[0052] Optionally, U services correspond to U rows of the matrix, and the first C neighbors are selected according to the degree centrality DC. The matrix is ​​[U×C], and the adjacency matrix of entities and relationships is constructed through software modules or programming.

[0053] In an exemplary embodiment, the target node selects a neighbor sample with a neighbor number of 2 as shown in the attached Figure 3 As shown, the two dotted lines represent two layers of information transmission, where DC i It is the sorted degree centrality, for example: DC1>DC2>DC3>DC4.

[0054] Step S3: Input the manufacturing service neighbor data obtained in steps S1-S2 into the manufacturing service neighborhood information representation unit, and calculate the manufacturing service relationship score through preference propagation And normalized, according to the normalized manufacturing service relationship score Get the weighted manufacturing service neighbor vector For the weighted manufacturing service neighbor vector Use the self-attention mechanism to obtain the linear combination μ of the manufacturing service neighborhood representation vector s(z) ;

[0055] The service relationship score is generated in step S3 The formula is as follows:

[0056]

[0057] In the formula, z0 represents the initial vector of manufacturing service, r0 represents the initial vector of relationship, and g:R d ×R d →R is the score calculated by the inner product function, and d is the vector dimension;

[0058] When the model is trained, an initialization vector is assigned to each entity and relationship based on the training.

[0059] Normalized manufacturing service relationship score The calculation is as follows:

[0060]

[0061] Where N(z) is the neighbor set of the manufacturing service z target node after degree centrality processing;

[0062] Weighted manufacturing service neighbor vector after normalized score calculation The calculation is as follows:

[0063]

[0064] where e z To make the initial vector of the neighbors of service z, the vector is read and vectorized in step S1;

[0065] The linear combination calculation formula of the manufacturing service neighborhood representation vector is as follows:

[0066]

[0067] in is the self-attention score of the neighborhood vector of service z, where d is the vector dimension.

[0068] Step S4: input the task neighbor data obtained in steps S1-S2 into the task neighborhood information representation unit, and obtain a weighted linear combination of task neighborhood representation vectors according to the self-attention mechanism;

[0069] The linear combination of the weighted task neighborhood representation vectors in step S4 is calculated as follows:

[0070]

[0071] in is the self-attention score of the neighbor vector of task t, e t is the initial neighbor vector of task t, which is the vectorized vector after being read in step S1, and d is the vector dimension; N(t) is the neighbor set of the target node of task t after degree centrality processing.

[0072] Step S5: Input the linear combination of the manufacturing service neighborhood representation vector obtained in steps S3-S4 and the manufacturing service initial vector into the graph convolutional neural network model, input the weighted task neighborhood representation vector linear combination and the task initial vector into the graph convolutional neural network model, and use the summation aggregation method to extract features to obtain the final representation vector of the manufacturing service and the final representation vector of the task;

[0073] Optionally, the graph convolution in step S5 is a 2-layer neural network structure, in which the calculation method of obtaining the final representation vector of the manufacturing service and task by summing and aggregating is as follows:

[0074] agg z =σ(W Z ·(Z0+u k s(z) )+b z )

[0075] agg t =σ(W t ·(Z t +u k s(t) )+b t )

[0076] where agg z The final representation vector for manufacturing services, agg t is the final representation vector of the task, σ is a nonlinear function such as ReLU, and W Z Transform weight for service aggregation, W t is the task aggregation transformation weight, b z Aggregate deviation for manufacturing services, b t is the task aggregation deviation, u k s(z) is the linear combination of the k-th manufacturing service neighborhood representation vectors of manufacturing service z, u k s(t) is the linear combination of the k-th layer weighted task neighborhood representation vectors of task t, z0 represents the initial vector of the manufacturing service, z t Represents the initial vector of the task.

[0077] Step S6: Predict the interaction probability by inner product of the final representation vector of the manufacturing service obtained in step S5 and the final representation vector of the task Recommend manufacturing services based on interaction probability;

[0078] Optionally, the top K manufacturing services are recommended according to the interaction probability, namely, Top-K recommendation, so as to obtain a reliable recommendation effect.

[0079] In step S6, the interaction probability between the manufacturing service and the task is calculated, and the top K manufacturing services are recommended to the task from high to low according to the interaction probability; K can be 1, 2, 5, 10, 20, 50, 100.

[0080] In order to verify the effect of the method of the present invention, the method of the present invention is compared with other conventional methods in terms of click-through rate (CTR) and Top-K recommendation.

[0081] The experimental environment of the method of the present invention is a 64-bit Windows 10 operating system, using an Intel (R) Core (TM) i5-9400 CPU, an NVIDIA GeForce GTX 1650 GPU, 24GB memory, and 4GB video memory; the number of model training rounds is 15, the entity embedding vector dimension is 32, the number of graph convolution layers is 2, the learning rate is 0.01, the regularization term coefficient is 0.0005, the batch size is 128, and the number of neighbor sampling is 20;

[0082] The methods used for comparison are: a recommendation method combining knowledge graph and graph convolution (KGCN), a knowledge graph convolution network and label smoothness regularization for recommendation systems (KGNN-LS), a recommendation method based on collaborative knowledge-aware attention network (CKAN), and a knowledge-aware graph convolution network method with collaborative guidance for personalized recommendation (CG-KGR); the click-through rate CTR uses AUC and F1 indicators, and Top-K (Top-20) uses Precision@20 and Recall@20 for comparison, where AUC is the area under the ROC curve; the F1 indicator is the harmonic mean of precision (Precision) and recall (Recall), which is used to comprehensively evaluate the classification performance of the model; Precision@20 refers to the proportion of actual positive samples in the top 20 samples with the highest probability in the model prediction results; Recall@20 refers to the proportion of the number of actual positive samples in the top 20 samples with the highest probability in the model prediction results to the number of all actual positive samples. This method has better results than conventional methods, and the comparison results are shown in Table 2.

[0083] Table 2 Comparison of click rate and Top-20

[0084]

[0085] The present invention proposes a manufacturing service recommendation method based on knowledge graph and graph convolutional network. When selecting the neighborhood, the degree centrality index for measuring the importance of nodes in network analysis is introduced to divide the importance of each neighbor in the neighborhood, and the neighbor is selected according to the degree centrality, thereby reducing noise, optimizing data quality, and improving input accuracy. On the basis of the existing service party neighborhood information representation unit, a task neighborhood information representation unit is added, and the graph convolutional network and the self-attention mechanism are combined to selectively and biasedly aggregate the manufacturing service neighbor information and the task neighborhood information, so as to fully learn the structural information and semantic information of the knowledge graph. By selectively and biasedly aggregating the manufacturing service neighbor information and the task neighborhood information, a dual-end modeling representation is realized, thereby improving the accuracy of the recommendation results.

[0086] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A manufacturing service recommendation method based on knowledge graph and graph convolutional network, characterized in that: The following steps are involved: Step S1: construct a manufacturing service-task information knowledge graph G through manufacturing service history data and historical task data; and read data information from the manufacturing service-task information knowledge graph G; Step S2: Calculate the degree centrality DC of each node in the manufacturing service-task information knowledge graph G, perform a set number of neighbor sampling according to the calculation result, construct the entity and relationship adjacency matrix through the sampling information, and initialize the node and relationship vectors; Step S3: Input the manufacturing service neighbor data obtained in steps S1-S2 into the manufacturing service neighborhood information representation unit, and calculate the manufacturing service relationship score through preference propagation And normalized, according to the normalized manufacturing service relationship score Get the weighted manufacturing service neighbor vector For the weighted manufacturing service neighbor vector Use the self-attention mechanism to obtain the linear combination μ of the manufacturing service neighborhood representation vector s(z) ; Step S4: input the task neighbor data obtained in steps S1-S2 into the task neighborhood information representation unit, and obtain a weighted linear combination of task neighborhood representation vectors according to the self-attention mechanism; The linear combination of the weighted task neighborhood representation vectors in step S4 is calculated as follows: in is the self-attention score of the neighbor vector of task t, e t is the initial vector of neighbors of task t, d is the vector dimension; N(t) is the neighbor set of the target node of task t after degree centrality processing; Step S5: Input the linear combination of the manufacturing service neighborhood representation vector obtained in steps S3-S4 and the manufacturing service initial vector into the graph convolutional neural network model, and input the weighted task neighborhood representation vector linear combination and the task initial vector into the graph convolutional neural network model, and use the summation aggregation method to extract features to obtain the final representation vector agg of the manufacturing service. z And the final representation vector agg of the task t ; Step S6: Predict the interaction probability by inner product of the final representation vector of the manufacturing service obtained in step S5 and the final representation vector of the task Manufacturing service recommendations based on interaction probability.

2. The manufacturing service recommendation method based on knowledge graph and graph convolutional network according to claim 1, characterized in that: The service relationship score is generated in step S3 The formula is as follows: In the formula, z0 represents the initial vector of manufacturing service, r0 represents the initial vector of relationship, and g:R d ×R d →R is the score calculated by the inner product function, and d is the vector dimension; Normalized manufacturing service relationship score The calculation is as follows: Where N(z) is the neighbor set of the manufacturing service z target node after degree centrality processing; Weighted manufacturing service neighbor vector after normalized score calculation The calculation is as follows: where e z The initial vector of neighbors for manufacturing service z; The linear combination calculation formula of the manufacturing service neighborhood representation vector is as follows: in is the self-attention score of the neighborhood vector of service z, where d is the vector dimension.

3. The manufacturing service recommendation method based on knowledge graph and graph convolutional network according to claim 2 is characterized by: The method of obtaining the final representation vector of the manufacturing service and task by summing and aggregating in step S5 is as follows: agg z =σ(W Z ·(Z0+u k s(z) )+b z ) agg t =σ(W t ·(Z t +u k s(t) )+b t ) where agg z The final representation vector for manufacturing services, agg t is the final representation vector of the task, σ is a nonlinear function such as ReLU, and W Z Transform weight for service aggregation, W t is the task aggregation transformation weight, b z Aggregate deviation for manufacturing services, b t is the task aggregation deviation, u k s(z) is the k-layer neighborhood combination vector of manufacturing service z, u k s(t) is the k-layer neighborhood combination vector of task t, z0 represents the initial vector of manufacturing service, z t Represents the initial vector of the task.

Images