Enterprise expert information recommendation method and device based on non-local graph perception

Abstract

This invention discloses a method and device for recommending enterprise expert information based on non-local graph perception. The method labels and defines expert information and project information to obtain a label space and label set. The expert social network is derived from the expert's basic information, research direction, and project associations. Label weights are calculated using PMI based on the expert information label set and the project basic information label set, respectively, to construct an expert association graph and an expert-project association graph. This is then input into a GNN to extract neighborhood aggregation features, and non-local attention is used to extract non-local node centroids to obtain mixed information features. Graph perception weights are calculated, compared with mixed information feature judgments, and input into a GCN using the new weights to obtain key feature representations. The method then selects and recommends appropriate experts through classification. This invention uses graph neural networks and graph convolutional neural networks combined with non-local attention and graph structure perception for classification, and is generally applicable to enterprises in selecting expert information through project matching to obtain optimal recommendations.

Description

Technical Field

[0001] The present invention belongs to the technical fields of graph neural networks, graph convolutional networks and enterprise expert recommendation, and in particular relates to a method and device for recommending enterprise expert information based on non-local graph perception. Background Art

[0002] Expert recommendation integrates and analyzes multiple perspectives, including experts' academic research fields and expert relationships, and presents selection lists to businesses and government departments in various formats. Experts' professional fields are usually not limited to a single discipline, and scientific and technological projects do not accurately recommend experts in specific fields. Expert data in the scientific and technological field involves many complex professional fields and terminology, which increases the difficulty of data processing, feature extraction, and text graph construction. Existing expert recommendation technologies only consider single features, ignoring the connections between features and the issue of expanding domains. They do not construct feature and relationship text graphs for multi-source information, and do not deeply explore the relationship features between entities. The accuracy of data processing and the reasonable extraction and construction of multi-feature text graphs directly affect the accuracy and application effectiveness of recommendation methods. The accuracy of traditional expert recommendation technologies needs to be improved. Summary of the Invention

[0003] Purpose of the invention: In response to the above problems, the present invention provides a method and device for recommending enterprise expert information based on non-local graph perception, which can effectively extract features of graph structure, mine hidden features, and update nodes based on graph perception, thereby solving the difficulty of finding corresponding experts in enterprises.

[0004] Technical solution: The present invention discloses a method for recommending enterprise expert information based on non-local graph perception, comprising the following steps:

[0005] Step 1: After deduplication, de-duplication, cleaning, and basic information division of the expert information dataset D and the scientific research project dataset D1, we obtain the expert basic information label dataset D2, the expert research direction information label dataset D3, and the expert-project information label dataset D4;

[0006] Step 2: Build an expert social network based on the expert basic information label dataset D2, the expert research direction information label dataset D3, and the expert-project information label dataset D4.

[0007] Step 3: Use PMI to construct the expert association graph and the expert-project association graph. By extracting the corresponding keywords from the statistical expert information label dataset D2, the project basic information dataset label D3, and the expert-project information label dataset D4, and combining them with the expert social network, the edge weights of the association graph are calculated to obtain the expert association graph G i =(V i ,E i ) and expert-item association graph Gj =(V j ,E j );

[0008] Step 4: Use GNN to extract local neighborhood aggregation features of graph nodes and use non-local attention to the expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) Extract the node centroid non-locally and obtain the mixed information feature e″ (l) ;

[0009] Step 5: Expert association graph G i =(V i1 ,E i1 ) and expert-item association graph G j =(V j1 ,E j1 ) Extract graph perception features, calculate graph perception weights based on the expert research direction association and project association in the expert social network The updated expert association graph and expert-item association graph are input into GCN, combined with the graph-aware weights. The average weight e″ of the neighbor set of the original node (l) And the original node weight is combined to get the new weight representation c i , use the new weight convolution to obtain feature representation, use the Softmax layer to calculate and select different experts, and finally recommend the appropriate expert.

[0010] Furthermore, the specific method of step 1 is:

[0011] Step 1.1: Define Text as a single expert text information set to be cleaned, define id1, content1, content2, content3, and content4 as the expert's serial number, basic information, research direction, number of published papers, and achievements, respectively, and satisfy the relationship Text = {id1, content1, content2, content3, content4};

[0012] Step 1.2: Define Text1 as the research project text to be cleaned information set, define id2, content5, content6, content7 as research serial number, research field, project expert and member, and expandable experts in the same field, and satisfy the relationship Text1 = {id, content5, content6, content7};

[0013] Step 1.3: Define D as the expert information data set to be cleaned, D1 as the scientific research project information data set to be cleaned, D={Text1, Text2,…, Text a ,…,Text len(D)}, Text a is the ath expert data to be cleaned in D, D1={Text11,Text12,…,Text1 b ,…,Text1 len(D1)}, Text1 b For the bth research project data to be cleaned in D1, remove duplicates and blanks from the text in the dataset and extract labels from the content in the information set to be cleaned;

[0014] Step 1.4: After data cleaning and label extraction of datasets D and D1, we can obtain expert information label dataset D2, scientific research project information label dataset D3 and expert-project information label dataset D4, where D2 = {T1, T2, ..., T c ,…,T len(D2)}, T c is the cth text data in D2, where variable c∈[1,len(D2)], D3={T11,T12,…,T1 d ,…,T1 len(D3)}, T1 d is the dth text data in D3, where variable d∈[1,len(D3)], D4={T21,T22,…,T1 e ,…,T2 len(D4)}, T2 e is the e-th text data in D4, where variable e∈[1,len(D4)];

[0015] Step 1.5: Define the expert association graph G i =(V i ,E i ), node v 1i ∈V i , side e 1i ∈E i , expert-item association graph G j =(V j ,E j ), node v 2j ∈V j , side e 2j ∈E j .

[0016] Furthermore, the specific method of step 2 is:

[0017] Step 2.1: Count the labels in the label datasets D2 and D3, and summarize the label datasets such as the expert's current employment unit and research direction, which is defined as D5;

[0018] Step 2.2: The single text information set in the label dataset D5 is Text2 = {id3, label1, label2, label3, label4}, where id3, label1, label2, label3, and label4 are the text information sequence number, the expert's current employment unit label, the research direction label, the label of whether they have participated in a joint scientific research project, and the intersection label of the published paper authors, respectively;

[0019] Step 2.3: Define the expert social network G u =(U,S), where S represents the edge and node u i ∈U i , Bian s i ∈S i If two experts a and b have an intersection, then node u i ∈U i , side e i ∈E i

[0020] Step 2.4: Define the sliding window a as 2;

[0021] Step 2.5: Define the loop variable l, assign l to 0, and start the loop;

[0022] Step 2.6: If l + 1 ≤ len(D2), proceed to step 2.7. Otherwise, jump to step 2.9, where len(D2) is the amount of data in dataset D2.

[0023] Step 2.7: Define experts a and b, where a∈D2 and b∈D2, and determine whether a and b have an intersection based on the expert’s detailed label data;

[0024] Step 2.8: If the experts have an intersection, create an edge s between experts a and b i ,

[0025] Step 2.9: Otherwise, experts a and b do not build an edge.

[0026] Furthermore, the specific method of step 3 is:

[0027] Step 3.1: Use a fixed-size sliding window to calculate word co-occurrence information across the text for D2, D3, and D4, and use PMI to calculate the connection weight between two words;

[0028] Step 3.2: Pre-set C win (i) and C win(i,j) sliding window threshold, C win is the total number of sliding windows in the dataset;

[0029] Step 3.3: Use and Calculate the proportion of a single label in the total as the basis for calculating the edge weight;

[0030] Step 3.4: Exploitation Calculate different types of edges and edge weights between words i and j;

[0031] Step 3.5: Determine whether the PMI value is positive. If it is a positive PMI value, the two words are determined to be related, otherwise they are not related.

[0032] Step 3.6: Use the weight partition threshold T calculated by PMI, where T≥0. If the PMI value between words i and j is greater than T, they are considered highly correlated. If 0≤PMI≤T, they are considered normally correlated. This is used to construct the expert association graph G. i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ).

[0033] Furthermore, the specific method of step 4 is:

[0034] Step 4.1: Expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) node v i ∈V i and node v j ∈V j , use GNN to extract local neighborhood features;;

[0035] Step 4.2: Definition and represents the initial user that can be found through embedding, u and i represent the serial numbers of experts and projects, and represents the embedding of expert u and item i,

[0036] Step 4.3: Use GNN to perform convolution to aggregate neighborhood features of nodes, using and and represents the aggregated neighborhood embedding at layer l;

[0037] Step 4.4: Use pooling technology to embed node neighborhood features. and The final neighborhood feature is expressed as and

[0038] Step 4.5: Use spatial aggregation under non-local conditions to extract non-local node centroids and use Perform K-Means clustering on the node embeddings, g ki represents the cluster membership of the node, For the centroid c k When g ki =1, then Assigned to cluster c k ;

[0039] Step 4.6: Extract the centroid features of non-local nodes. The vector obtained after clustering the l-th layer nodes is represented as

[0040] Step 4.7: Mix the local neighborhood features and non-local node centroid features to obtain and

[0041] Furthermore, the specific method of step 5 is:

[0042] Step 5.1: Associate the experts into the graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) is input into GCN for feature extraction, defined as matrix H;

[0043] Step 5.2: Define weights f i and f j are the weights between experts and between experts and projects, respectively. F is a constant used for normalization.

[0044] Step 5.3: Perform graph perception operations on each relationship type of the node to obtain the average aggregation value of the node N c (i) represents the neighborhood set of node i and relationship c, h j represents the hidden vector of node j, C represents the number of neighboring nodes of node i, and W c and b c represents the graph convolution parameters, represents the normalization factor;

[0045] Step 5.4: Define k as a loop variable and assign its initial value to 1, then start the loop.

[0046] Step 5.5: If k ≤ len(Max(D2, D3, D4)), go to step 5.6; otherwise, jump to step 5.9;

[0047] Step 5.6: Use weight c to calculate e″ (l) and Make judgments;

[0048] Step 5.7: If c ≥ e″ (l) ,but n1 is the number of members in the cluster, otherwise c2=c+e″ (l) ;

[0049] Step 5.8: If the result of step 5.7 is but n2 is the number of neighborhood sensing nodes, otherwise c4=c i +e c ;

[0050] Step 5.9: The new weighted result is input into GCN to obtain the new neighborhood node feature X;

[0051] Step 5.10: Input the enhanced matrix X into the Softmax layer, output the probabilistic recommendation results, and select appropriate experts for recommendation based on the results.

[0052] The present invention also discloses a device for recommending enterprise expert information based on non-local graph perception, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the above-mentioned method for recommending enterprise expert information based on non-local graph perception are executed.

[0053] Beneficial effects:

[0054] 1. The present invention constructs an expert association graph and an expert-project association graph. When constructing edges between labels, the co-occurrence information between labels is utilized so that each node not only carries its own information but also represents neighborhood characteristics. When recommending and matching enterprise experts and projects, the association graph can be used to deeply explore the hidden relationships between different labels, obtain features more comprehensively, and ultimately obtain more accurate classification probabilities.

[0055] 2. The present invention uses GNN to extract local neighborhood aggregation features of graph nodes from the association graph, mainly based on the key features of the expert association graph and supplemented by the key features of domain aggregation. The weights of the nodes in the expert-project association graph are judged and updated, and important nodes are strengthened and secondary nodes are weakened in the final classification, making the output results more hierarchical and clear.

[0056] 3. The present invention uses non-local attention on the association graph, extracts the non-local node centroid for weight screening and judgment, and makes the final graph structure weight judgment, making the enhancement effect more purposeful and accurate.

[0057] 4. The present invention adopts a graph perception method to extract the weights of the graph neighbor set, combines neighborhood features and non-local attention features to take average weights, perceives and obtains node hidden features to the greatest extent, and assigns important data information after perception to the original node, making the hidden relationship between recommendation results and captured labels more comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is the overall flow chart of the present invention;

[0059] Figure 2 Flowchart for data cleaning;

[0060] Figure 3 Flowchart for selecting basic datasets for building association graphs;

[0061] Figure 4 To build a flowchart of the association diagram;

[0062] Figure 5 Flowchart for obtaining mixed new features using GNN;

[0063] Figure 6 Experts recommend a flowchart for companies using GCN. DETAILED DESCRIPTION

[0064] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0065] The present invention discloses a method for recommending enterprise expert information based on non-local graph perception, see the attached Figure 1 To the attached Figure 6 , including the following steps:

[0066] Step 1: De-duplicate, remove, clean, and divide the expert information dataset D and the scientific research project dataset D1 into basic information label datasets D2, research direction information label dataset D3, and expert-project information label dataset D4.

[0067] Step 1.1: Define Text as a single expert text information set to be cleaned, define id1, content1, content2, content3, and content4 as the expert serial number, basic information, research direction, number of published papers, and achievements, respectively, and satisfy the relationship Text = {id1, content1, content2, content3, content4}.

[0068] Step 1.2: Define Text1 as the information set of the scientific research project text to be cleaned, define id2, content5, content6, content7 as the scientific research serial number, the research field to which it belongs, the project experts and members, and the scalable experts in the same field, and satisfy the relationship Text1 = {id, content5, content6, content7}.

[0069] Step 1.3: Define D as the expert information data set to be cleaned, D1 as the scientific research project information data set to be cleaned, D={Text1, Text2,…, Text a ,…,Text len(D)}, Text a is the ath expert data to be cleaned in D, D1={Text11,Text12,…,Text1 b ,…,Text1 len(D1)}, Text1 b For the bth scientific research project data to be cleaned in D1, the text in the data set is deduplicated and de-nulled, and labels are extracted for the content in the information set to be cleaned.

[0070] Step 1.4: After data cleaning and label extraction of datasets D and D1, we can obtain expert information label dataset D2, scientific research project information label dataset D3 and expert-project information label dataset D4, where D2 = {T1, T2, ..., T c ,…,T len(D2)}, T c is the cth text data in D2, where variable c∈[1,len(D2)], D3={T11,T12,…,T1 d ,…,T1 len(D3)}, T1 d is the dth text data in D3, where variable d∈[1,len(D3)], D4={T21,T22,…,T1 e ,…,T2 len(D4)}, T2 e is the e-th text data in D4, where the variable e∈[1,len(D4)].

[0071] Step 1.5: Define the expert association graph G i =(V i ,E i ), node v 1i ∈V i , side e 1i ∈E i , expert-item association graph G j =(V j ,E j ), node v 2j ∈V j , side e 2j ∈E j .

[0072] Step 2: Build an expert social network based on the expert basic information label dataset D2, the expert research direction information label dataset D3, and the expert-project information label dataset D4.

[0073] Step 2.1: Count the labels in the label datasets D2 and D3, and summarize the label datasets such as the expert's current employment unit, research direction, etc., which is defined as D5.

[0074] Step 2.2: The single text information set in the label dataset D5 is Text2 = {id3, label1, label2, label3, label4}, where id3, label1, label2, label3, and label4 are the text information serial number, the expert's current employment unit label, the research direction label, the label of whether they have jointly participated in scientific research projects, and the intersection label of the published paper authors.

[0075] Step 2.3: Define the expert social network G u =(U,S), where S represents the edge and node u i ∈U i , Bian s i ∈S i If two experts a and b have an intersection, then node u i ∈U i , side e i ∈E i .

[0076] Step 2.4: Define the sliding window a as 2.

[0077] Step 2.5: Define the loop variable l, assign l to 0, and start the loop.

[0078] Step 2.6: If l + 1 ≤ len(D2), proceed to step 2.7; otherwise, jump to step 2.9, where len(D2) is the amount of data in dataset D2.

[0079] Step 2.7: Define experts a and b, where a∈D2 and b∈D2, and determine whether a and b have an intersection based on the expert's detailed label data.

[0080] Step 2.8: If the experts have an intersection, create an edge s between experts a and b i .

[0081] Step 2.9: Otherwise, experts a and b do not build an edge.

[0082] Step 3: Use PMI to construct the expert association graph and the expert-project association graph. By extracting the corresponding keywords from the statistical expert information label dataset D2, the project basic information dataset label D3, and the expert-project information label dataset D4, and combining them with the expert social network, the edge weights of the association graph are calculated to obtain the expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ).

[0083] Step 3.1: Use a fixed-size sliding window to calculate word co-occurrence information of D2, D3 and D4 through the text, and use PMI to calculate the connection weight between two words.

[0084] Step 3.2: Pre-set C win (i) and C win (i,j) sliding window threshold, C win is the total number of sliding windows in the dataset.

[0085] Step 3.3: Use and Calculate the proportion of a single label in the total as the basis for calculating the edge weight.

[0086] Step 3.4: Exploitation Calculate the different types of edges and edge weights between words i and j.

[0087] Step 3.5: Determine whether the PMI value is positive. If it is a positive PMI value, the two words are determined to be related, otherwise they are not related.

[0088] Step 3.6: Use the weight partition threshold T calculated by PMI, where T≥0. If the PMI value between words i and j is greater than T, they are considered highly correlated. If 0≤PMI≤T, they are considered normally correlated. This is used to construct the expert association graph G. i =(V i ,E i ) and expert-item association graph G j =(Vj ,E j ).

[0089] Step 4: Use GNN to extract local neighborhood aggregation features of graph nodes and use non-local attention to the expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) Extract the node centroid non-locally and obtain the mixed information feature e″ (l) .

[0090] Step 4.1: Expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) node v i ∈V i and node v j ∈V j , using GNN to extract local neighborhood features.

[0091] Step 4.2: Definition and represents the initial user that can be found through embedding, u and i represent the serial numbers of experts and projects, and represents the embedding of expert u and item i,

[0092] Step 4.3: Use GNN to perform convolution to aggregate neighborhood features of nodes, using and and represents the aggregated neighborhood embedding at layer l.

[0093] Step 4.4: Use pooling technology to embed node neighborhood features. and The final neighborhood feature is expressed as and

[0094] Step 4.5: Use spatial aggregation under non-local conditions to extract non-local node centroids and use Perform K-Means clustering on the node embeddings, g ki represents the cluster membership of the node, For the centroid c k When g ki =1, then Assigned to cluster c k .

[0095] Step 4.6: Extract the centroid features of non-local nodes. The vector obtained after clustering the l-th layer nodes is represented as

[0096] Step 4.7: Mix the local neighborhood features and non-local node centroid features to obtain and

[0097] Step 5: Expert association graph G i =(V i1 ,E i1 ) and expert-item association graph G j =(V j1 ,E j1 ) Extract graph perception features, calculate graph perception weights based on the expert research direction association and project association in the expert social network The updated expert association graph and expert-item association graph are input into GCN, combined with the graph-aware weights. The average weight e″ of the neighbor set of the original node (l) And the original node weight is combined to get the new weight representation c i , use the new weight convolution to obtain feature representation, use the Softmax layer to calculate and select different experts, and finally recommend the appropriate expert.

[0098] Step 5.1: Associate the experts into the graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) is input into GCN for feature extraction and is defined as matrix H.

[0099] Step 5.2: Define weights f i and f j are the weights between experts and between experts and projects, respectively. F is a constant used for normalization.

[0100] Step 5.3: Perform graph perception operations on each relationship type of the node to obtain the average aggregation value of the node N c (i) represents the neighborhood set of node i and relationship c, h j represents the hidden vector of node j. C represents the number of neighboring nodes of node i, W c and b c represents the graph convolution parameters, represents the normalization factor.

[0101] Step 5.4: Define k as the loop variable, assign its initial value to 1, and start the loop.

[0102] Step 5.5: If k ≤ len(Max(D2, D3, D4)), proceed to Step 5.6; otherwise, jump to Step 5.9.

[0103] Step 5.6: Use weight c to calculate e″ (l) and Make a judgment.

[0104] Step 5.7: If c ≥ e″ (l) ,but n1 is the number of members in the cluster, otherwise c2=c+e″ (l) .

[0105] Step 5.8: If the result of step 5.7 is but n2 is the number of neighborhood sensing nodes, otherwise c4=c i +e c .

[0106] Step 5.9: The new weight result is input into GCN to obtain the new neighborhood node feature X.

[0107] Step 5.10: Input the enhanced matrix X into the Softmax layer, output the probabilistic recommendation results, and select appropriate experts for recommendation based on the results.

[0108] The following table explains the meanings of the relevant letters involved in the present invention:

[0109]

[0110]

[0111] The present invention can be combined with a computer system to become an enterprise expert information recommendation device based on non-local graph perception. The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the above-mentioned enterprise expert information recommendation method based on non-local graph perception is implemented.

[0112] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A method for recommending enterprise expert information based on non-local graph perception, characterized in that: The steps include: Step 1: After deduplication, de-duplication, cleaning, and basic information division of the expert information dataset D and the scientific research project dataset D1, we obtain the expert basic information label dataset D2, the expert research direction information label dataset D3, and the expert-project information label dataset D4; Step 2: Build an expert social network based on the expert basic information label dataset D2, the expert research direction information label dataset D3, and the expert-project information label dataset D4. Step 3: Use PMI to construct the expert association graph and the expert-project association graph. By extracting the corresponding keywords from the statistical expert information label dataset D2, the project basic information dataset label D3, and the expert-project information label dataset D4, and combining them with the expert social network, the edge weights of the association graph are calculated to obtain the expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ); Step 4: Use GNN to extract local neighborhood aggregation features of graph nodes and use non-local attention to the expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) Extract the node centroid non-locally and obtain the mixed information feature e" (l) ; Step 5: Expert association graph G i =(V i1 ,E i1 ) and expert-item association graph G j =(V j1 ,E j1 ) Extract graph perception features, calculate graph perception weights based on the expert research direction association and project association in the expert social network The updated expert association graph and expert-item association graph are input into GCN, combined with the graph-aware weights. The average weight e" of the neighbor set of the original node (l) And the original node weight is combined to get the new weight representation c i , use the new weight convolution to obtain feature representation, use the Softmax layer to calculate and select different experts, and finally recommend the appropriate expert; Step 5.1: Associate the experts into the graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) is input into GCN for feature extraction, defined as matrix H; Step 5.2: Define weights f i and f j are the weights between experts and between experts and projects, respectively. F is a constant used for normalization. Step 5.3: Perform graph perception operations on each relationship type of the node to obtain the average aggregation value of the node N c (i) represents the neighborhood set of node i and relationship c, h j represents the hidden vector of node j, C represents the number of neighboring nodes of node i, and W c and b c represents the graph convolution parameters, represents the normalization factor; Step 5.4: Define k as a loop variable and assign its initial value to 1, then start the loop. Step 5.5: If k ≤ len(Max(D2, D3, D4)), go to step 5.6; otherwise, jump to step 5.9; Step 5.6: Use weight c to evaluate e" (l) and Make judgments; Step 5.7: If c ≥ e” (l) ,but n1 is the number of members in the cluster, otherwise c2=c+e" (l) ; Step 5.8: If the result of step 5.7 is but n2 is the number of neighborhood sensing nodes, otherwise c4=c i +e c ; Step 5.9: The new weighted result is input into GCN to obtain the new neighborhood node feature X; Step 5.10: Input the enhanced matrix X into the Softmax layer, output the probabilistic recommendation results, and select appropriate experts for recommendation based on the results.

2. The enterprise expert information recommendation method based on non-local graph perception according to claim 1 is characterized in that: The specific method of step 1 is: Step 1.1: Define Text as a single expert text information set to be cleaned, define id1, content1, content2, content3, and content4 as the expert's serial number, basic information, research direction, number of published papers, and achievements, respectively, and satisfy the relationship Text = {id1, content1, content2, content3, content4}; Step 1.2: Define Text1 as the research project text to be cleaned information set, define id2, content5, content6, content7 as research serial number, research field, project expert and member, and expandable experts in the same field, and satisfy the relationship Text1 = {id, content5, content6, content7}; Step 1.3: Define D as the expert information data set to be cleaned, D1 as the scientific research project information data set to be cleaned, D={Text1, Text2,…, Text a ,…,Text len(D) }, Text a is the ath expert data to be cleaned in D, D1={Text11,Text12,…,Text1 b ,…,Text1 len(D1) }, Text1 b For the bth research project data to be cleaned in D1, remove duplicates and blanks from the text in the dataset and extract labels from the content in the information set to be cleaned; Step 1.4: After data cleaning and label extraction of datasets D and D1, we can obtain expert information label dataset D2, scientific research project information label dataset D3 and expert-project information label dataset D4, where D2 = {T1, T2, ..., T c ,…,T len(D2) }, T c is the cth text data in D2, where variable c∈[1,len(D2)], D3={T11,T12,…,T1 d ,…,T1 len(D3) }, T1 d is the dth text data in D3, where variable d∈[1,len(D3)], D4={T21,T22,…,T1 e ,…,T2 len(D4) }, T2 e is the e-th text data in D4, where variable e∈[1,len(D4)]; Step 1.5: Define the expert association graph G i =(V i ,E i ), node v 1i ∈V i , side e 1i ∈E i , expert-item association graph G j =(V j ,E j ), node v 2j ∈V j , side e 2j ∈E j .

3. The enterprise expert information recommendation method based on non-local graph perception according to claim 1 is characterized in that: The specific method of constructing the expert social network in step 2 is: Step 2.1: Count the labels in the label datasets D2 and D3, and summarize the label datasets such as the expert's current employment unit and research direction, which is defined as D5; Step 2.2: The single text information set in the label dataset D5 is Text2 = {id3, label1, label2, label3, label4}, where id3, label1, label2, label3, and label4 are the text information sequence number, the expert's current employment unit label, the research direction label, the label of whether they have participated in a joint scientific research project, and the intersection label of the published paper authors, respectively; Step 2.3: Define the expert social network G u =(U,S), where S represents the edge and node u i ∈U i , Bian s i ∈S i , if the two experts are a and b, then the node u i ∈U i , side e i ∈E i ; Step 2.4: Define the sliding window a as 2; Step 2.5: Define the loop variable l, assign l to 0, and start the loop; Step 2.6: If l + 1 ≤ len(D2), proceed to step 2.

7. Otherwise, jump to step 2.9, where len(D2) is the amount of data in dataset D2. Step 2.7: Define experts a and b, where a∈D2 and b∈D2, and determine whether a and b have an intersection based on the expert’s detailed label data; Step 2.8: If the experts have an intersection, create an edge s between experts a and b i , Step 2.9: Otherwise, experts a and b do not build an edge.

4. The enterprise expert information recommendation method based on non-local graph perception according to claim 1 is characterized in that: The specific method of step 3 is: Step 3.1: Use a fixed-size sliding window to calculate word co-occurrence information across the text for D2, D3, and D4, and use PMI to calculate the connection weight between two words; Step 3.2: Pre-set C win (i) and C win (i,j) sliding window threshold, C win is the total number of sliding windows in the dataset; Step 3.3: Use and Calculate the proportion of a single label in the total as the basis for calculating the edge weight; Step 3.4: Exploitation Calculate different types of edges and edge weights between words i and j; Step 3.5: Determine whether the PMI value is positive. If it is a positive PMI value, the two words are determined to be related, otherwise they are not related. Step 3.6: Use the weight partition threshold T calculated by PMI, where T≥0. If the PMI value between words i and j is greater than T, they are considered highly correlated. If 0≤PMI≤T, they are considered normally correlated. This is used to construct the expert association graph G. i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ).

5. The enterprise expert information recommendation method and device based on non-local graph perception according to claim 1 is characterized in that: The specific method of step 4 is: Step 4.1: Expert association graph G i =(V i ,E i ) and expert-item association graph G j =(V j ,E j ) node v i ∈V i and node v j ∈V j , use GNN to extract local neighborhood features; Step 4.2: Definition and represents the initial user that can be found through embedding, u and i represent the serial numbers of experts and projects, and represents the embedding of expert u and item i, Step 4.3: Use GNN to perform convolution to aggregate neighborhood features of nodes, using and and represents the aggregated neighborhood embedding at layer l; Step 4.4: Use pooling technology to embed node neighborhood features. and The final neighborhood feature is expressed as and Step 4.5: Use spatial aggregation under non-local conditions to extract non-local node centroids and use Perform K-Means clustering on the node embeddings, g ki represents the cluster membership of the node, For the centroid c k When g ki =1, then Assigned to cluster c k ; Step 4.6: Extract the centroid features of non-local nodes. The vector obtained after clustering the l-th layer nodes is represented as Step 4.7: Mix the local neighborhood features and non-local node centroid features to obtain and 6. A device for recommending enterprise expert information based on non-local graph perception, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, the steps of the enterprise expert information recommendation method and device based on non-local graph perception according to any one of claims 1 to 5 are executed.

Images