Academic paper reviewer recommendation method based on unsupervised pseudo-negative label strategy

Abstract

The invention provides an academic paper reviewer recommendation method based on an unsupervised pseudo-negative label strategy, and the method comprises the steps: obtaining and decomposing a heterogeneous academic knowledge graph, and obtaining a bigraph representing the interaction behavior of a paper and a reviewer and a knowledge graph describing node information; respectively encoding the bipartite graph and the knowledge graph by using a two-stage encoder to obtain a basic embedded representation of each node, and respectively fusing the basic embedded representations of each node through aggregation operation to obtain a fused embedded representation of each node; performing fine adjustment on the fusion embedding representation of each node through graph contrast learning, extracting a negative sample by using a pseudo tag obtained by clustering the nodes during graph contrast learning, and performing positive sample embedding representation learning to obtain a final embedding representation of each node; and selecting a to-be-predicted reviewer and a paper, and predicting the occurrence probability of review based on the final embedding representation of the corresponding node through the recommendation network. By adopting the scheme, effective and accurate academic reviewer recommendation is realized.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method for recommending academic paper reviewers based on an unsupervised pseudo-negative labeling strategy. Background Art

[0002] Academic reviewer recommendation is a key process for providing suitable peer reviewers for various academic documents (such as fund applications, research papers, technical reports, etc.). Existing reviewer recommendation research is usually based on a topic similarity assumption, that is, reviewers are more likely to be willing to review papers that are closely related to their research interests. However, due to potential reviewer preferences, this assumption may not always hold true. For example, reviewers' review decisions are not only influenced by the direct relevance to their professional field, but also by factors such as the reputation of the publication, the implicit relevance of the field, and the reviewer's current workload. In addition, in order to ensure the objectivity of the review process, most peer review records are confidential, a policy that makes information sharing between editorial offices and funding agencies almost impossible. Therefore, existing graph neural network models cannot directly adapt to the inherent graph sparsity of "reviewer-paper" and are difficult to be directly used for academic reviewer recommendation. Summary of the invention

[0003] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0004] To this end, the first purpose of this application is to propose an academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy, which solves the difficulty of capturing comprehensive contextual information and considering the complex association analysis between heterogeneous entities, and integrates the association between the dense knowledge information of the paper and the limited sparse behavioral labels between reviewers, thereby achieving effective and accurate recommendation of academic reviewers.

[0005] The second object of the present application is to provide a computer device.

[0006] To achieve the above-mentioned purpose, the first embodiment of the present application proposes a method for recommending academic paper reviewers based on an unsupervised pseudo-negative label strategy, including: obtaining a heterogeneous academic knowledge graph, wherein the nodes in the academic graph include reviewer nodes and paper nodes, and the relationships in the academic graph include review relationships; decomposing the academic knowledge graph to obtain a bipartite graph representing the interaction behavior between papers and reviewers and a knowledge graph describing the node information; using a two-stage encoder to encode the bipartite graph and the knowledge graph respectively to obtain a first basic embedding representation and a second basic embedding representation of each node, and fusing the first basic embedding representation and the second basic embedding representation of each node through an aggregation operation to obtain a fused embedding representation of each node; fine-tuning the fused embedding representation of each node through graph contrast learning, using the pseudo-labels obtained by clustering the nodes during graph contrast learning to extract negative samples, and performing positive sample embedding representation learning to obtain the final embedding representation of each node; selecting the reviewers and papers to be predicted, and predicting the probability of review occurrence based on the final embedding representation of the corresponding nodes through a recommendation network.

[0007] The academic paper reviewer recommendation method based on the unsupervised pseudo-negative labeling strategy of the embodiment of the present application adopts a two-stage encoder structure to learn the comprehensive embedding representation of reviewers and papers. In the first stage, a decoupled GNN is used to encode the behavioral preferences of reviewers, and a language model based on academic material pre-training is used to capture the prior scientific semantic knowledge of the paper; in the second stage, a pseudo-negative labeling strategy is introduced to solve the false negative problem caused by the extreme sparsity of the bipartite graph of "reviewer-paper" to improve the negative sampling performance in the graph contrast learning process. The embodiment of the present application uses an unsupervised pseudo-negative labeling strategy to enhance graph contrast learning, more effectively select nodes that are inconsistent with the semantic scope of the current node, and provide relative representation information support when used to recommend academic reviewers.

[0008] Optionally, in one embodiment of the present application, the academic knowledge graph is represented as:

[0009] G=(N,ε)

[0010] Where N is a set of nodes of different types, N = {N scholar ,N submission},N scholar is the reviewer node, N submission is the paper node, ε={E review}, E review For review relationship.

[0011] Optionally, in one embodiment of the present application, the two-stage encoder includes a behavior preference encoder and a knowledge encoder, the behavior preference encoder is a decoupled graph convolutional network, and the knowledge encoder is a pre-trained language encoding model OAG-BERT. The bipartite graph and the knowledge graph are encoded by the two-stage encoder respectively to obtain a first basic embedding representation and a second basic embedding representation of each node, including:

[0012] Encode the bipartite graph through the behavior preference encoder to obtain the first basic embedding representation of each node;

[0013] The bipartite graph is encoded by the knowledge encoder to obtain the second basic embedding representation of each node;

[0014] The above method further includes:

[0015] When encoding through the behavior preference encoder and the knowledge encoder, the parameters in the decoupled graph convolutional network are adjusted by optimizing the loss function of the behavior preference encoder.

[0016] Optionally, in one embodiment of the present application, the encoding process of the behavior preference encoder is expressed as:

[0017]

[0018] Among them, the embedding representation vector of node u in layer l, and the node representation matrix derived in layer l is

[0019] d represents the dimension of the embedded representation, is the normalized adjacency matrix, A is the adjacency matrix with self-loops, D is the diagonal matrix corresponding to A, and W b (l) is the learnable parameter matrix of the lth layer in the decoupled graph convolutional network;

[0020] The loss function of the behavior preference encoder is expressed as:

[0021]

[0022] Where (u,v)∈ε review Indicates that y u,v represents the edge observed between nodes u and v, is the probability of the edge between the paper to be reviewed u and the reviewer v, express, express,

[0023] Optionally, in one embodiment of the present application, the first basic embedding representation and the second basic embedding representation of each node are fused respectively through aggregation operation to obtain a fused embedding representation of each node, which is expressed as:

[0024]

[0025] in, is the first basic embedding representation of each node, is the second basic embedding representation of each node.

[0026] Optionally, in one embodiment of the present application, fine-tuning the fusion embedding representation of each node through graph comparison learning includes:

[0027] Graph contrast learning is performed through the first collaborative contrast encoder and the second collaborative contrast encoder to optimize the fused embedding representation of each node, wherein the first collaborative contrast encoder processes the data of the bipartite graph and the knowledge graph respectively through the behavior preference encoder and the knowledge encoder, and fuses the processed embedding representations to obtain the basic embedding representation of each node, and the second collaborative contrast encoder extracts negative samples using pseudo labels, and the first collaborative contrast encoder and the second collaborative encoder learn the embedding representations of positive samples and negative samples respectively;

[0028] The above method further includes:

[0029] When performing graph contrast learning through the encoder, nodes that are inconsistent with the current node pseudo-label are selected as negative samples, wherein the pseudo-label generation process of the node includes: clustering the nodes through the clustering layer and assigning an identifier to each node;

[0030] The above method further includes:

[0031] When performing graph contrastive learning through the encoder, positive samples are obtained by modifying the feature embedding representation on each node, and the embedding representation information is propagated on the graph through the graph convolution layer;

[0032] The parameters in the graph convolutional layer are adjusted by jointly optimizing the loss functions of the first collaborative contrastive encoder and the second collaborative contrastive encoder, and the contrastive mutual information is encoded into the embedded representation of the node.

[0033] Optionally, in one embodiment of the present application, clustering nodes by a clustering layer and assigning an identifier to each node includes:

[0034] Assume that there are C clusters, and the corresponding cluster centers In R d The space is randomly initialized, and each node is assigned a cluster ID as a pseudo label by minimizing the loss function of the first collaborative contrastive encoder;

[0035] The graph convolution layer is a graph convolution layer with shared weights, expressed as:

[0036]

[0037] in, is the normalized adjacency matrix, A is the adjacency matrix with self-loops, D is the diagonal matrix corresponding to A, and W c is the weight matrix in the graph convolution layer, H 1 is the embedding representation matrix of all nodes,

[0038] The loss function of the first collaborative contrastive encoder is expressed as:

[0039]

[0040] Among them, N schoiar is the reviewer node, N submission is the paper node, p u,i is the target distribution, q u,i is the embedding representation h u and cluster centers μ i The similarity between P u,i represents the target distribution;

[0041] The loss function of the second collaborative contrastive encoder is expressed as:

[0042]

[0043] in, represents the positive sample, It means that by randomly masking some features on a given node, W c is the weight matrix in the graph convolutional layer, represents the negative sampling set, c(u) represents the cluster ID of node u.

[0044] Optionally, in one embodiment of the present application, the final embedding representation of the corresponding node includes the final embedding representation of the reviewer node to be predicted, the final embedding representation of the paper node to be predicted, and the final embedding representation of the reviewer's historical review paper node;

[0045] The recommendation network is a decoder network, which predicts the probability of review occurrence based on the final embedding representation of the corresponding node through the recommendation network, including:

[0046] Calculate the interactive attention between the final embedding representation of the paper node to be predicted and the final embedding representation of the historical review paper node, and determine the attention weight of the historical review paper node;

[0047] The final embedding representation of the reviewer node to be predicted, the final embedding representation of the paper node to be predicted, and the weighted historical review paper nodes are concatenated, and the review probability between the reviewer to be predicted and the contributor is calculated through a multi-layer perceptron.

[0048] Optionally, in one embodiment of the present application, calculating the interactive attention between the final embedding representation of the paper node to be predicted and the final embedding representation of the historical review paper node, and determining the attention weight of the historical review paper node, includes:

[0049] The final embedding representation of each historical review paper node and the final embedding representation of the paper node to be predicted are calculated using the outer product and decoded through a multi-layer perceptron to obtain the attention weight of the historical review paper node, where

[0050] The attention weight of the historical review paper node is expressed as:

[0051]

[0052] in, is the bias matrix Dimension (d 2 ,η), (η,1), η is the size of the intermediate encoding dimension, is the final embedding representation of the historical review paper node, h v is the final embedding representation of the paper node to be predicted.

[0053] Optionally, in one embodiment of the present application, during the training process of the academic paper reviewer recommendation method, the loss function used is the BCE loss function, the training object is the edge probability between the paper and the reviewer, and during training, the loss gradient is backpropagated to the weight matrix in the decoupled graph convolutional network. During training, all loss functions are multiplied by coefficients that adjust them to the same numerical level to obtain the total loss function.

[0054] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0056] Figure 1A flowchart of a method for recommending academic paper reviewers based on an unsupervised pseudo-negative labeling strategy provided in Example 1 of the present application;

[0057] Figure 2 This is the overall framework diagram of the RevGNN model of the embodiment of the present application;

[0058] Figure 3 A flowchart of another method for recommending academic paper reviewers based on an unsupervised pseudo-negative labeling strategy according to an embodiment of the present application;

[0059] Figure 4 A schematic diagram of the process of generating two encoders for behavior preference and knowledge representation according to an embodiment of the present application;

[0060] Figure 5 A schematic diagram of the network structure of a decoder according to an embodiment of the present application. DETAILED DESCRIPTION

[0061] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0062] The following describes the academic paper reviewer recommendation method based on the unsupervised pseudo-negative label strategy of an embodiment of the present application with reference to the accompanying drawings.

[0063] Figure 1 A flowchart of an academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy provided in Example 1 of the present application.

[0064] like Figure 1 As shown in FIG. , the academic paper reviewer recommendation method based on the unsupervised pseudo-negative label strategy includes the following steps:

[0065] Step 101, obtaining a heterogeneous academic knowledge graph, wherein the nodes in the academic graph include reviewer nodes and paper nodes, and the relationships in the academic graph include review relationships;

[0066] Step 102, decomposing the academic knowledge graph to obtain a bipartite graph representing the interaction between the paper and the reviewer and a knowledge graph describing the node information;

[0067] Step 103, using a two-stage encoder to encode the bipartite graph and the knowledge graph respectively, to obtain a first basic embedding representation and a second basic embedding representation of each node, and to fuse the first basic embedding representation and the second basic embedding representation of each node respectively through an aggregation operation to obtain a fused embedding representation of each node;

[0068] Step 104, fine-tuning the fusion embedding representation of each node through graph contrast learning, using pseudo labels obtained by clustering nodes to extract negative samples during graph contrast learning, and performing positive sample embedding representation learning to obtain the final embedding representation of each node;

[0069] Step 105, select the reviewers and papers to be predicted, and predict the probability of review occurrence based on the final embedding representation of the corresponding nodes through the recommendation network.

[0070] The academic paper reviewer recommendation method based on the unsupervised pseudo-negative labeling strategy of the embodiment of the present application adopts a two-stage encoder structure to learn the comprehensive embedding representation of reviewers and papers. In the first stage, a decoupled GNN is used to encode the behavioral preferences of reviewers, and a language model based on academic material pre-training is used to capture the prior scientific semantic knowledge of the paper; in the second stage, a pseudo-negative labeling strategy is introduced to solve the false negative problem caused by the extreme sparsity of the bipartite graph of "reviewer-paper" to improve the negative sampling performance in the graph contrast learning process. The embodiment of the present application uses an unsupervised pseudo-negative labeling strategy to enhance graph contrast learning, more effectively select nodes that are inconsistent with the semantic scope of the current node, and provide relative representation information support when used to recommend academic reviewers.

[0071] The present application embodiment also provides another method for recommending academic paper reviewers based on an unsupervised pseudo-negative labeling strategy. The method models different types of components in the graph to form Figure 2 The RevGNN model shown in Figure 2 As shown in Figure 1, the RevGNN model includes an input layer, an event encoder, a graph neural network layer, a relational attention mechanism, an aggregation layer, and an output layer. Specifically,

[0072] (1) Input layer

[0073] Nodes: Represents entities participating in an event.

[0074] Edges: Represents the relationship event between two nodes.

[0075] Timestamps: Record the time when an event occurs.

[0076] (2) Event Encoder

[0077] Node Embeddings: Convert nodes into vector representations.

[0078] Edge Embeddings: Convert edges into vector representations, including participating nodes and timestamp information.

[0079] (3) Graph Neural Network Layers

[0080] Structured Graph Convolutional Layer: Captures local structural information between nodes.

[0081] Message Passing Mechanism: Pass information between nodes and update node features.

[0082] Temporal Convolutional Layer: Captures dynamic characteristics that change over time.

[0083] Temporal Encoding: Encode temporal information into a vector representation.

[0084] Temporal Convolution: Perform convolution operations on the time dimension to extract time series features.

[0085] (4) Relational Attention Mechanism

[0086] Attention Weight Calculation: Calculate the importance weights of different relationships.

[0087] Weighted Aggregation: Aggregate node features according to attention weights.

[0088] (5) Aggregation Layer

[0089] Feature Fusion: Integrate the features captured from different levels to generate the final representation of nodes and edges.

[0090] (6) Output Layer

[0091] Node Representation: Generates the final vector representation of the node.

[0092] Edge Representation: Generates the final vector representation of the edge.

[0093] Predictions: Perform task predictions based on the final representation, such as node classification, link prediction, etc.

[0094] (7) Training and Optimization

[0095] Loss Function: used to optimize model parameters and ensure the accuracy of the model on a specific task.

[0096] Optimization Algorithm: Commonly used optimization algorithms such as Adam or SGD are used to adjust model parameters to minimize the loss function.

[0097] The specific recommendation method for academic paper reviewers includes:

[0098] First, the input heterogeneous graph is decomposed into a bipartite graph representing the "paper-reviewer" interaction behavior and a knowledge graph describing the node knowledge. According to the type of graph, two encoders with specific structures, namely the behavior preference encoder and the knowledge encoder, are used to learn the basic embedding representation of each node, and they are fused again using aggregation operations. Then, the pseudo-labels obtained by clustering are used to extract negative samples, and the positive sample embedding representation learning is performed at the same time, and graph contrast learning is realized in two collaborative contrast encoders. Finally, based on the extraction operation, the interaction-based neural network structure is used to predict the edges based on the embedding of the "paper-reviewer" pair, thereby realizing academic reviewer recommendations. The entire calculation is carried out under a framework process. See the overall process. Figure 3 .

[0099] like Figure 3 As shown, the input is a heterogeneous academic graph G and its corresponding relationship matrix E.

[0100] Step 1 Generate two encoders: behavior preference and knowledge representation

[0101] The input heterogeneous graph is decomposed into two subgraphs: one is a bipartite graph representing the interaction behavior of "paper-reviewer", which only includes papers, reviewers and the review relationship between the two, and provides information about the potential preferences and habits of reviewers in the past review process. The other subgraph describes node knowledge, which is equivalent to the complement of the "paper-reviewer" bipartite graph and provides text topic information related to papers and reviewers.

[0102] Step 2 Aggregation and pooling of two encoders

[0103] After the reviewer preference behavior and paper knowledge information are encoded by their respective encoders, a specific strategy is adopted to aggregate them together as the basis for the next stage of encoding. The goal is to transfer the information of heterogeneous paper and reviewer nodes through association relationships, and then compress the representation vectors of papers and reviewers into a fixed low-dimensional space, depending on the type of graph.

[0104] Step 3 uses the pseudo labels obtained by clustering to extract negative samples and simultaneously performs positive sample embedding representation learning

[0105] Using the reviewer and paper representation vectors learned in step 1 and step 2, we further fine-tune them based on their relative position relationship to obtain more accurate representation vectors. We use two collaborative contrast encoders to implement graph contrast learning.

[0106] Step 4: Use the interactive neural network structure to complete the edge prediction

[0107] This is done based on the embedding of the paper-reviewer pair. After step 3, the paper and reviewer information is fully represented by the model using an embedding vector of fixed dimension. Subsequently, it is only necessary to decode it through a relatively simple decoder to predict whether there is an edge between any given pair of paper-reviewers.

[0108] This embodiment also proposes an academic review recommendation method of an interactive neural network. Taking the interactive neural network as an example, Figure 4 The main description is the phenomenon that there are sparse and semantically ambiguous edges between papers and reviewers. In order to overcome the objective problem, this embodiment encodes the knowledge of the papers and the behavior preference of "paper-reviewers". Specifically, the process includes:

[0109] Step 1: Generate two encoders for behavior preference and knowledge representation

[0110] In the first stage of this embodiment, the behavior preference encoder and knowledge encoder are used to learn the "paper-reviewer" bipartite graph and the academic knowledge graph respectively, and the loss L is optimized. behavior To adjust the parameters W in the decoupled GCN layer b , and uses a pre-trained academic language model to encode interactions and prior knowledge into node embedding representation vectors.

[0111] First, in the "paper-reviewer" bipartite graph, the present invention adopts a decoupled graph convolutional network. This decoupling method allows different types of nodes to have different influences in the representation learning process, and better captures the complex relationship between different nodes in the heterogeneous graph. By independently learning the weight parameters of each node, the decoupled graph convolutional network can better adapt to the feature distribution of different nodes and improve the generalization ability of the model. This design makes the model more capable of coping with complex graph structures, especially when heterogeneous graphs contain multiple types of nodes.

[0112] By aggregating embeddings from their neighboring nodes, we effectively learn embeddings for papers and reviewers. Formalize this encoding process, where the nodes derived from layer l are represented by It means that given A is an adjacency matrix with self-loops, D is the corresponding diagonal matrix, is the normalized adjacency matrix. The learnable parameter matrix of the first layer is The dimension of the embedded representation is denoted by d. The probability of the edge from the paper to be reviewed u to the reviewer v can be obtained by a non-parametric residual product: in Using binary cross entropy as loss function based on the observed edges:

[0113]

[0114] Among them, y u,v represents the edge observed between nodes u and v.

[0115] Then, in knowledge encoding, this embodiment introduces the pre-trained language encoding model OAG-BERT to generate the initial embedding representation of each paper. The design of OAG-BERT includes a unique two-dimensional position encoding, which allows heterogeneous entity relationship types that have been entity-masked to be input into BERT. Through the additional dimension, information such as entity type can be encoded, and the embedded representation of this information can also be input into OAG-BERT as prior knowledge.

[0116] Step 2: Aggregation and pooling of two encoders

[0117] In this aggregation process, specific strategies are adopted to ensure that the information of reviewers and papers can be effectively combined. The encoding purpose of this stage is to transfer the node information of heterogeneous papers and reviewers by considering the association relationship between them, and compress their representation vectors into a low-dimensional fixed space. This low-dimensional space is designed to retain information directly related to the review goal, so that only important information is retained in subsequent processing, improving the efficiency and performance of the model. This process helps to extract and retain important features related to the review task, so as to better complete the reviewer recommendation task. Specifically, when a node is a reviewer, its knowledge embedding representation comes from the average pooling operation of all paper neighbors directly connected to the reviewer. The reason for this practice is that the research topic of the reviewer can often be reflected by the papers he has reviewed before. Therefore, by averaging all papers directly related to the reviewer, a more comprehensive and accurate knowledge embedding representation can be obtained, which better captures the research direction and interest area of ​​the reviewer. This method helps to improve the representation quality of the reviewer node, thereby enhancing the model's understanding and prediction of the reviewer's behavior. This process is described in detail as follows: That is, using vector set join operations and Two vectors are concatenated. Among them, It is the embedding representation vector generated by OAG-BERT or the behavior graph convolution layer as the knowledge encoder.

[0118] Step 3: Use the pseudo-labels obtained by clustering to extract negative samples and perform positive sample embedding representation learning at the same time

[0119] In graph contrastive learning, the core idea is to model data points through relative representation learning. The key to this method is to optimize the contrast loss function so that the representations of similar data pairs (positive pairs) are closer and the representations of dissimilar data pairs (negative pairs) are farther apart (or in some cases, more orthogonal). Therefore, developing an effective strategy to extract positive and negative samples from the current node to establish a contrast relationship between similar and dissimilar data pairs becomes a key step in graph contrastive learning.

[0120] Currently, recommendation models based on graph contrastive learning mainly focus on the construction of positive samples. Although many existing methods focus on how to construct negative samples, their main goal is to improve the model to learn more features in difficult negative samples to improve the discrimination between positive and negative samples. In the "paper-reviewer" recommendation process, this embodiment must be careful to avoid negative samples that may lead to false negatives. This means that this embodiment must be careful not to mistakenly mark data that may actually be positive samples as negative samples, so as not to affect the accuracy of the recommendation model.

[0121] Therefore, this embodiment proposes a method for constructing pseudo labels by selecting nodes that are inconsistent with the current node pseudo labels as negative samples to adjust the embedded representation of the nodes. At this stage, the encoder uses graph contrastive learning to optimize the representation of the nodes. At the same time, this embodiment also designs a pseudo negative label sampling strategy to avoid selecting nodes that may cause false negatives for the current node, ensuring the quality of the negative example pairs.

[0122] First, we obtain positive samples by modifying the feature embedding representation at each node and propagating the embedding representation information on the graph using a graph convolutional layer. Then, we jointly optimize L contrast and L cluster To adjust the parameters W in the graph convolution layer c , and encodes the contrastive mutual information into the embedding representation of the node.

[0123] In order to solve the problem of false negative samples, this embodiment introduces an unsupervised learning clustering layer on the embedded representation of nodes and jointly optimizes it with the graph contrast learning framework. Specifically, the present invention first uses a graph convolution layer with shared weights, which is the same as the graph convolution layer used in positive sampling and can be expressed as H 1 is the embedding representation matrix of all nodes This layer is designed to better optimize the embedding representation of nodes through the clustering process of unsupervised learning, thereby reducing the impact of false negative samples. The clustering layer clusters the nodes and assigns a unique identifier to each node, thereby helping the model to better learn the relationship and features between nodes and improve model performance. Assume that there are C clusters and their cluster centers are In R d The space is randomly initialized, i represents the clustering of the i-th layer. By minimizing L cluster , a cluster ID can be assigned to each node and this ID can be regarded as a pseudo label. cluster The calculation is expressed as:

[0124]

[0125] in, is the target distribution, is the embedding representation h u and cluster centers μ i The similarity between .

[0126] This embodiment uses represents the positive sample, in, It means that by randomly masking some features on a given node, W c is the weight matrix, σ() is the activation function. represents the negative sampling set, Where c(u) represents the cluster ID of node u.

[0127] This embodiment is based on the derived positive sample Negative sample set and the original encoding embedding h u , constructing the InfoNEC-like contrast loss L contrast , the calculation process is as follows:

[0128]

[0129] Step 4: Use the interaction-based neural network structure to complete the edge prediction

[0130] After the two-stage encoding included in the above steps, this embodiment can learn the information of the paper and the reviewer, and obtain an embedded vector representation with a fixed dimension. Next, these embedded representations are extracted through a simple decoder operation to predict whether there is an edge for any given "paper-reviewer" pair. Considering that the embedded representation includes behavior, prior knowledge, and comparative information, this embodiment designs a ranking-based recommendation network that can calculate the interactive attention between the current paper and the papers reviewed by the reviewer in the past, thereby providing more accurate recommendation results. Figure 5 As shown in Figure 1, interactive attention is actually a variant of deep interest network. A reviewer's review history may include multiple papers. For ease of understanding, the figure only shows the corresponding scenario with only one paper.

[0131] At the beginning of decoding, the decoder network is provided with an embedding representation vector h of a potential reviewer. u 、Papers that the reviewer has reviewed H p And the representation vector h of the candidate paper to be submitted v The decoder network predicts the probability of a review by considering the attention between the candidate submission and the papers that the reviewer has reviewed in the past.

[0132] First, in the decoder stage, this embodiment calculates the interactive attention between the candidate paper and the historical review paper. and h v The outer product is used for calculation and then decoded through a multi-layer perceptron (MLP): The obtained i It is regarded as an attention weight to measure the influence of historically reviewed papers. and is the bias matrix and The dimension is (d 2 ,η) and (η,1), η is the size of the intermediate encoding dimension. The activation function σ is selected as PReLU according to DIN.

[0133] Then, in the recommendation phase, this embodiment connects the reviewers, submitted papers, and weighted historical review papers, and uses MLP again to calculate the review probability between the reviewers and the submitted papers. This embodiment uses a similar BCE loss function and back-propagates the loss gradient to the weight matrix in the encoder to obtain L supervised .

[0134] Finally, the total loss function of the present invention is obtained by multiplying the loss function of each part by the coefficient adjusted to the same numerical level and summing them up.

[0135] The experiment was based on the contrast enhancement model of sparse behavior. The relevant review information was collected from Frontiers Publishing House, which has open paper review information, and a large-scale public review dataset was constructed for the evaluation of relevant models. The experiment selected 5 journals with an impact factor greater than 6.0 and retained the corresponding 40,578 paper data. The statistical data is shown in Table 1.

[0136] Table 1 Statistical data of article sources collected in the experiment

[0137]

[0138] Among them, 396 articles have only one reviewer, 34,166 articles have two reviewers, 5,435 articles have three reviewers, and 505 articles have four reviewers. On this basis, the experiment divides the data to generate training sets, validation sets, and test sets. In the end, there are 72,360 interaction records between papers and reviewers in the training set, and 7,344 interaction records in the validation set and test set. The experiment sorted and summarized the collected data and named it Frontiers-4k. The experiment further divided the Frontiers-4k data set, generating training sets, validation sets, and test sets with the interaction records in a ratio of 8:1:1. When dividing, it is necessary to ensure that all paper IDs and reviewer IDs appear at least once in the training set. Under this condition, the remaining data is randomly stored. When testing the recommendation performance, for each paper, a reviewer is randomly selected as the test set, and the rest are used as the training set.

[0139] The experiment uses the RecBole toolset to load all experimental comparison models. In addition, all experiments are conducted on an Ubuntu 18.04.2LTS server, which uses an AMD EPYC 7642 48-core processor, 1008GB random access memory and an NVIDIA A100 graphics processor. The present invention is implemented based on PyTorch1.9.1 and Python3.9.7. In the experiment, the model selects ReLU as H 2 The activation function is Set η=36. middle and The dimensions are (3d, 200) and (200, 80), respectively, where d = 932. In addition, for the first stage encoder, the experiment uses 3 layers of GCN for encoding; for the second stage encoder, 1 layer of graph convolution layer is used for encoding. In addition, the model uses ADAM with different learning rates to optimize the different structures of RevGNN. Specifically, for the first stage encoder and recommendation decoder, the model sets the learning rate to 0.001; and for the second stage encoder, the model sets the learning rate to 0.0001. The experimental results are shown in Table 2. This embodiment achieves the best performance in all evaluation indicators on both datasets.

[0140] Table 2 Performance comparison of the proposed model and the comparative model on the reviewer recommended dataset

[0141]

[0142] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0143] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0144] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0145] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0146] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0147] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0148] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0149] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for recommending academic paper reviewers based on an unsupervised pseudo-negative labeling strategy, characterized in that: include: Obtaining a heterogeneous academic knowledge graph, wherein the nodes in the academic graph include reviewer nodes and paper nodes, and the relationships in the academic graph include review relationships; Decomposing the academic knowledge graph to obtain a bipartite graph representing the interaction between papers and reviewers and a knowledge graph describing node information; Using the two-stage encoder to encode the bipartite graph and the knowledge graph respectively, obtaining a first basic embedding representation and a second basic embedding representation of each node, and fusing the first basic embedding representation and the second basic embedding representation of each node respectively through an aggregation operation to obtain a fused embedding representation of each node; The fused embedding representation of each node is fine-tuned through graph contrast learning. During graph contrast learning, pseudo labels obtained by clustering nodes are used to extract negative samples, and positive sample embedding representation learning is performed to obtain the final embedding representation of each node. Select the reviewers and papers to be predicted, and use the recommendation network to predict the probability of review occurrence based on the final embedding representation of the corresponding nodes.

2. The method according to claim 1, characterized in that The academic knowledge graph is represented as: G=(N,ε) Where N is a set of nodes of different types, N = {N scholar , N submission },N scholar is the reviewer node, N submission is the paper node, ε={E review }, E review For review relationship.

3. The method according to claim 1, characterized in that The two-stage encoder includes a behavior preference encoder and a knowledge encoder, the behavior preference encoder is a decoupled graph convolutional network, the knowledge encoder is a pre-trained language encoding model OAG-BERT, and the two-stage encoder is used to encode the bipartite graph and the knowledge graph respectively to obtain a first basic embedding representation and a second basic embedding representation of each node, including: Encoding the bipartite graph through a behavior preference encoder to obtain a first basic embedding representation of each node; Encoding the bipartite graph through a knowledge encoder to obtain a second basic embedding representation of each node; The method further comprises: When encoding through the behavior preference encoder and the knowledge encoder, the parameters in the decoupled graph convolutional network are adjusted by optimizing the loss function of the behavior preference encoder.

4. The method according to claim 3, characterized in that The encoding process of the behavior preference encoder is expressed as: in, represents the embedding representation vector of node u at layer l, and the node representation matrix derived at layer l is d represents the dimension of the embedded representation, is the normalized adjacency matrix, A is the adjacency matrix with self-loops, D is the diagonal matrix corresponding to A, and W b ( l ) is the learnable parameter matrix of the lth layer in the decoupled graph convolutional network; The loss function of the behavior preference encoder is expressed as: Where (u, v)∈ε review Indicates that y u，v represents the edge observed between nodes u and v, is the probability of the edge between the paper to be reviewed u and the reviewer v, express, express, 5. The method according to claim 1, characterized in that The first basic embedding representation and the second basic embedding representation of each node are respectively fused by aggregation operation to obtain a fused embedding representation of each node, which is expressed as: in, is the first basic embedding representation of each node, is the second basic embedding representation of each node.

6. The method according to any one of claims 2 to 5, characterized in that: The fusion embedding representation of each node is fine-tuned through graph contrast learning, including: Graph contrast learning is performed through the first collaborative contrast encoder and the second collaborative contrast encoder to optimize the fused embedding representation of each node, wherein the first collaborative contrast encoder processes the data of the bipartite graph and the knowledge graph respectively through the behavior preference encoder and the knowledge encoder, and fuses the processed embedding representations to obtain the basic embedding representation of each node, and the second collaborative contrast encoder extracts negative samples using pseudo labels, and the first collaborative contrast encoder and the second collaborative encoder learn the embedding representations of positive samples and negative samples respectively; The method further comprises: When performing graph contrast learning through the encoder, nodes that are inconsistent with the current node pseudo-label are selected as negative samples, wherein the pseudo-label generation process of the node includes: clustering the nodes through the clustering layer and assigning an identifier to each node; The method further comprises: When performing graph contrastive learning through the encoder, positive samples are obtained by modifying the feature embedding representation on each node, and the embedding representation information is propagated on the graph through the graph convolution layer; The parameters in the graph convolutional layer are adjusted by jointly optimizing the loss functions of the first collaborative contrastive encoder and the second collaborative contrastive encoder, and the contrastive mutual information is encoded into the embedded representation of the node.

7. The method according to claim 6, characterized in that Clustering the nodes through the clustering layer and assigning an identifier to each node includes: Assume that there are C clusters, and the corresponding cluster centers In R d The space is randomly initialized, and each node is assigned a cluster ID as a pseudo label by minimizing the loss function of the first collaborative contrastive encoder; The graph convolution layer is a graph convolution layer with shared weights, which is expressed as: in, is the normalized adjacency matrix, A is the adjacency matrix with self-loops, D is the diagonal matrix corresponding to A, and W c is the weight matrix in the graph convolution layer, H 1 is the embedding representation matrix of all nodes, The loss function of the first collaborative contrastive encoder is expressed as: Among them, N scholar is the reviewer node, N submission is the paper node, p u，i is the target distribution, q u，i is the embedding representation h u and cluster centers μ i The similarity between P u，i represents the target distribution; The loss function of the second collaborative contrastive encoder is expressed as: in, represents the positive sample, It means that by randomly masking some features on a given node, W c is the weight matrix in the graph convolution layer, represents the negative sampling set, c(u) represents the cluster ID of node u.

8. The method according to claim 1, characterized in that The final embedding representation of the corresponding node includes the final embedding representation of the reviewer node to be predicted, the final embedding representation of the paper node to be predicted, and the final embedding representation of the reviewer's historical review paper node; The recommendation network is a decoder network, and the method of predicting the probability of review occurrence based on the final embedding representation of the corresponding node through the recommendation network includes: Calculate the interactive attention between the final embedding representation of the paper node to be predicted and the final embedding representation of the historical review paper node, and determine the attention weight of the historical review paper node; The final embedding representation of the reviewer node to be predicted, the final embedding representation of the paper node to be predicted, and the weighted historical review paper nodes are concatenated, and the review probability between the reviewer to be predicted and the contributor is calculated through a multi-layer perceptron.

9. The method according to claim 8, characterized in that The step of calculating the interactive attention between the final embedding representation of the paper node to be predicted and the final embedding representation of the historical review paper node, and determining the attention weight of the historical review paper node, includes: The final embedding representation of each historical review paper node and the final embedding representation of the paper node to be predicted are calculated using the outer product and decoded through a multi-layer perceptron to obtain the attention weight of the historical review paper node, where The attention weight of the historical review paper node is expressed as: in, is the bias matrix Dimension (d 2 ,η), (η,1), η is the size of the intermediate encoding dimension, is the final embedding representation of the historical review paper node, h v is the final embedding representation of the paper node to be predicted.

10. The method according to any one of claims 2 to 9, characterized in that: During the training process of the academic paper reviewer recommendation method, the loss function used is the BCE loss function, and the training object is the edge probability between the paper and the reviewer. During training, the loss gradient is backpropagated to the weight matrix in the decoupled graph convolutional network. During training, all loss functions are multiplied by coefficients that are adjusted to the same numerical level to obtain the total loss function.

Images