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

Abstract

The application provides a method for recommending academic paper reviewers based on an unsupervised pseudo-negative label strategy, which comprises the following steps: obtaining a heterogeneous academic knowledge graph and decomposing the same to obtain a two-part graph representing the interaction behavior of papers and reviewers and a knowledge graph describing node information; using a two-stage encoder to encode the two-part graph and the knowledge graph respectively to obtain basic embedding representations of nodes, and fusing the basic embedding representations of each node through aggregation operation to obtain fusion embedding representations of the nodes; fine-tuning the fusion embedding representations of the nodes through graph contrast learning, using pseudo-labels obtained by clustering the nodes to extract negative samples during the graph contrast learning, and performing positive sample embedding representation learning to obtain final embedding representations of the nodes; selecting a reviewer and a paper to be predicted, and predicting the probability of review occurrence based on the final embedding representations of the corresponding nodes through a recommendation network. The method of the application realizes effective and accurate recommendation of academic reviewers.

Description

Technical Field

[0001] This 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 Technology

[0002] Academic reviewer recommendations are a crucial process for providing suitable peer reviewers for various academic documents, such as grant applications, research papers, and technical reports. Existing reviewer recommendation research is typically based on the assumption of topic similarity: reviewers are more likely to review papers closely related to their research interests. However, this assumption may not always hold true due to potential reviewer preferences. For example, reviewers' decisions are influenced not only by direct relevance to their field of expertise but also by factors such as the reputation of the publication, implicit relevance within the field, and the reviewer's current workload. Furthermore, to ensure objectivity, most peer review records are kept confidential, making information sharing between editorial offices and funding agencies virtually impossible. Therefore, existing graph neural network models cannot directly adapt to the inherent graph sparsity of the "reviewer-paper" relationship and are difficult to directly apply to academic reviewer recommendations. Summary of the Invention

[0003] This application aims to at least partially address one of the technical problems in the related art.

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

[0005] The second objective of this application is to propose a computer device.

[0006] To achieve the above objectives, the first aspect of this application proposes an academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy, comprising: acquiring 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 of papers and reviewers and a knowledge graph describing 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 aggregation operations to obtain a fused embedding representation of each node; fine-tuning the fused embedding representation of each node through graph contrastive learning, in which pseudo-labels obtained by clustering nodes are used to extract negative samples during graph contrastive learning, and positive sample embedding representation learning is performed to obtain the final embedding representation of each node; selecting reviewers and papers to be predicted, and predicting the probability of review occurrence through a recommendation network based on the final embedding representation of the corresponding nodes.

[0007] This application's embodiment of the academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy employs a two-stage encoder structure to learn comprehensive embedding representations of reviewers and papers. In the first stage, a decoupled GNN is used to encode reviewers' behavioral preferences, while a language model pre-trained based on academic corpora captures prior scientific semantic knowledge of the paper. In the second stage, a pseudo-negative labeling strategy is introduced to address the false negative problem caused by the extreme sparsity of the bipartite graph of the "reviewer-paper" relationship, thereby improving the negative sampling performance during graph contrastive learning. This application's embodiment utilizes an unsupervised pseudo-negative labeling strategy to enhance graph contrastive learning, more effectively selecting nodes whose semantic range is inconsistent with the current node, and providing relative representation information support when recommending academic reviewers.

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

[0009] G=(N,ε)

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

[0011] Optionally, in one embodiment of this 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 coding model OAG-BERT. The two-stage encoder encodes the bipartite graph and the knowledge graph respectively to obtain the first basic embedding representation and the second basic embedding representation of each node, including:

[0012] The bipartite graph is encoded using a behavior preference encoder to obtain the first basic embedding representation of each node;

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

[0014] The above methods also include:

[0015] When encoding using a behavior preference encoder and a 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 this application, the encoding process of the behavior preference encoder is represented as follows:

[0017]

[0018] Wherein, the embedding representation vector of node u in layer l, and the node representation matrix derived in layer l are:

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

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

[0021]

[0022] Where (u,v)∈ε review It means that y u,v This represents the edge observed between nodes u and v. Let be the probability of an edge connecting the paper u to the reviewer v. express, express,

[0023] Optionally, in one embodiment of this application, the first basic embedding representation and the second basic embedding representation of each node are fused through aggregation operations to obtain the fused embedding representation of each node, as follows:

[0024]

[0025] in, This represents the first basic embedding representation for each node. This is the second basic embedding representation for each node.

[0026] Optionally, in one embodiment of this application, fine-tuning the fused embedding representation of each node through graph contrastive learning includes:

[0027] The graph comparison learning is performed by the first and second collaborative comparison encoders to optimize the fusion embedding representation of each node. The first collaborative comparison encoder processes the data of the bipartite graph and the knowledge graph through the behavior preference encoder and the knowledge encoder respectively, and fuses the processed embedding representations to obtain the basic embedding representation of each node. The second collaborative comparison encoder uses pseudo-labels to extract negative samples. The first and second collaborative comparison encoders learn the embedding representations of positive and negative samples respectively.

[0028] The above methods also include:

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

[0030] The above methods also include:

[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 graph convolutional layers;

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

[0033] Optionally, in one embodiment of this application, nodes are clustered through a clustering layer, and an identifier is assigned to each node, including:

[0034] Suppose there are C clusters, with 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 co-contrast encoder;

[0035] A graph convolutional layer is a graph convolutional layer with shared weights, represented as follows:

[0036]

[0037] in, Let A be a normalized adjacency matrix, D be an adjacency matrix with self-loops, and W be the diagonal matrix corresponding to A. c H is the weight matrix in the graph convolutional layer. 1 Let be the embedding representation matrix for all nodes.

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

[0039]

[0040] Where, N schoiar For the reviewer node, N submission For the paper node, p u,i For the target distribution, q u,i For embedding representation h u and cluster center μ i The similarity between them P u,i Indicates the target distribution;

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

[0042]

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

[0044] Optionally, in one embodiment of this 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 reviewed paper nodes.

[0045] The recommender network is a decoder network. It predicts the probability of a review occurring based on the final embedding representation of the corresponding node, including:

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

[0047] The final embedded representation of the reviewer node to be predicted, the final embedded representation of the paper node to be predicted, and the weighted historical review paper nodes are concatenated, and the review probability between the reviewer and the submitter is calculated through a multilayer perceptron.

[0048] Optionally, in one embodiment of this application, calculating the interaction attention between the final embedded representation of the paper node to be predicted and the final embedded representation of the historical review paper nodes, and determining the attention weights of the historical review paper nodes, 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 computed using an outer product, and then decoded using a multilayer perceptron to obtain the attention weights of the historical review paper nodes.

[0050] The attention weights for historical review nodes are represented as follows:

[0051]

[0052] in, It is the bias matrix Dimension (d) 2 The weight matrix is ​​(η,1), where η is the size of the intermediate encoding dimension. h is the final embedded representation of the historical review paper nodes. v This represents the final embedding representation of the paper node to be predicted.

[0053] Optionally, in one embodiment of this 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 connection probability between the paper and the reviewer, the loss gradient is backpropagated to the weight matrix in the decoupled graph convolutional network during training, and all loss functions are multiplied by a coefficient that adjusts them to the same numerical level during training to obtain the total loss function.

[0054] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

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

[0056] Figure 1This is a flowchart illustrating an academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy, as provided in Embodiment 1 of this application.

[0057] Figure 2 This is a diagram illustrating the overall framework of the RevGNN model in an embodiment of this application.

[0058] Figure 3 This is a flowchart illustrating another academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy, as an embodiment of this application.

[0059] Figure 4 This is a schematic diagram illustrating the process of generating two encoders for behavioral preferences and knowledge representations in an embodiment of this application.

[0060] Figure 5 This is a schematic diagram of the network structure of the decoder in an embodiment of this application. Detailed Implementation

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

[0062] The following describes, with reference to the accompanying drawings, an academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy, according to embodiments of this application.

[0063] Figure 1 This is a flowchart illustrating an academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy, as provided in Embodiment 1 of this application.

[0064] like Figure 1 As shown, this academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy includes the following steps:

[0065] Step 101: Obtain 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: Decompose the academic knowledge graph to obtain a bipartite graph representing the interaction between the paper and the reviewers and a knowledge graph describing node information.

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

[0068] Step 104: Fine-tune the fusion embedding representation of each node through graph contrastive learning. In graph contrastive learning, negative samples are extracted using pseudo-labels obtained from node clustering, and positive sample embedding representation learning is performed 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 occurring based on the final embedding representation of the corresponding nodes using the recommendation network.

[0070] This application's embodiment of the academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy employs a two-stage encoder structure to learn comprehensive embedding representations of reviewers and papers. In the first stage, a decoupled GNN is used to encode reviewers' behavioral preferences, while a language model pre-trained based on academic corpora captures prior scientific semantic knowledge of the paper. In the second stage, a pseudo-negative labeling strategy is introduced to address the false negative problem caused by the extreme sparsity of the bipartite graph of the "reviewer-paper" relationship, thereby improving the negative sampling performance during graph contrastive learning. This application's embodiment utilizes an unsupervised pseudo-negative labeling strategy to enhance graph contrastive learning, more effectively selecting nodes whose semantic range is inconsistent with the current node, and providing relative representation information support when recommending academic reviewers.

[0071] This application also provides another academic paper reviewer recommendation method based on an unsupervised pseudo-negative labeling strategy. This method models different types of components in the graph, forming a structure such as... Figure 2 The RevGNN model shown is as follows: Figure 2 As shown, 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: Represent entities that participate in an event.

[0074] Edges: Represent the relationship between two nodes.

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

[0076] (2) Event Encoder

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

[0078] Edge Embeddings: Convert edges into vector representations that include 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: Passing information between nodes and updating node characteristics.

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

[0083] Temporal encoding: Encoding time information into a vector representation.

[0084] Temporal Convolution: Performs convolution operations in the time dimension to extract temporal features.

[0085] (4) Relational Attention Mechanism

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

[0087] Weighted Aggregation: Aggregates node features based on attention weights.

[0088] (5) Aggregation Layer

[0089] Feature fusion integrates features captured from different levels to generate the final representation of nodes and edges.

[0090] (6) Output Layer

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

[0092] Edge Representation: The final vector representation of the generated edges.

[0093] Predictions: Based on the final representation, perform task predictions, such as node classification and link prediction.

[0094] (7) Training and optimization

[0095] Loss function: Used to optimize model parameters and ensure the model's accuracy 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 methods for recommending academic papers by reviewers include:

[0098] First, the input heterogeneous graph is decomposed into a bipartite graph representing the "paper-reviewer" interaction behavior and a knowledge graph describing node knowledge. Depending on the graph type, two specific encoder structures—a behavior preference encoder and a knowledge encoder—are used to learn the basic embedding representations of each node, and then fused together using aggregation operations. Next, pseudo-labels obtained from clustering are used to extract negative samples, while simultaneously learning the embedding representations of positive samples, achieving graph comparison learning in two collaborative comparison encoders. Finally, based on the extraction operations, an interaction-based neural network structure is used to predict edges based on the embeddings of the "paper-reviewer" pairs, thereby achieving academic reviewer recommendations. The entire computation is performed within a single framework. See the overall workflow below. Figure 3 .

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

[0100] Step 1 generates two encoders: behavioral preferences and knowledge representations.

[0101] The input heterogeneous graph is decomposed into two subgraphs: a bipartite graph representing the "paper-reviewer" interaction, which includes only the paper, the reviewer, and the review relationship between them, providing information on the reviewer's potential preferences and habits in past review processes. The other subgraph describes node knowledge, equivalent to the complement of the "paper-reviewer" bipartite graph, providing textual topic information related to the paper and the reviewer.

[0102] Step 2: Aggregation and Pooling of the Two Encoders

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

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

[0105] Using the reviewer and paper representation vectors learned in steps 1 and 2, further refinement is performed based on their relative positions to obtain more accurate representation vectors. Graph contrastive learning is implemented using two collaborative contrastive encoders.

[0106] Step 4 uses an interaction-based neural network structure to predict the connections.

[0107] This is achieved based on the embedding of "paper-reviewer" pairs. After step 3, the paper and reviewer information are fully represented by the model using fixed-dimensional embedding vectors. Subsequently, a relatively simple decoder is used to predict whether there is an edge between any given pair of "paper-reviewer" pairs.

[0108] This embodiment also proposes an academic review and recommendation method using an interactive neural network, taking the interactive neural network as an example. Figure 4 The main issue described is the sparse and semantically ambiguous edge between papers and reviewers. To overcome this problem, this embodiment encodes the knowledge of the paper and the behavioral preferences of the "paper-reviewer" relationship. Specifically, the process includes:

[0109] Step 1: Generate two encoders: behavioral preference and knowledge representation.

[0110] In the first stage of this embodiment, a behavioral preference encoder and a knowledge encoder are used to learn the "paper-reviewer" bipartite graph and the academic knowledge graph, respectively, by optimizing the loss L. behavior To adjust the parameter W in the decoupled GCN layer b Furthermore, it utilizes 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, this invention employs a decoupled graph convolutional network. This decoupling method allows different types of nodes to have different influences during representation learning, better capturing the complex relationships between different nodes in heterogeneous graphs. By independently learning the weight parameters of each node, the decoupled graph convolutional network can better adapt to the feature distribution of different nodes, improving the model's generalization ability. This design makes the model more capable of handling complex graph structures, especially in heterogeneous graphs containing multiple types of nodes.

[0112] By aggregating embedding representations from their neighboring nodes, embedding representation vectors about papers and reviewers are effectively learned. Formalize this encoding process, where the nodes derived from layer l are used This means that, given A is an adjacency matrix with self-loops, and D is the corresponding diagonal matrix, It is a normalized adjacency matrix. The learnable parameter matrix of the first layer is used. Let d represent the dimension of the embedded representation. The probability of an edge connecting the paper to be reviewed (u) to the reviewer (v) can then be obtained through a non-parametric residual product: in Based on the observed edges, the binary cross-entropy is used as the loss function:

[0113]

[0114] Among them, y u,v This 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 for each paper. OAG-BERT's design incorporates a unique two-dimensional positional encoding, allowing heterogeneous entity relation types, masked by entity representations, to be input into BERT. This additional dimension allows for the encoding of information such as entity types, and the embedding representations of this information can also be input into OAG-BERT as prior knowledge.

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

[0117] In this aggregation process, specific strategies are employed to ensure that the information of reviewers and papers is effectively combined. The encoding at this stage aims to convey node information from heterogeneous papers and reviewers by considering their relationships, compressing their representation vectors into a low-dimensional fixed space. This low-dimensional space is designed to retain information directly relevant to the review objective, ensuring that only important information is preserved in subsequent processing, thus improving model efficiency and performance. This process helps extract and retain important features relevant to the review task, thereby better fulfilling the reviewer's 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 that reviewer. This is because a reviewer's research topic can often be reflected through the papers they have previously reviewed. Therefore, by performing average pooling on all papers directly related to the reviewer, a more comprehensive and accurate knowledge embedding representation can be obtained, better capturing the reviewer's research direction and areas of interest. This method helps improve the representation quality of reviewer nodes, thereby enhancing the model's understanding and prediction of reviewer behavior. This process is described in detail below: That is, using vector set concatenation operation to and Two vectors are concatenated. It is the embedded representation vector generated by OAG-BERT or behavioral graph convolutional layers as a knowledge encoder.

[0118] Step 3: Use the pseudo-labels obtained from clustering to extract negative samples, and simultaneously perform positive sample embedding representation learning.

[0119] In graph contrastive learning, the core concept lies in modeling data points through relative representation learning. The key to this method is optimizing the contrastive loss function to make the representations of similar data pairs (positive pairs) closer together, while the representations of dissimilar data pairs (negative pairs) further apart (or, in some cases, more orthogonal). Therefore, developing an effective strategy to extract positive and negative samples from the current node to establish the contrastive relationship between similar and dissimilar data pairs becomes a crucial step in graph contrastive learning.

[0120] Currently, recommendation models based on graph contrastive learning primarily focus on constructing positive samples. While many existing methods focus on constructing negative samples, their main goal is to improve the model's ability to learn more features from challenging negative samples, thereby enhancing the discriminative power between positive and negative samples. In the "paper-reviewer" recommendation process, this embodiment must be careful to avoid negative samples that could lead to false negatives. This means that this embodiment must be careful not to incorrectly label data that could 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 whose pseudo-labels are inconsistent with the current node's, negative samples are used to adjust the node's embedding representation. At this stage, the encoder uses graph contrastive learning to optimize the node representation. Simultaneously, this embodiment also designs a pseudo-negative label sampling strategy to avoid selecting nodes that might lead to false negatives for the current node, ensuring the quality of the formed negative pairs.

[0122] First, positive samples are obtained by modifying the feature embedding representation at each node, and the embedding representation information is propagated on the graph using a graph convolutional layer. Then, L is jointly optimized. contrast and L cluster To adjust the parameter W in the graph convolutional layer c The comparison mutual information is then encoded into the embedded representation of the node.

[0123] To address the false negative negative sample problem, this embodiment introduces an unsupervised learning clustering layer on the node embedding representation and performs joint optimization with a graph contrastive learning framework. Specifically, this invention first uses a graph convolutional layer with shared weights, which is the same as the graph convolutional layer used in positive sampling, and can be represented as follows: H 1 It is the embedding representation matrix of all nodes. This layer is designed to optimize node embeddings through unsupervised learning and clustering, thereby reducing the impact of false negatives. The clustering layer clusters nodes and assigns a unique identifier to each node, helping the model better learn the relationships and features between nodes and improving model performance. Assume there are C clusters, and their cluster centers... In R d The space is randomly initialized, where i represents the clustering at the i-th level. By minimizing L... cluster Each node can be assigned a cluster ID, which can be considered a pseudo-label. cluster The calculation is expressed as:

[0124]

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

[0126] This embodiment uses Represents positive samples. in, W represents the random masking of some features on a given node. c This is the weight matrix, and σ() is the activation function. (Using...) Represents the negative sample set, Where c(u) represents the cluster ID of node u.

[0127] This embodiment is based on the derived positive samples. negative sample set And the original encoded embedding h u A contrastive loss L, similar to InfoNEC, was constructed. contrast The calculation process is as follows:

[0128]

[0129] Step 4: Predict the edges using an interaction-based neural network structure.

[0130] Through the two-stage encoding steps described above, this embodiment learns information about the paper and reviewers, obtaining embedded vector representations with fixed dimensions. Next, a simple decoder extracts these embedded representations to predict whether any given "paper-reviewer" pair has an edge. Considering that the embedded representations include behavior, prior knowledge, and comparative information, this embodiment designs a ranking-based recommendation network. This network can calculate the interactive attention between the current paper and papers previously reviewed by the reviewer, thereby providing more accurate recommendation results. Figure 5 As shown, interactive attention is actually a variant of deep interest networks. A reviewer's review history may include multiple papers; for ease of understanding, the diagram only shows the scenario with only one paper.

[0131] At the beginning of the decoding phase, the decoder network is provided with an embedding representation vector h of a potential reviewer. u The reviewer has already reviewed the paper H. p and the representation vector h of the candidate papers to be submitted. v The decoder network predicts the probability of a review occurring by considering the attention between candidate submissions and papers previously reviewed by reviewers.

[0132] First, in the decoder stage, this embodiment calculates the interactive attention between candidate papers and papers reviewed in the past. For each paper reviewed in the past... and h v The calculation is performed using the outer product, followed by decoding using a multilayer perceptron (MLP): The obtained a i It is considered an attention weight for measuring the influence of papers that have been reviewed in the past. and It is the bias matrix and It is a dimension of (d) 2 The weight matrices are (η,1) and (η,2), where η is the size of the intermediate encoding dimension. The activation function σ is PReLU, chosen with reference to DIN.

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

[0134] Finally, the total loss function of this invention is obtained by multiplying the loss functions of each part by the coefficients that adjust them to the same numerical level and summing them.

[0135] The experiment tested a contrast-enhanced model based on sparse behavior. Relevant peer review information was collected from Frontiers, a publisher with open peer review data, and a large-scale publicly available peer review dataset was constructed for evaluating the model. Five journals with an impact factor greater than 6.0 were selected, and data from 40,578 corresponding papers were retained. The statistical data is shown in Table 1.

[0136] Table 1. Statistical data on the sources of articles collected in the experiment.

[0137]

[0138] Of the 396 articles, 34,166 had only one reviewer, 5,435 had three reviewers, and 505 had four reviewers. Based on this, the experiment split the data into training, validation, and test sets. Ultimately, the training set contained 72,360 interaction records between papers and reviewers, while the validation and test sets each contained 7,344. The collected data was then organized and named Frontiers-4k. Further, the Frontiers-4k dataset was split, with interaction records generated in an 8:1:1 ratio for training, validation, and test sets. During the split, it was ensured that all paper IDs and reviewer IDs appeared at least once in the training set; under this constraint, the remaining data was stored randomly. When testing recommendation performance, for each paper, a reviewer was randomly selected for the test set, and the rest were used as the training set.

[0139] The experiments used the RecBole toolset to load the comparison models for all experiments. Furthermore, all experiments were conducted on an Ubuntu 18.04.2LTS server, equipped with an AMD EPYC 7642 48-core processor, 1008GB of random access memory, and one NVIDIA A100 graphics processor. This invention is implemented using PyTorch 1.9.1 and Python 3.9.7. In the experiments, ReLU was selected as the model... H 2 The activation function, and for Set η = 36. middle and The dimensions are (3d, 200) and (200, 80), respectively, where d = 932. Furthermore, for the first-stage encoder, a 3-layer GCN was used for encoding; for the second-stage encoder, a 1-layer graph convolutional layer was used. Additionally, the model used ADAM with different learning rates to optimize different RevGNN structures. Specifically, for the first-stage encoder and recommendation decoder, the learning rate was set to 0.001; while for the second-stage encoder, the learning rate was set to 0.0001. Experimental results are shown in Table 2. This embodiment achieved state-of-the-art performance on all evaluation metrics across both datasets.

[0140] Table 2 Performance comparison of the present invention and the comparative model on the reviewer recommendation dataset.

[0141]

[0142] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0143] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0144] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0145] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0146] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0147] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0148] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0149] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for recommending academic paper reviewers based on an unsupervised pseudo-negative labeling strategy, characterized in that, include: Obtain a heterogeneous academic knowledge graph, wherein the nodes in the academic knowledge graph include reviewer nodes and paper nodes, and the relationships in the academic graph include review relationships; The academic knowledge graph is decomposed to obtain a bipartite graph representing the interaction between the paper and the reviewers, and a knowledge graph describing node information; The bipartite graph and the knowledge graph are encoded using a two-stage encoder to obtain the first basic embedding representation and the second basic embedding representation of each node. The first basic embedding representation and the second basic embedding representation of each node are then fused together by an aggregation operation to obtain the fused embedding representation of each node. The fusion embedding representation of each node is fine-tuned through graph contrastive learning. During graph contrastive learning, negative samples are extracted using pseudo-labels obtained from node clustering, 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 a recommendation network to predict the probability of review occurring based on the final embedding representation of the corresponding nodes; 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 coding model OAG-BERT. The two-stage encoder encodes the bipartite graph and the knowledge graph respectively to obtain the first basic embedding representation and the second basic embedding representation of each node, including: The bipartite graph is encoded using a behavior preference encoder to obtain the first basic embedding representation of each node; The bipartite graph is encoded using a knowledge encoder to obtain the second basic embedding representation of each node; The method further includes: When encoding using a behavior preference encoder and a knowledge encoder, the parameters in the decoupled graph convolutional network are adjusted by optimizing the loss function of the behavior preference encoder.

2. The method as described in claim 1, characterized in that, The academic knowledge graph is represented as follows: in, A collection of nodes of different types. , For the reviewer's node, For the paper node, , This is a review relationship.

3. The method as described in claim 1, characterized in that, The encoding process of the behavior preference encoder is represented as follows: in, This represents the embedding representation vector of node u in layer l, the first... The node representation matrix derived from the layer is: , , The dimension representing the embedded representation. , The normalized adjacency matrix, Let A be an adjacency matrix with self-loops, and D be the diagonal matrix corresponding to A. In a decoupled graph convolutional network, the first... The learnable parameter matrix of the layer; The loss function of the behavior preference encoder is expressed as: in, Indicates at node and The edges observed between, Papers awaiting review To the reviewer The probability of connecting edges between them. , , .

4. The method as described in claim 1, characterized in that, The first and second basic embedding representations of each node are fused through aggregation operations to obtain the fused embedding representation of each node, as follows: in, This represents the first basic embedding representation for each node. This is the second basic embedding representation for each node.

5. The method as described in any one of claims 2-4, characterized in that, Fine-tuning of the fused embedding representation of each node through graph contrastive learning includes: The graph comparison learning is performed by the first and second collaborative comparison encoders to optimize the fusion embedding representation of each node. The first collaborative comparison encoder processes the data of the bipartite graph and the knowledge graph through the behavior preference encoder and the knowledge encoder respectively, and fuses the processed embedding representations to obtain the basic embedding representation of each node. The second collaborative comparison encoder uses pseudo-labels to extract negative samples. The first and second collaborative comparison encoders learn the embedding representations of positive and negative samples respectively. The method further includes: When performing graph contrast learning through the encoder, nodes that are inconsistent with the pseudo-label of the current node are selected as negative samples. The process of generating pseudo-labels for nodes includes: clustering nodes through a clustering layer and assigning an identifier to each node. The method further includes: 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 graph convolutional layers; The parameters in the graph convolutional layer are adjusted by jointly optimizing the loss functions of the first and second co-contrastive encoders, and the contrastive mutual information is encoded into the embedded representation of the nodes.

6. The method as described in claim 5, characterized in that, The process of clustering nodes through a clustering layer and assigning an identifier to each node includes: Set to exist There are 1 cluster, and the corresponding cluster center is 2. exist 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 co-contrast encoder; The graph convolutional layer is a shared-weight graph convolutional layer, represented as follows: in, , The normalized adjacency matrix, Let A be an adjacency matrix with self-loops, and D be the diagonal matrix corresponding to A. This is the weight matrix in the graph convolutional layer. Let be the embedding representation matrix for all nodes. ; The loss function of the first collaborative comparison encoder is expressed as: in, For the reviewer's node, For the paper node, For the target distribution, , For embedded representation and cluster center Similarity between them , Indicates the target distribution; The loss function of the second collaborative contrast encoder is expressed as: in, Represents positive samples. , This means randomly masking some features on a given node. It is the weight matrix in the graph convolutional layer. Represents the negative sample set, , Represents a node Cluster ID.

7. The method as described in 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 reviewed paper nodes. The recommender network is a decoder network, and the prediction of the probability of review occurring based on the final embedding representation of the corresponding node through the recommender network includes: Calculate the interaction attention between the final embedding representation of the paper node to be predicted and the final embedding representation of the historical review paper nodes, and determine the attention weight of the historical review paper nodes. The final embedded representation of the reviewer node to be predicted, the final embedded representation of the paper node to be predicted, and the weighted historical review paper nodes are concatenated, and the review probability between the reviewer and the submitter is calculated through a multilayer perceptron.

8. The method as described in claim 7, characterized in that, The calculation of the interaction attention between the final embedded representation of the paper node to be predicted and the final embedded representation of the historical review paper nodes, and the determination of the attention weights of the historical review paper nodes, includes: The final embedding representation of each historical review paper node and the final embedding representation of the paper node to be predicted are computed using an outer product, and then decoded using a multilayer perceptron to obtain the attention weights of the historical review paper nodes. The attention weights of the historical review paper nodes are represented as follows: in, , It is the bias matrix. , It is a weight matrix. , The dimensions are respectively , , The size of the intermediate encoding dimension. This is the final embedded representation of the nodes of historical reviewed papers. This represents the final embedding representation of the paper node to be predicted.

9. The method as described in claim 1, 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 connection 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 a coefficient that adjusts them to the same numerical level to obtain the total loss function.

Images