Academic paper reviewer recommendation method based on dynamic balance edge graph

Abstract

The invention discloses an academic paper reviewer recommendation method based on a dynamic balance edge graph, which accurately matches a reviewer through a dynamic balance edge graph technology, and efficiently predicts the reviewer in a complex scene through centroid contrast learning and a dynamic synthesis and deletion mechanism of a tail node. According to the method, through a dynamic balance edge graph technology, the defect that reviewer selection is inaccurate due to link category imbalance is effectively overcome, and the appropriate reviewer can be accurately matched; through centroid contrast learning and a dynamic synthesis and deletion mechanism of the tail node, the robustness of the prediction model in processing the class imbalance problem is significantly improved, efficient reviewer prediction in a complex scene is ensured, and the stability and the overall performance of the prediction model are enhanced at the same time.

Description

Technical Field

[0001] This invention relates to the fields of graph neural networks, link symbol prediction, and recommender systems, and in particular to a peer review recommendation method for academic papers based on a dynamically balanced edge graph, i.e., a method for selecting and recommending reviewers. Background Technology

[0002] With the rapid development of online platforms such as social networks and recommender systems, studying the interaction relationships between users and items has become increasingly important. Existing work has explored these complex relationships in depth using the tool of signed bipartite graphs. In signed bipartite graphs, links with positive or negative signs represent positive or negative relationships, respectively. The analysis tasks cover node classification, node ranking, and symbolic link prediction, with symbolic link prediction being particularly crucial. The diversity of signed bipartite graph analysis has contributed to the advancement of graph representation learning.

[0003] In the field of academic paper reviewer recommendation, there is a similar need to recommend suitable expert reviewers for submitted papers to evaluate their content and quality and provide feedback. Early methods, such as random walks and matrix factorization, primarily focused on the structural properties of graphs. For example, DeepWalk samples node sequences in a graph through random walks as a way to learn node embeddings, while Node2Vec is an extension of DeepWalk. With the development of deep learning techniques, these methods have been applied to learning symbolic graphs. For example, SiNE uses balance theory to construct its objective function to learn embeddings of symbolic networks. SGCN introduces graph convolutional networks to handle symbolic graphs. SNEA introduces a graph attention mechanism to capture complex interactions between nodes. Meanwhile, SGCL uses contrastive learning to optimize sample similarity, learning robust node representations that can cope with noise and data sparsity challenges. Furthermore, some methods are specifically designed for analyzing symbolic bipartite graphs and can be used for paper reviewer prediction tasks. For example, Derr et al. introduced the concept of symbolic butterfly isomorphism. SBGNN addresses the balance problem in symbolic bipartite networks by modifying the message passing mechanism. SBGCL employs an innovative two-layer graph data augmentation method to improve the representation learning performance of signed bipartite graphs. However, when using signed bipartite graphs to predict paper reviewers, existing methods still face the link class imbalance problem, which has not yet been effectively resolved.

[0004] In graph representation learning, class imbalance is a common problem. In real-world graph data, researchers often observe that the number of nodes in the head class is significantly greater than the number of nodes in the tail class. This imbalance causes models to perform well in recognizing head classes but poorly in recognizing tail classes. To address this issue, two strategies are typically employed. The first strategy is to adjust the loss function to increase the focus on tail classes during model training. For example, the reweight method improves the model's ability to recognize tail classes by assigning different weights to samples from different classes. The second strategy is to use generative methods to balance the dataset by synthesizing nodes with fewer samples in the tail class. For example, GraphSmote synthesizes tail class nodes using interpolation techniques and generates new edges to achieve data balance. DR-GCN and ImGAGN utilize generative adversarial networks to synthesize tail class nodes. GraphENS attempts to extend the decision boundary of tail classes by synthesizing subnetworks, considering the features of all samples in the graph. TailGNN enhances the feature representation of tail nodes by changing the neighborhood structure and transferring knowledge from head class nodes to tail class nodes. GraphSHA expands the decision boundary by synthesizing more challenging tail class samples and ensures that this boundary information propagates only within the tail class, avoiding interference with the data space of adjacent classes. Although the above methods address the class imbalance problem to some extent, they do not effectively solve the extreme imbalance phenomenon in the side graph. This leads to the tail class node's feature space being over-compressed during paper reviewers' predictions, thus affecting the accuracy of predictions and recommendations.

[0005] To address the need for peer review prediction and recommendation in academic research, existing work typically utilizes bipartite graph models, particularly signed bipartite graphs. In a signed bipartite graph, nodes are divided into two disjoint sets, such as users and products, or reviewers and papers. Connections between sets are represented by edges with positive and negative signs; positive edges represent positive reviews, and negative edges represent negative reviews. Link sign prediction is a crucial task in the study of signed bipartite graphs. Although existing methods have made some progress, the problem of link class imbalance persists. In certain datasets, the number of links of one type (head class) far exceeds that of another (tail class). This imbalance causes the model to favor the head class during training, leading to inaccurate predictions. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this invention provides a Line-Graph-Based Dynamic Balancing Prediction (LDBP) method for predicting academic paper reviewers. This method selects reviewers by accurately matching them using Balanced Line Graph (BaLG) technology. Furthermore, through centroid contrastive learning and dynamic synthesis and deletion mechanisms of tail nodes included in the dynamic balanced line graph, it achieves efficient prediction of reviewers in complex scenarios and improves robustness.

[0007] The peer review prediction method for academic papers based on dynamic equilibrium edge graphs proposed in this invention includes the following steps:

[0008] 1) Construct a signed bipartite graph using existing academic papers and corresponding reviewer datasets. In the bipartite graph, nodes represent papers and reviewers, forming two disjoint sets of nodes. Signed edges represent the evaluation relationships between the two sets of nodes, with positive signs indicating positive evaluations and negative signs indicating negative evaluations.

[0009] 2) Node Embedding Pre-training: The signed bipartite graph and its positive and negative subgraphs are input into a shared-weight graph neural network (GNN) model, generating feature representations for each node in both the signed bipartite graph and its positive and negative subgraphs. These feature representations are then concatenated along the column direction. A learnable linear transformation and bias are used to further generate the feature representation for each node. Finally, a single-layer perceptron (MLP) is used to predict the edge signs. A binary cross-entropy loss function is used to measure the difference between the probability distribution of the edge signs predicted by the single-layer perceptron model and the actual edge signs (actual link signs) in the bipartite graph, thereby optimizing the vectors encoded by nodes in high-dimensional space and the learnable linear transformation and bias. This invention's graph neural network (GNN) model is based on existing publicly available technologies; however, this invention introduces unique improvements and innovative processing, including joint modeling of the bipartite graph and its positive and negative subgraphs and the introduction of learnable linear transformations and biases.

[0010] 3) Edge Graph Transformation and Node Feature Initialization: Based on the generated signed bipartite graph representation, it is transformed into the corresponding edge graph representation. Each node category in the edge graph includes a head class and a tail class. The generated node embeddings are transferred to the edge graph as features for initializing node embeddings. This invention naturally transforms the link class imbalance in the bipartite graph into the node class imbalance in the edge graph.

[0011] 4) Centroid Contrastive Learning: For each node category (head and tail) in the edge graph, the mean of its node feature vectors is calculated and used as the centroid of that category. A uniform intra-class loss function is designed to ensure that all node feature vectors within the same category converge as closely as possible to the centroid of that category. Simultaneously, an inter-class separation loss is used to push the centroids of different categories apart in the feature space, maximizing the distance between them. During training, the model parameters are adjusted to minimize the comprehensive loss function, thereby improving the model's accuracy in predicting node categories. The centroid contrastive learning method can more effectively adjust the decision boundary and avoid bias caused by class imbalance.

[0012] 5) Dynamic Synthesis and Deletion of Tail Nodes: Based on sampling probabilities, N nodes are extracted from both the head and tail classes to form node pairs. New nodes are then generated through feature fusion. These new nodes are randomly connected to other nodes in the edge graph, updating the adjacency matrix. The original node features and the new node features are then input into the graph neural network model. Subsequently, the centroids of the head and tail classes are recalculated, and the distance between the synthesized node and the centroids of the head and tail classes is used to determine if the new node has been misclassified as a head node. If misclassified, it is deleted, and the node synthesis process is repeated to dynamically adjust the node distribution in the edge graph. This process is executed once per training cycle.

[0013] 6) Inference and Reviewer Prediction: During the inference phase, only the original set of nodes in the side graph is used for inference, without considering synthesized nodes. The model calculates a score for each possible reviewer-paper pair, based on the node's embedding representation and patterns learned from the training data. Then, the candidates are ranked according to this score, and the top-ranked candidate is selected as the reviewer recommended to the paper, ultimately achieving accurate reviewer recommendations for academic papers.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0015] This invention provides a method for recommending academic paper reviewers based on a dynamic equilibrium edge graph, which has significant technical advantages including:

[0016] (i) This invention effectively overcomes the problem of inaccurate reviewer selection caused by the imbalance of link categories in traditional methods by using dynamic balanced edge graph technology. It can accurately match suitable reviewers and significantly improve the fairness of the paper review process and the reliability of the review results.

[0017] (ii) By using centroid contrastive learning and dynamic synthesis and deletion mechanism of tail nodes, the robustness of the model in dealing with class imbalance problem is significantly improved, ensuring efficient reviewer prediction in complex scenarios, while enhancing the stability and overall performance of the model. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method procedure of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the implementation process of an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of the invention is not limited in any way.

[0021] The selection of peer reviewers for academic papers is a crucial step in the publication process. While existing recommendation systems and artificial intelligence technologies have improved the efficiency of reviewer selection to some extent, these methods typically rely on historical data of known reviewers and matching of their expertise in their respective domains, thus limiting the selection pool. This invention, however, proposes a new generation of reviewer selection methods that focus on generating and identifying suitable potential reviewers through more advanced graph neural networks to better address complex review needs. The selection of peer reviewers for academic papers has broad application prospects, particularly in ensuring the quality and impartiality of the review process. On the one hand, traditional reviewer selection methods rely on human judgment and subjective experience, easily leading to low review efficiency and inconsistent review quality. The reviewer selection method based on dynamic equilibrium side graph technology provided in this invention can effectively improve the scientific rigor and accuracy of reviewer selection by analyzing a large amount of historical review data and paper citation relationships.

[0022] Computer-aided reviewer selection methods not only play a crucial role in academic paper review but can also be applied to other fields requiring expert review and recommendations, such as patent examination, research project review, and academic conference submission review. The new generation reviewer selection system provided by this invention utilizes dynamic equilibrium edge graph technology and centroid contrastive learning, effectively addressing class imbalance issues and adapting to the needs of different review tasks through dynamic strategy adjustments, ensuring the fairness and accuracy of the review process.

[0023] This invention proposes a peer review prediction method for academic papers based on a dynamically balanced edge graph. The core idea is to transform links in a bipartite graph into nodes in an edge graph, and to solve the link class imbalance problem through dynamic edge graph technology, thereby achieving accurate selection of reviewers. In this invention, the dynamically balanced edge graph includes two dynamic processes for edge graph operations: centroid comparison learning and dynamic synthesis and deletion of tail nodes.

[0024] Figure 1The diagram illustrates the flowchart of the peer review prediction model (LDBP) for academic papers based on dynamically balanced edge graphs provided by this invention. This method comprises three modules: node embedding pre-training, centroid contrastive learning, and dynamic synthesis and deletion of tail nodes. First, a bipartite graph of papers and experts is constructed, and the original bipartite graph is pre-trained using a graph neural network to generate high-dimensional node embeddings, capturing the interaction patterns between nodes. Second, the links in the bipartite graph are transformed into nodes in the edge graph, and the centroid of each category is calculated using the centroid contrastive learning module. Simultaneously, intra-class uniformity loss and inter-class separation loss are introduced to enhance intra-class cohesion and inter-class discriminability. Finally, through a dynamic synthesis and deletion strategy, the feature space of tail category nodes is expanded to reduce model bias caused by class imbalance, thereby obtaining a dynamically balanced edge graph (BaLG). In paper reviewer prediction, only the original set of nodes in the edge graph is inferred; synthesized nodes are not considered. The inference result is the candidate paper reviewer predicted by the model.

[0025] This invention effectively solves the problem of reviewer selection in peer review of academic papers by introducing dynamic equilibrium edge graph technology and centroid contrastive learning. This method not only improves the accuracy and efficiency of reviewer selection but also enhances the work efficiency of paper administrators and the quality of paper review.

[0026] The specific implementation process of the dynamic equilibrium edge graph-based peer review prediction model (LDBP) method for academic papers proposed in this invention is as follows: Figure 2 As shown, it includes the following steps.

[0027] 1) Construction of a Signed Bipartite Graph Model for Papers and Reviewers

[0028] Collect text datasets of academic papers and their corresponding reviewers, and construct a symbolic bipartite graph.

[0029] Specifically, based on the text dataset of academic papers and their corresponding reviewers, the relationship between papers, reviewers, and review comments is represented as a signed bipartite graph, where papers and reviewers are two distinct sets of nodes. Positively marked edges represent positive evaluations or recommendations from reviewers, while negatively marked edges represent negative evaluations or rejections. In this way, the bipartite graph clearly reflects the interaction between each reviewer and the papers they review. This signed bipartite graph can be represented as follows:

[0030] G = (U, V, E, X) b )

[0031] Here, the graph G consists of two disjoint sets of nodes, U and V, representing the two parts of the bipartite graph. Set U includes the paper nodes (u1, u2, ..., u...). |U|}, and the set V includes reviewer nodes {v1, v2, ..., v}. |V| Let E represent a signed edge between a paper node and a reviewer node. Different symbols represent two different evaluation opinions from the reviewers. A positive edge indicates a positive evaluation or a good match between the reviewer and the paper. For example, in practical applications, if a reviewer has given a positive review opinion to a paper (such as "accept" or "accept after revision"), the edge is marked as a positive edge. A negative edge indicates a negative evaluation or a mismatch between the reviewer and the paper. If a reviewer believes that the paper is of poor quality or does not conform to their research field (such as "reject"), the edge is marked as a negative edge. In a bipartite graph, edges are considered undirected, indicating that the relationships they represent are bidirectional and directionality is not considered. X b Then, X represents the feature matrix associated with these nodes and edges. b Each row in the array represents a node or an edge, while the columns correspond to different characteristics of the nodes or edges.

[0032] 2) Pre-training of node embeddings for signed bipartite graphs

[0033] After representing the relationship between papers and reviewers as a signed bipartite graph, node embedding pre-training is performed. The purpose of pre-training is to generate high-dimensional representations of each node, which can capture the interaction patterns between nodes and provide guidance for subsequent edge graph training. In the first constructed positive and negative subgraphs, for the bipartite graph G containing positive and negative links, the graph is divided into two subgraphs according to the sign of the links: a positive subgraph G+ (containing positive links (positive sign edges)) and a negative subgraph G- (containing negative links (negative sign edges)), specifically represented as follows:

[0034] G + = (U, V, E) + X b )

[0035] G - = (U, V, E) - X b )

[0036] Among them, E + E represents the set of all positively marked edges, where each positively marked edge represents a reviewer's positive evaluation or matching relationship to the paper. - This represents the set of all negative edge symbols, where each negative edge represents a reviewer's negative evaluation of the paper or a mismatch.

[0037] Next, the bipartite graph G and the positive subgraph G + and negative subgraph G - This is input into the graph neural network (GNN) model. Specifically, it is represented as:

[0038]

[0039] in This represents the feature representation of node v in the l-th layer of a graph neural network (GNN). Through interaction with the features of its neighboring nodes, it gradually integrates more graph structure information. It can capture the local relationships and contextual semantic information of nodes in the graph structure. It represents the feature representation of node v at layer l-1. v θ represents the set of direct neighbors of node v. l θ represents the learnable parameters of the l-th layer of a graph neural network, including the weight matrix and bias terms. It controls the linear and nonlinear transformations of the node features in the l-th layer. By training the model, the loss function can be minimized, and the optimal θ can be learned. l This allows for the generation of more effective node feature representations.

[0040] In this invention, for the bipartite graph G and the positive subgraph G + and negative subgraph G - The input GNNs for each graph share parameters during parameter updates. The GNN processes the bipartite graph G and the subgraph G... + and negative subgraph G - Feature representations are generated separately for G+ and G-. These feature representations are then concatenated with the column representations of the original graph G to form a comprehensive feature matrix. The new feature matrix is ​​represented as follows:

[0041] H o ∈R (|U|+|V|)×d′

[0042] Among them, H o Let H be the new feature matrix, where the i-th row represents the embedding representation of the i-th node in the graph (whether it's a paper node or a reviewer node) in the new feature space. Each node's embedding representation contains comprehensive information about the node's own features and the features of its neighbors; d′ represents the dimension of the node's embedding representation in the high-dimensional space, a new and more expressive feature space dimension generated after learning by the GNN model, determined by the network structure and output dimension of the GNN model; |U| and |V| represent the number of nodes in the set of paper nodes U and the set of reviewer nodes V in the bipartite graph, respectively. The entire matrix H... o The number of rows is the sum of the number of all paper and reviewer nodes, |U|+|V|, representing all nodes in the bipartite graph G.

[0043] We will use the same method as for Figure G to represent Figure G. + And Figure G - Obtain model output (H) o (+)∈R (|U|+|V|)×d′ ) and (Ho (-) ∈ R (|U|+|V|)×d′ )。

[0044] Based on the spliced feature matrix, through a learnable linear transformation and bias, the edge features are further generated, specifically represented as:

[0045] feature representation, specifically:

[0046] H b = [H o || H o (+) || H o (-)] W o + b o

[0047] where, [||] represents the column-wise splicing operation, and b o ∈ R d′ are learnable parameters.

[0048] We generate a feature representation for each edge, and then use a single-layer perceptron, Single-Layer Perceptron (SLP) to predict the sign of the edge. Taking the edge (e k = (u i , v j )) as an example, we can calculate the prediction probability of the edge sign through the following formula:

[0049]

[0050] where, z k is used to represent the prediction probability that the edge e k is a specific sign (such as a positive sign or a negative sign). Specifically, z k is a two-dimensional vector, and the first component in z k represents the probability that e k is a positive sign, while the second component represents the probability that e k is a negative sign; represents splicing the feature representations of nodes i and j to form a new feature vector. W z is a learnable weight matrix that maps the spliced node feature representations to a two-dimensional space (i.e., two sign categories) to calculate the score of the relationship between two nodes. b z is a learnable bias term used to adjust the transformation result to ensure that the model can better fit the distribution of the training data. <XXX

[0051] Simultaneously, the binary cross-entropy loss function is used to measure the difference between the distribution of the model's predicted probabilities and the actual link symbols, thereby optimizing the high-dimensional feature representation and learnable parameters of the bipartite graph nodes. More specifically, the high-dimensional feature representation of the paper nodes and the reviewer nodes is optimized. Specifically, this is expressed as follows:

[0052]

[0053] Among them, L pre A loss function that measures the difference between the edge symbols predicted by the model and the actual edge symbols; y k ∈Y represents edge e k The actual symbol of ∈E, while Indicates that the model predicts edge e k The probability value of a positive sign. The model predicts edge e k The probability value for a negative sign. L pre By examining all edges e k The difference between the predicted and actual symbols is calculated to guide model optimization.

[0054] After pre-training, the generated high-dimensional feature representations of the nodes are transferred to the edge graph and used as features to initialize node embeddings. Through this pre-training step, the LDBP method can provide a rich feature representation foundation for the nodes of the edge graph. These feature representations can capture the complex relationships between nodes in the original bipartite graph, laying a solid foundation for subsequent link symbol prediction tasks.

[0055] 3) Construction of edge graphs and comparison of their centroids

[0056] For the construction of the edge graph, based on our generated signed bipartite graph representation, it is converted into the corresponding edge graph representation; specifically, the high-dimensional feature representation of the nodes generated in the previous pre-training step is transferred to the edge graph as the high-dimensional feature representation of the nodes used to initialize the edge graph. This invention naturally transforms the link class imbalance in the bipartite graph into node class imbalance, specifically represented as follows:

[0057] G = (V, E, X)

[0058] Each node corresponds to an edge in the original bipartite graph, so the set of nodes in the edge graph is V = {N(e) | e∈E}. Here, N(·) represents the transformation of an edge in the bipartite graph into a node in the edge graph. The class of a node in the edge graph is determined by the sign of the edge in the original bipartite graph. The head and tail classes of nodes are represented as Vhead and Vtail, respectively. h and V t If the two edges e in the original graph i and e j If nodes share a common node, then in the edge graph G, node N(e i) and N(e j An edge will be created between them.

[0059] To address the misclassification problem caused by excessive compression of the feature space of tail-class nodes due to class imbalance, this invention calculates the centroid of each class and introduces intra-class uniformity loss and inter-class separation loss to enhance intra-class cohesion and inter-class discriminability. For each node class (head class and tail class) in the edge graph, the mean of its node feature vector is calculated and used as the centroid of that class. Specifically, it is expressed as follows:

[0060]

[0061] in and Representing the head class node N(e) i ) and tail-type node N(e j The high-dimensional eigenvector representation of ).

[0062] For the intra-class uniform loss function, it ensures that the feature vectors of all nodes within the same class converge as closely as possible to the centroid of that class. This measure is used to assess the distance between the high-dimensional feature representation of nodes within the same class and the centroid of that class. This helps to strengthen the cohesion of nodes within the class, that is, to make nodes of the same class cluster more tightly in the feature space, specifically expressed as:

[0063]

[0064] Where |V| represents the total number of nodes in the graph. To ensure the loss function value is independent of the number of nodes and prevent loss value shift due to differences in the number of nodes; k∈{h,t}, k represents the class index, h represents the head class, and t represents the tail class. This formula iterates over and accumulates all classes (head and tail classes) to calculate the loss of all nodes within a class; N(e i )∈V k ,N(e i ) represents edge e i The set of connected nodes, V k Represents the set of nodes in category k; This represents the high-dimensional representation of the i-th node in category k in the feature space; This distance measures how close node i is to the centroid of the class; the smaller the distance, the closer the node is to the centroid of the class. This represents the relative similarity of node i to category k, and measures the centroid C between node i and category k. k The distance between them is the proportion of the centroid distances of all categories, and scale compression is performed by taking the logarithm.

[0065] For inter-class separation loss, the centroids of different classes are pushed apart in the feature space to maximize the distance between them. This helps improve inter-class discriminability, making nodes of different classes easier to distinguish in the feature space, specifically expressed as:

[0066] L inter =exp(-d(C h C t ))

[0067] This term represents the centroid C of the head type. h The center of mass of the tail class C t The distance d(C) between them h C t The exponential transformation of d(C). h C t The Euclidean distance is used to measure the difference between the two centroids.

[0068] During training, optimization techniques such as backpropagation and gradient descent are used to adjust model parameters to minimize the overall loss function. This helps improve the model's accuracy in predicting node categories. The centroid contrastive learning method can more effectively adjust the decision boundary, avoiding bias caused by class imbalance. Especially when the number of head class nodes far exceeds the number of tail class nodes, centroid contrastive learning helps reduce excessive preference for the head class.

[0069] By employing centroid contrastive learning, the LDBP method effectively addresses the class imbalance problem in edge graphs, improving the accuracy of link symbol prediction, especially when dealing with extremely imbalanced data distributions. This approach provides a novel perspective for handling complex graph data and can be extended to other scenarios requiring the resolution of class imbalance issues.

[0070] 4) To address the class balance problem, dynamically combine and delete tail nodes.

[0071] We need to pay special attention to tail category nodes because they are easily compressed in the feature space, leading to a decrease in classification accuracy. To avoid exacerbating the class imbalance problem, this invention designs a dynamic synthesis and deletion strategy for tail category nodes to obtain a balanced edge graph. Specifically, the centroids of the head and tail classes are recalculated, and it is determined whether a new node is misclassified as a head class. If correctly classified, the node is retained; otherwise, it is deleted to dynamically adjust the node distribution in the edge graph. This process is performed once in each training cycle.

[0072] Specifically, firstly, based on the similarity between nodes and their respective class centroids, the sampling probabilities of head and tail class nodes are calculated. The sampling probability is exponentially related to the distance from the node to the centroid, specifically expressed as follows:

[0073]

[0074] This represents the sampling probability that the head node i is selected. This represents the sampling probability of tail-class node j being selected. The sampling probability is positively correlated with the distance between the node and its class centroid: the greater the distance, the higher the sampling probability. This strategy helps to select nodes that are far from the centroid, i.e., nodes closer to the decision boundary, thus generating synthetic nodes that are more challenging for the model. Represents the high-dimensional feature vector of head node i. With the center of mass C of the head h The distance between them; Represents the embedding feature vector of tail class node j With tail-like centroid C t The distance between nodes. A larger distance indicates that the node is closer to the decision boundary and has a higher sampling probability. The denominator represents the exponential sum of the distances between all nodes in the head or tail class and the centroid, used to normalize the sampling probability of the nodes to ensure that the sum of the sampling probabilities of all nodes is 1.

[0075] Then, based on the sampling probability, N nodes are extracted from the head class and the tail class respectively, forming N pairs of nodes. For each pair of nodes, the features of the new node are synthesized through a mixing operation in the feature space, specifically represented as follows:

[0076]

[0077] in, Represents the composite node N syn (e k )∈V syn The characteristic is that δ is a value sampled from a uniform distribution from 0 to 1.

[0078] The synthesized nodes are randomly connected to other nodes in the edge graph to construct a new adjacency matrix, thus integrating the synthesized nodes into the edge graph. Then, feature updates are performed by inputting the edge graph with the added synthesized nodes into the graph neural network model to obtain the updated node features.

[0079] Then, the centroids of the head and tail classes are recalculated using the updated node features. Based on the distance between the synthesized node and the centroids of the head and tail classes, it is determined whether the synthesized node has been incorrectly classified as the head class, specifically as follows:

[0080]

[0081] Here, λ is a very large positive constant. When f k When = 1, node N syn (e k )∈V synNodes that are misclassified as head nodes are removed. We then re-synthesize nodes to maintain a total number of N nodes. This process is repeated once per training cycle.

[0082] Through this dynamic synthesis and deletion strategy, the LDBP method can effectively expand the feature space of tail class nodes, reduce model bias caused by class imbalance, and thus improve the accuracy of link symbol prediction.

[0083] 5) Overall training of the LDBP prediction model

[0084] The model is trained in two phases: the first phase uses pre-trained node embeddings as features for training, and the second phase uses intra-class uniform loss and inter-class separation loss for training.

[0085] In the first phase of training, the model uses pre-trained node embeddings as input features. These features are high-dimensional representations generated by analyzing node interaction patterns in the original bipartite graph and utilizing a graph neural network. The goal of this phase is to teach the model how to use these node embeddings to capture the complex structures and relationships between nodes in the graph. By inputting the pre-trained node embeddings into the model, it gains a deep understanding of node characteristics and their interactions early in the training process, laying a solid foundation for subsequent training. The completion of the pre-training phase provides the model with a good starting point, enabling it to more accurately capture and understand the positive and negative relationships between nodes when handling more complex link symbol prediction tasks.

[0086] In the second phase of training, the cross-entropy loss function is first used for node classification:

[0087]

[0088] Among them, L rw To measure the model's performance on the symbolic edge classification task, it addresses the class imbalance problem by weighting the positive and negative classes separately, thereby reducing the model's bias towards the head class (majority class) nodes; k ∈Y represents node N(e k The actual symbol for ) is , where 0 represents a negative node and 1 represents a positive node. p k The i-th element. pos and w neg These represent the weights of the positive and negative classes, respectively.

[0089] Intraclass uniform loss L intraThe goal is to encourage nodes of the same category to cluster around their centroids in the feature space, thereby enhancing the model's cohesion towards its internal nodes. In this way, the model learns to better distinguish between nodes of different categories, even if they are very close together in the feature space.

[0090] Inter-class separation loss L inter This process promotes the separation of centroids of different classes in the feature space, increasing the distance between them. This helps the model to more clearly distinguish between different classes during classification, especially when there is overlap between classes. By maximizing the distance between centroids, the model can create a more defined decision boundary, reducing classification errors.

[0091] In the second stage, the model is trained end-to-end using all three loss functions. By adjusting the weight coefficients in the loss functions, the model can balance the contributions of each to the training process to achieve optimal classification results. This combined approach not only improves the model's ability to identify tail categories but also maintains high accuracy for head categories, thereby improving the overall performance of the model in predicting link symbols in bipartite graphs constructed by the paper and reviewers. Specifically:

[0092] L = L rw +αL inter +βL intra

[0093] Where α and β are loss coefficients.

[0094] 6) Inferences and the reviewers' predictions

[0095] When this model is applied to a dataset of peer review papers, the inference results are the candidate reviewers predicted by the model.

[0096] During the inference phase, inference is performed only on the original set of nodes in the edge graph, without considering synthetic nodes. The model calculates a score for each possible reviewer-paper pair, based on the node's embedding representation and patterns learned from the training data. Then, the candidates are ranked according to this score, and the top-ranked candidate is selected as the reviewer to recommend to the paper.

[0097] The LDBP model focuses on analyzing the original set of nodes in the edge graph, excluding nodes synthesized during training to ensure the accuracy and reliability of inferences. At this point, the model has undergone two phases of training and possesses the ability to effectively predict node link symbols. Applying this model to a peer review dataset means that it can be used to predict who is most likely to become a reviewer for a particular paper. Based on the node features and link patterns learned during training, the model assesses the relevance of potential reviewers to the paper's topic and the interactions between them.

[0098] In practice, the model calculates a score for each possible reviewer-paper pair, based on the node's embedding representation and patterns learned from the training data. Then, it ranks the candidates according to this score, selecting the top-ranked candidate as the recommended reviewer for the paper. This approach not only improves the efficiency of the review process by quickly selecting the most suitable reviewer from a large number of possible candidates, but also enhances the quality of the review process because the recommendation system is based on deep data understanding and complex graph structure analysis. Ultimately, the model's inference results in a list of candidate paper reviewers, which can serve as an important reference for conference organizers or journal editorial departments in assigning paper reviews.

[0099] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A method for recommending academic paper reviewers based on a dynamic equilibrium edge graph, characterized in that, By employing dynamic balanced edge graph technology to accurately match reviewers, and through centroid contrastive learning and a dynamic synthesis and deletion mechanism for tail nodes, efficient reviewer prediction is achieved in complex scenarios. The specific steps include: 1) Using academic papers and their corresponding reviewer datasets, construct a signed bipartite graph model of paper-reviewer to represent the evaluation interaction between each reviewer and the papers they review; the signed bipartite graph includes a positive subgraph and a negative subgraph; the nodes in the bipartite graph represent papers and reviewers respectively, and the sets of paper and reviewer nodes are two disjoint sets; the signed edges in the bipartite graph represent the evaluation connections between nodes of different classes, and the signs include positive and negative signs. 2) Perform node embedding pre-training to generate a high-dimensional representation of each node; A signed bipartite graph and its positive and negative subgraphs are input into a graph neural network with shared weights to generate feature representations for each node in the signed bipartite graph and its positive and negative subgraphs. The feature representations are then concatenated along the column direction, and edge feature representations are further generated through a learnable linear transformation and bias. A single-layer perceptron is used to predict the sign of the edges, and a binary cross-entropy loss function is used to measure the difference between the probability distribution predicted by the model and the actual link sign, thereby optimizing the high-dimensional representation of nodes and the learnable parameters. 3) Based on the signed bipartite graph, perform edge graph transformation and initialize the node features of the edge graph: Based on the generated signed bipartite graph representation, it is converted into the corresponding edge graph representation; each node in the edge graph includes a head class and a tail class; the generated node embeddings are transferred to the edge graph as features for initializing node embeddings; the imbalance of link classes in the bipartite graph is transformed into imbalance of node classes in the edge graph. 4) Adjusting decision boundaries through centroid comparison learning: For each node category in the edge graph, calculate the mean of the node feature vectors and use it as the centroid of that category; The design incorporates a uniform intra-class loss function, causing all node feature vectors within the same class to converge towards the centroid of that class. Simultaneously, an inter-class separation loss is used to push the centroids of different classes apart in the feature space, maximizing the distance between them. During training, the model parameters are adjusted to minimize the comprehensive loss function, thereby improving the model's accuracy in predicting node classes. 5) Dynamic composition and deletion of tail nodes: Based on the sampling probability, N nodes are extracted from the head class and the tail class respectively to form node pairs, and new nodes are generated through feature mixing operations; The new node is randomly connected to other nodes in the edge graph, the adjacency matrix is ​​updated, and the original node features and the new node features are input into the graph neural network model. Subsequently, the centroids of the head and tail classes are recalculated, and the distance between the synthesized node and the centroids of the head and tail classes is used to determine whether the new node is incorrectly classified as the head class. If a node is misclassified, it is removed and the node synthesis process is repeated to dynamically adjust the node distribution in the side graph; this process is performed once per training cycle. 6) Predicting reviewers: During the inference phase, inference is performed using only the original set of nodes in the side graph, without considering synthesized nodes; for each possible reviewer-paper pair, a score is calculated based on the node embedding representation and the pattern learned from the training data. The papers are ranked according to their scores, and the top candidates are selected as reviewers to recommend them to academic papers, thus achieving accurate reviewer recommendations.

2. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 1, characterized in that, Positive signs indicate positive evaluations, and negative signs indicate negative evaluations; edges with positive signs indicate positive evaluations or recommendations from reviewers, while edges with negative signs indicate negative evaluations or rejections.

3. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 2, characterized in that, A marked bipartite graph is represented as: G = (U, V, E, X) b ); where graph G consists of two disjoint sets of nodes U and V, representing the two parts of the bipartite graph respectively; set U includes the paper nodes {u1, u2, ..., u...} |U| The set V includes reviewer nodes {v1, v2, ..., v}. |V| }, E represents the signed edge between the paper node and the reviewer node; X b X represents the feature matrix associated with nodes and edges. b Each row in the array represents a node or an edge, and the columns correspond to different characteristics of the nodes or edges.

4. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 3, characterized in that, In a signed bipartite graph, the positive subgraph G+ containing positive links and the negative subgraph G- containing negative links are represented as follows: G + =(U,V,E + ,X b ) G - =(U,V,E - ,X b ) Among them, E + E represents the set of all positively signed edges; - This represents the set of all edges with negative signs.

5. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 4, characterized in that, Let the bipartite graph G and the normal subgraph G be... + and negative subgraph G - When input into a graph neural network model, it is represented as: in, This represents the feature representation of node v in the l-th layer of the graph neural network; N represents the feature representation of node v at layer l-1; v θ represents the set of direct neighbors of node v; l θ represents the learnable parameters of the l-th layer of the graph neural network, including the weight matrix and bias terms; the model is trained to minimize the loss function and learn the optimal θ. l Values ​​are used to generate valid node feature representations. For a bipartite graph G and a normal subgraph G + and negative subgraph G - Each input graph neural network model shares parameters during parameter updates.

6. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 5, characterized in that, Using a graph neural network model to analyze bipartite graph G and subgraph G + and negative subgraph G - The feature representations of G+ and G- are generated separately, and then concatenated with the feature representations of the original graph G along the column direction to form a comprehensive feature matrix, represented as follows: H o ∈R (|U|+|V|)×d′ Among them, H o The new feature matrix has its i-th row representing the embedding representation of the i-th node in the graph within the new feature space; each node embedding representation contains combined information about the node's own features and the features of its neighbors; d ′ is the dimension of the node embedding representation in the high-dimensional space, which is the dimension of the feature space generated after model learning; |U| and |V| represent the number of nodes in the paper node set U and the reviewer node set V in the bipartite graph, respectively; matrix H o The number of rows is the sum of the number of all paper and reviewer nodes |U|+|V|, representing all nodes in the bipartite graph G; Figure G + And Figure G - The model output is obtained and represented as: (H) o (+)∈R (|U|+|V|)×d′ ) and (H o (-)∈R (|U|+|V|)×d′ ).

7. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 6, characterized in that, The feature representation of the generated edges is as follows: H b =[H o ||H o (+)||H o (-)]W o +b o Where [||] indicates column concatenation. and b o ∈R d′ These are learnable parameters; To predict the sign of an edge using a single-layer perceptron, the predicted probability of the edge sign is calculated using the following formula: Among them, z k It is a two-dimensional vector used to represent edge e. k The predicted probability of a positive or negative sign; e k =(u i ,v j ); vector z k The first component in the equation represents e. k The probability of being positive, the second component represents e. k The probability of having a negative sign; This indicates that the feature representations of node i and node j are concatenated to form a new feature vector; W z It is a learnable weight matrix; b z It is a learnable bias term; By examining all edges e k Calculate the difference between the predicted symbol and the actual symbol, and optimize the high-dimensional feature representation of the paper node and the reviewer node, as follows: Where, y k ∈Y represents edge e k The actual symbol of ∈E, Indicates that the model predicts edge e k The probability value of a positive sign, 1- The model predicts edge e k The probability value of a negative sign; L pre By examining all edges e k The difference between the predicted and actual symbols is calculated to guide model optimization.

8. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 7, characterized in that, Step 3) involves constructing the edge graph, which includes: The high-dimensional feature representations of nodes generated during pre-training are transferred to the edge graph and used as the initial high-dimensional feature representations of nodes in the edge graph; the imbalance of link classes in the bipartite graph is transformed into imbalance of node classes; the edge graph is represented as follows: G = (V, E, X) The set of nodes in the edge graph is V = {N(e) | e ∈ E}; N(·) represents converting the edges in the bipartite graph into nodes in the edge graph; the node categories in the edge graph include head and tail classes; Calculate the mean of the node feature vectors for each node category in the edge graph, and use it as the centroid of that category, expressed as: in, and Representing the head class node N(e) i ) and tail-type node N(e j High-dimensional eigenvector representation of ). The intra-class uniform loss function is expressed as: Where |V| represents the total number of nodes in the graph; k∈{h,t}, k represents the category index, h represents the head class, and t represents the tail class; N(e i )∈V k ,N(e i ) represents edge e i The set of connected nodes, V k Represents the set of nodes in category k; This represents the high-dimensional representation of the i-th node in category k in the feature space; This is the distance used to measure the proximity of node i to the class centroid; This represents the relative similarity of node i belonging to category k; The inter-class separation loss is expressed as: L inter =exp(-d(C h ,C t )) Wherein d(C h C t () represents the Euclidean distance, used to measure the difference between two centroids.

9. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 8, characterized in that, Step 5) designs a dynamic composition and deletion strategy for tail category nodes, including: Based on the similarity between nodes and their respective class centroids, the sampling probabilities of head and tail class nodes are calculated; the sampling probability is exponentially related to the distance from the node to the centroid as follows: in, This represents the sampling probability that the head node i is selected. This represents the sampling probability that tail node j is selected; Represents the high-dimensional feature vector of head node i. With the center of mass C of the head h The distance between them; Represents the embedding feature vector of tail class node j With tail-like centroid C t The distance between them; Based on the sampling probability, N nodes are extracted from the head class and the tail class respectively, forming N pairs of nodes. For each pair of nodes, the features of the new node are synthesized through a mixing operation in the feature space, represented as: in, Represents the composite node N syn (e k )∈V syn The characteristic is that δ is a value sampled from a uniform distribution ranging from 0 to 1; The synthesized node is randomly connected to other nodes in the dynamic equilibrium edge graph to construct a new adjacency matrix, thus integrating the synthesized node into the dynamic equilibrium edge graph. Then, feature updates are performed by inputting the high-dimensional features of the original edge graph nodes and the high-dimensional features of the synthesized node into the graph neural network model to obtain the updated node features. The centroids of the head and tail classes are recalculated using the updated node features. The distance between the synthesized node and the centroids of the head and tail classes is used to determine whether the synthesized node was incorrectly classified as the head class; this is represented as: Where λ is a very large positive constant; when f k When = 1, node N syn (e k )∈V syn Nodes that are misclassified as head nodes are deleted and re-synthesized, maintaining a total number of synthesized nodes of N. This process is repeated once in each training cycle.

10. The academic paper reviewer recommendation method based on dynamic equilibrium edge graph as described in claim 9, characterized in that, Training the prediction model involves two stages: the first stage uses pre-trained node embeddings as features for training, and the second stage uses intra-class uniform loss and inter-class separation loss for training. In the first stage of training, the model uses pre-trained node embeddings as input features, with the goal of enabling the model to learn to use node embeddings to capture complex structures and relationships between nodes in the graph. The loss function L used in the second stage of end-to-end training is expressed as: L=L rw +αL inter +βL intra Where α and β are loss coefficients; L rw L is the cross-entropy loss function; intra For uniform loss within the class; L inter This is the inter-class separation loss; The node classification using the cross-entropy loss function is expressed as follows: Among them, L rw To measure the model's performance in the symbolic edge classification task; y k ∈Y represents node N(e k The actual symbol for ), where 0 represents a negative node and 1 represents a positive node; p k The i-th element; w pos and w neg These represent the weights of the positive and negative classes, respectively.

Images