Method and apparatus for identifying key nodes in heterogeneous networks based on decoupled causal interactions

Abstract

The application relates to a heterogeneous network key node identification method and device based on decoupling causal interaction, which comprises the following steps: designing an intention decoupling graph encoder, introducing a generation mechanism of a variational graph autoencoder, mapping node features of a heterogeneous network into a plurality of mutually independent Gaussian distributions, performing differentiable sampling on the node features through reparameterization sampling, and physically separating the node features; based on a contrast learning mechanism of mutual information maximization, regarding each layer of the heterogeneous network as a view, using the contrast learning mechanism to compare and learn mutual information of decoupled representations of the same node under different views, and filtering random noise edges appearing only in a single layer; constructing a structural causal model, identifying mixed factors based on the structural causal model, estimating the weight of the edge by adopting an inverse propensity weighting strategy, and performing causal readjustment on the weight of the edge to identify key nodes that truly have structural control power. The application has the effects of realizing semantic decoupling, structural denoising and causal correction, and improving the accuracy and robustness of network key node identification.

Description

Technical Field

[0001] This application relates to the field of information network analysis, and in particular to a method and apparatus for identifying key nodes in heterogeneous networks based on decoupled causal interactions. Background Technology

[0002] Heterogeneous networks are a common and important data model in everyday life, widely existing in fields such as social networks, bioinformatics networks, and knowledge graphs. Heterogeneous networks contain various node and edge types, with different nodes undertaking different structural functions and information dissemination roles. For example, in social networks, users, posts, and topics are core node types, while following, liking, and forwarding are typical edge types. In bioinformatics networks, genes, proteins, and metabolites constitute the node system, with interactions and regulatory relationships being the main edge types. Identifying key nodes in heterogeneous networks is a core research direction in network science. The goal is to extract nodes from a massive number of network nodes that play a decisive role in the stability of the network structure, the efficiency of information dissemination, and the effectiveness of functional implementation. This is of great significance for understanding network structure, analyzing network behavior, and solving practical problems, such as identifying influential nodes in social networks and discovering key genes in bioinformatics networks.

[0003] However, with the exponential growth in the scale of heterogeneous networks, the complexity and diversity of network structures have significantly increased. Existing identification methods are struggling to meet the needs of practical applications and are facing three major challenges: First, semantic entanglement. Existing methods treat node features as a single whole, failing to distinguish the potential intentions of nodes in different interaction scenarios (such as academic, commercial, and social), leading to ambiguous feature representations. Second, inter-layer noise interference. In multi-heterogeneous networks, the data quality of different views varies greatly. Simple aggregation often amplifies accidental connections or random noise, obscuring the true structural information. Third, popularity bias and structural mismatch. Existing methods are mostly based on statistical correlation for model learning, which generally suffers from severe "confounding bias." They tend to misclassify nodes with high degrees but lacking real propagation power as key nodes, resulting in a serious mismatch between the identification results and the actual dynamic characteristics of the network. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies and achieve semantic decoupling, structural denoising, and causal correction, thereby improving the accuracy and robustness of identifying key nodes in heterogeneous networks, this application provides a method and apparatus for identifying key nodes in heterogeneous networks based on decoupled causal interactions. It addresses the three core problems of existing methods from three levels: node attributes, network view, and causal logic. This method can be widely applied to various practical scenarios such as mining influential nodes in social networks, identifying key genes in bioinformatics networks, locating core network modules in software systems, and positioning public opinion in the field of public safety.

[0005] The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions provided in this application adopts the following technical solution: The design intent is to decouple the graph encoder by introducing the generation mechanism of the variational graph autoencoder, which maps the node features of the heterogeneous network into multiple independent Gaussian distributions, and performs differentiable sampling of the node features through reparameterized sampling, thereby physically separating the node features. Based on the contrastive learning mechanism that maximizes mutual information, each layer of the heterogeneous network is regarded as a view. The contrastive learning mechanism is used to compare and learn the mutual information of the decoupled representation of the same node under different views, and filter out random noise edges that only appear in a single layer. A structural causal model is constructed, confounding factors are identified based on the structural causal model, an inverse bias weighting strategy is adopted to predict the weights of the edges, and the weights of the edges are causally re-adjusted to identify the key nodes that truly have structural control.

[0006] By adopting the above technical solutions, the three major technical problems existing in the current methods for identifying key nodes in heterogeneous networks—semantic entanglement, inter-layer noise interference, popularity bias, and structural mismatch—are solved, and the accuracy and robustness of key node identification are greatly improved. First, the above scheme is designed to decouple the graph encoder. By using the generation mechanism of the variational graph autoencoder, it improves the traditional single vector feature mapping mode, maps the node features of heterogeneous networks into multiple independent Gaussian distributions, each of which corresponds to an independent semantic intent. Through reparameterized sampling, the physical separation of features is achieved, which can explicitly decouple the multiple semantic intents of nodes under different interaction scenarios and different relationship types. This allows each intent channel to have a clear physical meaning and semantic direction, thereby breaking the limitation of existing methods that confuse the multiple identities of nodes into a single vector. This makes the node feature representation more distinguishable and physically meaningful, and greatly improves the semantic entanglement problem. Secondly, a cross-view structural consistency learning module is designed. By using contrastive learning to maximize the mutual information of the decoupled representation of the same node in different views, it can effectively filter out random noise edges and false connections that only appear in a single layer, retain the real network structure information shared across views, and at the same time use complementary information between views to calibrate node features, avoid noise amplification from interfering with the recognition results, and significantly enhance the robustness and stability of the model in complex heterogeneous networks. Finally, a causal interaction decontamination module is constructed to accurately identify contamination factors such as popularity based on the structural causal model, clarify their dual influence mechanism on node attributes and connection structure, generate interaction masks using inverse bias weighting technology and causally readjust the weights of edges, identify key nodes that truly have structural control, effectively correct the popularity bias and structural mismatch problem of "high degree equals key" in existing methods, and discover key nodes with structural causal influence. In summary, the above-mentioned scheme can fully explore the multidimensional intent of node features in heterogeneous networks. Even with extremely high-dimensionality and complex meanings of node attributes, it can clearly decompose independent semantic channels, improving the interpretability of features; through cross-view... Figure 1 Due to consistency constraints, this invention exhibits strong robustness when processing low-quality network data containing a large amount of noise or missing edges, and can automatically filter out spurious connections. Most importantly, the above scheme effectively removes the false halo brought about by "popularity" through a causal decongestion mechanism, avoiding the defect of traditional algorithms blindly chasing highly connected nodes. It can accurately identify those "hidden key nodes" that have few connections but are located at the throat of the network and have real propagation and control capabilities, significantly improving the actual effect of downstream tasks (such as public opinion control, academic evaluation, and product recommendation).

[0007] Optionally, the feature distribution of the node is mapped to K independent Gaussian distributions, and the intention decoupling graph encoder is used to infer the mean vector and variance vector of the K Gaussian distributions in parallel. The step of performing differentiable sampling of node features through reparameterized sampling to physically separate node features includes: formalizing the feature generation process as follows: , in, As auxiliary noise, k is the intended channel. For nodes, This represents element-wise product.

[0008] By adopting the above technical solution, the potential intent representation of a node is modeled as multiple independent Gaussian distributions, which can more accurately and reasonably characterize the multiple semantic identities and potential interaction intentions of nodes in heterogeneous networks, making different intent channels independent of each other and possessing clear physical meanings, thereby effectively avoiding the problem of ambiguous feature expression caused by the mixing of different semantic information. Meanwhile, the shared graph convolutional network fully captures the network topology and node attribute features, predicts the distribution parameters of each intent channel, and then explores the feature distribution patterns under different semantic scenarios, thereby improving the reliability and stability of feature learning. Based on this, the reparameterized sampling process transforms the non-differentiable random sampling process into a combination of deterministic transformation and fixed-distribution noise. While ensuring the feature distribution characteristics, it achieves the differentiability and optimizability of model training, and at the same time completes the physical isolation of features of different intent channels. This significantly improves the expression accuracy and interpretability of node features, provides a clean and accurate feature foundation for subsequent structural denoising and causal correction, and also enhances the adaptability and recognition accuracy of the method in complex heterogeneous networks. In summary, the expression accuracy and interpretability of node features are significantly improved, providing a clean and accurate feature foundation for subsequent structural denoising, causal correction, and key node identification. This makes the overall method more adaptable and accurate in complex heterogeneous networks, effectively improving the accuracy and reliability of key node identification results.

[0009] Optionally, a KL divergence regularization term is embedded in the variational graph autoencoder to force the K Gaussian distributions to remain mutually orthogonal and independent, and to force the K Gaussian distributions to maintain a preset similarity to the standard normal distribution. The specific mathematical form of the KL divergence regularization term is as follows:

[0010] in, and They represent the first The mean square and variance square of each intention channel.

[0011] By adopting the above technical solution, during the training process of the intention decoupling graph encoder, a KL divergence constraint term is embedded in the loss function of the variational graph autoencoder to align the Gaussian distribution of each intention channel with the standard normal distribution. Through normalization constraints, the features of different intention channels are forced to remain orthogonal and independent in the potential decoupling space, avoiding feature collapse and overlap, and further enhancing the decoupling effect of multiple semantic intentions of nodes. During model training, the KL divergence constraint penalizes intent channels that deviate from the standard normal distribution, prompting the model to automatically decompose complex node attributes into independent intents with clear semantic directions. This decomposes the features of academic network researchers into "theoretical research," "application development," and "academic social interaction," and the features of social network users into "content creation," "social interaction," and "information acquisition," giving node feature representations clear physical meaning and semantic interpretation, thereby eliminating semantic ambiguity. In addition, the introduction of the KL divergence constraint term improves the model's generalization ability and noise resistance, making the feature distribution of each intent channel more stable and effectively dealing with missing values, random perturbations and outliers in node attributes.

[0012] Optionally, each of the views includes an original view and an enhanced view; The filtering of random noise edges that occur only in a single layer includes: Randomly discard edges from each of the original views to generate the enhanced view, requiring that the decoupling similarity of the same node in each of the original views and the enhanced views be within a preset range.

[0013] By employing the above technical solution, random edges are discarded from the original heterogeneous network view to generate an enhanced view. The decoupled representations of the same node in different views are constrained to remain within a preset range, meaning that the decoupled representations of the same node in different views should be as similar as possible. This effectively identifies and eliminates accidental connections, redundant associations, and random noise edges that only appear in a single structural layer of the network, avoiding distortion of network structural features due to noise interference. This maximizes the similarity of the decoupled representations of the same node in different views and minimizes the similarity of the decoupled representations of different nodes. The random edge discard ratio is adaptively adjusted according to the view data quality and noise level, typically set to 10%-20%. The above scheme is based on the design of a cross-view structural consistency learning mechanism that maximizes mutual information. By generating enhanced views through random edge dropping, it simulates noise changes and structural perturbations in the view and effectively filters out accidental connections or noisy edges that only appear in a single view. This ensures that the features input to subsequent modules have high robustness. For noisy edges and spurious connections that only exist in a single view or a single enhanced version, their corresponding node feature representations will be marginalized in the contrastive learning. Meanwhile, real structural connections that exist stably across views can resist random edge dropping perturbations, and their corresponding node feature representations will be strengthened, thereby highlighting the real structural information of the network.

[0014] Optionally, the weights of the edges predicted by implementing the inverse bias weighting strategy include: A bias estimation network is trained using an inverse bias weighting strategy, and the probability of each edge existing is calculated using the following formula:

[0015] in, Indicate whether u and v have an edge. This indicates that u and v have an edge. 0 indicates that u and v have no boundaries. For nodes and The decoupling characteristics, Represents a node and splicing of decoupling features To estimate the learnable weight matrix of the first-layer linear transformation of the network, To estimate the learnable weight matrix of the first-layer linear transformation of the network, For the first layer of learnable bias terms, For the second-layer learnable bias term, , is the activation function used to map the output value to the 0-1 interval, representing the probability of the edge existing.

[0016] By adopting the above technical solution, using an inverse bias weighting strategy and training a bias estimation network, the system accurately identifies and quantifies the confounding effects of popularity in the network based on a structural causal model. This solves the problems of popularity bias and structural mismatch in the key node identification process, laying the foundation for subsequent causal reweighting and counterfactual intervention. When calculating the probability of each edge, the system concatenates the decoupled features of the nodes and uses them as input. It also uses a multi-layer neural network and nonlinear activation function to construct a probability prediction model, which can fully integrate the multiple semantic intentions of the nodes and the network structure information, allowing the bias estimation network to accurately learn the influence of popularity on the nodes' edges. The activation function maps the output to the 0-1 range, which can intuitively represent the probability that each edge is driven by popularity. Specifically, the closer the propensity score is to 1, the more the existence of the edge depends on popularity and the more likely it is to be a spurious connection; the closer the propensity score is to 0, the existence of the edge is unrelated to popularity. In summary, the above-mentioned scheme can effectively distinguish between valid connections formed by real structural control and false connections formed solely by popularity, providing a foundation for subsequent counterfactual intervention of graph structures and elimination of backdoor path interference. It enables key node identification to focus more on the real structural value and causal role of the nodes themselves, significantly improving the objectivity and accuracy of the identification results compared to traditional identification methods.

[0017] Optionally, the probability of the existence of the edge is the edge's tendency score; When graph convolution aggregates information about connections between multiple edges, the structural causal model uses the reciprocal of the propensity score ( The neighboring features are weighted and aggregated, and the weights of the connections between each edge are assigned according to the weights of the weighted aggregation. The weights are mapped one-to-one to the preset scores and input into the scoring network to output a list of key nodes.

[0018] By adopting the above technical solution, the probability of an edge's existence is used as the corresponding edge's bias score, and the reciprocal of the bias score is used as a causal interaction mask. This mask is applied to message passing aggregation in graph neural networks. During message passing aggregation, the reciprocal of the bias score is used to weight and aggregate the features of neighboring nodes. This achieves counterfactual intervention in the graph structure and causal readjustment of edge weights from a structural causal perspective, cutting off the backdoor path influence caused by popularity confounding factors from a structural causal perspective. The core logic is: for edges with extremely high bias scores ( →1), assign low weight ( →1), reducing the contribution of popularity-driven spurious connections to feature aggregation; for edges with extremely low propensity scores ( →0), assigning high weight ( →+∞), strengthening the contribution of truly effective connections. This mechanism weakens the traditional model's over-preference for high-popularity nodes, avoids misjudging key nodes due to popularity, and allows the model to focus more on the true structural control and information propagation capabilities of nodes in information aggregation; After completing the causal weighted aggregation, the final node representation is input into the scoring network. By training the model with the ListMLE ranking loss function, the ranking results of key nodes can be directly optimized, ensuring that the output ranking list truly reflects the actual importance of the nodes and improving the accuracy and credibility of the recognition results. In summary, by recalibrating the edge weights at the causal level, the model pays more attention to the true structure and function of nodes during information aggregation, thereby further improving the unbiasedness and reliability of node feature learning.

[0019] Optionally, for any user node v, the node features after weighted aggregation using the structural causal model are represented as follows:

[0020] Where N(v) is the set of neighboring nodes of user node v; The node features after weighted aggregation of user node v; The node features of each user are input into the scoring network, and the importance score of each user's node is output. The nodes are then sorted according to the importance scores.

[0021] By employing the above technical solution, the reciprocal of the bias score is used as an interaction mask to assign low weight to popularity-driven spurious links and high weight to genuine valid links, resulting in a weighted aggregation. By eliminating the interference of popularity confounding factors, the system purely reflects the structural causal influence of user nodes themselves, thereby ensuring the objectivity and authenticity of subsequent importance scoring. Regarding the above The mathematical representation explains the principle as follows: Using the set of neighboring nodes N(v) as the aggregation range, feature transformation is performed by combining the learnable parameter W and the bias term b. Then, nonlinear modeling capability is introduced through the activation function σ, allowing... While preserving the decoupled semantic features of the nodes themselves, it fully integrates effective causal relationship information at the network topology level, so that the node importance score has both semantic integrity and structural relevance, significantly improving the comprehensiveness and accuracy of the score results. Based on the above centralized set results, the node ranking based on the real and effective importance scores can accurately reflect the actual influence level of nodes in heterogeneous networks, avoid the ranking bias caused by shallow features (popularity) in traditional methods, and accurately identify key nodes with real structural control and information dissemination capabilities.

[0022] Optionally, the ListMLE loss function is used as the final loss function of the structural causal model, which optimizes the ranking result of the key user nodes. The specific form of the ListMLE loss function is as follows:

[0023] Wherein, rank(v) is the ranking of the importance score of user node v, and rank(u) is the ranking of the importance score of user node u. For sorting temperature coefficients, Score the importance of user node u.

[0024] Using the ListMLE loss function as the final loss function of the structural causal model can directly minimize the log-likelihood error between the predicted ranking and the actual ranking, allowing the training process of the structural causal model to focus on optimizing the relative ranking relationship of user nodes, thereby matching the core objective of identifying key nodes in heterogeneous networks. Furthermore, by constraining rank(u) ≤ rank(v), the model is forced to learn the true influence priority between nodes, reducing random fluctuations and biases in the ranking results, and a ranking temperature coefficient is introduced. Used to flexibly adjust the discrimination of the score to adapt to different scenarios; Because the structural causal model employs causal decontamination techniques and plays a role in the entire process of feature representation, scoring, and ranking, it identifies key nodes with real structural control and information propagation capabilities, including "hidden key nodes" with few connections but in key positions in the network. This makes the final key node ranking list highly consistent with the actual dynamic characteristics of the network. Therefore, the above technical solution can significantly improve the accuracy and robustness of key node identification in heterogeneous networks.

[0025] On the other hand, this application provides a heterogeneous network key node identification device based on decoupled causal interaction, which adopts the following technical solution: A device for identifying key nodes in heterogeneous networks based on decoupled causal interactions, the device comprising: At least one processor; and At least one memory storing a computer program; When the computer program is executed by the at least one processor, the device performs the steps of the heterogeneous network key node identification method based on decoupled causal interaction.

[0026] By adopting the above technical solution, the method for identifying key nodes in heterogeneous networks based on decoupled causal interaction is integrated into a hardware device in the form of a computer program. Through parameter adaptation, it meets the analysis needs of heterogeneous networks of different types and sizes, and has strong versatility and scalability.

[0027] On the other hand, this application provides a computer-readable storage medium storing a computer program, employing the following technical solution: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for identifying key nodes in heterogeneous networks based on decoupled causal interactions.

[0028] By adopting the above technical solution, the heterogeneous network key node identification method of this application is solidified in the form of a computer program in a computer-readable storage medium, and widely installed in various computing devices, servers and embedded platforms, fully carrying all technical processes such as intent decoupling, cross-view structure denoising, causal interaction decontamination, causal weighted aggregation and sorting optimization.

[0029] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention proposes a semantic decoupling method based on variational inference, which abandons the traditional point-to-point feature mapping mode and instead learns the probability distribution parameters of features, such as mean and variance. It uses reparameterization techniques to perform differentiable sampling and uses KL divergence regularization terms to force multiple intent channels to remain orthogonal and independent. This is different from the traditional capsule network dynamic routing method, and provides a more efficient semantic separation means, which significantly improves the interpretability and expression accuracy of node features. 2. This invention constructs a view enhancement strategy based on edge dropping, designs a contrastive loss function specific to the key node identification task, organically combines contrastive learning with ranking tasks, maximizes cross-view mutual information while retaining the relative ranking relationship between nodes, effectively filters noisy edges and false connections in single views, makes full use of complementary information from multiple views, significantly suppresses inter-layer noise interference, and improves the robustness of the model in complex low-quality heterogeneous networks. 3. This invention defines and identifies popularity confounding factors in graph structures, designs a bias score generator to quantify the generation probability of edges, and uses inverse bias weighting to dynamically adjust the weights of neighbor aggregation in graph neural networks. This mechanism substantially cuts off the backdoor path of popularity bias, protects the core algorithm logic of "graph aggregation based on causal weights", and enables the model to accurately identify key nodes with real structural causal influence. 4. The three-layer progressive framework of "semantic decoupling - structural denoising - causal correction" constructed in this application is supported by each module. Semantic decoupling provides a pure feature basis for structural denoising, structural denoising provides real structural information for causal correction, and causal correction provides unbiased evaluation basis for key node identification, forming a complete technical closed loop. It can adapt to heterogeneous networks of different types, scales and qualities, and has strong versatility and practicality. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the overall architecture of an embodiment of this application.

[0031] Figure 2 This is an architecture diagram of the intended decoupling graph encoder in the embodiments of this application.

[0032] Figure 3 This is an architecture diagram of the cross-view structure consistency learning module in this application embodiment.

[0033] Figure 4 This is an architecture diagram of the causal interaction deproliferation module in the embodiments of this application. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] The following is in conjunction with the appendix Figure 1-4 This application will be described in further detail.

[0037] This application discloses a method for identifying key nodes in heterogeneous networks based on decoupled causal interactions. (Refer to...) Figure 1A heterogeneous network joint recognition method based on decoupled causal interaction includes several sequential and progressive core steps: 1. A heterogeneous network intent decoupling method based on variational inference: By using an intent decoupling graph encoder, the features of social network user nodes are mapped to multiple independent Gaussian distributions, accurately capturing the multifaceted nature of nodes and solving the semantic entanglement problem. 2. Cross-view comparison enhancement mechanism oriented to node sorting: Through the cross-view structural consistency learning module, the noise edges and false connections in each view are filtered by comparison learning, so as to achieve cross-view purification of the network structure and solve the problem of inter-layer noise interference. 3. Based on the inverse bias weighting causal decontamination technique for graph structures, the inverse bias weighting technique is used to causally readjust the edge weights to achieve causal correction of node importance assessment and solve the problems of popularity bias and structural mismatch.

[0038] Step 1: Heterogeneous networks intend to decouple based on variational inference.

[0039] Reference Figure 1 and Figure 2 The heterogeneous network intent decoupling method based on variational inference relies on the construction of a decoupled variational graph autoencoder. It utilizes the generative decoupling mechanism of the variational graph autoencoder to integrate the structural feature capture capability of graph convolutional networks with the probabilistic generation capability of the variational graph autoencoder, thereby achieving physical separation and explicit decoupling of multiple semantic intents of nodes.

[0040] Specifically, the intent decoupled graph encoder uses a shared graph convolutional encoder as its core, employing a two-layer graph convolutional network architecture. The input consists of the original feature matrix X of social network user nodes (containing attributes such as user gender, age, content posting features, and interaction frequency) and the network fusion adjacency matrix A (integrating topological information from three edge types: follow, like, and share). Through graph convolution operations, it fully captures the attribute features of nodes and the topological structure features of the network, outputting an intermediate feature matrix. The output of the shared graph convolutional encoder is connected to K parallel distributed parameter prediction layers, each corresponding to a semantic intent channel (e.g., the preset number of semantic intent channels for a node is K=3, corresponding to the three core semantic intents of social network users: "content creation," "social interaction," and "information dissemination"). A fully connected layer maps the intermediate features to the Gaussian distribution parameters of that intent channel—the mean vector μ and the variance vector σ—and uses reparameterization techniques to sample and generate K independent feature vectors for each intent channel. For any user node v with k intent channels, the k feature vectors of node v are... ,in To provide auxiliary noise, This represents element-wise product (i.e., element-wise Hadamard).

[0041] The aforementioned reparameterized sampling transforms the original random sampling process, which relies on the mean and variance, into a deterministic transformation of fixed standard normal noise and distribution parameters. This makes the originally non-differentiable sampling operation differentiable, and the model can update and optimize all learnable parameters of the shared graph convolutional encoder and the distribution parameter prediction layer through the backpropagation algorithm. At the same time, this operation ensures that the features of each intent channel follow a preset Gaussian distribution, achieving physical isolation of features in the latent space.

[0042] Furthermore, to prevent the feature distributions of the K intent channels from collapsing and overlapping, which could lead to decoupling failure, this invention introduces a KL divergence regularization term in the training objective. This term forces the feature distributions of the K intent channels to be as close as possible to the standard normal distribution and to be orthogonal to each other in the potential decoupling space.

[0043] The mathematical form of the KL divergence regularization term is:

[0044] in, and They represent the first The mean square and variance square of each intention channel.

[0045] The overall loss function of the intent decoupling graph encoder is a weighted sum of the reconstruction loss and the KL divergence constraint term: the reconstruction loss uses mean squared error (MSE) loss to constrain the model to reconstruct the original node features and network topology based on the decoupling features generated by sampling, ensuring the effectiveness of feature decoupling; the KL divergence constraint term is used to constrain the distribution characteristics of each intent channel to achieve semantic decoupling. The overall loss function (L) is in the form of:

[0046] Wherein, λ is a hyperparameter used to balance the weights of the reconstruction loss and the KL divergence constraint term, and can be adaptively adjusted according to the actual network data.

[0047] The intent-decoupled graph encoder is trained using a stochastic gradient descent optimizer. The learning rate, batch size, and number of training epochs are set, and an early stopping strategy is employed to prevent overfitting. Model training stops when the reconstruction loss on the validation set continuously exceeds a preset value. After training, the model automatically decomposes the original mixed features of user nodes into three independent decoupled feature channels. Each channel corresponds to a specific semantic intent. For example, the "Content Creation" channel accurately represents the user's content production ability (e.g., post quality, originality, and professionalism); the "Social Interaction" channel represents the user's social activity (e.g., the frequency and scope of likes, comments, and reposts); and the "Information Dissemination" channel represents the user's information dissemination ability (e.g., the post's dissemination range and reposting level).

[0048] Step 2: Enhanced cross-view contrast for node-oriented sorting.

[0049] Reference Figure 2 and Figure 3 The core of this step is to design the Cross-View Structure Consistency Learning (CSCL) module. This module is based on the idea of ​​self-supervised contrastive learning. Through view enhancement and cross-view contrastive learning, it filters out noisy edges and false connections in each view, extracts the real structural information shared across views, realizes cross-view purification of network structure, and improves the robustness of node features.

[0050] To facilitate understanding, let's illustrate with an example. Consider the three edge types on a social network—follow, like, and share—as three independent network views. First, perform random edge discarding enhancement on each of the three original views (follow view, like view, and share view) to generate corresponding enhanced views. The core of random edge discarding is to randomly delete some edges from the view's adjacency matrix, simulating noise changes and structural perturbations in the view. Differential discarding ratios need to be set based on the noise level of each view. For any original view Gi=(V,Ei,X), the generation rule for its enhanced view Gi′=(V,Ei′,X) is as follows: for each edge in Ei, randomly delete it from Ei′ with a preset discard probability p, and retain it with a probability of 1-p, ultimately obtaining the adjacency matrix Ai′ of the enhanced view. The user node decoupling features generated in step one are input into the graph convolutional layers corresponding to the three views respectively, and the decoupling feature representations of each user node under the original view and the enhanced view are extracted. The original view decoupling representation is h{v,i}^k, and the enhanced view decoupling representation is h'{v,i}^k, where i=1,2,3 correspond to the follow, like, and share views respectively, and k=1,2,3 correspond to the intent channels of "content creation", "social interaction", and "information dissemination" respectively.

[0051] Then, a loss function is constructed by comparing the original view and the enhanced view. This function includes two parts: intra-view contrast loss and cross-view contrast loss. It ensures high similarity of the decoupled representations of the same node in both the original and enhanced views, as well as high similarity of the decoupled representations of the same node in different original views, while maintaining low similarity of the decoupled representations of different nodes. This maximizes cross-view mutual information while preserving the relative ranking relationship between nodes. The intra-view contrast loss is used to constrain the decoupled representations of the same user node in the original and enhanced versions of the same view to be as close as possible. For any user v, view i, and intent channel k, the intra-view contrast loss is:

[0052] in, The contrast loss of the intent channels of node v, view i, and k within the view, value simτ is the cosine similarity function used to calculate the similarity between two feature vectors; τ is the temperature coefficient used to adjust the discriminative power of the similarity; the denominator is the sum of the similarities between the enhanced view decoupled representation of all user nodes and the original view decoupled representation of the target user, thus enabling the comparison of positive and negative sample pairs.

[0053] The cross-view contrast loss is used to constrain the decoupled representations of the same user node in different original views to be as close as possible. For any user v and intent channel k, the view of interest is selected as the baseline view, and its cross-view contrast loss is:

[0054] The overall contrast loss, after considering intra-view contrast loss and cross-view contrast loss, is the average of the intra-view contrast loss and cross-view contrast loss for all user nodes, all intent channels, and all views.

[0055] Finally, the cross-view structure consistency learning module and the intent decoupling graph encoder are jointly trained, with the optimizer, learning rate, batch size and other hyperparameters remaining consistent with those in step one.

[0056] Step two effectively suppresses the interference of low-quality views on the final result. By maximizing cross-view mutual information, the model automatically captures the most essential and stable structural patterns in the network, ensuring that only structural information with cross-view robustness can enter the next stage.

[0057] Step 3: Decontamination of causal relationships based on inverse bias weighted graph structures.

[0058] Reference Figure 3 and Figure 4The graph structure causal decontamination technique based on inverse bias weighting is used to solve the problems of popularity bias and structural mismatch. The core is to build a causal interaction decontamination module to clarify the causal relationship between variables. The core variables defined in this embodiment include Z: popularity contamination factor, which represents the popularity of user nodes and is represented by the normalized value of user followers and post exposure; X: node attribute features, which is the decoupling feature representation of user nodes output in step two; A: network connection structure, which is the network adjacency matrix after cross-view denoising; Y: outcome variable, which represents the importance of user nodes (public opinion dissemination ability), which is the target that the model ultimately needs to predict. Intuitively, popularity Z is a confounding factor that directly influences both node attribute X and connection structure A (Z→X, Z→A); node attribute X and connection structure A jointly influence node importance Y (X→Y, A→Y); popularity Z forms a false backdoor path from Z to Y through X and A (Z→X→Y, Z→A→Y). The core objective of this application is to cut off this false backdoor path, allowing the model to learn the true causal influence of X and A on Y, rather than the false correlation brought about by the confounding factor Z.

[0059] To quantify the impact of popularity on the connections between nodes, an inverse bias weighting strategy is adopted to construct a lightweight multilayer perceptron to train a bias estimation network and calculate the probability of each edge's existence (i.e., bias score). The calculation formula is:

[0060] in, Represents a node and splicing of decoupling features and These are learnable parameters.

[0061] The physical meaning of propensity score is: The closer it is to 1, the more the connection between users u and v depends on the popularity confounding factor Z, and the more likely it is to be a spurious popularity-driven connection. The closer the value is to 0, the more independent the connection is from popularity; it is an effective connection driven by users' ability to spread real information and their social needs.

[0062] In the message passing aggregation phase of the graph neural network, the inverse of the propensity score is used. As a causal interaction mask, neighbor features are weighted and aggregated. This mathematical operation essentially performs a "counterfactual intervention" on the graph structure, reducing the connection weights that arise solely from high popularity and cutting off false backdoor paths caused by popularity.

[0063] For any user node v, the node features are obtained by weighted aggregation of the structural causal model. Represented as:

[0064] Where N(v) is the set of neighboring nodes of user node v; The node features are weighted and aggregated for user node v.

[0065] The core logic of the causal weighted aggregation mechanism is: for popularity-driven spurious connections ( →1) Assign low weight ( →1), reducing its contribution to node feature representation; for truly effective connections ( →0) assigns high weight ( →+∞), strengthen its contribution to the node feature representation, so that the final generated feature representation purely reflects the true structural causal influence of user node v, rather than the false correlation brought about by popularity.

[0066] Reference Figure 1 and Figure 4 Features after causal weighting The input is a scoring network (a single fully connected layer), and the output is an importance score s for each user node. To allow the model to directly optimize the ranking results of key nodes, the ListMLE loss function is used as the final loss function of the model.

[0067] The specific form of the ListMLE loss function is as follows:

[0068] Wherein, rank(v) is the ranking of the importance score of user node v, and rank(u) is the ranking of the importance score of user node u. For sorting temperature coefficients, Score the importance of user node u.

[0069] The causal interaction decontamination module, the Intent Decoupling Graph Encoder (IDGE), and the Cross-View Structure Consistency Learning (CSCL) module are jointly trained end-to-end. The loss function of the overall model is the weighted sum of the loss functions of the three modules.

[0070] Among them, α, β, and γ are hyperparameters that are adaptively adjusted according to the actual network data and recognition task.

[0071] Finally, output the final list of key nodes in sorted order.

[0072] In summary, the key points of the embodiments of this application can be summarized into the following three points: I. Semantic decoupling based on variational inference: It abandons the traditional point-to-point feature mapping and instead adopts probability distribution modeling. By imposing orthogonality constraints and KL divergence penalties on the latent space, it forces the K intent channels learned by the model to remain independent, thereby accurately capturing the multifaceted nature of nodes.

[0073] II. Cross-view denoising based on contrast mechanism: Different relation types in the network are regarded as different views. Through view enhancement and contrast loss optimization, the essential structural features shared across views are extracted, effectively suppressing data noise in single views.

[0074] 3. Causal bias correction based on counterfactual logic: Introducing the backdoor adjustment idea in causal inference, the influence of "popularity" on the connection is quantified by learning the propensity score, and this bias is eliminated by reverse weighting in the aggregation stage, directly optimizing the loss of ListMLE for sorting, and ensuring that the identified nodes have real causal influence.

[0075] In addition, this embodiment also discloses two alternative solutions.

[0076] Option 1: Use a dynamic routing algorithm from capsule networks to replace the variational graph autoencoder of this invention, and calculate the coupling coefficient iteratively. The lower-level features are "voted" for the higher-level intention capsule.

[0077] While this approach can also achieve decoupling, dynamic routing involves complex inner loop iterative calculations, resulting in huge memory consumption and extremely slow training speed on large-scale graph data (such as millions of nodes). In contrast, the variational graph autoencoder used in this embodiment involves only one forward propagation, achieving several times higher computational efficiency, and also exhibits better noise resistance through probabilistic modeling.

[0078] Option 2: Eliminate the causal interaction decontamination module and directly build a multi-task learning framework that simultaneously predicts node classification, link prediction, and key node ranking, thereby improving the performance of the main task through auxiliary tasks.

[0079] The limitation of this approach is that while multi-task learning can improve generalization, it cannot fundamentally solve the "popularity bias" problem; the model will still tend to use the "degree" shortcut for prediction. In contrast, the causal decontamination mechanism corrects the bias at the mechanistic level, rather than merely enhancing the data.

[0080] The core technical advantage of the heterogeneous network key node identification method based on decoupled causal interaction in this application embodiment is as follows: I. By using variational inference to achieve explicit and physical decoupling of multiple semantic intents of nodes, the complex iterative calculations of traditional decoupling methods are avoided, thus improving computational efficiency. At the same time, each intent channel has a clear semantic orientation, which significantly improves the interpretability of node features. Second, cross-view structural consistency learning based on contrastive learning can accurately filter noisy edges and spurious connections in a single view, make full use of complementary information between views, and enable the model to maintain high robustness in complex, low-quality heterogeneous networks. Third, by introducing causal inference into the identification of key nodes in heterogeneous networks, the popularity bias is identified and eliminated from a mechanistic perspective, cutting off false backdoor paths and enabling the model to accurately discover key nodes with real structural causal influence, including "hidden key nodes," which greatly improves the accuracy and rationality of the identification results.

[0081] Based on the above-mentioned method for identifying key nodes in heterogeneous networks based on decoupled causal interaction, this application discloses a device for identifying key nodes in heterogeneous networks based on decoupled causal interaction and a computer-readable storage medium, both of which are equipped with the above-mentioned method for identifying key nodes in heterogeneous networks based on decoupled causal interaction at the hardware level.

[0082] This embodiment provides a heterogeneous network key node identification device based on decoupled causal interaction. This device is used to implement the aforementioned heterogeneous network key node identification method based on decoupled causal interaction. Its hardware architecture includes at least one processor and at least one memory storing a computer program. The memory and processor are connected via a bus or other means to store the computer program for the processor to execute. The processor executes the computer program stored in the memory to implement the entire process of heterogeneous network key node identification.

[0083] This embodiment provides a computer-readable storage medium, which is a non-transitory computer-readable storage medium storing a computer program thereon. When the computer program is called and executed by a processor, it can implement all the steps of the above-described method for identifying key nodes in heterogeneous networks based on decoupled causal interactions.

[0084] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0085] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0086] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for identifying key nodes in a heterogeneous network based on decoupled causal interactions, the method comprising: The method includes: The design intent is to decouple the graph encoder by introducing the generation mechanism of the variational graph autoencoder, which maps the node features of the heterogeneous network into multiple independent Gaussian distributions, and performs differentiable sampling of the node features through reparameterized sampling, thereby physically separating the node features. Based on the contrastive learning mechanism that maximizes mutual information, each layer of the heterogeneous network is regarded as a view. The contrastive learning mechanism is used to compare and learn the mutual information of the decoupled representation of the same node under different views, and filter out random noise edges that only appear in a single layer. A structural causal model is constructed, confounding factors are identified based on the structural causal model, an inverse bias weighting strategy is implemented to predict the weights of edges, and the weights of edges are causally re-adjusted to identify the key nodes that truly possess structural control.

2. The method of claim 1, wherein, The feature distribution of the node is mapped to K independent Gaussian distributions, and the intention decoupling graph encoder is used to infer the mean vector and variance vector of the K Gaussian distributions in parallel. The step of performing differentiable sampling of node features through reparameterized sampling and physically separating node features includes: formalizing the feature generation process as follows: ， wherein, is the auxiliary noise, k is the intended channel, is the node, denotes the element-wise product.

3. The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to claim 2, characterized in that, The variational graph autoencoder embeds a KL divergence regularization term, which forces the K Gaussian distributions to remain mutually orthogonal and independent, and forces the K Gaussian distributions to maintain a preset similarity to the standard normal distribution. The specific mathematical form of the KL divergence regularization term is as follows: in, and They represent the first The mean square and variance square of each intention channel.

4. The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to claim 1, characterized in that, Each of the views includes an original view and an enhanced view; The filtering of random noise edges that occur only in a single layer includes: Randomly discard edges from each of the original views to generate the enhanced view, requiring that the decoupling similarity of the same node in each of the original views and the enhanced views be within a preset range.

5. The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to claim 1, characterized in that, The weights for predicting edges using the reverse bias weighting strategy include: A bias estimation network is trained using an inverse bias weighting strategy to calculate the probability of each edge existing: in, Indicate whether u and v have an edge. This indicates that u and v have an edge. 0 indicates that u and v have no boundaries. For nodes and The decoupling characteristics, Represents a node and splicing of decoupling features To estimate the learnable weight matrix of the first-layer linear transformation of the network, To estimate the learnable weight matrix of the first-layer linear transformation of the network, For the first layer of learnable bias terms, For the second-layer learnable bias term, is the activation function used to map the output value to the 0-1 interval, representing the probability of the edge existing.

6. The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to claim 5, characterized in that, The probability of the existence of an edge is the edge's tendency score; When graph convolution aggregates information about connections between multiple edges, the structural causal model uses the reciprocal of the propensity score: We perform weighted aggregation on neighboring features and assign weights to the connections between each edge based on the weights of the weighted aggregation. The weights are mapped one-to-one to the preset scores and input into the scoring network to output a list of key nodes.

7. The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to claim 6, characterized in that, For any user node v, the node features after weighted aggregation of the structural causal model are represented as follows: Where N(v) is the set of neighboring nodes of user node v; The node features after weighted aggregation of user node v; The node features of each user are input into the scoring network, and the importance score of each user's node is output. The nodes are then sorted according to the importance scores.

8. The method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to claim 7, characterized in that, The ListMLE loss function is used as the final loss function of the structural causal model, which optimizes the ranking result of the key user nodes. The specific form of the ListMLE loss function is as follows: Wherein, rank(v) is the ranking of the importance score of user node v, and rank(u) is the ranking of the importance score of user node u. For sorting temperature coefficients, Score the importance of user node u.

9. A device for identifying key nodes in heterogeneous networks based on decoupled causal interactions, characterized in that, The device includes: At least one processor; and At least one memory storing a computer program; When the computer program is executed by the at least one processor, the device performs the steps of the method for identifying key nodes in heterogeneous networks based on decoupled causal interactions according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method for identifying key nodes in heterogeneous networks based on decoupled causal interactions as described in any one of claims 1 to 8.

Images