Molecular property prediction robust heterograph classification method and device based on information bottleneck

Abstract

The invention relates to the field of molecular property prediction, in particular to a molecular property prediction robust heterogeneous graph classification method and device based on information bottleneck, and the method comprises the steps: obtaining a to-be-classified heterogeneous graph set which comprises at least one to-be-classified heterogeneous graph; the at least one to-be-classified heterogeneous graph is input into a pre-constructed robust heterogeneous graph classification model, the prediction probability of the category to which each to-be-classified heterogeneous graph belongs is obtained, and the pre-constructed robust heterogeneous graph classification model comprises a node semantic purifier and a global graph decoupling module; a node semantic purifier is used to generate robust heterostructures and node representations. Therefore, the problems that the node representation quality is reduced and the classification accuracy is relatively low under the confrontation attack in the related technology are solved, and the classification accuracy under the confrontation attack is improved.

Description

Technical Field

[0001] The present invention relates to the field of molecular property prediction, and in particular to a robust heterogeneous graph classification method and device for molecular property prediction based on information bottleneck. Background Art

[0002] Molecular property prediction plays a crucial role in drug discovery and virtual screening, automatically identifying candidates with desired properties from a large pool of molecules. With advances in deep learning, molecular property prediction has made significant progress in recent years. Because molecules can be modeled as heterogeneous graphs with various types of atoms and chemical bonds, heterogeneous graph neural networks (HGNNs) have achieved remarkable performance in molecular property prediction thanks to their ability to leverage rich intrinsic semantics. Despite their promising performance, HGNNs for molecular property prediction require modeling both global structure and heterogeneous semantics, which inevitably makes them vulnerable to adversarial attacks. For example, SiO2 and TiO2 have similar homogeneous structures and are both insoluble in water; however, replacing silicon with titanium alters their electrical properties: SiO2 is an insulator, while TiO2 is a semiconductor. Despite this, this vulnerability of HGNNs in molecular property prediction has remained unexplored in previous studies. Specifically, HGNNs for molecular property prediction face two robustness issues: global direct perturbations, which have a greater impact, and local intermediate perturbations, which are more vulnerable. Global direct perturbations applied to heterogeneous molecular graphs have a greater impact. Unlike homogeneous graphs and other heterogeneous graphs, heterogeneous molecular graphs contain implicit patterns that depend on heterogeneity, making them vulnerable to global direct perturbations. For example, combining the carbonyl carbon in an aldehyde group with another carbon instead of hydrogen to form a ketone group significantly changes the molecular properties, and local intermediate perturbations introduced by heterogeneity expose more vulnerabilities. The most popular variant of HGNNs, namely meta-path-based HGNNs, adopts a hierarchical aggregation method (i.e., including node level and semantic level) to transform the heterogeneous graph into multiple intermediate graphs through meta-paths. Adversaries can easily design strategies targeting a single intermediate graph when seeking the optimal perturbation, thereby destroying the semantics of the molecular graph. With these key robustness issues, there is currently no robust solution for heterogeneous graph classification.

[0003] In related technologies, a series of new methods using heterogeneous graph neural networks (HGNNs) to model the rich semantics of molecular graphs have achieved state-of-the-art performance. The robustness of graph classification models is crucial to ensuring high reliability in applications. Given its importance, researchers have worked hard to address the robustness challenge through various approaches, including adopting data augmentation strategies, improving graph representation learning, advancing graph structure learning, and introducing the graph information bottleneck principle.

[0004] However, despite the excellent performance of HGNNs, they exhibit serious robustness weaknesses against adversarial perturbations in molecular property prediction tasks. On the one hand, previous work has neglected to constrain Y-irrelevant information at the node and semantic levels when learning robust heterogeneous node representations. That is, existing methods follow the process of meta-path-based HGNNs, relying on the assumption that complementary information across meta-paths is predictive and noise-free. However, adversaries can inject noise to destroy this complementarity, thereby compromising the quality of node representations. On the other hand, previous work often faces information leakage problems when optimizing robust structures at the graph level. Moreover, existing IB-based methods rely on a trade-off hyperparameter that cannot achieve precise information control, resulting in Y-irrelevant data leaking into the final representation, thereby reducing performance and urgently needing to be addressed. Summary of the Invention

[0005] The present invention provides a robust heterogeneous graph classification method and device for molecular property prediction based on information bottleneck, so as to solve the problems of reduced node representation quality and low classification accuracy caused by adversarial attacks in related technologies, and improve the classification accuracy under adversarial attacks.

[0006] The first aspect of the present invention provides a robust heterogeneous graph classification method for molecular property prediction based on information bottleneck, wherein the method includes the following steps: obtaining a set of heterogeneous graphs to be classified, wherein the set of heterogeneous graphs to be classified includes at least one heterogeneous graph to be classified; inputting the at least one heterogeneous graph to be classified into a pre-constructed robust heterogeneous graph classification model to obtain a predicted probability of the category to which each heterogeneous graph to be classified belongs, wherein the pre-constructed robust heterogeneous graph classification model includes a node semantic purifier and a global graph decoupling module, and the node semantic purifier is used to generate robust heterogeneous structures and node representations.

[0007] Furthermore, in some embodiments, before the heterogeneous graph to be classified is input into the pre-constructed robust heterogeneous graph classification model, it also includes: constructing a target heterogeneous graph set for training; dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on a preset division ratio, and using the heterogeneous graph training set to train a heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model; verifying the initial robust heterogeneous graph classification model based on a preset loss function and the heterogeneous graph verification set; if the verification result meets the training condition, the initial robust heterogeneous graph classification model is used as the pre-constructed robust heterogeneous graph classification model; otherwise, adjusting the preset division ratio, and re-executing the step of dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on the preset division ratio, until the new verification result meets the training condition, and the robust heterogeneous graph classification model corresponding to the new verification result is used as the pre-constructed robust heterogeneous graph classification model.

[0008] Furthermore, in some embodiments, constructing a target heterogeneous atlas for training includes: determining multiple target molecules, node type mapping functions, and edge type mapping functions; based on the node type mapping function and the edge type mapping function, constructing a heterogeneous graph comprising a node set and an edge set for each target molecule to obtain the target heterogeneous atlas.

[0009] Furthermore, in some embodiments, the use of the heterogeneous graph training set to train the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model includes: inputting each target heterogeneous graph in the heterogeneous graph training set into the node semantic purifier, so that the node semantic purifier processes each target heterogeneous graph through graph random attention and Hilbert-Schmidt independence criterion to generate a robust node representation, a related adjacency matrix and an irrelevant adjacency matrix; inputting the robust node representation, the related adjacency matrix and the irrelevant adjacency matrix into the global graph decoupling module, optimizing each target heterogeneous graph by decoupling the graph information bottleneck, and obtaining a robust graph representation of each target heterogeneous graph; based on each target heterogeneous graph and the robust graph representation of each target heterogeneous graph, training the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model.

[0010] Furthermore, in some embodiments, the preset loss function is:

[0011]

[0012] Where λ is a constant, is the loss function of the NSP module, is the decoupling graph information bottleneck loss function of the GGD module, is the cross entropy loss function. Reconstruct the loss function for the graph, is the discriminator loss function.

[0013] According to an embodiment of the present invention, the robust heterogeneous graph classification method for molecular property prediction based on information bottlenecks uses graph random attention at the node level through a node semantic purifier to separate Y-related and Y-independent subgraphs. At the semantic level, the Hilbert-Schmidt independence criterion is used to enhance the complementarity of node representations from different meta-paths during fusion. Furthermore, global graph decoupling is used to optimize the robust graph structure at the graph level. This method solves the problems of reduced node representation quality and low classification accuracy caused by adversarial attacks in related technologies, thereby improving classification accuracy under adversarial attacks.

[0014] The second aspect of the present invention provides an information bottleneck-based robust heterogeneous graph classification device for molecular property prediction, wherein the device includes: an acquisition module for acquiring a set of heterogeneous graphs to be classified, wherein the set of heterogeneous graphs to be classified includes at least one heterogeneous graph to be classified; a training module for inputting the at least one heterogeneous graph to be classified into a pre-constructed robust heterogeneous graph classification model to obtain a predicted probability of the category to which each heterogeneous graph to be classified belongs, wherein the pre-constructed robust heterogeneous graph classification model includes a node semantic purifier and a global graph decoupling module, and the node semantic purifier is used to generate robust heterogeneous structures and node representations.

[0015] Furthermore, in some embodiments, before the heterogeneous graph to be classified is input into the pre-constructed robust heterogeneous graph classification model, the acquisition module is specifically used to: construct a target heterogeneous graph set for training; based on a preset division ratio, divide the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set, and use the heterogeneous graph training set to train a heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model; based on a preset loss function and the heterogeneous graph verification set, verify the initial robust heterogeneous graph classification model; if the verification result meets the training condition, the initial robust heterogeneous graph classification model is used as the pre-constructed robust heterogeneous graph classification model; otherwise, the preset division ratio is adjusted, and the step of dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on the preset division ratio is re-executed until the new verification result meets the training condition, and the robust heterogeneous graph classification model corresponding to the new verification result is used as the pre-constructed robust heterogeneous graph classification model.

[0016] Furthermore, in some embodiments, the acquisition module is also used to: determine multiple target molecules, node type mapping functions and edge type mapping functions; based on the node type mapping function and the edge type mapping function, for each target molecule, construct a heterogeneous graph containing a node set and an edge set to obtain the target heterogeneous graph set.

[0017] Furthermore, in some embodiments, the training module is specifically used to: input each target heterogeneous graph in the heterogeneous graph training set into the node semantic purifier, so that the node semantic purifier processes each target heterogeneous graph through graph random attention and Hilbert-Schmidt independence criterion to generate a robust node representation, a relevant adjacency matrix and an irrelevant adjacency matrix; input the robust node representation, the relevant adjacency matrix and the irrelevant adjacency matrix into the global graph decoupling module, optimize each target heterogeneous graph by decoupling the graph information bottleneck, and obtain a robust graph representation of each target heterogeneous graph; based on each target heterogeneous graph and the robust graph representation of each target heterogeneous graph, train the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model.

[0018] Furthermore, in some embodiments, the preset loss function is:

[0019]

[0020] Where λ is a constant, is the loss function of the NSP module, is the decoupling graph information bottleneck loss function of the GGD module, is the cross entropy loss function. Reconstruct the loss function for the graph, is the discriminator loss function.

[0021] According to the information bottleneck-based robust heterogeneous graph classification device for molecular property prediction provided by the embodiments of the present invention, a node semantic purifier uses graph random attention at the node level to separate Y-related and Y-independent subgraphs. At the semantic level, the Hilbert-Schmidt independence criterion is used to enhance the complementarity of node representations from different meta-paths during fusion. Global graph decoupling is then used to optimize the robust graph structure at the graph level. This solves the problems of reduced node representation quality and low classification accuracy caused by adversarial attacks in related technologies, thereby improving classification accuracy under adversarial attacks.

[0022] A third aspect of the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned robust heterogeneous graph classification method for molecular property prediction based on information bottleneck.

[0023] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck as described in the above embodiments.

[0024] A fifth aspect of the present invention provides a computer program product, comprising a computer program, which is executed to implement the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck as described in any one of the above items.

[0025] Therefore, the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck in the embodiment of the present invention has the following characteristics:

[0026] Beneficial effects:

[0027] (1) We first attempt to reformulate the molecular property prediction task as a robust heterogeneous graph classification problem, thereby extending the applicability of traditional methods in noisy scenarios.

[0028] (2) To address the problem of robust heterogeneous graph classification, RHGC is proposed, which is a holistic framework that for the first time simultaneously achieves node-level, semantic-level, and graph-level robustness in HGNNs by adopting a hierarchical mechanism based on the information bottleneck principle.

[0029] (3) Extensive experiments on three molecular datasets show that RHGC improves the classification accuracy by an average of 5.06% under all three attack settings and achieves a 4.33% performance improvement on clean data.

[0030] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0032] Figure 1 A schematic diagram comparing homogeneous and heterogeneous graph neural networks in related technologies under adversarial attacks;

[0033] Figure 2 A flow chart of a robust heterogeneous graph classification method for molecular property prediction based on information bottleneck provided in accordance with an embodiment of the present invention;

[0034] Figure 3 A schematic diagram of a robust heterogeneous graph classification framework structure provided according to a specific embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the accuracy of a graph classification task against non-targeted adversarial attacks on a real-world dataset provided by a specific embodiment of the present invention;

[0036] Figure 5 A schematic diagram of an evasion attack result provided according to a specific embodiment of the present invention;

[0037] Figure 6 A schematic diagram of ablation experiment results provided according to a specific embodiment of the present invention;

[0038] Figure 7 A schematic diagram of hyperparameter analysis results for different values of the hyperparameter λ on all data sets provided according to a specific embodiment of the present invention;

[0039] Figure 8 A block diagram of a robust heterogeneous graph classification device for molecular property prediction based on information bottleneck according to an embodiment of the present invention;

[0040] Figure 9 A schematic structural diagram of an electronic device provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention, but are not to be construed as limiting the present invention.

[0042] Before introducing the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck according to an embodiment of the present invention, a comparison is first made between homogeneous and heterogeneous graph neural networks under adversarial attacks.

[0043] Specifically, Figure 1 This is a comparative diagram of homogeneous and heterogeneous graph neural networks in related technologies under adversarial attacks. Figure 1 As shown, heterogeneous graph neural networks suffer from larger global direct perturbations because the implicit patterns that rely on heterogeneity are simultaneously destroyed. In addition, heterogeneity introduces additional local intermediate perturbations because it exposes more vulnerabilities that can be exploited by adversaries.

[0044] HGNNs (Hyper-Gate Neural Networks) in the related art exhibit severe robustness weaknesses against adversarial perturbations in molecular property prediction tasks. On the one hand, previous work has neglected to constrain Y-irrelevant information at the node and semantic levels when learning robust heterogeneous node representations. Specifically, existing methods follow the pipeline of meta-path-based HGNNs and rely on the assumption that complementary information across meta-paths is predictive and noise-free. However, adversaries can inject noise to disrupt this complementarity, thereby compromising the quality of node representations. On the other hand, previous work often faces information leakage issues when optimizing robust structures at the graph level. Specifically, in graph attacks and defenses, mining the most robust graph structure is crucial for classification. Information Bottleneck (IB) methods offer a principled approach to this problem and are receiving increasing attention. However, existing IB-based methods rely on a trade-off hyperparameter that cannot achieve precise information control, resulting in the leakage of Y-irrelevant data into the final representation, thereby degrading performance.

[0045] To address these two challenges, embodiments of the present invention propose a novel Robust Heterogeneous Graph Classification (RHGC) model for molecular property prediction. This model, for the first time, achieves robust graph-level representation learning specifically for heterogeneous molecular graphs. Notably, embodiments of the present invention achieve comprehensive robustness, encompassing node-level, semantic-level, and graph-level aspects. This is achieved by leveraging the information bottleneck principle, based on which a novel hierarchical mechanism is proposed to decouple Y-relevant from Y-irrelevant information. This mechanism aims to provide robustness simultaneously at the node-level, semantic-level, and graph-level, covering the entire design space of HGNNs. To achieve this goal, RHGC comprises two novel modules. First, a specialized Node Semantic Purifier (NSP) module is introduced to constrain Y-irrelevant information in node-level and semantic-level aggregation. In particular, to learn robust heterogeneous structures, the NSP module uses graph random attention at the node level to separate Y-relevant and Y-irrelevant subgraphs. Furthermore, the Hilbert-Schmidt Independence Criterion (HSIC) is used at the semantic level to enhance complementarity in the fusion of node representations from different meta-paths, resulting in more robust node representations. Secondly, a Global Graph Decoupling (GGD) module is proposed to optimize the robust graph structure at the graph level. To prevent Y-irrelevant information from leaking into the final graph-level representation, the core of GGD is a novel decoupling graph information bottleneck objective, which theoretically guarantees accurate information control and avoids information leakage.

[0046] Specifically, Figure 2 This is a flow chart of a robust heterogeneous graph classification method for molecular property prediction based on information bottleneck provided according to an embodiment of the present invention.

[0047] like Figure 2 As shown, the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck includes the following steps:

[0048] In step S201 , a heterogeneous atlas to be classified is obtained, wherein the heterogeneous atlas to be classified includes at least one heterogeneous atlas to be classified.

[0049] The heterogeneous atlas to be classified is the heterogeneous atlas used for classification.

[0050] Specifically, when there is a need to classify a heterogeneous atlas, the embodiment of the present invention will obtain a corresponding heterogeneous atlas, and the obtained heterogeneous atlas is the heterogeneous atlas to be classified. It should be noted that there are many methods for obtaining a heterogeneous atlas to be classified, which will not be described in detail here to avoid redundancy.

[0051] In step S202, at least one heterogeneous graph to be classified is input into a pre-built robust heterogeneous graph classification model to obtain a predicted probability of the category to which each heterogeneous graph to be classified belongs, wherein the pre-built robust heterogeneous graph classification model includes a node semantic purifier and a global graph decoupling module, and the node semantic purifier is used to generate robust heterogeneous structure and node representation.

[0052] Specifically, the pre-built robust heterogeneous graph classification model can be constructed based on the robust heterogeneous graph classification framework structure. Specifically, Figure 3 As shown, Figure 3 This figure shows a schematic diagram of the robust heterogeneous graph classification framework according to a specific embodiment of the present invention. The robust heterogeneous graph classification framework (RHGC) consists of two modules: the node semantic purifier (NSP) and the global graph decoupling (GGD). The NSP generates robust heterogeneous structures and node representations. At the node level, graph random attention is used to decouple Y-related and Y-independent subgraphs. At the semantic level, HSIC is used to optimize the complementarity of representations from different meta-paths. Global graph decoupling (GGD) is then used to mitigate information leakage at the graph level. A novel decoupled graph information bottleneck is used to optimize Y-related and Y-independent subgraphs. During training, an adversarial training algorithm is used to promote better decoupling.

[0053] Specifically, in some embodiments, at the node level, the adjacency matrix induced by a given meta-path In the above example, each node contains Y-related and Y-independent information, reflecting the prediction signal and noise respectively. Two different heterogeneous graph convolutional layers are used to obtain Y-related and Y-independent node representations respectively. and In addition, in order to Generate Y-dependent and Y-independent adjacency matrices, and NSP learns two subgraph extractors based on GSAT and Subgraph Extractor For example, for each edge (u, v), the concatenation of the source node and the target node is represented by Mapped to Finally, during the forward propagation of each training iteration, each edge is drawn from the Bernoulli distribution In order to make the The sampling process is differentiable, and the Gumbel-Softmax reparameterization technique is introduced.

[0054] Furthermore, we can guide the edge mask of attention Applied to the original adjacency matrix to extract the Y-correlation adjacency matrix Expressed as here Indicates that the edge (u,v) is in Φ i Importance of the induced Y-correlated subgraph, where ⊙ is the element-wise product.

[0055] Furthermore, for the Y-independent subgraph extractor The same process is performed. By applying the above process in each meta-path induced graph, it is possible to generate and This process helps reduce the propagation of noisy information at the node level.

[0056] Furthermore, in some embodiments, at the semantic level, before semantic fusion, it is very important to further enhance the complementarity of the representations of the same node from different meta-paths. Given that HSIC has been verified in theory and practice to be an appropriate measure for evaluating the (non)independence between two signals, HSIC is used as a regularization term to mine the complementary information between different meta-paths. By applying HSIC, the mutual dependencies between meta-paths are regulated, thereby promoting the adaptive extraction and fusion of information and significantly improving the prediction performance. The HSIC formula is as follows:

[0057]

[0058] in I(∈R N×N ) represents the identity matrix, and 1 represents a column vector whose elements are all 1. K X is the kernel matrix, whose elements Where σ is the bandwidth. The semantic HSIC objective of constructing the node semantic purifier using HSIC is as follows:

[0059]

[0060] In this way, noise information is eliminated at the semantic level. At the same time, semantic level attention is used to obtain the final Y-related and Y-irrelevant node representations, which are represented as Z T and Z S .

[0061] In summary, the robust node representation Z T (corresponding to Z S ) and Y-dependent (correspondingly Y-independent) adjacency matrix A T (corresponding to A S ) constitutes the Y-dependent (corresponding to Y-independent) subgraph G T (corresponding to G S ).

[0062] Secondly, it should be noted that for global graph decoupling, after extracting the robust node representation Z through the NSP module T ,Z S and the corresponding adjacency matrix AT ,A S Afterwards, the framework still faces the significant challenge of information leakage when learning robust structures for graph classification. The problem is that existing works inevitably lack precise information control at the graph level. For example, the work based on the information bottleneck (IB) identifies an information-rich but compressed subgraph G from the original graph G by optimizing the following objective: T :

[0063]

[0064] Among them G sub represents the set of all subgraphs of G, Y is the label of G, I(Y; G T ) and I(G;G T ) represent Y and G respectively T Between, G and G T The mutual information between them, β is the Lagrange multiplier, which is used to control the trade-off between these two items. In the traditional IB optimization process, multiple attempts are required to find the optimal value.

[0065] Specifically, the embodiment of the present invention extends the decoupling information bottleneck to the graph domain and defines the decoupling graph information bottleneck goal of the GGD module as follows:

[0066]

[0067] The above objectives can be divided into three tasks: First, maximize I(G T ; Y) to ensure that G T Accurately decode Y to ensure G T At least the information related to Y is captured (i.e., task 1). Secondly, maximize I(G; G S ,Y) to ensure (G S ,Y) represents the overall information of G, thus ensuring that G S Covering the information irrelevant to Y (Task 2). Finally, minimize I(G S ; G T ) to force G S and G T The decoupling between them effectively separates Y-related information from Y-irrelevant information, thereby tightening the information boundary between the two subgraphs (i.e., Task 3). It can be seen that modeling the Y-irrelevant part into the IB objective solves the problems of information leakage and multiple optimization of β at the same time.

[0068] Among them, for the optimization of Task 1 and Task 2 of the embodiment of the present invention, first derive I(G T ; Y) and I(G; G S ,Y) is as follows:

[0069]

[0070] Among them, c(y|G T ) and r(G|G S ,y) are two variational probability maps. The former can be obtained by G T Easily modeled as a cross-entropy classification loss

[0071]

[0072] Among them, the variational probability map c(y|G T ) is parameterized by a graph classifier f, Representation Subgraph Extractor The latter is achieved through G S and label y is modeled as graph reconstruction loss

[0073]

[0074] Among them, the variational probability mapping r(G|G S ,y) is parameterized by the graph reconstructor R, a and denote the adjacency matrices of the original graph and the reconstructed graph respectively.

[0075] Secondly, for the optimization of Task 3 of the embodiment of the present invention, directly optimize the target I(G S ; G T )=D KL [p(G S ,G T )∥pGSpGT is intractable because pGS, GT and pGSpGT all contain many correlated random variables, which makes it difficult to eliminate all of them. Inspired by generative adversarial networks, GGD reduces p(G S ,G T ) and p(G S )p(G T ) to minimize the gap between I(G S ; G T ).

[0076] Specifically, GGD is performed by uniformly sampling x randomly from the dataset and then extracting the x from the conditional distribution p(G S ,G T |x) to obtain the joint distribution p(G S ,G T ) samples. In order to approximate the product p(G S )p(G T ), we randomly shuffle p(G S ,G T) samples. Finally, GGD uses the density ratio trick to introduce a discriminator d to estimate its input from p(G S ,G T ) rather than p(F S )p(G T ) probability:

[0077]

[0078] It should be noted that the neural network training method in the embodiment of the present invention is set to adversarial training.

[0079] Furthermore, in some embodiments, the preset loss function is:

[0080]

[0081] Where λ is a constant, is the loss function of the NSP module, is the decoupling graph information bottleneck loss function of the GGD module, is the cross entropy loss function. Reconstruct the loss function for the graph, is the discriminator loss function. Among them, the loss from the NSP module and the decoupled graph information bottleneck loss from the GGD module are jointly optimized in adversarial training.

[0082] To facilitate those skilled in the art to further understand the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck according to an embodiment of the present invention, the pre-constructed robust heterogeneous graph classification model is described in detail below.

[0083] Optionally, in some embodiments, before the heterogeneous graph to be classified is input into a pre-constructed robust heterogeneous graph classification model, it also includes: constructing a target heterogeneous graph set for training; based on a preset division ratio, dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set, and using the heterogeneous graph training set to train the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model; based on a preset loss function and a heterogeneous graph verification set, verifying the initial robust heterogeneous graph classification model; if the verification result meets the training conditions, the initial robust heterogeneous graph classification model is used as the pre-constructed robust heterogeneous graph classification model; otherwise, the preset division ratio is adjusted, and the step of dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on the preset division ratio is re-executed until the new verification result meets the training conditions, and the robust heterogeneous graph classification model corresponding to the new verification result is used as the pre-constructed robust heterogeneous graph classification model.

[0084] In some embodiments, constructing a target heterogeneous atlas for training includes: determining multiple target molecules, node type mapping functions, and edge type mapping functions; based on the node type mapping function and the edge type mapping function, constructing a heterogeneous graph including a node set and an edge set for each target molecule to obtain a target heterogeneous atlas.

[0085] Specifically, the heterogeneous graph is represented as Contains node sets and edge set ε. The heterogeneous graph also has a node type mapping function φ: And edge type mapping function ψ: associated. and denote the predefined node type set and edge type set respectively, where For each edge type A R Denotes the corresponding binary adjacency matrix. The element path Φ is defined as describing the composite relation R = R1°R2°…°R l , where ° represents the combination operator of the relation. For each element path Φ i , represents the binary adjacency matrix induced by metapaths. Given a set of heterogeneous molecular graphs in Each G i are assigned a label C is the total number of categories. For pairs of heterogeneous graphs and labels The goal of robust heterogeneous graph classification is to learn a graph classifier f: The classifier is able to obtain more robust representations and perform downstream graph-level classification tasks. During testing, the classification performance of the learned graph classifier is expected to be able to resist both targeted and untargeted adversarial attacks.

[0086] In some embodiments, a heterogeneous graph training set is used to train a heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model, including: inputting each target heterogeneous graph in the heterogeneous graph training set into a node semantic purifier, so that the node semantic purifier processes each target heterogeneous graph through graph random attention and Hilbert-Schmidt independence criterion to generate a robust node representation, a related adjacency matrix and an irrelevant adjacency matrix; inputting the robust node representation, the related adjacency matrix and the irrelevant adjacency matrix into a global graph decoupling module, optimizing each target heterogeneous graph by decoupling the graph information bottleneck, and obtaining a robust graph representation of each target heterogeneous graph; based on each target heterogeneous graph and the robust graph representation of each target heterogeneous graph, training the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model.

[0087] First, it should be noted that for node purifiers (NSP), mainstream meta-path-based heterogeneous graph neural networks (HGNNs) learn node representations on intermediate graphs induced by different meta-paths (node-level aggregation) and then use semantic fusion to learn the final node representations (semantic-level aggregation). However, related technologies completely ignore the constraint of Y-irrelevant information in node-level and semantic-level aggregation. Different meta-paths describe semantics from different perspectives, and the true predictive information contained in different meta-paths is complementary, which is called the complementarity assumption. However, in practice, adversaries can perturb the intermediate graph induced by meta-paths, accumulating noise in node-level and semantic-level aggregation, thereby destroying complementarity.

[0088] Therefore, an embodiment of the present invention proposes to simultaneously achieve robustness at the node level and the semantic level to block Y-irrelevant information, that is, a node semantic purifier (NSP) based on node-level graph random attention is used to obtain a robust graph structure, and at the same time, the Hilbert-Schmidt independence criterion (HSIC) is used at the semantic level to further promote the complementarity of representations from different meta-paths when aggregated.

[0089] In order to enable relevant technical personnel in this field to better understand the effectiveness of the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck in the embodiment of the present invention, the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck in the embodiment of the present invention will be described in detail below in combination with specific embodiments.

[0090] It should be noted that the embodiments of the present invention are experimental settings including data sets, baseline methods, evaluation indicators, adversarial attack settings and implementation details.

[0091] First, the embodiment of the present invention uses three datasets to evaluate the performance of the model, including Open Graph Benchmark (OGB)-molbace, MUTAG and PROTEINS. These datasets are related to molecules or bioinformatics and are widely used for graph classification evaluation.

[0092] Secondly, baseline methods and evaluation indicators are selected. The baseline methods include heterogeneous graph classification methods and robust homogeneous graph classification methods. For heterogeneous graph classification methods, two baseline methods are selected, including muxGNN and HeGCL. For robust homogeneous graph classification methods, two data augmentation methods (including NodeSam and G-Mixup), two graph representation learning methods (including InfoGraph and MGRL), and three graph information bottleneck methods (including GIB, VIB-GSL and PGIB) are selected. For evaluation indicators, the accuracy of the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck provided by the embodiment of the present invention on all data sets will be reported.

[0093] Again, the embodiments of the present invention compare the baseline method and the proposed RHGC under two adversarial attack settings. For non-targeted adversarial attacks, synthetic datasets are generated by attacking the graph structure and node features respectively, including: (1) attacking the graph structure, selecting the nodes with the highest degree and randomly inserting / deleting an edge between them. For example, 20% of the nodes in the graph are selected for perturbation. (2) attacking the node features, adding random Gaussian noise m·r·∈ to each dimension of the node features of all nodes, where r is the reference amplitude of the original feature, ∈~N(0,I). m is used as a parameter to control the amplitude of feature adjustment. For example, the value of m is set to 2.0 in the experimental setting of the embodiments of the present invention.

[0094] Furthermore, for targeted adversarial attacks, we focus on evasion attacks, which are a typical type of targeted adversarial attacks that perturb the graph and guide the model to misclassify the target type graph during the test phase. The present embodiment generates perturbations based on GRABNEL because it can be easily adapted to perform various attacks on different meta-paths under the black-box evasion attack setting. The attack budget Δ (i.e., the maximum Δ edges allowed to be flipped from G) is set to Δ for each graph G. i 10%|ε i |, where |ε i | represents graph G i The number of edges in .

[0095] Finally, the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck provided by an embodiment of the present invention was implemented using the PyTorch 2.1 framework on Ubuntu 22.04. The maximum number of training rounds was set to 200, and an early stopping strategy was adopted. Each dataset was divided into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10%, respectively. The Adam optimizer was used, and the learning rate was selected from {1e-03, 3e-03, 1e-04, 3e-04}. The best performance was obtained through grid search using the validation set. To mitigate the impact of random noise, the results of 10 runs with different random seeds were reported in the experiment.

[0096] Thus, the experimental setting of the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck provided by the embodiment of the present invention is completed.

[0097] In the following, robustness experiments, ablation studies, and hyperparameter analysis against non-targeted and targeted adversarial attacks will be conducted in conjunction with specific embodiments to verify the RHGC method proposed in the embodiments of the present invention.

[0098] First, we will evaluate the performance of the model on heterogeneous graph classification tasks. Figure 4This is a schematic diagram of the accuracy of a graph classification task against non-targeted adversarial attacks on a real-world dataset provided according to a specific embodiment of the present invention, where the best result is shown in bold and the second best result is underlined.

[0099] like Figure 4 As shown, all models experience significant performance degradation under both feature and structural attacks, highlighting their common limitations. Adversarial attacks can significantly degrade the performance of heterogeneous graph classification baselines, with an average performance drop of 8.12%, indicating that HGNNs are more susceptible to global direct perturbations. Although robust homogeneous graph classification methods exhibit some resistance to adversarial attacks, they typically experience a 5.68% accuracy drop. Furthermore, these methods often fail to achieve precise information control and fail to address the additional vulnerabilities introduced by heterogeneity, resulting in subpar performance. In contrast, RHGC consistently outperforms other methods under all attack settings, with an average classification performance drop of only 4.94%. Furthermore, RHGC achieves an average improvement of 5.53% over the baseline under both structural and feature attacks. These results demonstrate RHGC's superior performance against both structural and feature attacks.

[0100] It should be understood that RHGC also outperforms all baseline methods on all unattacked datasets. Compared with the most advanced baseline methods, the method of the embodiment of the present invention improves the accuracy by 3.74%, 4.26% and 4.99% on PROTEINS, OGB-molbace and MUTAG, respectively. These results demonstrate the effectiveness of this method in molecular property prediction. This is mainly due to the NSP module and GGD module, which can effectively remove Y-irrelevant information and extract robust graph representations by decoupling the graph information bottleneck objective. Therefore, the method of the embodiment of the present invention can make full use of truly predictive information to enhance the performance of molecular property prediction.

[0101] Second, we consider the classification performance and robustness to the targeted adversarial attack to verify whether RHGC successfully defends against the attack. Figure 5 The figure is a schematic diagram of the results of evasion attacks provided according to a specific embodiment of the present invention, wherein the best result is shown in bold and the second best result is underlined.

[0102] like Figure 5 As shown in Figure 3, all models experience significant performance degradation under evasion attacks, with the baseline method experiencing an average drop of 18.40% in classification accuracy, indicating that considering meta-path perturbations may lead to more severe consequences. In contrast, RHGC experiences an average drop of 17.99% and achieves 4.89% higher accuracy under evasion attacks than the baseline method. This demonstrates the importance of enhancing robustness across all dimensions of the HGNNs design space.

[0103] To validate the effectiveness of the various modules in the proposed RHGC, ablation experiments were conducted on all datasets using different combinations of key modules. The baseline model is the GSAT backbone network, the NSP module is the baseline model, the GGD module is the baseline model, and the NSP+GGD module is the RHGC model proposed in this embodiment.

[0104] Figure 6 This is a schematic diagram of the ablation experiment results provided according to a specific embodiment of the present invention, wherein the baseline model is the basic model of the GSAT backbone network, NSP is the baseline model with the NSP module, GGD is the baseline model with the GGD module, and NSP+GGD is the RHGC model proposed in an embodiment of the present invention. The best result is shown in bold, and the second best result is underlined.

[0105] like Figure 6 As shown in the figure, the GSAT backbone network performs the worst, which is due to the neglect of heterogeneity and the lack of robust structure in modeling. Integrating the GGD module into the GSAT backbone network significantly improves the classification performance, which shows that the GGD module can effectively decouple global Y-irrelevant information from the robust graph-level representation. In addition, integrating the NSP module also improves the classification performance, which proves the effectiveness of constraining Y-irrelevant information in the node-level and semantic-level aggregation processes. The robust heterogeneous graph classification method for molecular property prediction based on information bottleneck in an embodiment of the present invention combines the RHGC of the NSP and GGD modules to fully utilize the above advantages and achieve optimal performance.

[0106] Finally, the hyperparameters are analyzed to evaluate the classification performance of the RHGC method according to the embodiment of the present invention when the key hyperparameter λ takes different values.

[0107] Figure 7 This is a schematic diagram of hyperparameter analysis results for different values of the hyperparameter λ on all data sets provided according to a specific embodiment of the present invention, where the values of λ are {0.01, 0.1, 0.3, 0.4, 1}.

[0108] like Figure 7 As shown, when λ=0.1, the method of the embodiment of the present invention achieves the best performance on all data sets. Therefore, the value of λ can be set to 0.1 in all experiments.

[0109] Therefore, the novel RHGC method proposed in this embodiment of the present invention is the first to achieve robust heterogeneous graph representation learning for molecular property prediction. It introduces a node semantic purifier and a global graph decoupling module to effectively decouple Y-dependent and Y-independent information at the node, semantic, and graph levels, thereby covering the full design space of HGNNs. Extensive experiments on three molecular benchmark datasets demonstrate that RHGC demonstrates significantly superior robustness to state-of-the-art baseline methods in molecular property prediction tasks, both on noisy and clean data.

[0110] According to an embodiment of the present invention, the robust heterogeneous graph classification method for molecular property prediction based on information bottlenecks uses graph random attention at the node level through a node semantic purifier to separate Y-related and Y-independent subgraphs. At the semantic level, the Hilbert-Schmidt independence criterion is used to enhance the complementarity of node representations from different meta-paths during fusion. Furthermore, global graph decoupling is used to optimize the robust graph structure at the graph level. This method solves the problems of reduced node representation quality and low classification accuracy caused by adversarial attacks in related technologies, thereby improving classification accuracy under adversarial attacks.

[0111] Next, a robust heterogeneous graph classification device for molecular property prediction based on information bottleneck proposed in an embodiment of the present invention will be described with reference to the accompanying drawings.

[0112] Figure 8 Schematic diagram of a block diagram of a robust heterogeneous graph classification device for molecular property prediction based on information bottleneck according to an embodiment of the present invention.

[0113] like Figure 8 As shown, the molecular property prediction robust heterogeneous graph classification device 10 based on information bottleneck includes: an acquisition module 100 and a training module 200.

[0114] Among them, the acquisition module 100 is used to obtain a set of heterogeneous graphs to be classified, wherein the set of heterogeneous graphs to be classified includes at least one heterogeneous graph to be classified; the training module 200 is used to input at least one heterogeneous graph to be classified into a pre-built robust heterogeneous graph classification model to obtain the predicted probability of the category to which each heterogeneous graph to be classified belongs, wherein the pre-built robust heterogeneous graph classification model includes a node semantic purifier and a global graph decoupling module, and the node semantic purifier is used to generate robust heterogeneous structure and node representation.

[0115] Furthermore, in some embodiments, before the heterogeneous graph to be classified is input into a pre-constructed robust heterogeneous graph classification model, the acquisition module 100 is specifically used to: construct a target heterogeneous graph set for training; based on a preset division ratio, divide the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set, and use the heterogeneous graph training set to train the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model; based on a preset loss function and a heterogeneous graph verification set, verify the initial robust heterogeneous graph classification model; if the verification result meets the training conditions, the initial robust heterogeneous graph classification model is used as the pre-constructed robust heterogeneous graph classification model; otherwise, the preset division ratio is adjusted, and the step of dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on the preset division ratio is re-executed until the new verification result meets the training conditions, and the robust heterogeneous graph classification model corresponding to the new verification result is used as the pre-constructed robust heterogeneous graph classification model.

[0116] Furthermore, in some embodiments, the acquisition module 100 is also used to: determine multiple target molecules, node type mapping functions, and edge type mapping functions; based on the node type mapping function and the edge type mapping function, for each target molecule, construct a heterogeneous graph including a node set and an edge set to obtain a target heterogeneous graph set.

[0117] Furthermore, in some embodiments, the training module 200 is specifically used to: input each target heterogeneous graph in the heterogeneous graph training set into the node semantic purifier, so that the node semantic purifier processes each target heterogeneous graph through graph random attention and Hilbert-Schmidt independence criterion to generate a robust node representation, a related adjacency matrix and an irrelevant adjacency matrix; input the robust node representation, the related adjacency matrix and the irrelevant adjacency matrix into the global graph decoupling module, optimize each target heterogeneous graph by decoupling the graph information bottleneck, and obtain a robust graph representation of each target heterogeneous graph; based on each target heterogeneous graph and the robust graph representation of each target heterogeneous graph, train the heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model.

[0118] Furthermore, in some embodiments, the preset loss function is:

[0119]

[0120] Where λ is a constant, is the loss function of the NSP module, is the decoupling graph information bottleneck loss function of the GGD module, is the cross entropy loss function. Reconstruct the loss function for the graph, is the discriminator loss function.

[0121] It should be noted that the above explanation of the embodiment of the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck is also applicable to the robust heterogeneous graph classification device for molecular property prediction based on information bottleneck in this embodiment, and will not be repeated here.

[0122] According to the information bottleneck-based robust heterogeneous graph classification device for molecular property prediction provided by the embodiments of the present invention, a node semantic purifier uses graph random attention at the node level to separate Y-related and Y-independent subgraphs. At the semantic level, the Hilbert-Schmidt independence criterion is used to enhance the complementarity of node representations from different meta-paths during fusion. Global graph decoupling is then used to optimize the robust graph structure at the graph level. This solves the problems of reduced node representation quality and low classification accuracy caused by adversarial attacks in related technologies, thereby improving classification accuracy under adversarial attacks.

[0123] Figure 9 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of the present invention. The electronic device may include:

[0124] A memory 901 , a processor 902 , and a computer program stored in the memory 901 and executable on the processor 902 .

[0125] When the processor 902 executes the program, the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck provided in the above embodiment is implemented.

[0126] Furthermore, the electronic device further includes:

[0127] The communication interface 903 is used for communication between the memory 901 and the processor 902 .

[0128] The memory 901 is used to store computer programs that can be run on the processor 902 .

[0129] The memory 901 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0130] If the memory 901, processor 902, and communication interface 903 are implemented independently, the communication interface 903, memory 901, and processor 902 can be connected to each other via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0131] Optionally, in a specific implementation, if the memory 901, the processor 902 and the communication interface 903 are integrated on a chip, the memory 901, the processor 902 and the communication interface 903 can communicate with each other through an internal interface.

[0132] The processor 902 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention.

[0133] In addition, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck is implemented as described above.

[0134] In addition, an embodiment of the present invention further provides a computer program product, including a computer program, which is executed to implement any of the above-mentioned robust heterogeneous graph classification methods for molecular property prediction based on information bottleneck.

[0135] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0136] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0137] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0138] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0139] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

Claims

1. A robust heterogeneous graph classification method for molecular property prediction based on information bottleneck, characterized by: The following steps are involved: Acquire a heterogeneous atlas to be classified, wherein the heterogeneous atlas to be classified includes at least one heterogeneous atlas to be classified; The at least one heterogeneous graph to be classified is input into a pre-built robust heterogeneous graph classification model to obtain a predicted probability of the category to which each heterogeneous graph to be classified belongs, wherein the pre-built robust heterogeneous graph classification model includes a node semantic purifier and a global graph decoupling module, and the node semantic purifier is used to generate a robust heterogeneous structure and node representation.

2. The method according to claim 1, characterized in that Before inputting the heterogeneous graph to be classified into the pre-built robust heterogeneous graph classification model, the method further includes: Construct a target heterogeneous atlas for training; Based on a preset division ratio, the target heterogeneous graph set is divided into a heterogeneous graph training set and a heterogeneous graph verification set, and the heterogeneous graph training set is used to train a heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model; Based on a preset loss function and the heterogeneous graph verification set, the initial robust heterogeneous graph classification model is verified. If the verification result meets the training conditions, the initial robust heterogeneous graph classification model is used as the pre-constructed robust heterogeneous graph classification model. Otherwise, the preset division ratio is adjusted, and the step of dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on the preset division ratio is re-executed until a new verification result meets the training conditions. The robust heterogeneous graph classification model corresponding to the new verification result is used as the pre-constructed robust heterogeneous graph classification model.

3. The method according to claim 2, characterized in that The construction of a target heterogeneous atlas for training includes: determining a plurality of target molecules, node type mapping functions, and edge type mapping functions; Based on the node type mapping function and the edge type mapping function, for each target molecule, a heterogeneous graph including a node set and an edge set is constructed to obtain the target heterogeneous graph set.

4. The method according to claim 3, characterized in that The method of using the heterogeneous graph training set to train a heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model includes: Inputting each target heterogeneous graph in the heterogeneous graph training set into the node semantic purifier, so that the node semantic purifier processes each target heterogeneous graph through graph random attention and Hilbert-Schmidt independence criterion to generate a robust node representation, a related adjacency matrix and an irrelevant adjacency matrix; Inputting the robust node representation, the relevant adjacency matrix, and the irrelevant adjacency matrix into the global graph decoupling module, optimizing each target heterogeneous graph by decoupling the graph information bottleneck, and obtaining a robust graph representation of each target heterogeneous graph; Based on each target heterogeneous graph and the robust graph representation of each target heterogeneous graph, the heterogeneous graph neural network is trained to obtain an initial robust heterogeneous graph classification model.

5. The method according to claim 3, characterized in that The preset loss function is: Where λ is a constant, is the loss function of the NSP module, is the decoupling graph information bottleneck loss function of the GGD module, is the cross entropy loss function, Reconstruct the loss function for the graph, is the discriminator loss function.

6. A robust heterogeneous graph classification device for molecular property prediction based on information bottleneck, characterized in that: include: An acquisition module, configured to acquire a heterogeneous atlas to be classified, wherein the heterogeneous atlas to be classified includes at least one heterogeneous atlas to be classified; A training module is used to input the at least one heterogeneous graph to be classified into a pre-built robust heterogeneous graph classification model to obtain the predicted probability of the category to which each heterogeneous graph to be classified belongs, wherein the pre-built robust heterogeneous graph classification model includes a node semantic purifier and a global graph decoupling module, and the node semantic purifier is used to generate robust heterogeneous structure and node representation.

7. The device according to claim 6, characterized in that Before inputting the heterogeneous graph to be classified into the pre-built robust heterogeneous graph classification model, the acquisition module is further configured to: Construct a target heterogeneous atlas for training; Based on a preset division ratio, the target heterogeneous graph set is divided into a heterogeneous graph training set and a heterogeneous graph verification set, and the heterogeneous graph training set is used to train a heterogeneous graph neural network to obtain an initial robust heterogeneous graph classification model; Based on a preset loss function and the heterogeneous graph verification set, the initial robust heterogeneous graph classification model is verified. If the verification result meets the training conditions, the initial robust heterogeneous graph classification model is used as the pre-constructed robust heterogeneous graph classification model. Otherwise, the preset division ratio is adjusted, and the step of dividing the target heterogeneous graph set into a heterogeneous graph training set and a heterogeneous graph verification set based on the preset division ratio is re-executed until a new verification result meets the training conditions. The robust heterogeneous graph classification model corresponding to the new verification result is used as the pre-constructed robust heterogeneous graph classification model.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck as described in any one of claims 1 to 5.

10. A computer program product comprising a computer program, characterized in that The computer program is executed to implement the robust heterogeneous graph classification method for molecular property prediction based on information bottleneck according to any one of claims 1 to 5.

Images