Graph representation learning method in government affair public service venue intelligent management based on generalized graph edge-to-vertex conversion and information bottleneck

Abstract

The invention relates to a graph representation learning method in government affair public service venue intelligent management based on generalized graph edge-to-vertex conversion and information bottleneck. A graph representation learning method in government affair public service venue intelligent management based on generalized graph edge-to-vertex conversion and information bottleneck comprises the following steps: generating a vertex graph # imgabs1 # and a vertex graph # imgabs2 # from an original graph # imgabs0 # through a view generator, and performing feature extraction; the formula of the feature extraction process of the vertex graph # imgabs3 # is # imgabs4 #, the vertex graph # imgabs5 # is subjected to generalized edge-to-vertex conversion, after a corresponding line graph # imgabs6 # is obtained, edge features are extracted, and the formula of the feature extraction process of the vertex graph # imgabs3 # is # imgabs7 #. The method has the advantages that the classification performance is good, the information loss is reduced, the method has obvious advantages in the aspects of online graph scale and conversion efficiency, and meanwhile, the method has more excellent performance in the aspects of graph classification and node classification tasks.

Description

Technical Field

[0001] The present invention belongs to the field of government data processing, and relates to the integration and analysis of multi-source heterogeneous data in government data. Specifically, it relates to a graph representation learning method in the intelligent management of government public service venues based on generalized graph edge-to-vertex conversion and information bottlenecks. Background Art

[0002] Graphs, as a common data structure, are widely used in fields such as social networks, multimedia analysis, and chemical molecules. In a graph, vertices typically represent descriptions of objects, while edges represent specific relationships between objects. Because vertices and edges are rich in information, graphs can describe almost any type of structure or system. For example, the properties of a compound can be determined by searching for functional groups in chemical molecules; human behaviors and events in videos can also be automatically identified and understood for video retrieval, monitoring, and recommendation. In recent years, graph-based representation learning techniques have received significant attention, especially unsupervised graph representation learning (UGRL). UGRL aims to extract high-level information from graphs and derive vectors that truly reflect the topological structure of the original space without manual annotation.

[0003] Existing UGRL methods are generally designed to minimize unimodal losses or maximize mutual information between node representations. Hou (2022) proposed the use of masking strategies and scaled cosine errors for feature reconstruction; Velickovic (2018) encouraged the learned representations to retain the key structural information of the graph by maximizing the mutual information of positive sample pairs, while minimizing the mutual information of negative sample pairs to distinguish between positive and negative samples. Sun (2021) proposed a new variational information bottleneck to guide graph structure learning from the perspective of information theory; Liu (2023) generated global and local views of different scales based on subgraph sampling, and constructed multiple contrast relationships based on their semantic associations. Although the above studies have achieved remarkable performance, the UGRL method still has certain limitations in effectiveness and transfer efficiency. Specifically:

[0004] On the one hand, the discriminative representations learned using a unimodal loss lack sufficient expressiveness and discrimination. Graph representation learning methods that use reconstruction as a learning objective capture latent features and structure through a priori graph encoders. This causes the graph encoder to focus on every detail of the reconstruction process, deviating from learning high-quality representation vectors.

[0005] On the other hand, learning graph representations based on mutual information can lead to information redundancy and loss. Without explicit constraints on node representation diversity during model learning, certain features in the representation space may be unnecessarily learned repeatedly, resulting in information redundancy. Furthermore, maximizing mutual information between nodes can lead to learned representations focusing too heavily on local similarities, trapping them in local optima. In this case, the model may fail to capture relevant information in the data, resulting in information loss.

[0006] The digital and intelligent operation of important social public service venues faces the problem of integrating and analyzing multi-source heterogeneous data. This data includes personnel information, vehicle information, equipment information, etc. The relationships between them are complex and diverse, forming a complex graph data structure. For example, the relationship between personnel and vehicles, and the usage relationship between equipment and personnel, etc., are of great significance to the operation management, security and service optimization of the venue. However, traditional data processing methods have difficulty in effectively mining the potential information in this data and cannot meet the needs of digital and intelligent operation of venues. Therefore, a new graph representation learning method is needed that can effectively process multi-source heterogeneous data and extract feature vectors with high expressiveness and discrimination to support the intelligent management of public service venues.

[0007] In view of this, the present invention proposes a new graph representation learning method for intelligent management of government public service venues, which is based on generalized graph edge-to-vertex conversion and information bottleneck, avoids information loss as much as possible, and obtains better graph classification and node classification effects. Summary of the Invention

[0008] The purpose of the present invention is to provide a graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex conversion and information bottlenecks, which performs well in graph classification and node classification with less information loss.

[0009] In order to achieve the above objectives, the technical solutions adopted are:

[0010] The graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex transformation and information bottleneck includes the following steps:

[0011] The original image Generating a vertex map via the View Builder and vertex graph Then, feature extraction is performed;

[0012] The vertex graph The formula for the feature extraction process is: In the formula represents the l-th level vertex features or embeddings, and

[0013] The vertex graph Perform generalized edge-to-vertex transformation to obtain the corresponding line graph After that, edge features are extracted, and the formula is: In the formula represents the l-th layer feature or embedding, and

[0014] Furthermore, the vertex graph In the feature extraction process, for graph-level tasks, the classic graph isomorphism network GIN is selected; for node-level tasks, the classic graph convolutional network GCN is selected.

[0015] Furthermore, the vertex graph After generating the connected graph and obtaining its corresponding non-zero degree connected graph, perform edge-to-vertex mapping f and convert it into the corresponding line graph. Extract edge features.

[0016] Furthermore, the connectivity process is as follows:

[0017] When the vertex graph For a directed graph, if there are i vertices with in-degree 0 and j vertices with out-degree 0 Then modify the graph; the modified graph Include a virtual vertex and i+j virtual edges; where the virtual vertex points to the vertex with in-degree 0 to form a virtual edge And pointed to by the vertex with out-degree 0 to form a virtual edge

[0018] When the vertex graph is an undirected graph, if There are no isolated vertices, no vertices with degree 0, and no operation is performed; if the graph There are i isolated vertices Then modify the graph; the modified graph Include a virtual vertex and 2*i virtual edges, where the virtual vertex passes through two virtual edges and Connect each isolated vertex

[0019] Further, after the feature extraction, the As an abstract representation of the original graph edge structure, information is added In the overall optimization, the formula is:

[0020] Furthermore, the mutual information and The following inequalities hold:

[0021]

[0022] Furthermore, in the overall optimization process, the overall training goal is:

[0023]

[0024]

[0025] Furthermore, the original graph is preprocessed data, and the process is: after collecting multi-source heterogeneous data, constructing a graph structure, and then performing cleaning and standardization processing to obtain the original graph.

[0026] Furthermore, the multi-source heterogeneous data includes: personnel information, vehicle information, equipment information and the interactive relationships between them.

[0027] Furthermore, in the process of constructing the graph structure, personnel information, vehicle information, and equipment information are used as vertices, and interaction relationships are used as edges.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] In response to the problems raised in the background technology, the present invention proposes a graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex transformation and information bottleneck, which is a graph representation learning network CONGIB based on the principle of supplementary graph information bottleneck. First, the problem of information loss in traditional edge-to-vertex transformation is discussed, and then a generalized graph edge-to-vertex transformation is proposed. CONGIB constructs a line graph as a supplementary modality of the original graph through the generalized graph edge-to-vertex transformation. On this basis, the original graph and the line graph are forced to filter redundant information, and the constraint model is guided by the expression of its own view to learn a more compact feature vector. At the same time, the line graph is used as a supplement to the abstract space transfer state of the original graph to make up for the non-redundant information lost when filtering information. Compared with the previous existing technology, the main innovations of the present invention are as follows:

[0030] 1. In the technical solution of the present invention, a special edge-to-vertex conversion process is proposed, which extends the edge-to-vertex conversion from connected graphs to generalized graph structures, and proves that the line graph obtained by this process is lossless.

[0031] 2. This technical solution proposes a graph representation learning network called CONGIB, based on the information bottleneck principle. This network maximizes the representation of line graphs by embedding features that complement the original graph feature vectors, thereby obtaining information-rich graph representations for downstream tasks. This is the first attempt to incorporate line graphs into graph representation learning.

[0032] 3. In the technical solution of the present invention, extensive experiments were conducted on graph classification tasks and node classification tasks. The experimental results show that CONGIB can produce performance that is superior to current advanced graph representation learning methods and achieve SOTA results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 In the figure, a is the traditional edge-to-vertex conversion process, and b is an example of information loss;

[0034] Figure 2 is the generalized graph edge to vertex conversion process; a is the directed graph conversion process, b is the undirected graph conversion process;

[0035] Figure 3 The proposed framework of the technical solution of the present invention;

[0036] Figure 4 The speed of edge-to-vertex conversion in each method and the size of the converted graph compared with the original graph;

[0037] Figure 5 For different training samples, the variance and information entropy evolve with the training period;

[0038] Figure 6 Supplementary information is affected by view type. DETAILED DESCRIPTION

[0039] In order to further illustrate the graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex conversion and information bottlenecks of the present invention and to achieve the intended purpose of the invention, the graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex conversion and information bottlenecks proposed by the present invention is described in detail below in combination with the preferred embodiments, as follows. In the following description, different "one embodiment" or "embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0040] The following is a detailed introduction to the graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex conversion and information bottlenecks, combined with specific embodiments:

[0041] Unsupervised graph representation learning has attracted widespread attention due to its excellent performance in modeling complex structures and high-dimensional features. However, the actual performance of existing graph representation learning is limited by the following facts: the insufficient expression ability of a single discriminant representation and the redundancy and loss in the information transmission process, which may cause the learned feature vectors to deviate from the learned high-quality representation or focus too much on local similarities and miss global information. In response to the limitations of existing unsupervised graph representation learning methods, the present invention starts from the edge-to-vertex transformation process and discusses a generalized graph edge-to-vertex lossless transformation process. Based on the line graph obtained by this transformation process, a supplementary graph information bottleneck network CONGIB is proposed. CONGIB encourages the compression of the original graph and line graph information to constrain the model to learn more compact feature vectors, and uses line graph features as a supplement to the abstract space transfer state of the original graph to obtain rich information mapping. A large number of experiments were carried out on 13 public datasets, and the results show that the method of the present invention outperforms the latest and most advanced methods on different downstream tasks. The specific embodiments are as follows:

[0042] Example 1.

[0043] The specific steps are as follows:

[0044] A. Preliminaries and Related Work

[0045] Given a graph in Representative diagram The vertex set of N represents the graph The number of vertices. E represents the edge set. X=[x1,x2,...,x N ]∈R N×D Represents the vertex feature matrix, x i Represents the D-dimensional feature vector of vertex i.

[0046] 1. Edge-to-vertex Transforms

[0047] Vertices and edges are information carriers in a graph. Vertices store more state information, while edges represent more abstract transitions between states. In graph theory, a graph is converted to a line graph by converting edges to vertices, where the original vertex attributes are stored in the edges, and the original edge attributes are stored in the vertices of the line graph, such as Figure 1 As shown in (a).

[0048] Edge to Vertex Conversion: Set is a connected graph The line graph (also called the dual graph) of is given by the bijection f: Get. Similarly, Represents the vertex set of graph H, EH =[e1,e2,...,e M ]∈R M represents the edge set, and M represents the number of edges in graph H. e =[x e1 ,x e2 ,...,x eM ]∈R M×D Represents the edge feature matrix (note that this is not the vertex feature matrix). i ∈E H With each edge are associated. And the two vertices If and only if the edge d=f(v i )∈E H The destination vertex is the edge e=f(v j )∈E H The source vertex is connected (v i ,v j ), then it is called a bijection f: For edge-to-vertex mapping, the process is called edge-to-vertex conversion.

[0049] There may be information loss in the edge-to-vertex conversion process. Figure 1 As shown in (b), lowercase letters represent vertices of the original graph (or edges of the corresponding line graph), and numbers represent edges of the original graph (or vertices of the corresponding line graph). It can be seen that after the traditional edge-to-vertex conversion, the information of vertices b and c in the original graph has been lost.

[0050] Over the past few years, some research and applications have been conducted on edge-to-vertex conversion: for example, Chen (2019) proposed using the line graph of the original graph to supplement the edge label information or structural information of the original graph omitted by Graph2vec (2017); to address the problem of information loss in the edge-to-vertex conversion process, Liu & Song (2022) proposed retaining the complete information of the vertex graph by introducing reverse edges with specific edge labels; Liu & Jiang (2022) proposed inserting a specific virtual vertex connected to all existing vertices to help establish an effective and lossless single-state edge-to-vertex conversion, and in this way promote graph structure learning. In this embodiment, the present invention adopts a simpler edge-to-vertex conversion process. While ensuring information losslessness, the process extends the concept of edge-to-vertex conversion from a connected graph to a generalized graph structure, and theoretically proves that the conversion process can ensure that the original graph information is completely preserved.

[0051] 2. Graph Representation Learning

[0052] Graph neural networks (GNNs) are neural network models specifically designed for processing graph data. Compared to traditional neural networks, which are suited for structured data but cannot describe structured data in Euclidean space, GNNs are capable of processing non-Euclidean structures, adaptive sampling, and learning feature propagation. As a unifying paradigm for graph neural networks, GNNs based on message passing mechanisms encompass several representative research areas, such as graph convolutional networks (GCNs), graph isomorphism networks (GINs), and edge-oriented graph convolutional networks (EdgeGCNs).

[0053] In recent years, a large number of GNN-based graph representation learning algorithms have been proposed. These methods are divided into supervised and unsupervised categories according to whether they have labels or not. Since labels are usually difficult to obtain in real scenarios, unsupervised graph representation learning has broader application prospects. For example, ARGVA encodes the topological structure and node features in the graph into a compact representation, and trains a decoder on this basis to reconstruct the graph structure. S2GAE proposes to randomly mask and perturb a part of the edges of the original graph and reconstruct the original graph through a cross-correlation decoder to train an encoder that can extract higher representation quality. DGI applies the idea of ​​mutual information to the graph field, comparing local node representation and global graph representation. GRLC maximizes the mutual information between the semantic information and structural information of the data, and designs a triple constraint to measure the representation learning goal and downstream tasks.

[0054] 3. Information Bottleneck

[0055] The Information Bottleneck (IB) method, based on information theory, is used to extract compact but important information. Given input data X, data representation Z, and predicted label Y, the IB method primarily maximizes the mutual information between the data representation Z and the predicted label Y, while limiting the mutual information between the data representation Z and the input data X. The formula used is as follows:

[0056] maxI(Y;Z)-βI(X;Z) (1)

[0057] where β is used to balance the mutual information I(Y; Z) with I(X; Z). It can be seen intuitively that formula (1) limits the amount of information carried by the code representation Z, forcing the more important codes that are sufficient for the target task to pass through the bottleneck.

[0058] Recent research has introduced information bottlenecks into graph-structured data, making the computational difficulties of mutual information even more challenging on irregular graphs. In recent years, many graph representation learning methods based on information bottlenecks have been proposed, such as Graph Information Bottleneck (GIB) (2020), Subgraph Information Bottleneck (SIB) (2021), and Comparative Graph Structure Learning (CGI) for Recommendation Tasks (2022). In this example, following the mutual information constraint in the information bottleneck, a new graph information bottleneck was developed for supplementing line graph information.

[0059] B. Methods

[0060] Combine Figure 3 The graph representation learning method proposed in this invention for intelligent management of government public service venues based on generalized graph edge-to-vertex transformation and information bottleneck specifically includes the following contents:

[0061] 1. Information Bottleneck

[0062] (1) View Generator

[0063] Recent work on graph representation learning has shown that enhancing the view generation for contrast graphs can enable the encoder to learn richer representations. In this embodiment, two widely used structures or enhancement methods are mainly studied: (1) adjacency matrix, which intuitively reflects the complete structure of the graph. (2) KNN graph, which uses original features to calculate the similarity between vertices to construct the graph structure. Compared with the original adjacency matrix, KNN graph can reduce the sparsity of the graph while retaining local connectivity. This embodiment uses two of these structures or enhancement methods to generate a vertex graph. and a vertex graph for conversion to a line graph Please note that since this embodiment aims to perform edge-to-vertex transformation on a structure with richer information, the adjacency matrix will not be used to generate the graph.

[0064]

[0065] (2) Generalized Edge to Vertex Transforms

[0066] The views generated by graph perturbations of the same sample have a low degree of differentiation and are not sufficiently distinguishable. Therefore, this embodiment introduces edge-to-vertex transformation to construct the opposite spatial representation of the original graph - a line graph. Since the traditional edge-to-vertex transformation process is not applicable to non-connected graphs (such as those with isolated vertices), and there is information loss in the edge-to-vertex transformation of partially connected graphs, this embodiment proposes the concept of generalized graph edge-to-vertex transformation.

[0067] Design It is a generalized graph structure. is the vertex set, ε is the edge set, and X is the vertex feature matrix. This embodiment first performs CGG (Connective Graph Generation), and connects the graph according to the directional properties of the edges in the graph in the following two cases:

[0068] picture For a directed graph (such as Figure 2 (a)): If there are i vertices with in-degree 0 and j vertices with out-degree 0 (including isolated vertices), then modify the graph, and the modified graph Include a virtual vertex and i+j virtual edges, where the virtual vertex points to the vertex with in-degree 0 to form a virtual edge And pointed to by the vertex with out-degree 0 to form a virtual edge

[0069] picture For an undirected graph: If the graph There are no isolated vertices (such as Figure 2 (b) Left), there is no vertex with degree 0, and no operation is performed; if the graph There are i isolated vertices (like Figure 2 (b) right), then modify the graph, the modified graph Include a virtual vertex and 2*i virtual edges, where the virtual vertex passes through two virtual edges and Connect each isolated vertex

[0070] picture After the connectivity processing, a connected graph is generated, and the corresponding non-zero degree connected graph is obtained. A non-zero degree connected graph does not have vertices with degree 0, but there may still be isolated graphs.

[0071] Generalized edge-to-vertex conversion: If the following process exists: is a generalized graph structure. By generating a connected graph, we can obtain its corresponding non-zero degree connected graph (Non-zero Degree Connected Graph). The edge-to-vertex mapping f on the non-zero degree connected graph can transform the non-zero degree connected graph into the corresponding line graph H. This process is called generalized edge-to-vertex transformation.

[0072] The line graph obtained by the generalized edge-to-vertex conversion process contains all the information of the original graph. After the generalized edge-to-vertex conversion process, the corresponding lossless line graph is obtained The process can be abstractly represented as:

[0073]

[0074] (3) Feature Extraction

[0075] In order to make full use of the information in the original graph and line graph, this embodiment proposes a supplementary graph information bottleneck CONGIB. The proposed framework is as follows Figure 3 The specific steps are as follows:

[0076] ① Existing original image As model input, generate vertex graph through View Generator and a vertex graph for conversion to a line graph

[0077] ② Feature extraction

[0078] 1) For the vertex graph For the picture The vertex set, E v For the picture The edge set of X v For the picture Vertex feature matrix. This embodiment directly extracts features from its vertices. Formally, the L-th layer of the GNN layer can be written as:

[0079]

[0080] in, represents the l-th level vertex features or embeddings, and For graph-level tasks, the classic graph isomorphism network GIN is selected as the main architecture of the model; for node-level tasks, the classic graph convolutional network GCN is selected as the main architecture of the model.

[0081] 2) For the vertex graph used to convert to a line graph In this embodiment, a generalized edge-to-vertex transformation is performed to obtain the corresponding line graph H, i.e., the method in (2) generalized edge-to-vertex transformation, and the formula is as follows:

[0082]

[0083] in, is the vertex set, E e is an edge set, X v For the picture The edge feature matrix, CGG represents the connectivity operation, and f is the edge-to-vertex mapping.

[0084] After conversion, the figure The vertex features of correspond to the edge features of the line graph H. Therefore, the extraction of the edge features of H and The vertex feature extraction is for different perspectives of the same object. In this embodiment, the edge features of the line graph H are encoded by EdgeGCN to obtain the vertex feature embedding under the line graph perspective. The process can be expressed as:

[0085]

[0086] in, represents the l-th layer feature or embedding, and

[0087] (4) Maximize self-representation

[0088] This embodiment aims to obtain more valuable information for downstream tasks while minimizing redundant information. That is, in order to retain more information and In order to maximize the vertex map Vertex Features and Line Graphs The above process can be intuitively represented as maximizing the vertex graph and The mutual information between the line graph H and Mutual information between:

[0089]

[0090] Vertex Graph Carries the original image The state information of , while the line graph H describes more of the abstract spatial transfer state. Formula (7) aims to compress data noise as much as possible and retain information that can fully express the two views.

[0091] (5) Supplementary information

[0092] Because it is difficult to extract structural embedding using only edge structure, it is difficult to directly maximize the mutual information between structural information and semantic information. Therefore, as a preferred solution, this embodiment uses the embedded image of the obtained line graph based on step (3) As an abstract representation of the original graph edge structure, it is indirectly added to the graph Vertex feature encoding In. Combined Figure 3 To achieve this process, this embodiment proposes an information supplement principle:

[0093] Information Complementation Principle: Given random variables X1 and X2, the bottleneck random variables Z1 and Z2 that describe X1 and X2 are used. The Information Complementation Principle aims to use X1 as the primary information carrier, compressing the bottleneck random variable Z2 that describes X2 while preserving information related to the description of X1.

[0094] maxI(X1;Z1)+I(X2;Z2)-βI(Z2;X1) (8)

[0095] Where β is the Lagrange multiplier that describes the compression rate. In other words, formula (8) uses the information that can describe the random variable X2 as a supplement to the information that can describe the random variable X1.

[0096] Based on formula (7) and the principle of information supplementation, this embodiment gives the overall optimization goal: given a graph Vertex graph perspective and line graph perspective and For the encoding representation of the two perspectives, the overall optimization goal can be expressed as:

[0097]

[0098] It can be seen that the first two items encourage maximizing encoding The third term is the compression term, which encodes the Optimized for vertex viewing angle More approximate compact information. In other words, Compressed information is used to supplement optimization This increases the amount of information used for prediction. Therefore, joint optimization of these two items can obtain more task-related information. In order to apply this principle to model objectives, this embodiment introduces the following lemma:

[0099] Bounds of the Information Supplementation Principle: For the overall optimization goal The mutual information and The following inequalities hold:

[0100]

[0101] The information complementation principle shows that the optimization objective of formula (8) has an upper or lower bound, which provides an optimization basis for the posterior distribution of the difficult mutual information calculation.

[0102] 2. Optimization

[0103] According to the lower and upper bounds described in formulas (10), (11), and (12), the overall training objective (i.e., loss function) is:

[0104]

[0105] Among them, the first and second terms are vertex graphs The third term is the upper bound loss of the variational contrast logarithm ratio. As the main framework, it will have been Line graph that expresses as much as possible Supplement the information In the latent representation Include as much view information as possible.

[0106] For the upper bound of the reconstruction target, in order to facilitate calculation, this embodiment adopts the Monte Carlo method for sampling estimation. The specific formula is as follows:

[0107]

[0108]

[0109] Where B represents the total number of sampled data. For node-level tasks, since the edge structure is too complex to be quantified, this embodiment adopts a resampling method to construct a subgraph for the line graph to reduce the amount of computation.

[0110] For the contrast log ratio upper bound target, we have:

[0111]

[0112] Where B represents the total number of sampled data.

[0113] This embodiment assumes that an additional graph neural network encoder By Gaussian approximation Parameterized, then for a given sample data We can get:

[0114]

[0115] Therefore, the upper bound of the contrast logarithm target can be calculated as follows:

[0116]

[0117] Where diag[·] represents a diagonal matrix.

[0118] In summary, the overall training objective (i.e., loss function) can be estimated as:

[0119]

[0120] The algorithm shown in Table 1 calculates the training data based on the back propagation technique.

[0121] Table 1

[0122]

[0123]

[0124] C. Experiment and Results Analysis

[0125] 1. Experimental Setup

[0126] [1] Datasets: This example conducts experiments on graph classification and node-level datasets, including molecular compound datasets MUTAG, PTC-MR, PROTEINS, NCI1, DD, movie collaboration datasets IMDB-BINARY, IMDB-MULTIfromTUDataset, and citation network datasets Cora, Citeseer, Pubmed from Planetoid, Amazon co-purchase graph Computers, Photo, and Coauthor CS.

[0127] [2] Baseline model: To verify the effectiveness of the proposed model, this example compares CONGIB (using the algorithm in Table 1) with the following three representative models from different categories: (1) graph kernel models: WL and DGK; (2) reconstruction models based on unimodal loss: ARGVA, GPT-GNN, GraphMAE, S2GAE, and MaskGAE; (3) comparison models based on mutual information: DGI, InforGraph, GIC, MVGRL, InfoGCL, GRACE, SimGRACE, MSSGCL, VMA, and GRLC.

[0128] [3] Evaluation Metrics: Accuracy (ACC) was used as the main evaluation metric on all models, and the proposed model was evaluated under a linear evaluation protocol. For the graph classification task, the graph representation was fed into the SVC, and the average 10-fold cross-validation accuracy and standard deviation were reported after 5 runs. For the node classification task, all datasets kept the same differentiation ratio (8:2), trained the LIBSVM classifier by learning the node representation, and applied 5-fold cross-validation to estimate the performance. The average 10-fold cross-validation accuracy and standard deviation were reported after 10 runs.

[0129] 2. Main experiments:

[0130] Table 2 shows the results of the image classification experiment under unsupervised learning, including the accuracy and standard deviation of all data sets, which are specifically divided into Table 2-1 and Table 2-2.

[0131] Table 2-1

[0132]

[0133] Table 2-2

[0134]

[0135]

[0136] Note: Avg.R in the table is the average ranking of the 7 data sets in Tables 2-1 and 2-2.

[0137] Table 3 shows the experimental results of node classification under unsupervised learning, including the accuracy and standard deviation of all data sets, which are specifically divided into Table 3-1 and Table 3-2.

[0138] Table 3-1

[0139]

[0140] Table 3-2

[0141]

[0142]

[0143] Note: Avg.R in the table is the average ranking of the six data sets in Tables 3-1 and 3-2.

[0144] Table 2 compares graph classification performance. In these seven datasets, vertex features are represented as one-hot vectors of vertex labels or degrees. Compared to reconstruction models based on unimodal loss, our method achieves the best results in all datasets. Furthermore, our model outperforms existing mutual information-based comparison models in six of the seven datasets and is highly competitive on the remaining datasets. Furthermore, CONGIB significantly outperforms the best existing kernel methods.

[0145] Table 3 shows the node classification performance. As can be seen, our method outperforms the comparison model based on mutual information across all datasets. Compared to the latest reconstruction model based on a unimodal loss, our method achieves the best performance in most datasets. Compared to MaskGAE, which also uses a multimodal loss, CONGIB maintains its advantage. This indicates that the line graph is more distinguishable than the perturbed views of the original graph.

[0146] In summary, our model achieves comparable performance to state-of-the-art graph representation learning in both graph-level and node-level tasks. This can be attributed to the fact that our approach considers line graphs as another scale view and compresses their information as a supplement to graph representation, resulting in a higher-quality feature space.

[0147] 3. Efficiency analysis

[0148] To evaluate the overall conversion efficiency, this example compares the number of vertices, number of edges, and conversion time of the line graph obtained by generalized edge-to-vertex conversion with those of traditional edge-to-vertex conversion methods and virtual vertex conversion methods on datasets of different scales and task levels, including graph-level task datasets MUTAG and DD and node-level task datasets Cora and Citeseer. For the sake of fairness, this example conducts experiments in the same environment and with the same basic settings.

[0149] Specific results such as Figure 4 As shown in the figure, compared with the traditional edge-to-vertex conversion method, the method of the present invention increases the online graph size by an average of 2.5 times and the running time by an average of 1.4 times. The expanded line graph is used to make up for the information loss that exists in the conversion process of the traditional method. Compared with the conversion method that inserts virtual vertices to avoid information loss in the conversion process, the method of the present invention has obvious advantages in online graph size and conversion efficiency, and has superior performance in both graph classification and node classification tasks. It can be considered that, under the premise of ensuring information integrity, the investment of moderately expanding the graph size and increasing the running time can be considered reasonable.

[0150] 4. Exploration of information flow

[0151] To explore whether line graph information can flow into the vertex graph, this example removes the supplementary line graph information and uses variance and information entropy to measure the amount of embedded information. Information entropy is an indicator used in information theory to measure the uncertainty of random variables. In the discrete case, for a random variable X, its information entropy H(X) is defined as:

[0152] H(X)=-∑p(X)*log2p(X) (20)

[0153] Here, p(X) is the probability that the random variable X takes a certain value x, and log2 represents the logarithm with base 2. The unit of information entropy is usually bit.

[0154] Table 4 shows the relevant experimental results.

[0155] Table 4 Information content and variance of the supplementary figures and original figures

[0156]

[0157] Table 4 shows that a greater variance in node embeddings indicates a higher degree of dispersion and is therefore considered to contain more information, while a lower information entropy indicates a relatively certain and richer information in the node embeddings. It can be seen that node embeddings supplemented with line graphs have significantly greater variance and lower information entropy. This suggests that line graphs have a positive effect on learning high-quality representations in graph representations.

[0158] Figure 5 The variance and information entropy calculated from the graph representations for different datasets are shown as they evolve over the training period. It can be seen that with increasing training rounds, the representations extracted by the model become increasingly discrete, while the information contained therein gradually stabilizes. The overall evolution exhibits a nearly symmetrical pattern. It is also observed that the representations often exhibit large inverse fluctuations in the early stages of training. This may be due to the model's need to adapt to the direction of information compression within the data during these initial training rounds.

[0159] 5. Ablation experiment:

[0160] (1) Impact of information supplementation

[0161] In order to verify the influence of information supplementation principle on model performance, this embodiment conducts ablation experiments on graph classification and node classification tasks. The experimental results are as follows: Figure 6 As shown in Figure 2, the information supplementation principle can be considered to enhance the original graph representation by compressing the line graph information. The information supplementation principle can compensate for any missing or incomplete information in the compressed original graph representation. Results show that performance significantly degrades without the line graph as supplementary information, demonstrating the importance of the information supplementation principle in avoiding information redundancy and loss.

[0162] (2) Influence of view type

[0163] Figure 6 We explored the impact of different view generation methods on model performance. The adjacency matrix + KNN method combines global graph information with local similarity information, potentially offering advantages in capturing relationships between data and learning more accurate and consistent feature representations. While the KNN + KNN method provides diverse information, it can lead to information conflicts or inconsistencies, and can struggle to integrate global information, leading to local optima. Therefore, it can be seen that in most cases, the adjacency matrix + KNN method performs more effectively than the KNN + KNN method.

[0164] Example 2.

[0165] Combined with the operating steps of Example 1, the specific operating steps are as follows:

[0166] (1) Data preprocessing: First, the collected multi-source heterogeneous data is preprocessed to obtain the original graph. This data includes personnel information, vehicle information, equipment information, and the interactions between them. By constructing a graph structure, personnel, vehicles, equipment, etc. are used as vertices, and the interactions between them (such as the relationship between personnel and vehicles, the usage relationship between equipment and personnel, etc.) are used as edges to form a complex graph data structure. The graph data is cleaned and standardized to remove noise data and ensure data quality, preparing for subsequent information extraction and analysis to obtain the original graph.

[0167] (2) Feature extraction and conversion: Based on the preprocessed graph data (i.e., the original graph), the Complementary Graph Information Bottleneck Network (CONGIB) is applied, that is, the method of "B method" in Example 1 is used for feature extraction. Specifically: First, a vertex graph is generated by the view generator. and vertex graph Finally, a generalized edge-to-vertex transformation is used to construct a line graph, treating the vertex features of the original graph and the edge features of the line graph as two complementary perspectives. By maximizing the mutual information between the vertex graph and the line graph, and leveraging the principle of information complementation to embed the line graph features into the feature vector of the original graph, a more compact and information-rich feature representation is obtained.

[0168] (3) Intelligent management application: The extracted feature representation is applied to the intelligent management system of public service venues. These features can better reflect the complex relationships and interaction patterns between entities in the venue, providing support for subsequent intelligent analysis and decision-making. For example:

[0169] ① Personnel management: Optimize personnel scheduling and management by analyzing the interaction between personnel, equipment, and vehicles.

[0170] ② Vehicle management: Utilize the relationship between vehicles, personnel and equipment to achieve intelligent vehicle scheduling and safety monitoring.

[0171] ③ Equipment management: Intelligent maintenance and management of equipment can be achieved through the use records of equipment and its association with personnel.

[0172] Effect:

[0173] 1. Improved Management Efficiency: Through generalized graph edge-to-vertex transformation and information bottleneck graph representation learning methods, we can more accurately capture the complex relationships between entities in the venue, providing higher-quality feature representations for the intelligent management system. This enables managers to more efficiently dispatch personnel, manage vehicles, and maintain equipment, significantly improving management efficiency.

[0174] 2. Enhanced security: In terms of security monitoring, this approach can more effectively identify abnormal behavior and potential risks. For example, by analyzing the interaction patterns between people and vehicles, unauthorized vehicle entry or abnormal human activity can be detected in a timely manner, thereby enhancing the security capabilities of the venue.

[0175] 3. Optimize resource allocation: Through intelligent analysis of equipment usage, resources can be allocated more rationally. For example, based on equipment usage frequency and personnel requirements, equipment layout and maintenance plans can be optimized to improve equipment utilization and reduce operating costs.

[0176] 4. Improve service quality: Through intelligent management of personnel, vehicles, and equipment, we can better meet public needs and improve service quality. For example, by optimizing personnel scheduling, we can reduce public waiting time; through intelligent vehicle management, we can improve traffic flow; and through intelligent equipment maintenance, we can ensure the normal operation of equipment.

[0177] The present invention is a special edge-to-vertex conversion process that generalizes edge-to-vertex conversion from connected graphs to generalized graphs. On this basis, the present invention proposes a supplementary graph information bottleneck network (CONGIB). By restricting the vertex graph representation learned by the model, the model is forced to capture the most important feature information in the data while filtering out redundant information. Finally, the line graph obtained by the generalized graph edge-to-vertex conversion is used as a supplementary perspective of the original graph abstract space, and the line graph information is supplemented into the vertex graph representation using the principle of information supplementation. The present invention has conducted a large number of experiments on a series of graph classification and node classification datasets to confirm the superiority of the proposed method.

[0178] The above is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the embodiments of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the embodiments of the present invention are still within the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A graph representation learning method for intelligent management of government public service venues based on generalized graph edge-to-vertex transformation and information bottleneck, characterized by: The following steps are involved: The original image Generating a vertex map via the View Builder and vertex graph Then, feature extraction is performed; The vertex graph The formula for the feature extraction process is: In the formula represents the l-th level vertex features or embeddings, and The vertex graph Perform generalized edge-to-vertex transformation to obtain the corresponding line graph After that, edge features are extracted, and the formula is: In the formula represents the l-th layer feature or embedding, and 2. The graph representation learning method according to claim 1, characterized in that The vertex graph In the feature extraction process, for graph-level tasks, the classic graph isomorphism network GIN is selected; for node-level tasks, the classic graph convolutional network GCN is selected.

3. The graph representation learning method according to claim 1, characterized in that The vertex graph After generating the connected graph and obtaining its corresponding non-zero degree connected graph, perform edge-to-vertex mapping f and convert it into the corresponding line graph. Extract edge features.

4. The graph representation learning method according to claim 3, characterized in that The connectivity process is as follows: When the vertex graph For a directed graph, if there are i vertices with in-degree 0 and j vertices with out-degree 0 Then modify the graph; the modified graph Include a virtual vertex and i+j virtual edges; The virtual vertex points to the vertex with in-degree 0 to form a virtual edge And pointed to by the vertex with out-degree 0 to form a virtual edge When the vertex graph is an undirected graph, if There are no isolated vertices, no vertices with degree 0, and no operation is performed; if the graph There are i isolated vertices Then modify the graph; the modified graph Include a virtual vertex and 2*i virtual edges, where the virtual vertex passes through two virtual edges and Connect each isolated vertex 5. The graph representation learning method according to claim 1, characterized in that After the feature extraction, the As an abstract representation of the original graph edge structure, information is added In the overall optimization, the formula is:

6. The graph representation learning method according to claim 5, characterized in that The mutual information and The following inequalities hold:

7. The graph representation learning method according to claim 5, characterized in that During the overall optimization process, the overall training goal is:

8. The graph representation learning method according to claim 1, characterized in that The original graph is preprocessed data, and the process is: after collecting multi-source heterogeneous data, constructing a graph structure, and then performing cleaning and standardization processing to obtain the original graph.

9. The graph representation learning method according to claim 8, characterized in that The multi-source heterogeneous data includes: personnel information, vehicle information, equipment information and the interactive relationships between them.

10. The graph representation learning method according to claim 9, characterized in that: In the process of constructing the graph structure, personnel information, vehicle information, and equipment information are used as vertices, and interaction relationships are used as edges.

Images