Economic policy uncertainty prediction method and system based on small sample graph classification

Abstract

The invention relates to the field of economic policy uncertainty prediction, discloses an economic policy uncertainty prediction method and system based on small sample graph classification, and aims to construct a graph structure sample and convert a target country EPU trend prediction problem into a graph classification problem based on a trade association relationship and an EPU index. A small sample graph classification model is constructed based on the multi-branch structure and the twin network, the small sample graph classification model comprises a first network and a second network, and each of the first network and the second network comprises a plurality of branches. Processing the first graph structure sample by using a plurality of branches of the first network to obtain a plurality of first branch features, and fusing the plurality of first branch features to obtain a first sample feature; the second graph structure sample is processed by using the plurality of branches of the second network to obtain the plurality of second branch features, and the plurality of second branch features are fused to obtain the second sample feature, so that the feature information in the graph is more efficiently and more comprehensively mined, the problem of small samples is solved, and the prediction accuracy of the EPU change trend is improved.

Description

Technical Field

[0001] The present application relates to the field of economic policy uncertainty prediction, and specifically to a method and system for predicting economic policy uncertainty based on small sample graph classification. Background Art

[0002] In international economic relations, economic policy uncertainty is an important factor, which refers to the uncertainty caused by the government's inability to clarify the direction and strength of economic policy expectations, policy implementation and policy stance changes. As a policy maker, the volatility of the government's policies is highly correlated with market risks, which will affect investors and consumers' expectations of the economic situation in the market, trigger emotional fluctuations of investors and consumers, and quickly affect their investment and consumption behaviors, causing changes in the overall market situation. Therefore, the prediction of economic policy uncertainty is of great significance to the judgment of market behavior. The prediction of economic policy uncertainty has difficulties such as complex and diverse influencing factors and great international political influence. At present, researchers measure and predict economic policy uncertainty through econometrics and mathematical methods. In order to accurately measure the level of government economic policy uncertainty, Scott R. Baker et al. constructed the Economic Policy Uncertainty Index (EPU) using relevant data such as economic policy-related texts, consumer price index and local expenditure level forecasts. Therefore, the change of EPU has become one of the perspectives for predicting market changes. That is, the increase of EPU corresponds to large volatility of economic policies, high instability of market sentiment and increased investment risks.

[0003] Existing research on EPU prediction mainly discusses the impact mechanism of EPU, focusing on studying the factors affecting EPU through econometric methods, verifying the existence of spillover effects under the influence of factors such as international trade, and laying a theoretical foundation for EPU prediction, but the prediction of changes is relatively neglected. In the field of economics, data is difficult to label or the labeling cost is high, resulting in limited labeled samples and a typical small sample problem, which makes it impossible to fully utilize limited sample information to predict changing trends. Summary of the invention

[0004] This application provides an economic policy uncertainty prediction method based on small sample graph classification to solve the problems in the prior art that the prediction of changes in EPU prediction is relatively neglected, and there is a typical small sample problem, which makes it impossible to fully utilize limited sample information to predict changing trends.

[0005] Correspondingly, the present application also provides an economic policy uncertainty prediction system based on small sample graph classification, an electronic device, and a computer-readable storage medium to ensure the implementation and application of the above method.

[0006] In order to solve the above technical problems, the present application discloses a method for predicting economic policy uncertainty based on small sample graph classification, the method comprising:

[0007] Based on trade associations and EPU index, a graph structure sample is constructed;

[0008] A small sample graph classification model is constructed based on a multi-branch structure and a twin network; the small sample graph classification model includes a first network and a second network, and the first network and the second network both include multiple branches;

[0009] Using multiple branches of the first network to process the first graph structure sample to obtain multiple first branch features, and fusing the multiple first branch features to obtain a first sample feature;

[0010] Using multiple branches of the second network to process the second graph structure sample to obtain multiple second branch features, and fusing the multiple second branch features to obtain a second sample feature;

[0011] The similarity between the first sample feature and the second sample feature is calculated based on a preset contrast loss function, and a prediction result is obtained based on the similarity.

[0012] The present application also discloses an economic policy uncertainty prediction system based on small sample graph classification, the system comprising:

[0013] Sample construction module, used to construct graph structure samples based on trade associations and EPU index;

[0014] A model building module, used to build a small sample image classification model based on a multi-branch structure and a twin network; the small sample image classification model includes a first network and a second network, and the first network and the second network both include multiple branches;

[0015] A feature extraction module, used to process the first graph structure sample using multiple branches of the first network to obtain multiple first branch features, and fuse the multiple first branch features to obtain a first sample feature;

[0016] The feature extraction module is further used to process the second graph structure sample using multiple branches of the second network to obtain multiple second branch features, and fuse the multiple second branch features to obtain a second sample feature;

[0017] The prediction module calculates the similarity between the first sample feature and the second sample feature based on a preset contrast loss function, and obtains a prediction result based on the similarity.

[0018] The present application also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, one or more methods described in the present application are implemented.

[0019] The present application also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, one or more methods described in the present application are implemented.

[0020] In this application, considering the EPU spillover effect under the influence of international trade, a graph structure sample is constructed based on trade associations and EPU index, and the EPU trend prediction problem of the target country is converted into a graph classification problem. A small sample graph classification model is constructed based on a multi-branch structure and a twin network. The small sample graph classification model includes a first network and a second network, and the first network and the second network both include multiple branches. The first graph structure sample is processed by multiple branches of the first network to obtain multiple first branch features, and the multiple first branch features are fused to obtain the first sample features; the second graph structure sample is processed by multiple branches of the second network to obtain multiple second branch features, and the multiple second branch features are fused to obtain the second sample features. By designing a multi-branch structure for the graph, the model can learn features at different levels and angles, and fuse multi-branch features to improve the utilization of graph topology information. The twin network is introduced, and the similarity between the first sample feature and the second sample feature is calculated based on a preset contrast loss function, and the prediction result is obtained based on the similarity. The accuracy and stability of the model can be improved by expanding the training set by pairing.

[0021] Additional aspects and advantages of the present application will be given in the following description, which will become apparent from the following description, or will be understood through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0023] Figure 1 A flowchart of an economic policy uncertainty prediction method based on small sample graph classification provided in an embodiment of the present application;

[0024] Figure 2 A schematic diagram of a twin network structure provided for an embodiment of the present application;

[0025] Figure 3 A schematic diagram of the process of constructing a line graph provided in an embodiment of the present application;

[0026] Figure 4 A schematic diagram of the Levi graph construction process provided in the embodiment of the present application;

[0027] Figure 5 A schematic diagram of predicting economic policy uncertainty using a small sample graph classification model provided in an embodiment of the present application;

[0028] Figure 6 A node graph constructed based on economic policy uncertainty provided in an embodiment of the present application;

[0029] Figure 7 A schematic diagram of DGCNN extracting graph features provided in an embodiment of the present application;

[0030] Figure 8 A schematic diagram of the structure of an economic policy uncertainty prediction system based on small sample graph classification provided in an embodiment of the present application;

[0031] Fig. 9 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] The embodiments of the present application are described in detail below, and 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 only used to explain the present application, and cannot be interpreted as limiting the present application.

[0033] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0034] Those skilled in the art will appreciate that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the field to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0035] The solution provided in the embodiments of the present application can be executed by any electronic device, such as a terminal device or a server, wherein the server can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application does not limit this. With regard to the technical problems existing in the prior art, the economic policy uncertainty prediction method and system based on small sample graph classification provided in this application is intended to solve at least one of the technical problems of the prior art.

[0036] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0037] The present application embodiment provides a possible implementation method, such as Figure 1 As shown, a flowchart of an economic policy uncertainty prediction method based on small sample graph classification is provided, which can be executed by any electronic device, and optionally, can be executed on a server or terminal device.

[0038] like Figure 1 As shown in , the method may include the following steps:

[0039] Step 101, constructing a graph structure sample based on trade associations and EPU index;

[0040] Although econometrics can deeply explore the changing mechanism of economic policy uncertainty, it cannot meet the needs of EPU trend judgment from a prediction perspective. EPU is essentially a complex system under the influence of multiple subjects. Economic policies have formed a national EPU action network in the economic activities and political games between national subjects, which contains the mutual correlation and interaction information between economic policy subjects. When predicting the trend of national EPU, considering that the EPU of the prediction subject is a node in a complex network, it is necessary to make fuller use of the EPU action relationship between subjects to make accurate predictions. The econometric model mainly explains the impact mechanism of EPU, and the deep neural network can improve the feature mining ability by learning the correlation relationship between data, thereby improving the prediction performance. Therefore, the embodiment of the present application proposes an inheritance model based on EPU spillover effect and multi-branch neural network to predict the trend of national EPU changes. In the embodiment of the present application, the national correlation relationship is abstracted into a network graph structure, and the past EPU of the associated countries in the network is mapped to the EPU change trend, thereby abstracting the classification problem.

[0041] Step 102, constructing a small sample image classification model based on a multi-branch structure and a twin network; the small sample image classification model includes a first network and a second network, and the first network and the second network both include multiple branches.

[0042] In the field of economics, data is difficult to label or the labeling cost is high, resulting in limited labeled samples and a typical small sample problem. Although there are a large number of mature technologies in the field of small sample research, such as data enhancement and knowledge transfer methods, data enhancement methods are based on transforming sample features (such as rotation, cropping, scaling, etc.) through auxiliary information to expand the sample training set, thereby fully extracting data features. However, the intersection of graph representation learning is still in its infancy. The graph classification model constructed in this case will overfit the training set, seriously affecting the generalization ability on unknown samples and failing to predict trends well.

[0043] In the embodiment of the present application, a small sample graph classification model is constructed using a twin network. The twin network (SiameseNetwork) can be considered as a structure reorganized on the basis of the basic network. Unlike the traditional end-to-end classification algorithm that directly uses labels to classify samples, the twin network provides a way of learning sample knowledge through similarity comparison. The twin network obtains the discriminant features required for downstream tasks by learning branch samples, and completes the classification matching problem by performing similarity learning on the discriminant features of the twin branches. The comparison idea of ​​the twin network enables it to be used for data enhancement of small samples. By generating comparative training pairs, more training data can be obtained than the original instance, which can improve the accuracy and robustness of solving small sample problems. The graph structure samples used for training in the data set are paired, and the twin network is trained one by one. The feature mapping function is obtained by training the shared weights of the model, so that the graph structure samples are classified using the similarity of low-dimensional space features.

[0044] Step 103, using multiple branches of the first network to process the first graph structure sample to obtain multiple first branch features, and fusing the multiple first branch features to obtain a first sample feature;

[0045] Step 104, using multiple branches of the second network to process the second graph structure sample to obtain multiple second branch features, and fusing the multiple second branch features to obtain a second sample feature;

[0046] The twin network also has the idea of ​​multiple branches, but its application of branch structure focuses on three aspects: weight sharing mechanism, number of branches, and interaction between branches. The basic twin network has two branches, weights are shared between branches, and interaction between branches is achieved through contrast loss. Structurally, the twin network consists of three components: feature mapping, sample comparison, and decision making.

[0047] like Figure 2 As shown, it is a typical Siamese framework, in which the two branches f 1 (·) and f 2 (·) share the same network structure and weight information w. Paired instances are input into the model, and the two branches f 1 (·) and f 2 (·) for instance x 1 and instance x 2 Processing, through the branch model F(x 1 ) and F(x 2 ) extracts the features of the instance and uses f 1 (x 1 ) and f 2 (x 2 ) represents the feature result. In addition, the label of the new paired sample is constructed based on the original sample label:

[0048]

[0049] Among them, y 1 Represents the label of sample 1, y 2 Represents the label of sample 2. If the two samples belong to the same category, the label of the sample pair is set to 1, otherwise the label of the sample pair is set to 0.

[0050] Based on the structure of the above-mentioned twin network, two branch networks are correspondingly set in the embodiment of the present application, namely the first network and the second network, and the characteristic results of the outputs of the first network and the second network are the first sample characteristics and the second sample characteristics respectively.

[0051] Regarding the processing of graph structures, most current models iteratively fuse graph structure information from the perspective of nodes, and do not fully utilize edge features, while edge information contains important information about the graph. Therefore, in an embodiment of the present application, a multi-branch feature extraction is proposed, and a multi-branch structure is set in the first network and the second network. The difference between different branches mainly lies in the observation angle of the input graph structure sample. The first graph structure sample is processed using multiple branches of the first network to obtain multiple first branch features, and then the multiple first branch features are fused to obtain the first sample feature; the second graph structure sample is processed using multiple branches of the second network to obtain multiple second branch features, and then the multiple second branch features are fused to obtain the second sample feature. Through the multi-branch structure, you can choose to extract features from aspects such as node topology information, edge topology information, and topological connections between nodes and edges, thereby improving the model's ability to extract features.

[0052] Step 105: Calculate the similarity between the first sample feature and the second sample feature based on a preset contrast loss function, and obtain a prediction result based on the similarity.

[0053] like Figure 2 The twin network shown in the figure, in the sample comparison stage, is h (·) to compare f 1 (x 1 ) and f 2 (x 2 ), the comparison layer results use the comparison score C score express.

[0054] Generally speaking, the comparison mechanism used by the comparison head mainly includes similarity comparison and ranking comparison. In similarity, it is usually calculated by calculating the feature result f 1 (x 1 ) and f 2 (x 2) to represent the similarity between input instances. The smaller the distance value D, the higher the similarity. Similarity comparison is more widely used in classification problems. Ranking comparison compares the priorities between samples, records the comparison results as ranking scores, and uses them to calculate model losses, which are used to continue to optimize the model. It is more widely used in regression problems. In an embodiment of the present application, the problem to be solved is a graph classification problem, so the idea of ​​similarity comparison is adopted. Similarity comparison is based on the idea that feature distribution is related to data labels, and emphasizes projecting similar samples into similar representations. Therefore, when performing feature mapping from high-dimensional space samples to low-dimensional vector space, instances with the same label will have obvious clustering characteristics for the corresponding features.

[0055] In the similarity comparison, we first need to map the paired instances to the target feature space through the explored feature mapping model in the previous step to obtain the instance feature f 1 (x 1 ) and f 2 (x 2 ), and then optimize the model by calculating the distance D. The distance can be measured by different methods such as Euclidean distance and cosine distance. In the embodiment of the present application, the first graph structure sample is mapped to the target feature space by the first network to obtain the first sample feature, and the second graph structure sample is mapped to the target feature space by the second network to obtain the second sample feature, and then the similarity between the first sample feature and the second sample feature of the first network is calculated by using a preset contrast loss function, and the final classification result, that is, or the final prediction result, is determined according to the similarity.

[0056] In the embodiment of the present application, considering the EPU spillover effect under the influence of international trade, a graph structure sample is constructed based on trade associations and EPU index, and the EPU trend prediction problem of the target country is converted into a graph classification problem. A small sample graph classification model is constructed based on a multi-branch structure and a twin network, and the small sample graph classification model includes a first network and a second network, and the first network and the second network both include multiple branches. The first graph structure sample is processed by multiple branches of the first network to obtain multiple first branch features, and the multiple first branch features are fused to obtain the first sample feature; the second graph structure sample is processed by multiple branches of the second network to obtain multiple second branch features, and the multiple second branch features are fused to obtain the second sample feature. By designing a multi-branch structure for the graph, the model can learn features at different levels and angles, and fuse multi-branch features to improve the utilization of graph topology information. The twin network is introduced, and the similarity between the first sample feature and the second sample feature is calculated based on a preset contrast loss function, and the prediction result is obtained based on the similarity, wherein the accuracy and stability of the model can be improved by expanding the training set by pairing.

[0057] In an optional embodiment, the graph structure sample includes a node graph, a line graph, and a Levi graph. Based on the trade association and the EPU index, the graph structure sample is constructed, including:

[0058] Construct a node graph with countries as nodes and economic and trade exchanges as edges;

[0059] Convert node graphs to line graphs and Levi graphs.

[0060] In an optional embodiment, the first network and the second network each include three branches for respectively extracting features of the node graph, the line graph and the Levi graph.

[0061] Graph convolutional networks (GCNs) are representative and powerful structures that can take graph nodes as microscopic observation subjects and aggregate local neighborhood information around nodes. Low-rank proximity and node features are successfully utilized in existing GCNs. However, the attributes that edges may carry are usually ignored because these models almost all simplify graph links into binary or scalar values ​​that describe node connectivity. Among the existing methods, some methods introduce tensor products as neighbor features, define graph kernels on this basis, and develop the GCN-LASE framework to improve model performance and prove the effectiveness of edge information. Other methods integrate graph neural network models, use new attention mechanisms to fuse multi-dimensional edge features, and design neural network structures to adapt to edge features at each layer of the neural network. These studies have noticed the effectiveness of edge information and tried to use it in the model. However, in the graph structure, there is a structural correlation between nodes and edges. When edge features are extracted and fused with node information, node information and edge information often affect each other during information transmission, which may lead to information loss or confusion, affecting the model effect.

[0062] In the embodiment of the present application, the difference between different branches in the first network and the second network mainly lies in the observation angle of the input graph structure sample, and the three angles of node topology information, edge topology information, and topological connection between nodes and edges are selected as the multi-branch perspective of the graph to better utilize edge features, including undirected edges and multi-dimensional features. Therefore, the construction of node graphs, line graphs, and Levi graphs is introduced through the following content to prepare for extracting graph features from different perspectives.

[0063] In the process of predicting the trend of changes in national economic policy uncertainty, it can be regarded as a complex network, the country can be regarded as a node of the complex network, and the trend of national EPU changes can be regarded as the label of the complex network graph, so as to fully consider the relationship between the changes in the EPU of the related countries and the EPU of the predicted country, thereby helping the prediction. However, it is an important issue to select which kind of relationship as the link between countries and how to construct the characteristics of the nodes. Similarly, in a social network, when exploring the attributes of a group, the individuals in the group can be used as nodes, and the edges can be constructed through the relationship between the nodes. By learning the graph features, the group characteristics can be identified. In the EPU study, according to the relevant exploration of the impact mechanism of economic policy uncertainty, the trade relationship between countries is an important consideration for the spillover of economic policy uncertainty. In the embodiment of the present application, the prediction subject and its important related countries are used as the predicted group, and the edges and edge attributes of the node graph are constructed accordingly. Considering the impact of past years, the country's past EPU is used as a node attribute. By learning the graph attributes, the graph classification problem is abstracted when predicting the trend of changes in national EPU.

[0064] Line graph is a kind of graph structure in graph theory. It is a kind of undirected graph that can reflect the connectivity between the edges in the graph. Given an original graph G = (V, E), its corresponding line graph is L(G) = (LV, LE), which can represent the connection relationship between the edges in the graph G. When constructing L(G), it is necessary to convert the edges in G into nodes in L(G), that is, LV = {V(e)|e∈E}. If two edges e in G are i and e j have a common endpoint, then the corresponding node v(e i ) and v(e j ) has a connecting edge. Figure 3 As shown in the figure, the left figure is the original graph G, and the right figure is the converted line graph L(G). The edges of the original graph correspond to the nodes converted to the line graph, so the number of nodes in the right graph is the same as the number of edges in the left graph. 1 ,v 2 ) and (v 1 ,v 4 ) have the same node v 1 , so in the line graph L(G), the node (v 1 ,v 2 ) and (v 1 ,v 4 ) has a connecting edge.

[0065] Classic graph neural network models are often based on graph embedding thinking and are unable to process edge attribute features or make insufficient use of edge information. They only use graph topology information and node features, and are unable to retain sufficient information to represent the graph structure. The advantage of line graphs is that the edge features in the original graph G can be converted into node features in the line graph L(G). Classic graph neural networks can be used to extract edge information, supplement edge labels or structural information, and improve model performance. In addition, since L(G) discards the node features of G during the construction process, it is more focused on processing edge attributes and edge topology information when obtaining edge information through L(G), which can avoid being interfered by node information.

[0066] Levi Graph is a bipartite graph associated with an association structure. In the process of construction, edges are abstracted into nodes and connected to the nodes at both ends to form new edges. Figure 4 As shown in the figure, there is an edge between node v1 and node v2. When the node graph is converted to a Levi graph, the edge is abstracted into a new node (v 1 ,v 2 ), and at the same time, the new node is connected to nodes v1 and v2 with new edges, the edge attributes are converted to the attributes of the new node, and new attributes are added to the node to indicate the node type.

[0067] Based on the above construction method, Figure 5 As shown, in the data collection stage, the embodiment of the present application obtains the trade data between countries from the website of the International Monetary Fund (IMF), obtains the EPU quarterly data of relevant countries from the Economic Policy Uncertainty website founded by Scott Ross Baker, a professor at Northwestern University in the United States, and other websites, and by constructing a national trade network, the EPU trend prediction problem is abstracted into a graph classification problem.

[0068] In the national association network, there are abundant connections between countries. These connections are reflected in the edges between countries and the attributes of the edges. At the same time, countries have their own attribute vectors, forming a complex network topology in the data. In the embodiment of the present application, national trade data and EPU data are used for data modeling to construct a node graph with countries as nodes and economic and trade exchanges as edges, which is used to predict the trend of changes in the national EPU index. Specifically, countries with higher trade exchanges are selected as graph nodes, and trade values ​​are used to measure the degree of economic association between countries, that is, as edge attributes, and past EPU data are used as node attributes, and EPU reductions are marked as positive samples.

[0069] More specifically, countries are regarded as network nodes, represented by n, N represents the set of national nodes, and the national past EPU index is used as the node attribute; the relationship between nodes is represented by e, E represents the set of relationships, and edges are used to represent the economic and trade relationship between countries. The strength of economic relationship is used as the characteristic attribute of the edge. The stronger the economic relationship, the stronger the mutual influence between countries, and the greater the impact of their EPU. Figure 6 As shown, in the embodiment of the present application, blue nodes are used to represent the currently predicted country (i.e., country 1), and green nodes are used to represent the main countries with which it has economic and trade relations (i.e., country 2, country 3, country 4, and country 5). Countries are linked together through trade exchanges to construct a node graph.

[0070] In order to make better use of edge features, the node graph is converted to generate line graph and Levi graph. In the classification of small sample graphs, the node graph, line graph and Levi graph contain rich topological information and attribute information, respectively, which provide solid data support for the extraction of original node information, edge information and node and edge topological connections. Using different branches to process the node graph, line graph and Levi graph respectively can better mine the graph features in the node graph, line graph and Levi graph. Figure 5 In the data analysis and decision-making stage shown, the input input1 of the first network and the input input2 of the second network both include a node graph, a line graph and a Levi graph, and the first network and the second network are each provided with corresponding three branches for processing the node graph, the line graph and the Levi graph respectively.

[0071] like Figure 5 As shown, in an optional embodiment, each branch in the first network and the second network adopts a DGCNN network.

[0072] DGCNN (Deep Graph Convolutional Neural Network) can accept graph structure data as input, fully retain node information, and learn the model of graph information from the global graph topology. The learning of DGCNN includes three stages: first, the graph convolution layer extracts the local substructure features of the vertex and defines a consistent vertex order. The graph convolution layer uses the GCN network, and multiple graph convolution layers (i.e., multiple GCNs) can be used to obtain neighborhood features at different levels; second, the SortPooling layer sorts the vertex features in a predefined order to unify the input size; third, the traditional convolution layer and dense layer are used to read the sorted graph representation and output the feature vector.

[0073] DGCNN implementation process and specific structure are as follows Figure 7Taking the node graph as an example, the embodiment of the present application uses different colors to represent node features. After the node graph is input, it first passes through multiple graph convolution layers to obtain neighborhood features at different levels. The nodes and node features are sorted according to the features in the SortPooling layer, and then passed to the one-dimensional convolution layer and the dense layer to further integrate the feature information and output the results.

[0074] In the graph convolution layer, a special spatial graph convolution layer is used to extract multi-scale vertex features. The convolution layer is expressed as:

[0075]

[0076] in is the graph adjacency matrix with self-loops added, I is the identity matrix, X is the node information matrix, W is the graph convolution parameter matrix to be trained, f is the nonlinear activation layer, and Z is the output activation matrix. As can be seen from the formula, the graph convolution stage generally includes four steps: linearly transforming the node information matrix, mapping the feature channel to the next layer, propagating the node information to adjacent nodes and the node itself, normalizing by row to maintain a fixed feature scale, applying the nonlinear activation function point by point, and outputting the graph convolution result. The model achieves node aggregation in the local neighborhood through graph convolution operations to obtain local structural information.

[0077] To obtain multi-scale substructure features, multiple graph convolutional layers are stacked and a layer is used to connect the outputs of each graph convolutional layer to output multi-scale local substructure information:

[0078]

[0079] Among them, Z 0 =X,Z t is the output result after the tth layer of graph convolution. After multiple convolutional layers, a connection layer is added to horizontally splice the outputs of each layer. In the spliced ​​output, each row can be regarded as a feature of a vertex, containing its neighborhood structure information at different scales:

[0080]

[0081] The main function of the SortPooling layer is to sort the node features in a consistent order to prepare for the input convolutional layer and dense layer. Unlike image data with neatly arranged pixels and text data with text order, the graph structure, as a kind of non-Euclidean data, cannot directly obtain the order of its nodes. In the embodiment of the present application, the vertices are sorted by the graphical annotation method.

[0082]

[0083] WL uses the graph labeling process to minimize the difference between the distance in the vector space and the expected value of the graph distance between two randomly drawn graphs from G. The SortPooling layer first introduces WL colors to group nodes. Nodes are sorted according to the structural roles they play in the graph to ensure that nodes with similar structural roles have consistent final sorting results.

[0084] After the SortPooling layer, the corresponding features are organized into feature tensors according to the node sorting results, with each row representing a vertex and each column representing a feature channel. The tensor is input into the traditional convolutional layer and dense layer to further extract the node graph information. The fully connected layer and softmax layer in the dense layer also help map the graph features to the low-dimensional vector space, ultimately realizing the transformation of the node graph into low-dimensional feature data.

[0085] In an optional embodiment, using multiple branches of the first network to process the first graph structure sample to obtain multiple first branch features, and fusing the multiple first branch features to obtain the first sample feature, includes:

[0086] Using multiple branches of the first network to process the node graph, line graph, and Levi graph of the first graph structure sample respectively, to obtain multiple first branch features;

[0087] Based on the linear layer, learn the weights corresponding to different branches;

[0088] A plurality of first branch features are weightedly fused based on the weights of the corresponding branches to obtain a first sample feature.

[0089] For the first network, there are three branches, which take the node graph, line graph and Levi graph as input respectively. After the three branches extract the three first branch features respectively, the features of each branch are spliced, and then the corresponding weights of different branches are learned by adding linear layers to perform weighted fusion of the first branch features, that is:

[0090] F = a 1 F 1 +a 2 F 2 +a 3 F 3

[0091] Among them, F is the first sample feature, F 1 、F 2 and F 3 are the three first branch features, a 1 、a 1 and a 1 They are the weights corresponding to the three branches respectively.

[0092] The structure of the second network is the same as that of the first network, and the only difference is that the two networks process different graph structure samples respectively.

[0093] Therefore, in the MSGNN model, the feature extraction process of the model is transformed from a single graph convolutional neural network model to a multi-branch model.

[0094] In an optional embodiment, calculating the similarity between the first sample feature and the second sample feature based on a preset contrast loss function includes:

[0095] The Euclidean distance is used as the contrast loss function to calculate the similarity between the first sample feature and the second sample feature.

[0096] exist Figure 2 In the twin network shown, the specific formula using Euclidean distance as the similarity measurement indicator is as follows:

[0097]

[0098] Among them, use (w 11 ,w 12 ,…,w 1n ) T ,(w 21 ,w 22 ,…,w 2n ) T Respectively represent the low-dimensional features of the two instance mappings in the n-dimensional space.

[0099] Based on the calculated instance distance, the similarity comparison loss is calculated, and the target model is gradually trained by minimizing the loss during the optimization process:

[0100] L(c 1 ,x 2 ,y)=α(1-y)D w 2 +βy max(0,mD w ) 2

[0101] Among them, D w =||f 1 (x 1 )-f 2 (x 2 )||

[0102] Among them, x 1 、x 2 is the input sample instance, y comes from the supervision information of the data label, which is the label of the paired sample. y = 1 means that the samples belong to the same category, otherwise they belong to different categories. α and β are weight coefficients. wis the Euclidean distance after the sample instance is mapped to the low-dimensional feature space, m>0, is the parameter used to control the penalty threshold, and L is the contrast loss in the current training stage.

[0103] In the small sample graph classification model constructed in the embodiment of the present application, the first graph structure sample G is input 1 And the second graph structure sample G 2 , it is necessary to introduce a multi-branch structure on the basic twin network structure, which is reflected in the formula of contrast loss by converting the feature extraction of a single-branch network into the comprehensive features extracted by a multi-branch network. Therefore, the formula of contrast loss is converted to:

[0104] L(G 1 ,G 2 ,y)=α(1-y)D w 2 +βy max(0,nD w ) 2

[0105] D w =||(a 1 f 11 (G 1 )+a 2 f 12 (L(G 1 ))+a 3 f 13 (Levi(G 1 )))-(a 1 f 21 (G 2 )+a 2 f 22 (L(G 2 )))+a 3 f 23 (Levi(G 2 ))||

[0106] Among them, G 1 and G 2 are the first graph structure sample and the second graph structure sample respectively, y is the label supervision information, α and β are the weight coefficients, D w is the Euclidean distance after the graph structure sample is mapped to the low-dimensional feature space, m>0, is the parameter used to control the penalty threshold, a 1 、a 2 and a 3 are the weights corresponding to each branch, f 11 (G 1 ), f 12 (L(G 1 )) and f 13 (Levi(G 1)) corresponds to each branch in the first network; f 21 (G 2 ), f 22 (L(G 2 )) and f 23 (Levi(G 2 )) is the second sample feature corresponding to each branch in the second website.

[0107] Based on the same principle as the method provided in the embodiment of the present application, the embodiment of the present application also provides an economic policy uncertainty prediction system based on small sample graph classification, such as Figure 8 As shown, the system comprises:

[0108] The sample construction module 801 is used to construct a graph structure sample based on the trade association relationship and the EPU index;

[0109] A model building module 802 is used to build a small sample image classification model based on a multi-branch structure and a twin network; the small sample image classification model includes a first network and a second network, and the first network and the second network both include multiple branches;

[0110] A feature extraction module 803 is used to process the first graph structure sample using multiple branches of the first network to obtain multiple first branch features, and fuse the multiple first branch features to obtain a first sample feature;

[0111] The feature extraction module 803 is further used to process the second graph structure sample using multiple branches of the second network to obtain multiple second branch features, and fuse the multiple second branch features to obtain a second sample feature;

[0112] The prediction module 804 calculates the similarity between the first sample feature and the second sample feature based on a preset contrast loss function, and obtains a prediction result based on the similarity.

[0113] In the embodiment of the present application, considering the EPU spillover effect under the influence of international trade, a graph structure sample is constructed based on trade associations and EPU index, and the EPU trend prediction problem of the target country is converted into a graph classification problem. A small sample graph classification model is constructed based on a multi-branch structure and a twin network, and the small sample graph classification model includes a first network and a second network, and the first network and the second network both include multiple branches. The first graph structure sample is processed by multiple branches of the first network to obtain multiple first branch features, and the multiple first branch features are fused to obtain the first sample feature; the second graph structure sample is processed by multiple branches of the second network to obtain multiple second branch features, and the multiple second branch features are fused to obtain the second sample feature. By designing a multi-branch structure for the graph, the model can learn features at different levels and angles, and fuse multi-branch features to improve the utilization of graph topology information. The twin network is introduced, and the similarity between the first sample feature and the second sample feature is calculated based on a preset contrast loss function, and the prediction result is obtained based on the similarity, wherein the accuracy and stability of the model can be improved by expanding the training set by pairing.

[0114] The economic policy uncertainty prediction system based on small sample graph classification provided in the embodiment of the present application can achieve Figures 1 to 7 To avoid repetition, the various processes implemented in the method embodiment are not described here.

[0115] The economic policy uncertainty prediction system based on small sample graph classification of the embodiment of the present application can execute the economic policy uncertainty prediction method based on small sample graph classification provided in the embodiment of the present application, and the implementation principle is similar. The actions performed by each module and unit in the economic policy uncertainty prediction system based on small sample graph classification in each embodiment of the present application correspond to the steps in the economic policy uncertainty prediction method based on small sample graph classification in each embodiment of the present application. For the detailed functional description of each module of the economic policy uncertainty prediction system based on small sample graph classification, please refer to the description of the corresponding economic policy uncertainty prediction method based on small sample graph classification shown in the previous text, which will not be repeated here.

[0116] Based on the same principle as the method shown in the embodiment of the present application, the embodiment of the present application also provides an electronic device, which may include but is not limited to: a processor and a memory; a memory for storing a computer program; a processor for executing the economic policy uncertainty prediction method based on small sample graph classification shown in any optional embodiment of the present application by calling a computer program. Compared with the prior art, the economic policy uncertainty prediction method based on small sample graph classification provided by the present application considers the EPU spillover effect under the influence of international trade, constructs a graph structure sample based on trade association and EPU index, and converts the target country EPU trend prediction problem into a graph classification problem. A small sample graph classification model is constructed based on a multi-branch structure and a twin network, and the small sample graph classification model includes a first network and a second network, and the first network and the second network both include multiple branches. The first graph structure sample is processed by using multiple branches of the first network to obtain multiple first branch features, and the multiple first branch features are fused to obtain the first sample feature; the second graph structure sample is processed by using multiple branches of the second network to obtain multiple second branch features, and the multiple second branch features are fused to obtain the second sample feature. By designing a multi-branch structure for the graph, the model can learn features at different levels and angles, and fuse multi-branch features to improve the utilization of graph topology information. A twin network is introduced to calculate the similarity between the first sample feature and the second sample feature based on a preset contrast loss function, and a prediction result is obtained based on the similarity. The accuracy and stability of the model can be improved by expanding the training set through pairing.

[0117] In an optional embodiment, an electronic device is also provided, such as Fig. 9 As shown, Fig. 9 The electronic device 900 shown may be a server, including: a processor 901 and a memory 903. The processor 901 and the memory 903 are connected, such as through a bus 902. Optionally, the electronic device 900 may further include a transceiver 904. It should be noted that in actual applications, the transceiver 904 is not limited to one, and the structure of the electronic device 900 does not constitute a limitation on the embodiments of the present application.

[0118] Processor 901 may be a CPU (Central Processing Unit), a general purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 901 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0119] The bus 902 may include a path to transmit information between the above components. The bus 902 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 902 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig. 9 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0120] The memory 903 can be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited to these.

[0121] The memory 903 is used to store the application code for executing the solution of the present application, and the execution is controlled by the processor 901. The processor 901 is used to execute the application code stored in the memory 903 to implement the contents shown in the above method embodiment.

[0122] Among them, electronic devices include but are not limited to: mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and fixed terminals such as digital TVs and desktop computers. Fig. 9 The electronic device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0123] The server provided in this application can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected by wired or wireless communication, and this application does not limit this.

[0124] An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer-readable storage medium is run on a computer, the computer can execute the corresponding content in the aforementioned method embodiment.

[0125] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0126] It should be noted that the computer-readable storage medium mentioned above in the present application can also be a computer-readable signal medium or a combination of a computer-readable storage medium and a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal can take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0127] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0128] The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device executes the method shown in the above embodiment.

[0129] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the economic policy uncertainty prediction method and system based on small sample graph classification provided in the above-mentioned various optional implementations.

[0130] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0131] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0132] The modules involved in the embodiments described in this application can be implemented by software or hardware. The name of the module does not limit the module itself in some cases. For example, the sample construction module can also be described as "a sample construction module for constructing a graph structure sample based on trade associations and EPU indexes".

[0133] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.

Claims

1. A method for predicting economic policy uncertainty based on small sample graph classification, characterized in that: The method comprises: Based on trade associations and EPU index, a graph structure sample is constructed; A small sample image classification model is constructed based on a multi-branch structure and a twin network; the small sample image classification model includes a first network and a second network, and the first network and the second network both include multiple branches; Processing the first graph structure sample using multiple branches of the first network to obtain multiple first branch features, and fusing the multiple first branch features to obtain a first sample feature; Processing the second graph structure sample using the multiple branches of the second network to obtain multiple second branch features, and fusing the multiple second branch features to obtain a second sample feature; The similarity between the first sample feature and the second sample feature is calculated based on a preset contrast loss function, and a prediction result is obtained based on the similarity.

2. The economic policy uncertainty prediction method based on small sample graph classification according to claim 1 is characterized in that: The graph structure samples include node graphs, line graphs and Levi graphs. The graph structure samples are constructed based on trade associations and EPU indexes, including: Construct a node graph with countries as nodes and economic and trade exchanges as edges; The node graph is converted into a line graph and a Levi graph.

3. The economic policy uncertainty prediction method based on small sample graph classification according to claim 2 is characterized in that: The first network and the second network each include three branches, which are used to extract features of the node graph, the line graph and the Levi graph respectively.

4. The economic policy uncertainty prediction method based on small sample graph classification according to claim 1 is characterized in that: Each branch in the first network and the second network adopts a DGCNN network.

5. The economic policy uncertainty prediction method based on small sample graph classification according to claim 3 is characterized in that: The step of using the multiple branches of the first network to process the first graph structure sample to obtain multiple first branch features, and fusing the multiple first branch features to obtain the first sample feature includes: Using the multiple branches of the first network to process the node graph, line graph and Levi graph of the first graph structure sample respectively, to obtain multiple first branch features; Based on the linear layer, learn the weights corresponding to different branches; The first sample feature is obtained by weighted fusion of multiple first branch features based on the weights of corresponding branches.

6. The economic policy uncertainty prediction method based on small sample graph classification according to claim 5 is characterized in that: The calculating the similarity between the first sample feature and the second sample feature based on a preset contrast loss function includes: The Euclidean distance is used as a comparison loss function to calculate the similarity between the first sample feature and the second sample feature.

7. The economic policy uncertainty prediction method based on small sample graph classification according to claim 6 is characterized in that: The formula of the contrast loss function is: L(G1,G2,y)=α(1-y)D w 2 +βy max(0,m-D w ) 2 <h2 style=";text-align:left;direction:ltr">D<h2 style=";text-align:left;direction:ltr"> w <h2 style=";text-align:left;direction:ltr"> =||(a1f<h2 style=";text-align:left;direction:ltr"> 11 <h2 style=";text-align:left;direction:ltr"> (G1)+a2f<h2 style=";text-align:left;direction:ltr"> 12 <h2 style=";text-align:left;direction:ltr"> (L(G1))+a3f<h2 style=";text-align:left;direction:ltr"> 13 <h2 style=";text-align:left;direction:ltr"> (Levi(G1)))-(a1f<h2 style=";text-align:left;direction:ltr"> 21 <h2 style=";text-align:left;direction:ltr"> (G2)+a2f<h2 style=";text-align:left;direction:ltr"> 22 <h2 style=";text-align:left;direction:ltr"> (L(G2)))+a3f<h2 style=";text-align:left;direction:ltr"> 23 <h2 style=";text-align:left;direction:ltr"> (Levi(G2))|| Among them, G1 and G2 are the first and second graph structure samples respectively, y is the label supervision information, α and β are weight coefficients, and D w is the Euclidean distance after the graph structure sample is mapped to the low-dimensional feature space, m>0, is the parameter used to control the penalty threshold, a1, a2 and a3 are the weights corresponding to each branch, and f 11 (G1), f 12 (L(G1)) and f 13 (Levi(G1)) is the first sample feature corresponding to each branch in the first network; f 21 (G2), f 22 (L(G2)) and f 23 (Levi(G2)) is the second sample feature corresponding to each branch in the second website.

8. An economic policy uncertainty prediction system based on small sample graph classification, characterized by: The system comprises: Sample construction module, used to construct graph structure samples based on trade associations and EPU index; A model building module, used to build a small sample image classification model based on a multi-branch structure and a twin network; the small sample image classification model includes a first network and a second network, and the first network and the second network each include multiple branches; A feature extraction module, configured to process the first graph structure sample using the multiple branches of the first network to obtain multiple first branch features, and fuse the multiple first branch features to obtain a first sample feature; The feature extraction module is further used to process the second graph structure sample using the multiple branches of the second network to obtain multiple second branch features, and fuse the multiple second branch features to obtain a second sample feature; A prediction module calculates the similarity between the first sample feature and the second sample feature based on a preset contrast loss function, and obtains a prediction result based on the similarity.

9. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images