WO2023065545A1

WO2023065545A1 - Risk prediction method and apparatus, and device and storage medium

Info

Publication number: WO2023065545A1
Application number: PCT/CN2022/071241
Authority: WO
Inventors: 肖京; 李娜; 王磊; 赵盟盟; 王媛; 谭韬; 陈又新
Original assignee: 平安科技（深圳）有限公司
Priority date: 2021-10-19
Filing date: 2022-01-11
Publication date: 2023-04-27
Also published as: CN113822494A; CN113822494B

Abstract

The present application relates to artificial intelligence technology. Disclosed is a risk prediction method, comprising: constructing a temporal knowledge graph on the basis of a risk perception factor set that is extracted from a multi-source information set, and performing implicit relationship supplementation and causal relationship supplementation on the temporal knowledge graph, so as to obtain a standard knowledge graph and an event evolutionary graph; performing prediction by using a risk prediction model that is constructed on the basis of a reinforcement learning algorithm, so as to obtain a target risk entity; and performing relationship quantification and degree quantification on the standard knowledge graph, so as to obtain a dependency closeness and an event hazard degree, performing training on the basis of the event evolutionary graph, the dependency closeness and the event hazard degree and in view of a graph neural network and a semi-supervised method, so as to obtain a macro prediction model, and performing prediction by using the macro prediction model, so as to obtain a risk industry corresponding to the target risk entity. In addition, the present application further relates to blockchain technology. The event evolutionary graph can be stored in a node of a blockchain. Also provided in the present application are a risk prediction apparatus, an electronic device and a storage medium. By means of the present application, the accuracy of performing risk prediction on an industry can be improved.

Description

Risk prediction method, device, equipment and storage medium

This application claims the priority of a Chinese patent application with application number CN202111216347.6 and titled "Risk Prediction Method, Device, Equipment, and Storage Medium" filed with the China Patent Office on October 19, 2021, the entire contents of which are incorporated by reference in this application.

technical field

The present application relates to the technical field of artificial intelligence, and in particular to a risk prediction method, device, electronic equipment, and computer-readable storage medium.

Background technique

With the development of science and technology and the progress of society, in order to ensure the stable development of the economy, it is very necessary to predict the financial risks of the industry in advance. The financial risk of the industry may endanger the stability of the entire financial system and cause serious negative effects on the real economy.

The existing risk prediction methods are mainly based on the relatively simple network structure between enterprises, and the generation of the network structure depends on the relationship between enterprises, such as credit relationship. The inventor realized that the relationships among enterprises are complex and diverse, such as cooperative research and development relationships, competitive relationships, supply chain relationships, etc., which are difficult to observe, which will lead to the lack of many unobservable implicit relationships in the network structure, which in turn leads to risk prediction for the industry The accuracy is lower.

Contents of the invention

A risk prediction method provided by this application includes:

Obtaining a multi-source information set, extracting a risk-aware factor set from the multi-source information set, and constructing a time-series knowledge map based on the risk-aware factor set;

Complementing the implicit relationship in the time-series knowledge graph by using a preset implicit relationship complement algorithm to obtain a standard knowledge graph;

Build a risk prediction model based on a preset reinforcement learning algorithm;

Using the risk prediction model to perform risk prediction on the entities in the standard knowledge graph to obtain a risk probability, and use the entity whose risk probability is greater than or equal to a preset probability threshold as a target risk entity;

Using a preset causality supplement algorithm to supplement the time series knowledge map with causality to obtain an event map;

Using a preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the degree of dependency, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the degree of event hazard;

Based on the event map, the dependency closeness and the event hazard degree combined with graph neural network and semi-supervised method training to obtain a macro-prediction model;

Predict the target risk entity by using the macro prediction model to obtain a macro risk probability, and determine that the industry type corresponding to the entity whose macro risk probability is greater than or equal to a preset macro threshold is a risk industry.

The present application also provides a risk prediction device, the device comprising:

The knowledge map construction module is used to obtain a multi-source information set, extract a risk-aware factor set from the multi-source information set, and construct a time-series knowledge map based on the risk-aware factor set, and use the preset implicit relationship to complement the algorithm The implicit relationship in the time series knowledge graph is obtained to obtain a standard knowledge graph;

The target risk entity prediction module is used to construct a risk prediction model based on a preset reinforcement learning algorithm, use the risk prediction model to perform risk prediction on entities in the standard knowledge map, obtain a risk probability, and set the risk probability to be greater than Or an entity equal to a preset probability threshold as a target risk entity;

An event map generation module, configured to use a preset causal relation supplementary algorithm to supplement the time-series knowledge map with causal relations to obtain an event map;

The graph quantification module is used to use the preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the closeness of dependence, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the event hazard degree;

The macro prediction module is used to obtain a macro prediction model based on the event map, the dependency closeness and the event hazard degree combined with graph neural network and semi-supervised method training, and use the macro prediction model to predict the target risk entity Prediction is performed to obtain the macro risk probability, and the industry type corresponding to the entity whose macro risk probability is greater than or equal to the preset macro threshold is determined as a risk industry.

The present application also provides an electronic device, the electronic device comprising:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory stores a computer program executable by the at least one processor, the computer program is executed by the at least one processor, so that the at least one processor can perform the following steps:

The present application also provides a computer-readable storage medium, at least one computer program is stored in the computer-readable storage medium, and the at least one computer program is executed by a processor in an electronic device to implement the following steps:

Description of drawings

Fig. 1 is a schematic flow chart of a risk prediction method provided by an embodiment of the present application;

FIG. 2 is a functional block diagram of a risk prediction device provided by an embodiment of the present application;

Fig. 3 is a schematic structural diagram of an electronic device implementing the risk prediction method provided by an embodiment of the present application.

The realization, functional features and advantages of the present application will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

Detailed ways

It should be understood that the specific embodiments described here are only used to explain the present application, and are not intended to limit the present application.

The embodiment of this application provides a risk prediction method. The subject of execution of the risk prediction method includes but is not limited to at least one of electronic devices such as a server and a terminal that can be configured to execute the method provided by the embodiment of the present application. In other words, the risk prediction method can be executed by software or hardware installed on the terminal device or server device, and the software can be a block chain platform. The server includes, but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, and the like. The server can be an independent server, or it can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (ContentDelivery Network) , CDN), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.

Referring to FIG. 1 , it is a schematic flowchart of a risk prediction method provided by an embodiment of the present application. In this embodiment, the risk prediction method includes:

S1. Acquire a multi-source information set, extract a risk perception factor set from the multi-source information set, and construct a time-series knowledge graph based on the risk perception factor set.

In the embodiment of this application, the multi-source information set can be obtained from multi-source heterogeneous information sources, wherein the multi-source heterogeneous information sources come from multiple channels such as Internet public data, business data and government data, including but not It is limited to various types such as video, electronic pictures, remote sensing images, and texts.

Specifically, the extracting the risk perception factor set from the multi-source information set includes:

identifying text information and image information in the multi-source information set;

performing factor extraction on the text information using a preset natural language processing technology to obtain a text-aware factor set;

performing factor extraction on the image information by using a preset image recognition technology to obtain an image perception factor set;

Summarizing the text perception factor set and the image perception factor set to obtain a risk perception factor set.

In detail, the information in the multi-source information set contains different types, and the text information and image information in the multi-source information set are screened out, and the text information is factor-extracted using a preset natural language processing technology, wherein, Described natural language processing technique can be to utilize Word2Vec algorithm to carry out word embedding, utilize FastText model or ELMO (Embedding from language models, two-way language model) model to carry out text feature encoding, utilize BERT (Bidirectional Encoder Representations from Transformer, two-way encoder representation) The model performs text feature extraction. Use preset image recognition technology to extract factors from the image information, wherein the image recognition technology can be target detection algorithms such as YOLO V2-V4 or SSD, target recognition algorithms such as AlexNet or ResNet, and weakly supervised semantic segmentation methods and other semantic segmentation algorithms.

In this solution, for image information, for example, satellite image recognition technology, optical character recognition (OCR) and NLP technology can be used to complete information extraction. For example, targets such as crops, shipping goods, and sea and land transportation can be identified from ultra-high-resolution satellite images, and then early warnings can be made for changes in important links of economic production. OCR technology can be used to extract useful information from non-standard information such as financial notes and transaction notes Important information for risk assessment, while nighttime light remote sensing data can be used to dynamically predict population density and urban expansion speed. For text information content, natural language processing (NLP) combined with machine learning and other technologies can be used to complete information extraction. For example, it is possible to identify financial entities in real time from text data such as news, public opinion, and forum information, discover the relationship between financial events, and extract relevant factors that describe economic uncertainty; from listed company annual reports, IPO prospectuses, and company forward-looking statements Text data, which can mine information such as enterprise income, business development scale, and company development strategy tendency; it can also extract event tendency scores, attention Degree index, risk volatility and other factors.

Further, the constructing a time series knowledge map based on the risk perception factor set includes:

extracting entities and entity relationships in the set of risk perception factors;

A graph is constructed based on the entity and the entity relationship to obtain a time series knowledge graph.

Among them, using Word2Vec or LSTM+CRF for entity extraction, the algorithm for entity relationship extraction includes but is not limited to the following: supervised learning (SVM, NN, naive Bayesian), semi-supervised learning (distance supervision, Bootstrapping), deep learning ( Pipeline such as Att-CNN&Att-BLSTM, Joint Model such as LSTM-RNNs). Use rule-based reasoning SWRL or graph-based reasoning Path Ranking to construct graphs to obtain time-series knowledge graphs.

For example, taking the PPP (Public-Private Partnership, government and social capital cooperation) cooperation model as an example of the bond issuance event of "Oriental Garden" in 2018, the entity mentioned can be Oriental Garden, 1 billion corporate bonds, etc. An entity relationship can be Announcement, Issue, etc.

S2. Completing the implicit relationship in the time-series knowledge graph by using the preset implicit relationship complement algorithm to obtain a standard knowledge graph.

In the embodiment of the present application, by completing the implicit relationship in the time-series knowledge graph, the relationship between entities in the time-series knowledge graph can be enriched and calibrated, and the application range of the time-series knowledge graph can be broadened.

Wherein, the implicit relationship refers to the entity relationship that is difficult to obtain directly between entities, and the time-series knowledge graph contains multiple entities and the relationship between entities, but the relationship between entities in the time-series knowledge graph is more obvious Therefore, it is necessary to deeply mine each entity to obtain the implicit relationship, and add the implicit relationship to the time-series knowledge graph to obtain a standard knowledge graph.

Specifically, the use of the preset implicit relationship complement algorithm to complement the implicit relationship in the time-series knowledge graph to obtain a standard knowledge graph includes:

performing graph sparse processing on the time-series knowledge graph based on a preset sparse graph convolutional network to obtain a sparse knowledge graph;

Using the trained relational graph convolutional network to predict the relationship of the sparse knowledge graph to obtain the hidden relationship;

Completing the implicit relationship in the time series knowledge graph to obtain a standard knowledge graph.

In detail, this solution uses the graph convolutional network model to complete the implicit relationship of the time-series knowledge graph, mainly in that the graph sparse processing is performed on the time-series knowledge graph first, and the sparse data obtained after the graph sparse Knowledge graphs are used as input to relational graph convolutional networks for relational prediction. This method can avoid the problem of low prediction accuracy of the model due to sparse features when the observed entity relationship samples are insufficient and the entity relationship types are diverse. At the same time, it can solve the problem of nodes on large graphs as the size of the graph increases. Functions such as classification and relation prediction can also be time and space intensive problems.

Among them, the graph convolutional network model converts the input knowledge map into a sparse knowledge map and performs the next step of relationship prediction.

Further, the preset sparse graph convolutional network performs graph sparse processing on the time series knowledge graph to obtain a sparse knowledge graph, including:

Determine the adjacency matrix and feature matrix corresponding to the time series knowledge graph, and obtain the preset weight matrix of the sparse graph convolutional network;

constructing a sparse output function based on the adjacency matrix, the feature matrix and the preset weight matrix;

Optimizing the sparse output function by using a multiplier-based alternating direction algorithm, updating variables in the optimized sparse output function on the time-series knowledge graph to obtain a sparse knowledge graph.

In detail, if the time-series knowledge graph is G=(V,E), where V={v ₁ ,...,v _n } is the set of entities in the time-series knowledge graph, E={e ₁ ,..., e _m } is a set of relationships between entities in the time-series knowledge graph, and the adjacency matrix corresponding to the time-series knowledge graph is a two-dimensional array storing the relationship between entities. The feature matrix is the feature of the node corresponding to each entity, expressed as X(v)=[x ₁ ,...,x _k ], when the sparse graph convolutional network contains two layers of sub-networks, the sparse The preset weight matrix of the graph convolutional network can be W ⁽⁰⁾ and W ⁽¹⁾ .

Wherein, for the time series knowledge map is G=(V, E), its corresponding adjacency matrix is:

Specifically, the construction of the sparse output function based on the adjacency matrix, the feature matrix and the preset weight matrix includes:

in,

is the sparse output function, W is the preset weight matrix,

To update the adjacency matrix, A is the adjacency matrix, _IN is a fixed parameter, diag is a diagonal matrix, W ⁽⁰⁾ and W ⁽¹⁾ are the preset weight matrices of the sparse graph convolutional network, and ReLU is linear rectification function, X is the feature matrix.

Further, the optimization of the sparse output function by using the multiplier-based alternating direction algorithm includes:

Simplifying the sparse output function to obtain a simplified output function;

An adaptive moment estimation optimizer is used to update the gradients of the variables in the simplified output function to obtain the updated gradients of the variables.

In detail, simplifying the sparse output function means that the output of the sparse graph convolutional network depends on

and W, but

can be expressed as a function of A, then the output can be expressed as a function of A and W, that is, Z(A,W), because W remains unchanged, Z(A,W) can be simplified to Z(A) .

Wherein, the simplified output function is:

Specifically, the adaptive moment estimation optimizer is used to update the gradient of the variables in the simplified output function, and the variables in the optimized sparse output function are updated to the time-series knowledge graph to obtain a sparse knowledge graph .

In detail, the optimization process using the multiplier-based Alternating Direction Algorithm (ADMN) can preserve the network backbone of the structural and hierarchical information in the temporal knowledge graph, and preserve the performance of the temporal knowledge graph while maintaining the performance of node classification prediction. edge information.

Further, before using the trained relational graph convolutional network to perform relation prediction on the sparse knowledge map, and obtain the hidden relation, the method further includes:

Using the entity encoder in the preset relational graph convolution network to perform feature prediction on the sparse knowledge graph, and obtain the potential features corresponding to the entities in the sparse knowledge graph;

Scoring the latent features corresponding to the entity based on the decoder in the relational graph convolutional network, and using the corresponding latent features whose scores are greater than or equal to a preset scoring threshold as target latent features;

Calculate and obtain a cross-entropy loss value according to the target latent feature and a preset cross-entropy loss function;

When the cross-entropy loss value is less than or equal to a preset loss threshold, the relational graph convolutional network is output as a trained relational graph convolutional network.

Wherein, the relational graph convolutional network includes an entity encoder and a decoder, the entity encoder is used to generate a latent feature representation of an entity, and the decoder is used to score the latent feature representation through a scoring function.

Specifically, the entity encoder in the preset relational graph convolutional network is used to perform feature prediction on the sparse knowledge graph, that is, the R-GCN is used as the encoder to generate a real vector representation e _i of each entity. Among them, the R-GCN model stacks L layers according to the preset method, the output of the previous layer in the R-GCN model will be used as the input of the next layer, and the entity encoder uses the output of R-GCN as the vector of each entity means that

is the hidden vector (hidden state) of node v _i in the l-layer neural network, and d ^(l) is the dimension represented by the vector of this layer.

Further, the scoring of the latent features corresponding to the entity based on the decoder in the relational graph convolutional network includes:

The latent features corresponding to the entities are scored using the DistMult factorization model in the decoder.

In detail, the DistMult factorization model is a kind of semantic matching model. The semantic matching model uses a similarity-based scoring function to measure the possibility of the existence or establishment of this triple by matching the latent semantics of entities and relations in the latent space.

Specifically, use the DistMult factorization model in the decoder to score the potential features corresponding to the entity, including:

in,

is the implicit vector representation of the head entity s,

is the implicit vector representation of the tail entity o,

is the adjacency matrix for relation type r, and d is the dimension of the entity vector.

Further, the cross-entropy loss value calculated according to the target potential feature and the preset cross-entropy loss function includes:

The cross-entropy loss function is:

in,

is the cross-entropy loss value,

Is the set of all positive and negative triplet samples, for each element (s, r, o, y) in the set,

are head entity and tail entity, respectively,

is the relationship type, y is an indicator, y=1 means a positive sample, and y=0 means a negative sample.

Specifically, use the trained relational graph convolutional network to predict the relationship of the sparse knowledge graph to obtain implicit relationships, and complete the implicit relationships in the time-series knowledge graph to obtain a standard knowledge graph.

For example, the relationship prediction is carried out on the sparse knowledge map, and the implicit relationship is obtained as "using the PPP mode to trigger the oriental garden event", then the implicit relationship "using the PPP mode to trigger the oriental garden event" is added to the time series knowledge map In the process, the standard knowledge graph is obtained.

S3. Construct a risk prediction model based on a preset reinforcement learning algorithm.

In the embodiment of the present application, the preset reinforcement learning algorithm is a framework that can be applied to sequential decision-making and control tasks, wherein the agent (Agent) in the reinforcement learning algorithm optimizes its behavior by interacting with the environment (Environment).

Specifically, the construction of a risk prediction model based on a preset reinforcement learning algorithm includes:

Obtaining original risk state data, performing sampling processing on the original risk state data, and obtaining pre-training data;

Using a preset deep neural network to perform fitting processing on the pre-training data to obtain state actions corresponding to the pre-training data;

Acquiring initial risk state data under execution of the state action, and calculating a time difference between the initial risk state data and the original risk state data;

The time difference is used as an objective function, and the reinforcement learning algorithm is used as a framework to train a risk prediction model.

Wherein, the original risk status data refers to the identification data of the risk situation to which the current data belongs. Pre-training data can be obtained by extracting data conforming to preset sampling standards, for example, extracting data whose risk profile satisfies high-risk and medium-risk conditions as pre-training data.

In detail, the reinforcement learning algorithm is the Actor-Critic algorithm. In the Actor-Critic method based on the policy gradient, the role of the agent is divided into a participant (Actor) and a critic (Critic). In essence, Actor and Critic respectively represent Policy and Value function. Given a current state x, an actor is only responsible for generating action u. The critic is responsible for processing the received reward r, i.e. evaluating the quality of the current policy by adjusting the value function. After the critic has performed multiple policy evaluation steps, the participants are updated by using information from the critic.

S4. Use the risk prediction model to perform risk prediction on the entities in the standard knowledge graph to obtain risk probabilities, and use entities whose risk probabilities are greater than or equal to a preset probability threshold as target risk entities.

In the embodiment of the present application, the risk prediction model can be used to perform risk prediction on a plurality of different entities in the standard knowledge map, and the entities are input into the risk prediction model to obtain the risk corresponding to the entity probability, and the entity whose risk probability is greater than or equal to the preset probability threshold is taken as the target risk entity.

Wherein, the entities in the standard knowledge graph include different types of enterprises.

For example, in this solution, an enterprise whose risk probability in the standard knowledge graph is predicted to be greater than or equal to the preset probability threshold according to the risk prediction model is used as a target risk enterprise.

S5. Using a preset causality supplement algorithm to supplement the time series knowledge graph with causality to obtain an event graph.

In the embodiment of the present application, the causal relationship supplement refers to supplementing the causal sequence and other relationships among entities in the time series knowledge graph.

Specifically, the use of the preset causality supplement algorithm to supplement the time-series knowledge map with causality to obtain the event map includes:

Obtain a training text set, perform event extraction and causal relationship induction on the training text set, and obtain multiple causal triplets;

Retaining a plurality of causal events that meet preset screening criteria in the causal triplet as a standard triplet;

Event fusion is performed on multiple standard triples to obtain fusion events, and the fusion events are added to the time series knowledge graph to obtain an event graph.

In detail, an event refers to an event or state change that occurs at a specific point in time or a period of time, or in a specific geographical area, and consists of one or more actions involving one or more roles. One event causes or causes or causes another event between two events. The causal relationship includes positive, negative, explicit and implicit relationships, and includes other things such as exclusion of turning, juxtaposition, etc., which can help event fusion and reasoning various relationships. The causality induction is to form a causal triplet in the data form of a triplet of "causal event-relationship-result event" from two events with causal relationship extracted from the training text set.

Specifically, event extraction and causality induction can be completed through a pre-training model, wherein the basic model in the pre-training model adopts the structural idea of BERT+CRF, and in fact, RoBERTa, an improved version of BERT, is used as the pre-training model. The input is a word embedding vector. After paragraph embedding, position embedding and punctuation embedding, multiple transformer modules are used to output multiple hidden vectors. Then the sequence label generation task is completed by the Seq2Seq model. In addition, because each input word will get a corresponding label (such as the beginning word of the event, the middle word of the event, the ending word of the event, other words, etc.), this model can also complete the standardized expression of events and remove adverbs, particles and other tasks. The joint model makes full use of the semantic information of the pre-trained model, and achieves good results in both event extraction and induction.

In detail, the preset screening criteria are pre-constructed rules or templates to filter event nodes that do not conform to expression habits or incomplete expressions, and after obtaining standardized event representation and causal relationship, use pre-constructed rules or templates to filter events Express custom or incomplete event nodes. Different event nodes after screening may refer to the same entity in the real world because they have the same meaning. Therefore, event fusion is required to obtain fusion events.

Further, performing event fusion on multiple standard triples to obtain a fusion event includes:

vectorizing multiple standard triples to obtain multiple triple vector groups;

calculating the similarity between events in a plurality of said triplet vector groups;

If the similarity is greater than a preset first threshold, the standard triplet corresponding to the similarity is divided into the first cluster of events;

If the similarity is less than a preset second threshold, dividing the standard triplet corresponding to the similarity into a second cluster of events;

If the similarity is less than the first threshold and greater than the second threshold, classify the standard triplet corresponding to the similarity into buffered cluster events.

In detail, the incremental clustering algorithm can obtain real-time clustering results, so after clustering some event samples, samples can be extracted from the clustering results to expand the training set and retrain the model so that the model can learn new Event and text features to enhance the clustering effect. Finally, after the above steps are completed, combine all the clusters with fewer samples in the clustering result with Buffer as uncertain samples, and use the model after multiple trainings to cluster these samples to obtain the final clustering result. Thus, the process of this part of event fusion is completed.

S6. Use the preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the dependency closeness, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the event hazard degree.

In the embodiment of the present application, the use of the preset social network analysis algorithm to quantify the relationship of the standard knowledge map to obtain the closeness of dependence includes:

Determine the degree centrality and modular classification metrics of the standard knowledge graph according to the definitions of degree centrality and modular classification metrics;

Calculate the proximity centrality of the standard knowledge graph using a preset proximity centrality calculation formula, and substitute the degree centrality, the modular classification measure, and the proximity centrality into the preset dependency closeness calculation formula , get the dependency closeness.

In detail, the degree centrality (Degree centrality) refers to the number of edges connected to a node, which is used to represent the connection degree of nodes, and the modularity class is used for community detection, and is used to measure community division Quality or stability, a modular classification measure equal to the number of edges within a group minus the expected number of edges in an equivalent network with randomly set edges. The Betweenness centrality measures how easy it is for a node to reach other nodes.

Specifically, the calculation of the proximity centrality of the standard knowledge graph using a preset proximity centrality calculation formula includes:

Among them, C _B (v) represents the proximity centrality value of node v, σ _st (v) represents the sum of the shortest paths from node s to node t passing through v, and σ _st represents the distance between node s and node t The sum of all the shortest path numbers of , v,s,t∈V.

Further, the preset calculation formula of dependence closeness includes:

Among them, T(v _i , v _j ) is the closeness of dependence between v _i and v _j , w _D , w _B , w _M ∈ (0,1) are the weights of each sub-indicator and w _D +w _B + w _M =1.

Specifically, using the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the degree of event hazard, the graph attention network (GAT) learns the attention coefficients on all neighbors of the node to perform feature aggregation, and can Improving the performance of many graph learning tasks. Use the graph attention network to score the degree of risk hazard. The scoring function depends on the attention coefficient of the network and the related entity feature vector, and then train the model with the goal of minimizing the mean square error loss, and finally output the relationship between each pair of entities in the form of a matrix. risk hazard score. Therefore, for different knowledge graphs and entity characteristics, it is possible to quantitatively evaluate entity relationships in the financial field, such as corporate credit relationships, supply chain relationships, and inter-industry input-output relationships.

S7. Obtain a macro forecasting model based on the event graph, the dependency closeness and the event hazard degree combined with graph neural network and semi-supervised training.

In the embodiment of this application, the S7 includes:

Summarizing the event map, the degree of dependency and the degree of hazard of the event into labeled data, and using the labeled data as a risk label to construct a supervised model;

Obtaining unlabeled data, constructing an unsupervised time series model based on the unlabeled data and the labeled data;

The supervised model and the unsupervised time series model are combined into a macro forecasting model by using a preset semi-supervised Bayesian algorithm.

S8. Use the macro prediction model to predict the target risk entity to obtain a macro risk probability, and determine that the industry type corresponding to the entity whose macro risk probability is greater than or equal to a preset macro threshold is a risk industry.

In the embodiment of the present application, the macro forecast model is used to predict the risk of the industry, and the macro forecast model is used to predict the target risk entity to obtain the macro risk probability, and the macro risk probability is greater than or equal to the predicted The industry type corresponding to the entity of the set macro threshold is a risk industry.

For example, the target risk entity is the construction and environmental protection sector, and the macro forecast model is used to predict the target risk entity, and the macro risk probability is 0.6. If the preset macro threshold is 0.5, the macro risk probability is greater than As for the macro threshold, the target risk entity is the construction industry corresponding to the construction and environmental protection sector as the risk industry.

The embodiment of the present application extracts the risk perception factor set from the pre-acquired multi-source information set, and builds a time-series knowledge map based on the risk-aware factor set. The relation complement algorithm completes the implicit relation in the time series knowledge map to obtain the standard knowledge map. Building a risk prediction model based on a preset reinforcement learning algorithm, using the reinforcement learning algorithm to construct the model can ensure the stability of the model, and using the risk prediction model to perform risk prediction on entities in the standard knowledge map, and obtain Target risk entity. Respectively quantify the relationship and degree of the standard knowledge map to obtain the closeness of dependence and the degree of hazard of the event, and combine the training of the event map obtained by causal supplementation to obtain a macro-prediction model, and use the macro-prediction model to predict the target risk entity , to get the corresponding risk prediction industry. The risk forecasting model can realize risk forecasting from the perspective of entities, and the macro forecasting model can predict the target risk entities that have undergone risk forecasting from an industry perspective, thereby improving the accuracy of risk forecasting for industries. Therefore, the risk prediction method proposed in this application can solve the problem that the accuracy of risk prediction for the industry is not high enough.

As shown in FIG. 2 , it is a functional block diagram of a risk prediction device provided by an embodiment of the present application.

The risk prediction device 100 described in this application can be installed in an electronic device. According to the realized functions, the risk prediction device 100 may include a knowledge graph construction module 101 , a target risk entity prediction module 102 , an event graph generation module 103 , a graph quantification module 104 and a macro prediction module 105 . The module described in this application can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of the electronic device and can complete fixed functions, and are stored in the memory of the electronic device.

In this embodiment, the functions of each module/unit are as follows:

The knowledge map construction module 101 is used to obtain a multi-source information set, extract a risk-aware factor set from the multi-source information set, and construct a time-series knowledge map based on the risk-aware factor set, and use preset implicit relationships to supplement The algorithm completes the implicit relationship in the time-series knowledge graph to obtain a standard knowledge graph;

The target risk entity prediction module 102 is configured to construct a risk prediction model based on a preset reinforcement learning algorithm, use the risk prediction model to perform risk prediction on entities in the standard knowledge graph, obtain a risk probability, and convert the Entities whose risk probability is greater than or equal to the preset probability threshold are regarded as target risk entities;

The event map generation module 103 is used to supplement the sequence knowledge map with causality by using a preset causal relation supplementary algorithm to obtain an event map;

The graph quantification module 104 is used to quantify the relationship of the standard knowledge graph by using a preset social network analysis algorithm to obtain the degree of dependency, and use the preset graph attention network to quantify the degree of the standard knowledge graph, Get the degree of hazard of the event;

The macro-prediction module 105 is used to obtain a macro-prediction model based on the event map, the dependency closeness and the degree of hazard of the event in combination with graph neural network and semi-supervised method training, and use the macro-prediction model to analyze the The target risk entity performs prediction to obtain the macro risk probability, and determines that the industry type corresponding to the entity whose macro risk probability is greater than or equal to the preset macro threshold is a risk industry.

In detail, each module described in the risk prediction device 100 in the embodiment of the present application uses the same technical means as the risk prediction method described in Figure 1 above, and can produce the same technical effect, and will not be repeated here. .

As shown in FIG. 3 , it is a schematic structural diagram of an electronic device implementing a risk prediction method provided by an embodiment of the present application.

The electronic device 1 may include a processor 10, a memory 11, a communication bus 12, and a communication interface 13, and may also include a computer program stored in the memory 11 and operable on the processor 10, such as a risk prediction program .

Wherein, the processor 10 may be composed of integrated circuits in some embodiments, for example, may be composed of a single packaged integrated circuit, or may be composed of multiple integrated circuits with the same function or different functions packaged, including one or A combination of multiple central processing units (Central Processing unit, CPU), microprocessors, digital processing chips, graphics processors and various control chips, etc. The processor 10 is the control core (Control Unit) of the electronic device, and uses various interfaces and lines to connect the various components of the entire electronic device, and runs or executes programs or modules stored in the memory 11 (such as executing risk prediction program, etc.), and call the data stored in the memory 11 to execute various functions of the electronic device and process data.

The memory 11 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card type memory (for example: SD or DX memory, etc.), magnetic memory, magnetic disk, optical disk, etc. . The storage 11 may be an internal storage unit of the electronic device in some embodiments, such as a mobile hard disk of the electronic device. The memory 11 can also be an external storage device of an electronic device in other embodiments, such as a plug-in mobile hard disk equipped on an electronic device, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD ) card, flash card (Flash Card), etc. Further, the memory 11 may also include both an internal storage unit of the electronic device and an external storage device. The memory 11 can not only be used to store application software and various data installed in electronic equipment, such as codes of risk prediction programs, but also can be used to temporarily store data that has been output or will be output.

The communication bus 12 may be a peripheral component interconnect (PCI for short) bus or an extended industry standard architecture (EISA for short) bus or the like. The bus can be divided into address bus, data bus, control bus and so on. The bus is configured to realize connection and communication between the memory 11 and at least one processor 10 and the like.

The communication interface 13 is used for communication between the electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and/or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), which are generally used to establish a communication connection between the electronic device and other electronic devices. The user interface may be a display (Display) or an input unit (such as a keyboard (Keyboard)). Optionally, the user interface may also be a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light-emitting diode) touch device, and the like. Wherein, the display may also be properly referred to as a display screen or a display unit, and is used for displaying information processed in the electronic device and for displaying a visualized user interface.

FIG. 3 only shows an electronic device with components. Those skilled in the art can understand that the structure shown in FIG. 3 does not constitute a limitation to the electronic device 1, and may include fewer or more components, or combinations of certain components, or different arrangements of components.

For example, although not shown, the electronic device may also include a power supply (such as a battery) for supplying power to various components. Preferably, the power supply may be logically connected to the at least one processor 10 through a power management device, so that Realize functions such as charge management, discharge management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components. The electronic device may also include various sensors, a Bluetooth module, a Wi-Fi module, etc., which will not be repeated here.

It should be understood that the embodiments are only for illustration, and are not limited by the structure in terms of the scope of the patent application.

The risk prediction program stored in the memory 11 in the electronic device 1 is a combination of multiple instructions. When running in the processor 10, it can realize:

Using the risk prediction model to perform risk prediction on the entities in the standard knowledge map to obtain a risk probability, and use the entity whose risk probability is greater than or equal to a preset probability threshold as a target risk entity;

Specifically, for the specific implementation method of the above instructions by the processor 10, reference may be made to the description of relevant steps in the corresponding embodiments in the drawings, and details are not repeated here.

Further, if the integrated modules/units of the electronic device 1 are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium may be volatile or non-volatile. For example, the computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, removable hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read- Only Memory).

The present application also provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor of an electronic device, it can realize:

In the several embodiments provided in this application, it should be understood that the disclosed devices, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional module in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware, or in the form of hardware plus software function modules.

It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above, but that the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

Therefore, the embodiments should be regarded as exemplary and not restrictive in all points of view, and the scope of the application is defined by the appended claims rather than the foregoing description, and it is intended that the scope of the present application be defined by the appended claims rather than by the foregoing description. All changes within the meaning and range of equivalents of the elements are embraced in this application. Any reference sign in a claim shall not be construed as limiting the claim concerned.

The blockchain referred to in this application is a new application mode of computer technologies such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm. Blockchain (Blockchain), essentially a decentralized database, is a series of data blocks associated with each other using cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify its Validity of information (anti-counterfeiting) and generation of the next block. The blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

The embodiments of the present application may acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. .

In addition, it is obvious that the word "comprising" does not exclude other elements or steps, and the singular does not exclude the plural. A plurality of units or devices stated in the system claims may also be realized by one unit or device through software or hardware. The terms first, second, etc. are used to denote names and do not imply any particular order.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application without limitation. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be Make modifications or equivalent replacements without departing from the spirit and scope of the technical solutions of the present application.

Claims

A risk prediction method, wherein the method comprises:

Obtaining a multi-source information set, extracting a risk-aware factor set from the multi-source information set, and constructing a time-series knowledge map based on the risk-aware factor set;

Complementing the implicit relationship in the time-series knowledge graph by using a preset implicit relationship complement algorithm to obtain a standard knowledge graph;

Build a risk prediction model based on a preset reinforcement learning algorithm;

Using the risk prediction model to perform risk prediction on the entities in the standard knowledge graph to obtain a risk probability, and use the entity whose risk probability is greater than or equal to a preset probability threshold as a target risk entity;

Using a preset causality supplement algorithm to supplement the time series knowledge map with causality to obtain an event map;

Using a preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the degree of dependency, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the degree of event hazard;

Based on the event map, the closeness of dependence and the degree of hazard of the event, combined with graph neural network and semi-supervised method training to obtain a macro-prediction model;

Predict the target risk entity by using the macro prediction model to obtain a macro risk probability, and determine that the industry type corresponding to the entity whose macro risk probability is greater than or equal to a preset macro threshold is a risk industry.
The risk prediction method according to claim 1, wherein the use of the preset implicit relationship supplement algorithm to complement the implicit relationship in the time-series knowledge graph to obtain a standard knowledge graph includes:

performing graph sparse processing on the time-series knowledge graph based on a preset sparse graph convolutional network to obtain a sparse knowledge graph;

Using the trained relational graph convolutional network to predict the relationship of the sparse knowledge graph to obtain the hidden relationship;

Completing the implicit relationship in the time series knowledge graph to obtain a standard knowledge graph.
The risk prediction method according to claim 2, wherein said sparse graph convolution network based on the preset sparse graph performs graph sparse processing on said time series knowledge graph to obtain a sparse knowledge graph, comprising:

Determine the adjacency matrix and feature matrix corresponding to the time series knowledge graph, and obtain the preset weight matrix of the sparse graph convolutional network;

constructing a sparse output function based on the adjacency matrix, the feature matrix and the preset weight matrix;

Optimizing the sparse output function by using a multiplier-based alternating direction algorithm, updating variables in the optimized sparse output function on the time-series knowledge graph to obtain a sparse knowledge graph.
The risk prediction method according to claim 2, wherein, using the trained relational graph convolutional network to predict the relationship of the sparse knowledge graph, before obtaining the implicit relationship, the method further includes:

Using the entity encoder in the preset relational graph convolution network to perform feature prediction on the sparse knowledge graph, and obtain the potential features corresponding to the entities in the sparse knowledge graph;

Scoring the latent features corresponding to the entity based on the decoder in the relational graph convolutional network, and using the corresponding latent features whose scores are greater than or equal to a preset scoring threshold as target latent features;

Calculate and obtain a cross-entropy loss value according to the target latent feature and a preset cross-entropy loss function;

When the cross-entropy loss value is less than or equal to a preset loss threshold, the relational graph convolutional network is output as a trained relational graph convolutional network.
The risk prediction method according to claim 1, wherein said constructing a time series knowledge graph based on said risk perception factor set comprises:

extracting entities and entity relationships in the set of risk perception factors;

A graph is constructed based on the entity and the entity relationship to obtain a time series knowledge graph.
The risk prediction method according to any one of claims 1 to 5, wherein said use of a preset causality supplementary algorithm to supplement the time-series knowledge map with causality to obtain an event map includes:

Obtain a training text set, perform event extraction and causal relationship induction on the training text set, and obtain multiple causal triplets;

Retaining a plurality of causal events that meet preset screening criteria in the causal triplet as a standard triplet;

Event fusion is performed on multiple standard triples to obtain fusion events, and the fusion events are added to the time series knowledge graph to obtain an event graph.
The risk prediction method according to any one of claims 1 to 5, wherein said extracting a risk perception factor set from said multi-source information set comprises:

identifying text information and image information in the multi-source information set;

performing factor extraction on the text information using a preset natural language processing technology to obtain a text-aware factor set;

performing factor extraction on the image information by using a preset image recognition technology to obtain an image perception factor set;

Summarizing the text perception factor set and the image perception factor set to obtain a risk perception factor set.
A risk prediction device, wherein the device comprises:

The knowledge map construction module is used to obtain a multi-source information set, extract a risk-aware factor set from the multi-source information set, and construct a time-series knowledge map based on the risk-aware factor set, and use the preset implicit relationship to complement the algorithm The implicit relationship in the time series knowledge graph is obtained to obtain a standard knowledge graph;

The target risk entity prediction module is used to construct a risk prediction model based on a preset reinforcement learning algorithm, use the risk prediction model to perform risk prediction on entities in the standard knowledge map, obtain a risk probability, and set the risk probability to be greater than Or an entity equal to a preset probability threshold as a target risk entity;

An event map generation module, configured to use a preset causal relation supplementary algorithm to supplement the time-series knowledge map with causal relations to obtain an event map;

The graph quantification module is used to use the preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the closeness of dependence, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the event hazard degree;

The macro prediction module is used to obtain a macro prediction model based on the event map, the dependency closeness and the event hazard degree combined with graph neural network and semi-supervised method training, and use the macro prediction model to predict the target risk entity Prediction is performed to obtain the macro risk probability, and the industry type corresponding to the entity whose macro risk probability is greater than or equal to the preset macro threshold is determined as a risk industry.
An electronic device, wherein the electronic device includes:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory stores a computer program executable by the at least one processor, the computer program is executed by the at least one processor, so that the at least one processor can perform the following steps:

Obtaining a multi-source information set, extracting a risk-aware factor set from the multi-source information set, and constructing a time-series knowledge map based on the risk-aware factor set;

Complementing the implicit relationship in the time-series knowledge graph by using a preset implicit relationship complement algorithm to obtain a standard knowledge graph;

Build a risk prediction model based on a preset reinforcement learning algorithm;

Using the risk prediction model to perform risk prediction on the entities in the standard knowledge graph to obtain a risk probability, and use the entity whose risk probability is greater than or equal to a preset probability threshold as a target risk entity;

Using a preset causality supplement algorithm to supplement the time series knowledge map with causality to obtain an event map;

Using a preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the degree of dependency, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the degree of event hazard;

Based on the event map, the closeness of dependence and the degree of hazard of the event, combined with graph neural network and semi-supervised method training to obtain a macro-prediction model;

Use the macro forecasting model to predict the target risk entity to obtain the macro risk probability, and determine that the industry type corresponding to the entity whose macro risk probability is greater than or equal to the preset macro threshold is a risk industry.
The electronic device according to claim 9, wherein the use of the preset implicit relationship complement algorithm to complement the implicit relationship in the time-series knowledge graph to obtain a standard knowledge graph includes:

performing graph sparse processing on the time-series knowledge graph based on a preset sparse graph convolutional network to obtain a sparse knowledge graph;

Using the trained relational graph convolutional network to predict the relationship of the sparse knowledge graph to obtain the hidden relationship;

Completing the implicit relationship in the time series knowledge graph to obtain a standard knowledge graph.
The electronic device according to claim 10, wherein said sparse graph convolutional network based on a preset performs graph sparse processing on said time-series knowledge graph to obtain a sparse knowledge graph, comprising:

Determine the adjacency matrix and feature matrix corresponding to the time series knowledge graph, and obtain the preset weight matrix of the sparse graph convolutional network;

constructing a sparse output function based on the adjacency matrix, the feature matrix and the preset weight matrix;

Optimizing the sparse output function by using a multiplier-based alternating direction algorithm, updating variables in the optimized sparse output function on the time-series knowledge graph to obtain a sparse knowledge graph.
The electronic device according to claim 10, wherein, the trained relational graph convolutional network is used to predict the relationship of the sparse knowledge graph, and before the implicit relationship is obtained, the computer program is executed by the at least one processor The following steps are also implemented during execution:

Using the entity encoder in the preset relational graph convolution network to perform feature prediction on the sparse knowledge graph, and obtain the potential features corresponding to the entities in the sparse knowledge graph;

Scoring the latent features corresponding to the entity based on the decoder in the relational graph convolutional network, and using the corresponding latent features whose scores are greater than or equal to a preset scoring threshold as target latent features;

Calculate and obtain a cross-entropy loss value according to the target latent feature and a preset cross-entropy loss function;

When the cross-entropy loss value is less than or equal to a preset loss threshold, the relational graph convolutional network is output as a trained relational graph convolutional network.
The electronic device according to claim 9, wherein said constructing a time series knowledge graph based on said risk perception factor set comprises:

extracting entities and entity relationships in the set of risk perception factors;

A graph is constructed based on the entity and the entity relationship to obtain a time series knowledge graph.
The electronic device according to any one of claims 9 to 13, wherein said use of a preset causality complement algorithm to supplement the sequence knowledge graph with causality to obtain an event graph includes:

Obtain a training text set, perform event extraction and causal relationship induction on the training text set, and obtain multiple causal triplets;

Retaining a plurality of causal events that meet preset screening criteria in the causal triplet as a standard triplet;

Event fusion is performed on multiple standard triples to obtain fusion events, and the fusion events are added to the time series knowledge graph to obtain an event graph.
The electronic device according to any one of claims 9 to 14, wherein said extracting a risk perception factor set from said multi-source information set comprises:

identifying text information and image information in the multi-source information set;

performing factor extraction on the text information using a preset natural language processing technology to obtain a text-aware factor set;

performing factor extraction on the image information by using a preset image recognition technology to obtain an image perception factor set;

Summarizing the text perception factor set and the image perception factor set to obtain a risk perception factor set.
A computer-readable storage medium storing a computer program, wherein the computer program implements the following steps when executed by a processor:

Obtaining a multi-source information set, extracting a risk-aware factor set from the multi-source information set, and constructing a time-series knowledge map based on the risk-aware factor set;

Complementing the implicit relationship in the time-series knowledge graph by using a preset implicit relationship complement algorithm to obtain a standard knowledge graph;

Build a risk prediction model based on a preset reinforcement learning algorithm;

Using the risk prediction model to perform risk prediction on the entities in the standard knowledge graph to obtain a risk probability, and use the entity whose risk probability is greater than or equal to a preset probability threshold as a target risk entity;

Using a preset causality supplement algorithm to supplement the time series knowledge map with causality to obtain an event map;

Using a preset social network analysis algorithm to quantify the relationship of the standard knowledge graph to obtain the degree of dependency, and use the preset graph attention network to quantify the degree of the standard knowledge graph to obtain the degree of event hazard;

Based on the event map, the closeness of dependence and the degree of hazard of the event, combined with graph neural network and semi-supervised method training to obtain a macro-prediction model;

Predict the target risk entity by using the macro prediction model to obtain a macro risk probability, and determine that the industry type corresponding to the entity whose macro risk probability is greater than or equal to a preset macro threshold is a risk industry.
The computer-readable storage medium according to claim 16, wherein said using a preset implicit relationship complement algorithm to complement the implicit relationship in the time-series knowledge graph to obtain a standard knowledge graph includes:

performing graph sparse processing on the time-series knowledge graph based on a preset sparse graph convolutional network to obtain a sparse knowledge graph;

Using the trained relational graph convolutional network to predict the relationship of the sparse knowledge graph to obtain the hidden relationship;

Completing the implicit relationship in the time series knowledge graph to obtain a standard knowledge graph.
The computer-readable storage medium according to claim 17, wherein said sparse graph convolutional network based on a preset performs graph sparse processing on said time-series knowledge graph to obtain a sparse knowledge graph, comprising:

Determine the adjacency matrix and feature matrix corresponding to the time series knowledge graph, and obtain the preset weight matrix of the sparse graph convolutional network;

constructing a sparse output function based on the adjacency matrix, the feature matrix and the preset weight matrix;

Optimizing the sparse output function by using a multiplier-based alternating direction algorithm, updating variables in the optimized sparse output function on the time-series knowledge graph to obtain a sparse knowledge graph.
The computer-readable storage medium according to claim 17, wherein, the trained relational graph convolutional network is used to predict the relationship of the sparse knowledge graph, and before the implicit relationship is obtained, the computer program is executed by the processor Also implement the following steps:

Using the entity encoder in the preset relational graph convolution network to perform feature prediction on the sparse knowledge graph, and obtain the potential features corresponding to the entities in the sparse knowledge graph;

Scoring the latent features corresponding to the entity based on the decoder in the relational graph convolutional network, and using the corresponding latent features whose scores are greater than or equal to a preset scoring threshold as target latent features;

Calculate and obtain a cross-entropy loss value according to the target latent feature and a preset cross-entropy loss function;

When the cross-entropy loss value is less than or equal to a preset loss threshold, the relational graph convolutional network is output as a trained relational graph convolutional network.
The computer-readable storage medium according to claim 16, wherein said constructing a time series knowledge graph based on said risk perception factor set comprises:

extracting entities and entity relationships in the set of risk perception factors;

A graph is constructed based on the entity and the entity relationship to obtain a time series knowledge graph.