Intelligent risk control system based on large model

Through the intelligent risk control system based on large models, using information collection and graph structure data pruning technology, the problems of information asymmetry and cognitive bias of investors in the rapidly changing market are solved, and more efficient risk prediction and investment decision-making are achieved. support.

CN118967326AInactive Publication Date: 2024-11-15TI RONG INTELLIGENT TECH (SUZHOU) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411019139.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The risk control of existing digital products relies on investor experience and is susceptible to information asymmetry and personal cognitive bias, resulting in poor investment decision-making, especially in rapidly changing markets, which is difficult to achieve the best investment choice.

Method used

Using an intelligent risk control system based on large models, a feature sequence is generated through the information collection module, a graph structure data construction module is used to construct graph structure data, a graph structure data pruning module is pruning according to investment preferences, and a risk score prediction module outputs a risk score and passes the Parallel computing layer and hidden layer generate investment recommendation lists to improve risk prediction accuracy.

Benefits of technology

By reducing the scale of graph structure data, improving calculation speed, enhancing risk prediction accuracy, providing investors with the best digital product choices and reducing investment risks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967326A_ABST
    Figure CN118967326A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of intelligent risk control, and discloses an intelligent risk control system based on a large model, and the system comprises an information collection module which collects the basic information and transaction information of a digital product, and generates a feature sequence; the graph structure data construction module is used for constructing first graph structure data; the graph structure data pruning module is used for pruning the first graph structure data according to the investment preference of the investor to obtain second graph structure data; the risk score prediction module is used for inputting the second graph structure data into a risk prediction model and outputting a risk score corresponding to each node of the second graph structure data; the graph structure data is pruned through the word vector model according to the investment preference of the investor, the calculation speed of the risk prediction model can be improved, the risk prediction model performs information aggregation on the transaction information of the digital product in the time dimension, the risk prediction precision of the digital product can be improved, and the risk prediction accuracy of the digital product is improved. Therefore, the best digital product selection is provided for investors.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent risk control technology, and more specifically, to an intelligent risk control system based on a large model. Background Art

[0002] Existing digital products, including but not limited to stocks, digital artworks, cryptocurrencies, etc., have gradually become an important part of today's financial market. Their high liquidity and technological innovation are accompanied by significant market risks, such as drastic and frequent price fluctuations.

[0003] Currently, risk control of digital products mainly relies on the investment experience of investors. However, over-reliance on the investment experience of investors may lead to information asymmetry problems, especially in the rapidly changing digital market. Investors have limited ability to obtain and interpret market information. Moreover, investors’ investment decisions are often affected by personal cognitive biases, which may result in investors’ investment decisions not being optimal. Summary of the invention

[0004] The present invention provides an intelligent wind control system based on a large model to solve the technical problems in the above-mentioned background technology.

[0005] The present invention provides an intelligent risk control system based on a large model, comprising:

[0006] The information collection module is used to collect basic information and transaction information of digital products and generate feature sequences; basic information includes: product type and unique identifier; transaction information includes: opening price, closing price, highest price, lowest price, transaction volume and transaction amount of the past N trading days;

[0007] The feature sequence includes N sequence units, each of which is represented by a combination vector. The combination vector of the t-th sequence unit is obtained by concatenating the opening price, closing price, highest price, lowest price, trading volume, and transaction amount of the digital product on the t-th trading day with the basic information, where 1≤t≤N;

[0008] A graph structure data construction module is used to construct first graph structure data according to a feature sequence of a digital product; the first graph structure data includes: nodes, node features, and edges between nodes;

[0009] Establish data mapping between nodes and digital products;

[0010] Feature sequence representation of node features and digital products to establish data mapping;

[0011] The edges between nodes include: the product types of the digital products mapped by the node data are the same; the total transaction amount of the digital products mapped by the node data in the past N trading days is lower than the preset total amount;

[0012] A graph structure data pruning module, which is used to prune the first graph structure data according to the investment preference of the investor to obtain the second graph structure data; the second graph structure data has the same representation as the first graph structure data; the number of nodes of the second graph structure data is M;

[0013] A risk score prediction module, which is used to input the second graph structure data into the risk prediction model, and the output value represents the risk score corresponding to each node of the second graph structure data;

[0014] The risk score ranges from 1 to 100. A higher risk score indicates a higher investment risk for digital products.

[0015] An investment recommendation list generating module is used to generate an investment recommendation list based on the risk score corresponding to each node of the second graph structure data.

[0016] Furthermore, N, M and the preset total amount are all custom parameters.

[0017] Further, product types include: finance, industry, consumer products, information technology, healthcare, energy, raw materials, real estate, communication services, and utilities, and are represented by real number coding.

[0018] Furthermore, the same word vector model is used to convert investors’ investment preferences and product types of digital products into word vector representations respectively. The cosine similarity between the two is calculated by cosine similarity, and the nodes whose cosine similarity is lower than the similarity threshold are removed. When the number of nodes is greater than M, the similarity threshold is increased. When the number of nodes is less than M, the similarity threshold is lowered until M nodes are obtained.

[0019] Further, the risk prediction model includes: M parallel computing layers, a first extraction layer, a hidden layer, a second extraction layer and a first classifier;

[0020] M parallel computing layers share weight parameters and bias parameters;

[0021] Each parallel computing layer includes N hidden units;

[0022] The i-th hidden unit inputs the i-th sequence unit of the feature sequence of a node of the second graph structure data, and outputs the first updated feature, where 1≤i≤N;

[0023] The first extraction layer is used to extract the first updated feature output by the Nth hidden unit of each parallel computing layer, and convert the feature sequence of each node of the second graph structure data into the first updated feature representation;

[0024] The hidden layer inputs the second graph structure data and outputs the graph update matrix;

[0025] A row vector of the graph update matrix corresponds to the second updated feature of a node;

[0026] The second updated feature of each node is sequentially input into the first classifier, and the classification space of the first classifier represents the risk score corresponding to each node.

[0027] Furthermore, the calculation formula of the i-th hidden unit of the parallel computing layer includes:

[0028]

[0029] in represents the first updated feature output by the i-th hidden unit, X i represents the i-th sequence unit of the feature sequence input to the i-th hidden unit, and denote the forward reset gate, forward update gate, forward first hidden state, and forward second hidden state of the i-th hidden unit, respectively. represents the forward second hidden state of the i-1th hidden unit, Assign a value of 0, and denote the first weight parameter, the second weight parameter and the bias parameter of the forward reset gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the forward update gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the forward first hidden state of the i-th hidden unit, respectively. and They represent the backward reset gate, backward update gate, backward first hidden state and backward second hidden state of the i-th hidden unit respectively. represents the backward second hidden state of the i-1th hidden unit, Assign a value of 0, and They represent the first weight parameter, the second weight parameter and the bias parameter of the backward reset gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the backward update gate of the i-th hidden unit, respectively. and They respectively represent the first weight parameter, the second weight parameter and the bias parameter of the backward first hidden state of the i-th hidden unit, ⊙ represents the Hadamard product, concat represents the concatenation operation, and sigmoid represents the sigmoid activation function.

[0030] Furthermore, the calculation formula of the hidden layer includes:

[0031]

[0032] Where 1≤u≤M, M represents the number of nodes in the second graph structure data, S represents the graph update matrix output by the hidden layer, and N u represents the set of nodes that have edges connected to the u-th node, represents the second updated feature of the u-th node, h u and h v Represents the first updated feature of the u-th node and the v-th node, Nattion uv represents the attention score between the u-th node and the v-th node, D uv represents the number of edges between the u-th node and the v-th node, and Respectively represent the first weight parameter, the second weight parameter and the third weight parameter between the uth node and the vth node, W u represents the weight parameter corresponding to the u-th node, b uv represents the bias parameter between the u-th node and the v-th node, pile represents the stacking operation, T represents the transposition operation, and sigmoid represents the sigmoid activation function.

[0033] Furthermore, the parallel computing layer is pre-trained before the risk prediction model is trained. During the pre-training of the parallel computing layer, the first updated feature output by the Nth hidden unit is input into the second classifier and the third classifier. The classification space of the second classifier represents the trading volume of the next trading day, and the classification space of the third classifier represents the trading amount of the next trading day.

[0034] Furthermore, the sample labels of the training samples used to train the risk prediction model are obtained by taking the average value of the results simultaneously labeled by multiple experts.

[0035] Furthermore, the risk scores corresponding to each node of the second graph structure data are sorted in ascending order, and the digital products mapped by the data of the first K nodes with the smallest risk scores are obtained as an investment recommendation list, where K is a custom parameter and 1≤K≤M.

[0036] The beneficial effects of the present invention are as follows: the present invention prunes graph structure data through a word vector model according to the investment preferences of investors, thereby reducing the scale of graph structure data and improving the calculation speed of the risk prediction model. The risk prediction model aggregates transaction information of digital products in the time dimension, thereby improving the risk prediction accuracy of digital products, thereby providing investors with the best digital product selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of an intelligent wind control system based on a large model of the present invention.

[0038] In the figure: information collection module 101, graph structure data construction module 102, graph structure data pruning module 103, risk score prediction module 104, investment recommendation list generation module 105. DETAILED DESCRIPTION

[0039] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0040] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in one or more embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0041] like Figure 1 As shown, an intelligent risk control system based on a large model includes:

[0042] The information collection module 101 is used to collect basic information and transaction information of digital products and generate feature sequences;

[0043] Basic information includes: product type and unique identifier;

[0044] Transaction information includes: opening price, closing price, highest price, lowest price, trading volume and transaction amount of the past N trading days;

[0045] The feature sequence includes N sequence units, each of which is represented by a combination vector;

[0046] The combination vector of the t-th sequence unit is obtained by concatenating the opening price, closing price, highest price, lowest price, trading volume, and transaction amount of the digital product on the t-th trading day with the basic information, where 1≤t≤N;

[0047] A graph structure data construction module 102, which is used to construct first graph structure data according to a feature sequence of a digital product;

[0048] The first graph structure data includes: nodes, node features, and edges between nodes;

[0049] Establish data mapping between nodes and digital products;

[0050] Feature sequence representation of node features and digital products to establish data mapping;

[0051] The edges between nodes include: the product types of the digital products mapped by the node data are the same; the total transaction amount of the digital products mapped by the node data in the past N trading days is lower than the preset total amount;

[0052] A graph structure data pruning module 103, which is used to prune the first graph structure data according to the investment preference of the investor to obtain the second graph structure data;

[0053] The second graph structure data has the same representation as the first graph structure data;

[0054] The number of nodes of the second graph structure data is M;

[0055] A risk score prediction module 104, which is used to input the second graph structure data into the risk prediction model, and the output value represents the risk score corresponding to each node of the second graph structure data;

[0056] The risk score ranges from 1 to 100. A higher risk score indicates a higher investment risk for digital products.

[0057] The investment recommendation list generating module 105 is used to generate an investment recommendation list according to the risk score corresponding to each node of the second graph structure data.

[0058] In one embodiment of the present invention, N, M and the preset total amount are all custom parameters. Preferably, N is set to 30, M is set to 50, and the preset total amount is set based on the evaluation of relevant technical experts, for example, 100,000 yuan.

[0059] In one embodiment of the present invention, the basic information may also include: the name of the digital product and the financial indicators of the digital product publisher, such as revenue, net profit, total assets, and total liabilities.

[0060] In one embodiment of the present invention, the transaction information may also include: the opening price, closing price, highest price, lowest price, trading volume and transaction amount of the digital product on each trading day and the percentage change between the opening price, closing price, highest price, lowest price, trading volume and transaction amount of the previous trading day.

[0061] In one embodiment of the present invention, product types include: finance, industry, consumer goods, information technology, healthcare, energy, raw materials, real estate, communication services and utilities, and are represented by real number codes. For example, the product type of finance is represented by the real number 1.1, and the product type of industry is represented by the real number 1.2.

[0062] In one embodiment of the present invention, the unique identifier of the digital product is represented by binary code, and may also be represented by generating the unique identifier by a snowflake algorithm.

[0063] For example, a stock's unique identifier is represented by its stock symbol on the exchange where it is listed.

[0064] In one embodiment of the present invention, the investment preferences of investors and the product types of digital products are respectively converted into word vector representations through the same word vector model, the cosine similarity of the two is calculated through cosine similarity, and the nodes whose cosine similarity is lower than the similarity threshold are removed. When the number of nodes is greater than M, the similarity threshold is increased, and when the number of nodes is less than M, the similarity threshold is lowered until M nodes are obtained.

[0065] In one embodiment of the present invention, the word vector model may be a Word2Vec, FastText or GPT model.

[0066] It should be noted that by dynamically adjusting the similarity threshold to ensure that the number of nodes in the second graph structure data is M, unnecessary pruning of nodes can be avoided, thereby retaining digital products that are most closely associated with the investor's investment preferences.

[0067] In one embodiment of the present invention, the edge between nodes may also include: the Euclidean distance similarity of the node features between the nodes is greater than or equal to a preset similarity threshold, wherein the preset similarity threshold is a custom parameter, and preferably, the preset similarity threshold is set to 0.8.

[0068] In one embodiment of the present invention, the risk prediction model includes: M parallel computing layers, a first extraction layer, a hidden layer, a second extraction layer and a first classifier;

[0069] M parallel computing layers share weight parameters and bias parameters;

[0070] Each parallel computing layer includes N hidden units;

[0071] The i-th hidden unit inputs the i-th sequence unit of the feature sequence of a node of the second graph structure data, and outputs the first updated feature, where 1≤i≤N;

[0072] The first extraction layer is used to extract the first updated feature output by the Nth hidden unit of each parallel computing layer, and convert the feature sequence of each node of the second graph structure data into the first updated feature representation;

[0073] The hidden layer inputs the second graph structure data and outputs the graph update matrix;

[0074] A row vector of the graph update matrix corresponds to the second updated feature of a node;

[0075] The second updated feature of each node is sequentially input into the first classifier, and the classification space of the first classifier represents the risk score corresponding to each node.

[0076] In one embodiment of the present invention, the calculation formula of the i-th hidden unit of the parallel computing layer includes:

[0077]

[0078] in represents the first updated feature output by the i-th hidden unit, X i represents the i-th sequence unit of the feature sequence input to the i-th hidden unit, and denote the forward reset gate, forward update gate, forward first hidden state, and forward second hidden state of the i-th hidden unit, respectively. represents the forward second hidden state of the i-1th hidden unit, Assign a value of 0, and denote the first weight parameter, the second weight parameter and the bias parameter of the forward reset gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the forward update gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the forward first hidden state of the i-th hidden unit, respectively. and They represent the backward reset gate, backward update gate, backward first hidden state and backward second hidden state of the i-th hidden unit respectively. represents the backward second hidden state of the i-1th hidden unit, Assign a value of 0, and They represent the first weight parameter, the second weight parameter and the bias parameter of the backward reset gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the backward update gate of the i-th hidden unit, respectively. and They respectively represent the first weight parameter, the second weight parameter and the bias parameter of the backward first hidden state of the i-th hidden unit, ⊙ represents the Hadamard product, concat represents the concatenation operation, and sigmoid represents the sigmoid activation function.

[0079] In one embodiment of the present invention, the parallel computing layer can also be constructed based on RNN (recurrent neural network model), LSTM (long short-term memory neural network model) or GRU (gated neural network model), which will not be described in detail here.

[0080] In one embodiment of the present invention, the calculation formula of the hidden layer includes:

[0081]

[0082] Where 1≤u≤M, M represents the number of nodes in the second graph structure data, S represents the graph update matrix output by the hidden layer, and N u represents the set of nodes that have edges connected to the u-th node, represents the second updated feature of the u-th node, h u and h v Represents the first updated feature of the u-th node and the v-th node, Nattion uv represents the attention score between the u-th node and the v-th node, D uv represents the number of edges between the u-th node and the v-th node, and Respectively represent the first weight parameter, the second weight parameter and the third weight parameter between the uth node and the vth node, W u represents the weight parameter corresponding to the u-th node, b uv represents the bias parameter between the u-th node and the v-th node, pile represents the stacking operation, T represents the transposition operation, and sigmoid represents the sigmoid activation function.

[0083] In one embodiment of the present invention, the parallel computing layer is pre-trained before the risk prediction model is trained. During the pre-training of the parallel computing layer, the first updated feature output by the Nth hidden unit is input into the second classifier and the third classifier. The classification space of the second classifier represents the trading volume of the next trading day, and the classification space of the third classifier represents the trading amount of the next trading day.

[0084] It should be noted that pre-training can improve the model's ability to extract information. Fine-tuning the model parameters after pre-training and then training can speed up the model's training and, to a certain extent, prevent the model from overfitting and improve the model's generalization ability.

[0085] In one embodiment of the present invention, the sample labels of the training samples used to train the risk prediction model are obtained by taking the average value of the results simultaneously labeled by multiple experts.

[0086] It should be noted that during the training process of the risk prediction model, the difference between the value output by the risk prediction model in each iteration and the sample label of the training sample is specified as the loss function. The weight parameters and bias parameters of the risk prediction model are updated through back propagation and gradient descent algorithms. The training of the neural network model is a conventional technical means and will not be elaborated here.

[0087] In one embodiment of the present invention, the risk scores corresponding to each node of the second graph structure data are sorted in ascending order, and the digital products mapped by the data of the first K nodes with the smallest risk scores are obtained as an investment recommendation list, where K is a custom parameter and 1≤K≤M. Preferably, K is set to 10.

[0088] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. An intelligent risk control system based on a large model, It is characterized in that include: An information collection module, which is used to collect basic information and transaction information of digital products and generate feature sequences; Basic information includes: product type and unique identifier; transaction information includes: opening price, closing price, highest price, lowest price, transaction volume and transaction amount of the past N trading days; The feature sequence includes N sequence units, each of which is represented by a combination vector. The combination vector of the t-th sequence unit is obtained by concatenating the opening price, closing price, highest price, lowest price, trading volume, and transaction amount of the digital product on the t-th trading day with the basic information, where 1≤t≤N; A graph structure data construction module is used to construct first graph structure data according to a feature sequence of a digital product; the first graph structure data includes: nodes, node features, and edges between nodes; Establish data mapping between nodes and digital products; Feature sequence representation of node features and digital products to establish data mapping; The edges between nodes include: the product types of the digital products mapped by the node data are the same; the total transaction amount of the digital products mapped by the node data in the past N trading days is lower than the preset total amount; A graph structure data pruning module, which is used to prune the first graph structure data according to the investment preference of the investor to obtain the second graph structure data; the second graph structure data has the same representation as the first graph structure data; the number of nodes of the second graph structure data is M; A risk score prediction module, which is used to input the second graph structure data into the risk prediction model, and the output value represents the risk score corresponding to each node of the second graph structure data; The risk score ranges from 1 to 100. A higher risk score indicates a higher investment risk for digital products. An investment recommendation list generating module is used to generate an investment recommendation list based on the risk score corresponding to each node of the second graph structure data.

2. According to the intelligent risk control system based on a large model as described in claim 1, It is characterized in that N, M and the preset total amount are all custom parameters.

3. According to the intelligent risk control system based on a large model as described in claim 1, It is characterized in that Product types include: financial, industrial, consumer products, information technology, healthcare, energy, raw materials, real estate, communication services and utilities, and are represented by real number codes.

4. According to the intelligent risk control system based on a large model as described in claim 1, It is characterized in that The same word vector model is used to convert investors’ investment preferences and digital product types into word vector representations. The cosine similarity between the two is calculated through cosine similarity, and the nodes whose cosine similarity is lower than the similarity threshold are removed. When the number of nodes is greater than M, the similarity threshold is increased. When the number of nodes is less than M, the similarity threshold is lowered until M nodes are obtained.

5. According to the intelligent risk control system based on a large model as described in claim 1, It is characterized in that The risk prediction model includes: M parallel computing layers, a first extraction layer, a hidden layer, a second extraction layer and a first classifier; M parallel computing layers share weight parameters and bias parameters; Each parallel computing layer includes N hidden units; The i-th hidden unit inputs the i-th sequence unit of the feature sequence of a node of the second graph structure data, and outputs the first updated feature, where 1≤i≤N; The first extraction layer is used to extract the first updated feature output by the Nth hidden unit of each parallel computing layer, and convert the feature sequence of each node of the second graph structure data into the first updated feature representation; The hidden layer inputs the second graph structure data and outputs the graph update matrix; A row vector of the graph update matrix corresponds to the second updated feature of a node; The second updated feature of each node is sequentially input into the first classifier, and the classification space of the first classifier represents the risk score corresponding to each node.

6. According to the intelligent risk control system based on a large model as described in claim 5, It is characterized in that The calculation formula for the i-th hidden unit of the parallel computing layer includes: in represents the first updated feature output by the i-th hidden unit, X i represents the i-th sequence unit of the feature sequence input to the i-th hidden unit, and denote the forward reset gate, forward update gate, forward first hidden state, and forward second hidden state of the i-th hidden unit, respectively. represents the forward second hidden state of the i-1th hidden unit, Assign a value of 0, and denote the first weight parameter, the second weight parameter and the bias parameter of the forward reset gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the forward update gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the forward first hidden state of the i-th hidden unit, respectively. and They represent the backward reset gate, backward update gate, backward first hidden state and backward second hidden state of the i-th hidden unit respectively. represents the backward second hidden state of the i-1th hidden unit, Assign a value of 0, and They represent the first weight parameter, the second weight parameter and the bias parameter of the backward reset gate of the i-th hidden unit, respectively. and They represent the first weight parameter, the second weight parameter and the bias parameter of the backward update gate of the i-th hidden unit, respectively. and They respectively represent the first weight parameter, the second weight parameter and the bias parameter of the backward first hidden state of the i-th hidden unit, ⊙ represents the Hadamard product, concat represents the concatenation operation, and sigmoid represents the sigmoid activation function.

7. According to the intelligent risk control system based on a large model as described in claim 5, It is characterized in that The calculation formula for the hidden layer includes: Where 1≤u≤M, M represents the number of nodes in the second graph structure data, S represents the graph update matrix output by the hidden layer, and N u represents the set of nodes that have edges connected to the u-th node, represents the second updated feature of the u-th node, h u and h v Represents the first updated feature of the u-th node and the v-th node, Nattion uv represents the attention score between the u-th node and the v-th node, D uv represents the number of edges between the u-th node and the v-th node, and Respectively represent the first weight parameter, the second weight parameter and the third weight parameter between the uth node and the vth node, W u represents the weight parameter corresponding to the u-th node, b uv represents the bias parameter between the u-th node and the v-th node, pile represents the stacking operation, T represents the transposition operation, and sigmoid represents the sigmoid activation function.

8. According to the intelligent risk control system based on a large model as described in claim 5, It is characterized in that The parallel computing layer is pre-trained before the risk prediction model is trained. During the pre-training of the parallel computing layer, the first updated feature output by the Nth hidden unit is input into the second classifier and the third classifier. The classification space of the second classifier represents the trading volume of the next trading day, and the classification space of the third classifier represents the trading amount of the next trading day.

9. According to the intelligent risk control system based on a large model as described in claim 1, It is characterized in that The sample labels of the training samples used to train the risk prediction model are obtained by taking the average of the results simultaneously labeled by multiple experts.

10. According to claim 1, a smart risk control system based on a large model, It is characterized in that The risk scores corresponding to each node of the second graph structure data are sorted in ascending order, and the digital products mapped by the data of the first K nodes with the smallest risk scores are obtained as an investment recommendation list, where K is a custom parameter and 1≤K≤M.