Market risk cross infection prevention and control method based on SIRS and Q-learning collaborative optimization

Abstract

The invention discloses a market risk cross infection prevention and control method based on SIRS and Q-learning collaborative optimization, and the method comprises the steps: building a subject network and the weight of the subject network through a capital market subject association knowledge graph, building a risk cross infection SIRS model, enabling the node states to be susceptible, infected and recovered, and enabling the core parameters alpha, beta and gamma to be dynamically and adaptively adjusted through multi-agent Q-learning. And further identifying a risk infection path in real time, screening key nodes based on three-degree influence, and realizing rapid intervention in combination with breadth-first search and pruning. The scheme comprises the following steps: S1, constructing a knowledge graph; s2, constructing a network model; s3, constructing an SIRS infection model; and S4, learning an optimal strategy to dynamically optimize parameters. According to the method, a four-dimensional dynamic prevention and control system of a knowledge graph layer, an SIRS layer, a Q-learning layer and a key node layer is formed, an intervention window is effectively expanded, and system stability and risk coverage efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of market risk cross-contagion prevention and control method based on SIRS and Q-learning synergistic optimization. Background Technology

[0002] Analysis of pain points in existing technologies

[0003] (1) Technical bottlenecks in the construction of financial knowledge graphs

[0004] Data heterogeneity challenge: Traditional methods use general word segmentation (such as HanLP and jieba) to process financial text, but in scenarios such as institutional abbreviations and product nesting, the entity recognition accuracy is less than 68%;

[0005] Limitations of dynamic relationship modeling: Existing D2R / R2RML frameworks cannot handle implicit associations in unstructured data;

[0006] (2) Limitations of the risk contagion model

[0007] Static parameter limitations: The internationally accepted Battiston SIRS model uses a fixed β value, which cannot adapt to the policy interventions unique to China's capital market;

[0008] The problem of missing dimensions: Existing studies only consider direct relationships between institutions, ignoring the three-dimensional indirect influences;

[0009] (3) Lagging optimization of prevention and control strategies

[0010] The drawback of traditional PID control is that the adjustment delay of the recovery rate α is too long, causing the optimal intervention window to be missed.

[0011] Limitations of single-dimensional optimization: Coordinated adjustment of the three parameters α / β / γ was not achieved;

[0012] (4) Key node identification deviation

[0013] Centrality indicator distortion: The traditional PageRank algorithm suffers from the "degree centrality trap" in financial networks;

[0014] Lack of dynamic response: Existing studies use static thresholds, which cannot adapt to real-time risk evolution. Summary of the Invention

[0015] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0016] In view of the problems existing in the current market risk cross-infection prevention and control methods, this invention is proposed.

[0017] Therefore, the technical problem solved by this invention is to address the following issues in existing market risk cross-contagion prevention and control methods: insufficient dynamism and accuracy in the construction of financial knowledge graphs, poor adaptability due to static risk contagion model parameters, lag and uniformity in risk prevention and control strategy optimization, and insufficient accuracy in identifying key nodes and lack of dynamic response.

[0018] To address the aforementioned technical problems, this invention provides the following technical solution: a market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization, comprising the following steps: S1: constructing a knowledge graph of capital market entity relationships; S2: constructing a capital market entity network structure model based on the knowledge graph of capital market entity relationships; S3: constructing a capital market risk cross-contagion model based on SIRS infectious disease dynamics; S4: dynamically optimizing the key parameters of the SIRS model by adjusting risk prevention and control strategies through learning the optimal solution.

[0019] As a preferred embodiment of the market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization described in this invention, the construction of the capital market entity association knowledge graph in S1 specifically includes the following steps: S11: Obtain capital market entity data; S12: Use recognition technology to convert the data into text, and then into knowledge; S13: For structured data stored in relational databases, directly extract entities and relationships through D2R conversion; for text generated from unstructured databases, generate custom word segmentation based on the professional corpus of capital market entities, and segment the text; S14: Use CRF and bidirectional LSTM for entity recognition, and extract relationships based on SDP and LSTM to form preliminary knowledge; S15: Perform disambiguation, synonym substitution, and coreference resolution on entities; S16: After quality assessment, form correct knowledge and store it in the knowledge graph to generate the capital market entity association knowledge graph.

[0020] As a preferred embodiment of the market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization described in this invention, in step S2, it is assumed that there are N nodes in the capital market, and the subject association structure is described by a directed graph G=(V,E), where V represents the set of nodes and E represents the edge.

[0021] Therefore, the constructed capital market entity network structure model is expressed as a... A square matrix of order A:

[0022]

[0023] Among them, Aij This indicates whether there is a connection and the strength of the connection between any two capital market entities i and j; 0 indicates that nodes i and j are not related; the weight W i，j (0 < W) i，j <1) represents the weight of any associated edge <i,j>∈E in a directed graph, that is, the association strength between node i and node j.

[0024] As a preferred embodiment of the market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization described in this invention, the constructed capital market risk cross-contagion model is specifically as follows:

[0025]

[0026] The state space of the capital market entity network can be defined as a combination of states for each node:

[0027]

[0028] Among them, S i (t), I i (t), R i (t) represents the probability that node i is in a susceptible, infected, or recovered state at time t, respectively; β is the recovery rate, representing the probability that a node recovers from an infected state to a recovered state; β is the infection rate, representing the probability of risk transmission between nodes; and γ is the reinfection rate, representing the probability that a node reverts from a recovered state to a susceptible state.

[0029] As a preferred embodiment of the market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization described in this invention, the specific steps of dynamically optimizing the key parameters of the SIRS model by adjusting the risk prevention and control strategy through learning the optimal solution include the following:

[0030] H1: Create table Q table Let c*a be a two-dimensional matrix. Then each element Q(i,j) represents the expected reward Q value of the agent taking action j in state i. Initially, Q(i,j) = 0 for all c and a.

[0031] but:

[0032]

[0033] H2: Updated according to the following formula Updates and adjustments for β and γ:

[0034]

[0035] H3: According to the Bellman equation, the Q table is updated as follows:

[0036]

[0037] Where Q(c) t a t () indicates that at time t, in state c t Take action a t The Q-value; r1 represents the learning rate; r2 represents the discount factor; r t+1 Indicates the execution of action a t The immediate reward that the agent can obtain is calculated using the following formula; c t+1 and a t+1 This represents the new state and new action that follows the current action; max a Q(c) t+1 a t+1 ) indicates the next state c t+1 The maximum Q value among all possible actions taken;

[0038] in:

[0039] in, To perform action a t The number of infected individuals after I sum To perform action a t The number of infected individuals previously;

[0040] H4: Real-time detection of risk transmission paths, calculation and discovery of key nodes in the network;

[0041] H5: Based on the capital market entity network association model, and following the three-dimensional influence principle, a breadth-first traversal is used to search for risk contagion paths and calculate the influence on each node.

[0042] in:

[0043]

[0044] Among them, a certain node When credit risk occurs, its impact on itself... =1; For single-path continuous transmission nodes For nodes The impact generated;

[0045] in:

[0046]

[0047] in, for Any direct predecessor node, The source of credit risk For nodes The influence For multi-path composite transmission nodes For nodes The impact it generates.

[0048] As a preferred embodiment of the market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization described in this invention, in H4, the most influential key node is selected by using a web crawler combined with breadth-first search and pruning strategies.

[0049] This invention provides a market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization, which has the following beneficial effects:

[0050] 1. Four-dimensional dynamic prevention and control system

[0051] Knowledge graph layer: Enables minute-level updates of institutions, products, and transactions;

[0052] SIRS model layer: coordinated regulation of α / β / γ three parameters;

[0053] Q-learning layer: Multi-agent policy optimization (41% efficiency improvement compared to single agent).

[0054] Key Node Layer: Three-Dimensional Influence Identification (Covering 98.3% of Risk Transmission Paths);

[0055] 2. Improved risk management efficiency

[0056] Extended intervention window: By adjusting dynamic parameters, the effective intervention time window is extended from 3 hours in the traditional model to 7.2 hours;

[0057] System stability is enhanced: the depth of the risk propagation tree is reduced from 3.5 layers to 1.2 layers (mean). Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0059] Figure 1 The overall flowchart of the market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization provided by the present invention is shown.

[0060] Figure 2 This is a schematic diagram illustrating the changes in the state of network nodes at different times, starting from the initial time T0, for a capital market entity network provided by the present invention.

[0061] Figure 3 The infection curve provided by this invention is an example of an infection curve without optimized prevention and control strategies.

[0062] Figure 4 The parameter value variation curve provided for this invention.

[0063] Figure 5 Infection curves under the optimized risk control strategy provided by this invention. Detailed Implementation

[0064] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0065] Analysis of pain points in existing technologies:

[0066] (1) Technical bottlenecks in the construction of financial knowledge graphs

[0067] Data heterogeneity challenge: Traditional methods use general word segmentation (such as HanLP and jieba) to process financial text, but in scenarios such as institutional abbreviations and product nesting, the entity recognition accuracy is less than 68%;

[0068] Limitations of dynamic relationship modeling: Existing D2R / R2RML frameworks cannot handle implicit associations in unstructured data;

[0069] (2) Limitations of the risk contagion model

[0070] Static parameter limitations: The internationally accepted Battiston SIRS model uses a fixed β value, which cannot adapt to the policy interventions unique to China's capital market;

[0071] The problem of missing dimensions: Existing studies only consider direct relationships between institutions, ignoring the three-dimensional indirect influences;

[0072] (3) Lagging optimization of prevention and control strategies

[0073] The drawback of traditional PID control is that the adjustment delay of the recovery rate α is too long, causing the optimal intervention window to be missed.

[0074] Limitations of single-dimensional optimization: Coordinated adjustment of the three parameters α / β / γ was not achieved;

[0075] (4) Key node identification deviation

[0076] Centrality indicator distortion: The traditional PageRank algorithm suffers from the "degree centrality trap" in financial networks;

[0077] Lack of dynamic response: Existing studies use static thresholds, which cannot adapt to real-time risk evolution.

[0078] Therefore, please refer to Figure 1 This invention provides a market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization, comprising the following steps:

[0079] S1: Construct a knowledge graph of the relationships between capital market entities;

[0080] S2: Constructing a network structure model of capital market entities based on a knowledge graph of capital market entity relationships;

[0081] S3: Construct a capital market risk cross-contagion model based on SIRS infectious disease dynamics;

[0082] S4: Adjust risk control strategies by learning the optimal solution and dynamically optimize the key parameters of the SIRS model.

[0083] Furthermore, S1 constructs a knowledge graph of capital market entity relationships, which involves preprocessing the acquired capital market entity data, identifying entities and extracting relationships, aligning entities, and generating the knowledge graph. Specifically, this includes the following steps:

[0084] S11: Obtain data on capital market entities;

[0085] It should be noted that the data on capital market entities obtained were mainly acquired through publicly available databases and web crawlers, which will not be elaborated upon here.

[0086] S12: Use recognition technology to convert data into text, and then into knowledge;

[0087] S13: For structured data stored in relational databases, entities and relationships are extracted directly through D2R transformation; for text generated from unstructured databases, custom word segmentation is generated based on the professional corpus of capital market entities, and the text is segmented.

[0088] It should be noted that data extracted from public databases is considered structured data; while unstructured databases are mainly used to store data that is not suitable for neat representation using the rows and columns of traditional two-dimensional tables, such as content crawled from the web.

[0089] S14: Entity recognition is performed using CRF and bidirectional LSTM, and relationship extraction is performed based on SDP and LSTM to form preliminary knowledge;

[0090] S15: Perform disambiguation, synonym substitution, and coreference resolution on entity relationships;

[0091] S16: After conducting a quality assessment, the correct knowledge is formed and stored in the knowledge graph to generate a knowledge graph of the relationships between capital market entities.

[0092] Furthermore, in step S2, let the capital market have N nodes (in the form of a graph, each node in the graph represents an entity, such as Company A). The entity relationship structure is described by a directed graph G=(V,E), where V represents the set of nodes (capital market entities), E represents the edge (relationship), V is the edge of the graph, and N is the node in the graph.

[0093] Therefore, the constructed capital market entity network structure model is expressed as a... A square matrix of order A:

[0094]

[0095] Among them, A ij This indicates whether there is a connection and the strength of the connection between any two capital market entities i and j; 0 indicates that nodes i and j are not related; the weight W i，j (0 < W) i，j <1) represents the weight of any associated edge <i,j>∈E in the directed graph, that is, the association strength between node i and node j. This value is obtained through deep learning of the relationships between capital market entities.

[0096] Specifically, with increased inter-entity linkages among capital market participants, under systemic risk shocks, risk contagion within the market exhibits a cross-spreading trend. Existing research indicates that its risk contagion characteristics are similar to those of infectious disease models. The SIRS model, a classic model in infectious disease dynamics, describes the dynamic changes among susceptible individuals, infected individuals, and recovered individuals, and is more consistent with the risk contagion pattern in the capital market. In a capital market participant network with N nodes, each node can be in one of the following three states:

[0097] S (Susceptible): The node has not yet been infected by the risk, but may be infected.

[0098] I (Infected): Infected state, the node has been infected and may infect other nodes.

[0099] R (Recovered): The node has recovered from the risk and will not be infected again for the time being.

[0100] Therefore, the state space of the capital market entity network can be defined as the state combination of each node:

[0101]

[0102] The constructed capital market risk cross-contagion model is as follows:

[0103]

[0104] Among them, S i (t), I i (t), R i (t) represents the probability that node i is in a susceptible, infected, or recovered state at time t, respectively; β is the recovery rate, representing the probability that a node recovers from an infected state to a recovered state; β is the infection rate, representing the probability of risk transmission between nodes; and γ is the reinfection rate, representing the probability that a node reverts from a recovered state to a susceptible state.

[0105] It should be noted that the three parameters α, β, and γ are inherent parameters in the SIRS infectious disease model, and their values ​​are all real numbers between 0 and 1.

[0106] It should be noted that in the interconnected network of entities in the capital market, nodes influence each other through the network structure, and risks propagate across nodes. The SIRS model can be used to describe the transitions in node states between susceptible, infected, and recovered individuals. Figure 2 This is a network of capital market participants, illustrating the contagion of risk among them over time. Specifically, it shows the risk contagion at times T0, T1, T2, T3, T4, and T5. Starting from the initial moment... The diagram illustrates the changing states of network nodes at different times, with red nodes representing infected individuals, blue nodes representing susceptible individuals, and green nodes representing recovered individuals. When systemic risks occur, different prevention and control strategies have a significant impact on the scope and duration of network risks for capital market participants.

[0107] Furthermore, by learning the optimal solution (the currently found optimal or near-optimal solution) to adjust risk control strategies, the key parameters of the SIRS model (recovery rate α, infection rate β, reinfection rate γ) are dynamically optimized to adapt to dynamic changes in complex networks, control risk propagation, and enhance system stability. Specifically, this includes the following steps:

[0108] H1: Create Q tableLet Q(c, a) be a two-dimensional matrix of c*a, representing that the agent has c states and can choose a actions. Then each element Q(i,j) represents the expected reward Q value of the agent taking action j in state i (state refers to the agent's current solution or value; while action refers to the steps or optimization directions that the agent may take next). When initialized, the expected reward value of each element is 0, so Q(c, a) is set as a zero matrix, as shown in formula (7).

[0109] but:

[0110]

[0111] Regarding the optimization of risk prevention and control strategies for capital market entities, state i represents the current state of a node in the entity network, i.e., the degree of susceptibility, infection, or recovery of the node, i.e., the current values ​​of α, β, and γ, which are standardized to... There are several states. The action space a={0,1,2,3,4,5} represents the risk control strategy, namely, adjusting the infection rate β by strengthening supervision or isolating high-risk nodes, adjusting the recovery rate α by providing liquidity support or relief measures, and adjusting the reinfection rate γ by optimizing policies to prevent the risk from recurring.

[0112] H2: Updated according to the following formula Updates and adjustments for β and γ:

[0113]

[0114] Taking formula (8) as an example, the variables are implemented. The cycle is performed with increments of 0.015 between 0.1 and 0.5. This represents the i-th action (i=0,1,2,...5). Loop strategy: If... If it is positive, then The value increases by 0.02, and vice versa. The value decreased by 0.02.

[0115] H3: According to the Bellman equation, the Q table is updated as follows:

[0116]

[0117] Where Q(c) t a t () indicates that at time t, in state c t Take action a t The Q-value; r1 represents the learning rate (controlling the weights for each update), ranging from 0 to 1; r2 represents the discount factor (controlling the weights of future rewards, indicating the degree of importance attached to future rewards), ranging from 0 to 1; r t+1Indicates the execution of action a t The immediate reward that the agent can obtain is calculated using the following formula (12); c t+1 and a t+1 This represents the new state and new action that follows the current action; max a Q(c) t+1 a t+1 ) indicates the next state c t+1 The maximum Q value among all possible actions taken;

[0118] in:

[0119] in, To perform action a t The number of infected individuals after I sum To perform action a t The number of infected individuals previously;

[0120] It should be noted that by introducing Q-learning reinforcement learning, the system selects the optimal action based on the current state at each time step and evaluates the new state through a reward function. The reward function is designed as the difference in the number of infected individuals, with the goal of minimizing risk and maximizing stability.

[0121] H4: Real-time detection of risk transmission paths, calculation and discovery of key nodes in the network;

[0122] It should be noted that, as mentioned above, in the capital market entity network, when a risk occurs among the risk entities at a certain moment, the risk will rapidly spread among other entities. Real-time identification of risk transmission paths, calculation and discovery of key nodes in the network, and strengthening the supervision or isolation of key nodes are of great significance for risk prevention and control.

[0123] H5: Based on the capital market entity network correlation model, and following the three-dimensional influence principle, a breadth-first traversal search is used to find the risk contagion path and calculate the influence of each node. The specific definitions of influence are shown in formulas (13) and (14).

[0124] in:

[0125]

[0126] in, β represents the influence of node i on node j in a single-path continuous transmission; β is the infection rate; and β represents the influence of node i on itself when it experiences credit risk. =1; This represents the product of influence weights along the path from node i to node j.

[0127] in:

[0128]

[0129] in, for Any direct predecessor node; β is the infection rate; The source of credit risk For nodes Its influence; It is the edge Weighting on.

[0130] It's important to note that, according to the three-degree-of-effect principle in social networks, a node can influence not only its neighbors, but also their neighbors' neighbors, and even their neighbors' neighbors' neighbors. As long as the influence is within three degrees, it's considered a strong connection and has the potential to trigger action. Beyond three degrees, the influence between nodes weakens to the point of being negligible or even disappears. The degree centrality metric based on topological networks indicates the degree of a node's centrality in the graph; that is, the more frequently and closely a node is connected to other nodes, the higher its centralization index. Nodes closer to other nodes with higher centralization indices generally also have higher centralization indices. A node's degree centrality reflects its influence within the overall interconnected structure of the capital market.

[0131] Based on the capital market entity network association model, and following the three-dimensional influence principle, a breadth-first traversal is used to search for risk contagion paths and calculate the influence of each node. The specific definitions of influence are shown in formulas (13) and (14).

[0132] Furthermore, based on the above definition, a greedy strategy is used to select nodes that meet certain conditions as key nodes with potential influence. The basic idea is to measure the degree centrality of a node using its in-degree and out-degree, calculate the sum of the influence of each node on other nodes in the network, and identify the key nodes as those that satisfy the conditions of maximum influence and having the largest number of influential nodes. A web crawler combined with breadth-first search and pruning strategies is used to select the key nodes with the greatest influence using the `Find_Keynodes()` algorithm. The algorithm framework is as follows:

[0133] Algorithm 1: Key node search algorithm using web crawling combined with breadth-first search and pruning strategies <![CDATA[Input: G(V, E) Output: The set S of nodes with the greatest influence * > <![CDATA[1. S0 = Φ2. For each i, i ∈ V 3. Calculate its global influence factor μ]]> i = k1 * ID(i) + k2 * OD(i) 4. Endfor 5. Sort each node in descending order of μ]]> i 6. Select the top a% of the nodes and add them to S0 to obtain S]]> * . 7. For each i, i ∈ S]]> * 8. Set several crawler breadth - first search strategies according to its out - degree 9. Calculate the influence of each node j by i based on formula (13) or (14) 10. For any node j in the search path 11. If ρ]]> j > threshold θ 12. Use the pruning algorithm to delete all incoming edges of j 13. S]]> * = S]]> * ∪{j} 14. Endif 15. Endfor 16. For the newly added nodes after update, repeat steps 8 - 13 until no new nodes are added. 17. Endfor]]>

[0134] This algorithm employs an optimized prevention and control strategy that deeply integrates the Q-learning mechanism with the SIRS infectious disease dynamics model. Through dynamic perception of risk transmission paths and identification of key nodes, it strengthens in-depth monitoring of critical nodes. Its core idea is:

[0135] 1) Construction of Capital Market Entity Network: Based on the knowledge graph of capital market entity relationships, construct a capital market entity relationship network with N nodes.

[0136] 2) Risk control strategy optimization: Adjust the SIRS model parameters (α,β,γ) through the Q-learning algorithm to optimize the control strategy.

[0137] 3) Risk transmission and key node search: Based on the three-degree influence principle, risk influence is calculated. Web crawlers are used in combination with breadth-first search and pruning strategies to find the most influential key nodes in the network. Nodes with risk values ​​exceeding the threshold are identified in real time and targeted interventions are triggered.

[0138] Algorithm 2: Risk Control Strategy Optimization Algorithm Based on Q-learning Input: Capital market network size N, maximum time step Time_maxStep Output: Risk status of capital market participants in the network<s,i,r> <![CDATA[1. Initialize the capital market entity association matrix A2. Initialize the Q-table: Q(c, a) ← 0 3. Initialize the parameters: the number of susceptible individuals s, the number of infected individuals i, and the number of recovered individuals r 4. Initialize the risk recovery rate α, infection rate β, and reinfection rate γ of the capital market entity network 5. For t = 1 to Time_maxStep do 5.1 Use the ε-greedy strategy to select an action a, that is, adjust the SIRS model parameters (α, β, γ) according to formulas (8), (9), and (10) 5.2 Execute the selected prevention and control strategy, and update the SIRS model according to formulas (4), (5), and (6) 5.3 Update the Q-table according to formula (11), that is, the reward value of this prevention and control strategy 5.4 Use Find_Keynodes() to find the risk path and key nodes 5.5 Output the current network state <s, i, r> and the key node set S * 6. Endfor 7. Return the final network state <s, i, r>]]>

[0139] Example running result:

[0140] A network of capital market entities with 1000 nodes is set up, with an initial infection rate β of 0.5, an initial recovery rate α of 0.3, and an initial reinfection rate γ of 0.2. Through 300 iterations, the system can achieve the following prevention and control effects:

[0141] 1) Infection control: The number of infected nodes decreased from the initial 30 to less than 5 (a reduction of 83%).

[0142] 2) Parameter optimization:

[0143] The recovery rate α gradually increased from 0.3 to the range of 0.65-0.75;

[0144] The infection rate β decreased from 0.5 to the range of 0.2-0.25;

[0145] The reinfection rate γ remained stable below 0.1.

[0146] 3) System stability: The depth of the risk propagation tree decreased from 3 layers to 1.2 layers (mean);

[0147] See Figure 3 , Figure 3 This indicates that, assuming constant recovery, infection, and reinfection rates (i.e., without optimized prevention and control strategies), the number of susceptible individuals will rapidly decrease over time in the initial stage, while the number of infected and recovered individuals will also increase relatively quickly. Within a short period, the number of susceptible individuals will decrease to a minimum, while the number of infected and recovered individuals will reach a maximum, meaning that most nodes in the network will be infected. Subsequently, as time progresses, the numbers of susceptible, infected, and recovered individuals will all tend to stabilize.

[0148] Figure 4This indicates that by dynamically adjusting the recovery rate, infection rate, and reinfection rate using the Q-learning algorithm, in the early and mid-stages, the infection rate and reinfection rate gradually decrease, while the recovery rate gradually increases. In the mid and late stages, the recovery rate increases slightly and then tends to stabilize, while the infection rate and reinfection rate decrease slightly and eventually also tend to stabilize. The diagram showing the state changes of the main network nodes in the capital market is as follows. Figure 5 As shown.

[0149] Figure 5 This indicates that although the number of infected and recovered individuals gradually increases while the number of susceptible individuals decreases in the early stages of infection, over time, by implementing dynamically optimized prevention and control strategies, the number of infected and recovered individuals gradually decreases while the number of susceptible individuals gradually increases. This demonstrates that introducing Q-learning algorithms to optimize risk control strategies is effective for controlling network risks in the capital market.

[0150] This invention provides a market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization, which has the following beneficial effects:

[0151] 1. Four-dimensional dynamic prevention and control system

[0152] Knowledge graph layer: Enables minute-level updates of institutions, products, and transactions;

[0153] SIRS model layer: coordinated regulation of α / β / γ three parameters;

[0154] Q-learning layer: Multi-agent policy optimization (41% efficiency improvement compared to single agent).

[0155] Key Node Layer: Three-Dimensional Influence Identification (Covering 98.3% of Risk Transmission Paths);

[0156] 2. Improved risk management efficiency

[0157] Extended intervention window: By adjusting dynamic parameters, the effective intervention time window is extended from 3 hours in the traditional model to 7.2 hours;

[0158] System stability is enhanced: the depth of the risk propagation tree is reduced from 3.5 layers to 1.2 layers (mean).

[0159] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A market risk cross-contagion prevention and control method based on SIRS and Q-learning synergistic optimization, characterized in that, Includes the following steps: S1: Construct a knowledge graph of the relationships between capital market entities; S2: Construct a network structure model of capital market entities based on the aforementioned knowledge graph of capital market entity relationships; S3: Construct a capital market risk cross-contagion model based on SIRS infectious disease dynamics; S4: Adjust risk control strategies by learning the optimal solution and dynamically optimize the key parameters of the SIRS model.

2. The market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization according to claim 1, characterized in that, The construction of a knowledge graph of capital market entity relationships in S1 includes the following steps: S11: Obtain data on capital market entities; S12: Use recognition technology to convert data into text, and then into knowledge; S13: For structured data stored in relational databases, entities and relationships are extracted directly through D2R transformation; for text generated from unstructured databases, custom word segmentation is generated based on the professional corpus of capital market entities, and the text is segmented. S14: Entity recognition is performed using CRF and bidirectional LSTM, and relationship extraction is performed based on SDP and LSTM to form preliminary knowledge; S15: Perform disambiguation, synonym substitution, and coreference resolution on entity relationships; S16: After conducting a quality assessment, the correct knowledge is formed and stored in the knowledge graph to generate a knowledge graph of the relationships between capital market entities.

3. The market risk cross-contagion prevention and control method based on SIRS and Q-learning synergistic optimization according to claim 2, characterized in that: In step S2, the capital market has N nodes. The main entity relationship structure is described by a directed graph G=(V,E), where V represents the set of nodes and E represents the edge. Therefore, the constructed capital market entity network structure model is expressed as a... A square matrix of order A: ； Among them, A ij This indicates whether there is a connection and the strength of the connection between any two capital market entities i and j; 0 indicates that nodes i and j are not related; the weight W i，j (0 < W) i，j <1) represents the weight of any associated edge <i,j>∈E in a directed graph, that is, the association strength between node i and node j.

4. The market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization according to claim 3, characterized in that, The constructed capital market risk cross-contagion model is specifically as follows: The state space of the capital market entity network can be defined as a combination of states for each node: Among them, S i (t), I i (t), R i (t) represents the probability that node i is in a susceptible, infected, or recovered state at time t, respectively; β is the recovery rate, representing the probability that a node recovers from an infected state to a recovered state; β is the infection rate, representing the probability of risk transmission between nodes; and γ is the reinfection rate, representing the probability that a node reverts from a recovered state to a susceptible state.

5. The market risk cross-contagion prevention and control method based on SIRS and Q-learning synergistic optimization according to claim 4, characterized in that, By learning the optimal solution and adjusting risk control strategies, the key parameters of the SIRS model are dynamically optimized through the following steps: H1: Create table Q table Let c*a be a two-dimensional matrix. Then each element Q(i,j) represents the expected reward Q value of the agent taking action j in state i. Initially, Q(i,j) = 0 for all c and a. but: H2: Updated according to the following formula Updates and adjustments for β and γ: H3: According to the Bellman equation, the Q table is updated as follows: Where Q(c) t a t () indicates that at time t, in state c t Take action a t The Q-value; r1 represents the learning rate; r2 represents the discount factor; r t+1 Indicates the execution of action a t The immediate reward that the agent can obtain is calculated using the following formula; c t+1 and a t+1 This represents the new state and new action that follows the current action; max a Q(c) t+1 a t+1 ) indicates the next state c t+1 The maximum Q value among all possible actions taken; in: in, To perform action a t The number of infected individuals after I sum To perform action a t The number of infected individuals previously; H4: Real-time detection of risk transmission paths, calculation and discovery of key nodes in the network; H5: Based on the capital market entity network association model, and following the three-dimensional influence principle, a breadth-first traversal is used to search for risk contagion paths and calculate the influence on each node. in: Among them, a certain node When credit risk occurs, its impact on itself... =1; For single-path continuous transmission nodes For nodes The impact generated; in: in, for Any direct predecessor node, The source of credit risk For nodes The influence For multi-path composite transmission nodes For nodes The impact it generates.

6. The market risk cross-contagion prevention and control method based on SIRS and Q-learning collaborative optimization according to claim 5, characterized in that: In H4, a crawler combined with breadth-first search and pruning strategies is used to select the most influential key nodes.

Images