Investment risk decision support method and system based on evidence reasoning and deep learning

Abstract

The invention discloses an investment risk decision support method and system based on evidence reasoning and deep learning, and the method comprises the steps: obtaining investment risk data, and determining a risk tag; constructing a Widamp based on evidence reasoning; the Deep risk early warning model comprises a Wide component, a Deep component, a full connection layer and an output layer; the Wide component comprises a plurality of basic precaution devices, investment risk data output by each basic precaution device is classified into confidence of each risk label, and basic probability distribution is calculated by combining the weight of each basic precaution device; combining the basic probability distribution by using an evidence reasoning combination rule to obtain a first classification probability that the investment risk data is classified into each risk tag; the Deep component comprises a deep neural network model, and outputs a second classification probability that the investment risk data is classified into each risk tag; the two classification probabilities are connected with the output layer through the full connection layer, and the final classification probability is output. According to the method, investment risk prediction can be carried out, and the result has interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of investment decision-making, and in particular to an investment risk decision support method and system based on evidence reasoning and deep learning. Background Art

[0002] Artificial intelligence and machine learning technologies play an important role in investment decisions. They help process large amounts of multi-dimensional information about market conditions, company fundamentals, financial statements, etc., discover hidden patterns and trends, and help investors more accurately predict key indicators such as market trends, stock prices, and corporate performance. They enhance decision makers' data analysis and precise prediction capabilities, and provide strong support for making correct investment decisions.

[0003] However, current machine learning models have defects in decision-making applications. The existing artificial intelligence models and their scope of application, model characteristics, advantages and disadvantages are shown in Table 1. In addition, the black box problem is a challenge faced by many advanced machine learning models (such as deep learning models) in their applications. In some industries (such as finance, medical care, etc.), decision makers and regulators require that model decisions must be explainable and traceable. Due to their complexity and black box nature, these models are difficult to meet the interpretability requirements.

[0004] Table 1 Related existing technology status

[0005]

[0006] Summary of the invention

[0007] The purpose of the present invention is to provide an investment risk decision support method and system based on evidential reasoning and deep learning, which can ensure the high accuracy of risk prediction results while making the risk prediction results explainable.

[0008] To achieve the above object, according to a first aspect of the present invention, a method for supporting investment risk decision-making based on evidence reasoning and deep learning is provided, the method comprising:

[0009] Obtain investment risk data and determine risk labels;

[0010] Construct a Wide & Deep risk warning model based on evidence reasoning, including Wide component, Deep component, fully connected layer and output layer;

[0011] The Wide component includes multiple base early warning devices, each of which outputs the confidence level of the investment risk data being classified into each risk label. Combined with the weight of each base early warning device, the basic probability distribution is calculated. Then, the basic probability distribution is combined using the evidence reasoning combination rule to obtain the first classification probability of the investment risk data being classified into each risk label.

[0012] The Deep component includes a deep neural network model, which outputs the second classification probability of the investment risk data being classified into various risk labels;

[0013] The fully connected layer is fully connected with the first classification probability of the Wide component and the second classification probability of the Deep component;

[0014] The output layer is connected to the fully connected layer, and outputs the final classification probability of the investment risk data being classified into various risk labels;

[0015] Using investment risk data and corresponding risk labels, we train a Wide & Deep risk warning model based on evidence reasoning, and use the trained Wide & Deep risk warning model based on evidence reasoning to predict investment risks.

[0016] In the above scheme, the basic probability distribution is calculated, including:

[0017] Assume that the identification frame Θ = {H1, H2...H n ...H N}; where Hn is the confidence rating result and N is the total number of risk labels;

[0018] Taking investment risk data as evidence, then:

[0019] S(e i )={(H n , β n，i )|n=1, 2, ..., N; i=1, 2, ..., L}

[0020] In the formula, e i is the i-th evidence; β n,i For evidence i Rated H n The confidence level is the confidence level that the i-th investment risk data is classified as the n-th risk label by the base early warning device; L is the total number of evidences;

[0021] Calculate the basic probability distribution as follows:

[0022] m n,i =w i β n,i

[0023]

[0024] In the formula, m n,i Rated H n Evidence i The basic probability distribution of m H,i Not rated H n Evidence i The basic probability distribution of w i is the weight of the base early warning device and ∑w i =1.

[0025] In the above scheme, the weight of the base early warning device is calculated as follows:

[0026]

[0027] In the formula, is the weight of the k-th base early warning device, accuracy k is the accuracy of the kth base early warning device, and K is the total number of base early warning devices.

[0028] In the above scheme, the basic probability distribution is combined by using the evidence reasoning combination rule to obtain the first classification probability of the investment risk data being classified into various risk labels, including:

[0029] M H,i It is divided into two parts:

[0030]

[0031] Assume m n For evidence i and evidence j The ER rule is as follows:

[0032]

[0033] {H n}:m n =K[m n,i m n,j +m H,i m n,j +m n,i m H,j ]

[0034]

[0035] In the formula, K represents the degree of conflict between the evidences;

[0036] Calculate the first classification probability of the investment risk data being classified into each risk label:

[0037]

[0038] In the formula, β n is the first classification probability that the investment risk data is classified as the nth risk label.

[0039] In the above scheme, the base early warning devices include LR, Adaboost and SVM.

[0040] In the above scheme, the deep neural network model is a DNN deep neural network.

[0041] In the above scheme, the investment risk data are predetermined indicator data related to investment risks, and the risk labels include no risk warning and risk warning.

[0042] According to a second aspect of the present invention, there is provided an investment risk decision support system based on evidence reasoning and deep learning for implementing any one of the above-mentioned investment risk decision support methods based on evidence reasoning and deep learning, the system comprising: a database subsystem and a model library subsystem;

[0043] The database subsystem is used to store and manage investment risk data and its corresponding risk labels;

[0044] The model library subsystem is used to store Wide & Deep risk warning models based on evidence reasoning;

[0045] The system also has a human-computer dialogue subsystem, which includes a human-computer interaction interface; the human-computer dialogue subsystem obtains investment risk data to be decided through the human-computer interaction interface, and processes it into computer language and transmits it to the database subsystem; the database subsystem extracts risk decision data and hands it over to the model library subsystem for investment risk prediction, and finally displays the results through the human-computer interaction interface of the human-computer dialogue subsystem.

[0046] In the above scheme, the database subsystem includes a database and a database management system;

[0047] The database is used to store investment risk data and its corresponding risk labels, and to achieve effective data access through ETL data development and task engines. After data access, the data is cleaned and converted to ensure data quality and consistency.

[0048] The database management system is used to complete the operation, maintenance and management functions of data extraction, realize the creation of data objects, query, load, add, backup, modify, delete operations of stored data, and user management and permission management of data.

[0049] According to a third aspect of the present invention, a terminal is provided, which adopts the investment risk decision support method based on evidence reasoning and deep learning as described in any one of the above.

[0050] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0051] The present invention proposes an investment risk decision support method and system based on evidence reasoning and deep learning, and realizes investment risk prediction through a Wide & Deep risk warning model based on evidence reasoning. The Wide & Deep risk warning model based on evidence reasoning includes a Wide component, a Deep component, a fully connected layer and an output layer.

[0052] Among them, the Wide component adopts a multi-warning detector integration model based on evidence reasoning, which integrates three different machine learning algorithms: LR, Adaboost and SVM. After the data is input into the model, the three basic warning detectors will output their respective warning results; then the ER combination rule will be used to infer and fuse the warning results of the three basic warning detectors, and the Wide part warning results presented in the form of confidence distribution will be output.

[0053] The Deep component uses deep neural network (DNN) technology to capture complex correlation patterns in data. After the data is input, the Deep part undergoes multiple iterative learning processes and outputs corresponding risk warning results. Compared with traditional shallow neural network models, it shows excellent performance in feature learning and complex function fitting, can deeply analyze the internal laws of data, and has achieved good application results in various tasks such as regression, classification, and clustering.

[0054] Finally, the Wide & Deep model deeply integrates the risk warning results generated by the Wide part and the Deep part through a fully connected method. In the continuous self-learning and iterative process, the overall model continuously adjusts various parameters to achieve a balance between memory and generalization capabilities within a model framework, and finally outputs the best warning results. This process is essentially completed by jointly training a linear model unit and a neural network unit.

[0055] Therefore, investment risk prediction can be carried out through investment risk data, and the process of rule reasoning based on the Wide & Deep risk warning model of evidence reasoning can be completely realized through a transparent and explainable rule interpretation program, avoiding the uncertainty of the "black box" and providing decision makers with a reliable and explainable decision-making basis for the processing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the framework of the investment risk decision support system;

[0057] Figure 2 Schematic diagram of the logical structure of data extraction;

[0058] Figure 3 This is the overall architecture diagram of the Wide&Deep risk warning model based on evidence reasoning;

[0059] Figure 4 This is the Wide component architecture diagram;

[0060] Figure 5 This is a diagram showing the distribution of extreme cash flow risks for real estate companies. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0062] The present invention provides an investment risk decision support system based on intelligent reasoning, specifically an investment risk intelligent decision support system based on evidence reasoning and deep learning technology. The system specifically relates to a Wide and Deep decision model based on evidence reasoning, which ensures the high accuracy of risk prediction results while making the risk prediction results interpretable. The model intends to adopt a hybrid reasoning strategy, which can be repeatedly reasoned. As long as one of the investment risk factors and the conclusion or fact of the decision result is known, two-way reasoning can be performed. The model performs evidence reasoning and fusion on the results of multiple classifiers, and organically combines with deep learning to output decision results, which is applicable to decisions with incomplete, uncertain and high accuracy information. The process of the model performing rule reasoning can be completely realized by a transparent and interpretable rule interpretation program, that is, according to the conditional part of the input target rule, the decision rules matching it in the investment risk intelligent decision support system database are activated and extracted, and the corresponding operations are performed on the matched activation rules, avoiding the uncertainty of the "black box", and providing decision makers with a reliable and interpretable decision basis for the processing process.

[0063] Among them, Dempster-Shafer evidential reasoning (DS theory) is a method for dealing with uncertainty and fuzzy reasoning, which aims to extract evidence from multiple uncertain sources and make reasonable synthesis and reasoning. The core of this theory lies in its flexibility. Unlike traditional probability theory, it does not rely on precise probability distribution, but describes the influence of different sources of evidence through measures such as "belief" and "doubt", and provides a systematic framework to deal with incomplete or conflicting information. Its interpretability is not only reflected in how to deal with the transparency of information, but also in how to integrate conflicting evidence, express ambiguity, and help decision makers make more confident decisions. The following three points make the results of evidential reasoning interpretable.

[0064] 1. Clarity of trust and suspicion

[0065] In DS theory, trust and suspicion are key metrics for dealing with uncertain information. Trust represents the degree of support for a proposition, reflecting the strong support of the evidence for a hypothesis. Suspicion, on the other hand, represents the degree to which a proposition lacks evidence support, expressing opposition to the proposition or doubt about its authenticity. Through these two concepts, DS theory enables the decision-making process to not only provide a clear assessment of "support for a proposition", but also clearly reveal areas of "uncertain evidence".

[0066] The advantage of this pair of concepts is that they can handle incomplete information and ambiguous interpretations of evidence. For example, suppose you are doing a product quality assessment and rely on evidence from different sources to determine whether the product meets the quality standards. Under the DS theory framework, not only will the confidence that the quality meets the standards be calculated, but the suspicion that the quality standards are not met will also be reflected, thereby fully demonstrating the uncertainty of the information. Compared with traditional probability assessments, DS theory can clearly indicate "how much we trust the current information" and "how uncertain we are about other possibilities", which provides decision makers with more understanding and transparency.

[0067] Real-world example: Suppose a pharmaceutical company is evaluating a marketing decision for a new drug. The company has collected three main pieces of evidence:

[0068] Evidence 1 (market research): Consumers respond well to new drugs, with a trust rating of 0.6 and a suspicion rating of 0.3;

[0069] Evidence 2 (clinical trials): The drug performed well in clinical trials, with a confidence level of 0.8 and a suspicion level of 0.15;

[0070] Evidence 3 (Competitor Products): The sales data of similar drugs performed relatively weakly, with a trust level of 0.5 and a suspicion level of 0.4.

[0071] In traditional probabilistic reasoning methods, only a weighted average method may be used, but DS theory can help decision makers understand the impact of each piece of evidence and how they interact by combining the trust and suspicion of different evidence. For example, market research shows that trust is low but suspicion is high, which may indicate that market feedback is not stable, while clinical trials provide strong support, indicating that new drugs have higher credibility in medical effects. The integration of this information can provide pharmaceutical companies with a more comprehensive and clear decision support framework to help companies better identify potential market risks.

[0072] 2. Rules for integrating different pieces of evidence

[0073] A notable feature of the DS theory is its evidence fusion mechanism, which is to synthesize the trust from multiple evidence sources through the Dempster combination rule. The Dempster combination rule is different from the weighted average in probability theory. It does not simply calculate the linear sum of evidence, but adjusts the trust according to the conflict between evidence to ensure that the combined result can integrate the influence of multiple evidences and minimize the deviation caused by a single evidence.

[0074] This is of great value in decision support systems. Many real-world decisions are not based on a single piece of evidence, especially in complex environments where various sources of evidence are often conflicting. Dempster's combination rule can integrate these conflicting evidence into a more balanced result, thereby helping decision makers make more appropriate decisions in situations of asymmetric information and incomplete evidence.

[0075] Practical example: suppose we are analyzing whether a natural disaster (such as a flood) has occurred in a certain area. Evidence 1 comes from the weather forecast, which predicts that heavy rain is coming, with a confidence level of 0.7; Evidence 2 comes from the local government's real-time monitoring data, which monitors signs of flooding in some areas, with a confidence level of 0.6; Evidence 3 comes from remote sensing data from neighboring areas, which shows signs of flooding with a lower credibility, with a confidence level of 0.4.

[0076] In DS theory, the Dempster combination rule combines the confidence of these evidences and finally obtains a comprehensive evaluation. In this process, the Dempster combination rule takes into account the degree of conflict between different evidences and reduces the impact of low-confidence evidence, thereby helping us make more reasonable disaster response decisions. For example, although the confidence of remote sensing data is low, the support of other evidence is strong, and a more reliable judgment can be made by weighting the combination rule. This synthesis result helps government decision makers make more accurate disaster response decisions.

[0077] 3. Transparency in dealing with uncertainty

[0078] Compared with traditional probability methods, DS theory has higher transparency in dealing with uncertainty, allowing decision makers to clearly see the reasons behind the uncertainty of evidence. This is particularly important because in many real-world decisions, complete probability information is often unavailable. DS theory can explicitly display this uncertainty through the quantification of "trust" and "suspicion".

[0079] In this way, DS theory not only gives the probability of an event occurring, but also shows our degree of confidence in this event and our suspicion of other possible events. This way of expression can provide decision makers with more intuitive information and help them understand the inevitable uncertainty in decision-making.

[0080] Practical example: In financial risk management, risk assessment often faces incomplete information and uncertainty. For example, when a financial institution evaluates the risk of an investment project, it gets different opinions from different market analysts. DS theory can help them clearly distinguish between "the trust of different analysts" and "the degree of suspicion of project investment failure", so as to not only obtain a "risk assessment" of a project investment, but also clearly see the inconsistency and ambiguity between analysts, helping decision makers make more informed decisions in uncertainty.

[0081] The investment risk intelligent decision-making system of the embodiment of the present invention is mainly composed of two subsystems: a database and a model library. The subsystems can operate independently and are interconnected, so as to provide better and more intelligent decision support functions for investment decision makers. The data of the database comes from multi-source heterogeneous databases, and the models of the model library mainly store intelligent decision models, such as Figure 1 As shown in the figure. The investment decision maker inputs the decision information into the investment risk decision support system through the human-computer interaction interface. Taking the extreme cash flow risk warning of real estate enterprises as an example, the decision maker inputs the key risk indicators of extreme cash flow of real estate enterprises as the information for decision-making. The human-computer dialogue system processes the decision information into computer language and transmits it to the risk decision system. Then the database extracts the risk decision data and the model library performs risk data reasoning. Finally, the decision result is returned to the human-computer interaction interface to provide a visual decision solution for the investment decision maker. It can process qualitative, approximate or imprecise knowledge.

[0082] 1. Investment risk decision database

[0083] The investment risk decision database system mainly consists of two parts: the database and the database management system. The database stores a large amount of basic data and investment risk data from multi-source heterogeneous databases. The database management system can effectively organize and manage the data stored in the database. The investment-related information department inputs the real-time information of the risks faced by the enterprise into the system database, and updates the data according to the evolution and contagion of the risks. The database then screens the risk information and enters the risk data into the intelligent decision support system, which is sorted and summarized by the system to ensure that the information transmission mechanism can avoid problems such as untimely information updates, slow information transmission, and low information aggregation efficiency to a certain extent, saving time costs.

[0084] Taking the extreme cash flow risk warning task of real estate enterprises as an example, the investment department of the enterprise entered a total of 78 secondary indicators including economic environment risk sources, industry environment risk sources, and enterprise environment risk sources into the investment risk decision database, and measured the importance of each indicator of the initial extreme cash flow risk warning indicator system based on random forests, and obtained the importance coefficient of each indicator. Based on the size of the importance coefficient, the irrelevant, redundant or noisy indicators with low importance were eliminated, and 20 indicators that were highly correlated with the extreme cash flow risk of real estate were selected as the basis for the model to make inference decisions.

[0085] (1) Database

[0086] The data stored in the investment risk decision support system database is mainly case data, covering relevant internal and external data on the risks of each investment, including historical data on public opinion, politics, economy, law, competition, market, technology, etc.

[0087] The database construction in the investment risk decision support system is divided into four steps: data collection, database management, database and information output. Its logical structure is as follows: Figure 2 As shown. First of all, from the perspective of data structure, the data sources of multi-source heterogeneous databases include relational databases, non-relational databases, API interface data, file data, etc.; from the perspective of data content, the data of multi-source heterogeneous databases include quantitative indicators of macroeconomic environment, market competition, policies and regulations, and the operating conditions of the enterprise itself. Then, through flexible ETL (extraction, transformation, loading) data development and task engines, effective data access can be achieved. After accessing the data, it is usually necessary to clean and convert the data. Through data cleaning, data merging, data association and other operations, the quality and consistency of the data are ensured to facilitate subsequent analysis and processing. After the data processing is completed, it is output to the database for storage to facilitate subsequent online analysis and processing.

[0088] (2) Construction of data extraction management system

[0089] The data extraction management system (DBMS) is the core component of the database system. It mainly completes the operation, maintenance and management functions of data extraction, realizes the creation of data objects, query, load, add, backup, modify, delete and other operations of stored data, as well as user management and permission management of data.

[0090] The triggers and stored procedures built into the database management system can automatically update data. When the data in the database changes, these triggers and stored procedures will be automatically triggered and then perform the predetermined operations, such as updating, deleting or inserting data. This method does not require manual intervention by the database administrator, thus ensuring the consistency and accuracy of the data.

[0091] 2. Investment risk decision model library

[0092] The models stored in the model library of the investment risk decision support system are the Wide & Deep model (IEWD) that integrates evidence reasoning and deep neural networks and other related mathematical models. The risk factors faced by real estate companies are complex and diverse, and traditional methods are difficult to fully consider multi-dimensional risks. In particular, the dynamic change characteristics of extreme cash flow risks make it limited to make early warnings based on expert judgment alone. Therefore, the construction of a Wide & Deep intelligent early warning model for extreme cash flow risks of real estate companies that integrates evidence reasoning and deep neural networks can not only effectively integrate multi-source heterogeneous data, including economic environment, industry environment and enterprise environment risk sources, to improve the accuracy of early warnings; it can also provide explainable risk alerts, make up for the shortcomings of expert judgment, and improve the "black box" problem.

[0093] The intelligent early warning model for extreme cash flow risk of real estate enterprises incorporates multi-dimensional risk factors and indicator data such as internal and external factors of the enterprise and the characteristics of the real estate enterprise itself into the model, and comprehensively considers multi-dimensional risks. The evidence-based reasoning and Wide & Deep model (IEWD) is applied to the intelligent early warning of extreme cash flow risk of real estate enterprises, using computer intelligence to learn the risk mapping relationship of known samples, establish early warning rules and continuously iterate dynamically to realize intelligent data feature extraction. The introduction of evidence reasoning makes the early warning results output by the model highly interpretable.

[0094] The outbreak of extreme cash flow risk in real estate enterprises is the result of the combined influence of multiple factors. Its early warning work requires in-depth analysis of massive risk sample information and comprehensive deconstruction of the internal and external environmental factors of the risk. Given the complex implicit relationship between data, traditional early warning models have limitations in accommodating rich early warning indicators and handling nonlinear relationships. These deficiencies have been effectively improved by introducing machine learning technology. However, machine learning is still insufficient in processing high-dimensional data, and its nonlinear transformation mainly remains at the shallow network algorithm level. The traditional feedforward neural network limits the output capacity of the model due to the unidirectional nature of information transmission.

[0095] As research deepens, risk warning models continue to evolve, from the initial qualitative analysis to quantitative analysis models, machine learning models, and now deep learning and integrated learning models. Each model has its own unique characteristics, advantages and limitations. For example, support vector machines are suitable for small samples and strong generalization performance, but not for multi-classification problems; artificial neural networks have strong fault tolerance and learning rule experience, but are highly random and take a long time to process.

[0096] In order to meet the early warning needs of extreme cash flow risks of real estate enterprises, the constructed model must have memory, interpretability and generalization capabilities. Memory means that the model can learn high-frequency co-occurring feature combinations from historical data; interpretability requires that the model can provide a reasonable explanation for the prediction results; generalization means that the model can still make effective predictions when faced with unseen feature combinations.

[0097] The Wide&Deep model can integrate the advantages of different models to achieve a perfect combination of memory and generalization capabilities. The Wide part identifies correlations based on historical data and exhibits good memory and interpretability; while the Deep part learns new feature combinations through a deep learning network and has strong generalization capabilities. This collaborative training method enables the Wide&Deep model to achieve both memory and generalization effects within the same framework, thereby improving the accuracy and effectiveness of early warnings.

[0098] Although the existing Wide & Deep model can provide a certain degree of interpretability through the Wide component, the interpretability is limited because the Wide component relies on designed features, while the Deep component behaves as a black box. Based on the research on evidence reasoning theory and explainable artificial intelligence, this paper proposes an interpretable Wide & Deep intelligent early warning model (IEWD) based on evidence reasoning. Specifically, an evidence-based multi-warning integrated model as the Wide component and a deep learning model as the Deep component are proposed to classify and predict the extreme cash flow risks of real estate companies. As an effective method for integrating evidence from multiple sources, the evidence reasoning theory can obtain complementary information while tracking the confidence level of each piece of evidence, thereby providing more interpretability, so that the final warning result is feasible while ensuring high accuracy. The overall architecture of the model is as follows Figure 3 shown.

[0099] On the left is the Wide component of the Wide&Deep-based extreme cash flow risk intelligent early warning model (IEWD), which uses a multi-alert integration model based on evidence reasoning. This part integrates three different machine learning algorithms, LR, Adaboost and SVM. After the data is input into the model, the three basic early warning devices will output their respective early warning results; then the ER combination rule is used to infer and fuse the early warning results of the three basic early warning devices, and the Wide part of the early warning results presented in the form of confidence distribution is output.

[0100] On the right is the Deep component, which uses deep neural network (DNN) technology to capture complex correlation patterns in data. After the data is input, the Deep part undergoes multiple iterative learning processes and outputs corresponding risk warning results. Compared with traditional shallow neural network models, it shows excellent performance in feature learning and complex function fitting, can deeply analyze the internal laws of data, and has achieved good application results in various tasks such as regression, classification, and clustering.

[0101] Finally, the Wide & Deep model deeply integrates the risk warning results generated by the Wide part and the Deep part through a fully connected method. In the continuous self-learning and iterative process, the overall model continuously adjusts various parameters to achieve a balance between memory and generalization capabilities within a model framework, and finally outputs the best warning results. This process is essentially completed by jointly training a linear model unit and a neural network unit.

[0102] 2.1Wide Component

[0103] In the wide component part, an evidence-based multi-classifier ensemble technology is adopted. It includes the following key steps. First, a recognition framework (FoD) is created to cover all predetermined categories. Then the basic probability assignment (BPA) is obtained. In order to generate BPA, the present invention uses degrees of belief, which provides valuable evidence to indicate different categories and their degree of support based on a subset of sample attributes. Three different ML models are used, including Adaboost, SVM, and LR (basis early warning). Finally, ER rules are applied to BPA to improve classification accuracy by integrating information from different sources. The algorithm principle is as follows Figure 4 shown.

[0104] In order to classify different prediction results using evidential reasoning, we first consider it as a recognition framework, let Θ = {H1, H2, ..., H N}, Θ is a finite set that includes all possibilities and does not repeat, H nrepresents the nth element in Θ. This embodiment is a binary classification (no risk warning and risk warning), so N = 2. The basic probability assignment result (BPA) of a subset A of Θ is marked as m. m satisfies the following conditions:

[0105]

[0106] Where m(A) is the basic probability distribution result of A, reflecting the credibility of hypothesis A in the identification framework.

[0107] The mathematical representation is given as:

[0108] S(e i )={(H n , β n，i )|n=1,2,...,N; i=1,2,...,L}

[0109] In the formula, β n,i is the evidence output by the base early warning device i Evaluated as a certain type of H n The confidence level. This is a binary classification, so the result of the base early warning binary classification is used as evidence. For example, the result of the base early warning A: the probability of classification as H1, the probability of assignment as H2.

[0110] Among them, ei is evidence, which is the output result of the base early warning device. n,i is the confidence level that ei is classified as a certain level Hn. n is the confidence rating result. is the weight / probability assignment of the classifier / each result, accuracy k is the classification accuracy of the kth classifier.

[0111] Next, the normalized accuracy is used to estimate the weight of each classifier and obtain the probability assignment for each outcome, represents the weight of the k-th piece of evidence, expressed as:

[0112]

[0113] Among them, accuracy k Represents the accuracy of the kth classifier. On this basis, use formulas (2-1) and (2-2) to calculate BPA.

[0114] Finally, the BPAs are combined using the evidence combination rules (Formulas 2-6 to 2-9) to obtain the final classification probability of each result.

[0115] Basic probability assignment (BPA) refers to the process of calculating the basic probability of each piece of evidence (A) in the identification framework Θ. Since evidence ei may not be rated as a certain category, m is used to represent the basic probability assignment.

[0116] m n,i is the basic probability distribution representation of the evidence ei rated as Hn; m H,i It is the basic probability distribution value of not being rated into a certain category. Its calculation formula is:

[0117] m n,i =w i β n,i (2-1)

[0118]

[0119] where w i Yes i The weight of , and satisfy w i ≥0,∑W i = 1. Here m H,i It can be divided into two parts:

[0120]

[0121] Assume m n for e i and e j The ER rule is as follows:

[0122]

[0123] {H n}:m n =K[m n,i m n,j +m H,i m n,j +m n,i m H,j ] (2-7)

[0124]

[0125] Among them, m i，j 、m t，i It is the product of the probability mass of two different pieces of evidence with different evaluation levels. K represents the degree of conflict between the pieces of evidence.

[0126] Table 2 demonstrates the intersection of the basic probability distributions of the rating results H1 and H2 for evidence e1 and evidence e2.

[0127] Table 2 ER rules table

[0128]

[0129] The existing evidence e1, e2 are rated as H1, H2, H3 and the empty set is not rated as any result, m n,1 represents the basic probability distribution of each rating of evidence e1, m 1,1 =0.18, m 2,1 =0.12, m 3,1 =0.30, m n，2 represents the basic probability distribution of each rating result of evidence e2, m 1,2 =0.08, m 2,2 =0.2, m 3,2 =0.12.

[0130] When all the evidence fragments are combined, m H Then it is allocated to each layer according to the following rules:

[0131]

[0132] Where βn represents the first classification probability that the investment risk data is classified as the nth risk label.

[0133] 2.2 Deep Component

[0134] Deep learning technology is developed and evolved on the basis of artificial neural network models. In the field of deep learning, various model structures have their own areas of expertise, are good at mining the characteristic relationship between risk feature data sets and risk label data sets, and are good at extracting the intrinsic connection between training data. When deep learning models are applied to risk intelligent decision-making, they can automatically extract the data features existing in the risk intelligent decision-making data set from the irregular risk intelligent decision-making data set using computers. Taking into account the risk indicator characteristics and risk characteristics, the risk intelligent decision-making system adopts a deep neural network model.

[0135] (1) Construction of deep neural network model

[0136] Deep neural networks perform well in processing high-dimensional risk data features and can improve the efficiency and accuracy of risk warnings. This is essentially to simulate the human brain's processing of complex information through computers, build a topological knowledge structure inside the neural network, and achieve feedback and understanding of external multi-source information. The basic deep learning model can be expressed as:

[0137]

[0138] In the formula, x is the neuron node, w is the indicator weight, b is the bias between the simulated value and the true value (i.e. the degree of deviation), and ∫(.) is the corresponding activation function.

[0139] The model's depth part designs a variety of risk sensors, covering the input layer, multiple hidden layers, and the output layer, which together point to the risk intelligent warning results. The model does not need to deeply explore the complex functional relationship between risk variables, nor does it need to make independent assumptions about the warning indicators. It can effectively reveal the mapping relationship between the warning indicators of different dimensions of the enterprise and the warning results, and deeply explore the intrinsic connection between the training data. In the actual warning process, the model can approximate any nonlinear risk intelligent warning model, thereby more accurately capturing the risk dynamics.

[0140] (2) Reduce the error of early warning results

[0141] The deep learning neural network itself has the ability to automatically extract risk features from the data set based on the correlation between risk data. The supervised deep neural network trains the risk intelligent decision-making model.

[0142] L represents the number of risk assessment deep network layers; nL represents the number of risk assessment feature values ​​in the Lth layer; fL(*) represents the activation function of the risk assessment feature value in the Lth layer; W(L)∈ represents the weight matrix from the (L-1)-layer risk warning network to the Lth layer; b(L)∈ represents the bias from the (L-1)-layer risk warning network to the Lth layer; z(L)∈ represents the state of the risk assessment feature value in the Lth layer; A(L)∈ represents the activity value of the risk assessment feature value in the Lth layer; then the deep network performs feedforward propagation of information through the following calculation formula:

[0143] Z (l) =W (l) ×a (l-1) +b (l) (2-12)

[0144] a (l) =f l (Z (l) ) (2-13)

[0145] Formula 2-12 and 2-12 can also be combined into:

[0146] Z (l) =W (l) ×f l (Z (l-1) )+b (l) (2-14)

[0147] Deep neural networks usually pass the information inside each layer of the neural network layer by layer, and finally achieve the result output (l) .

[0148] x=a (0) →z (1) →a (1) →z (2) →...→a (l-1)→z (l) →a (l) =y (2-15)

[0149] Given a set of samples, (x(i), y(i), l≤i≤N), x(i) is the feature, y(i) is the label, l is the number of layers, the output of the network after feedforward propagation is f(x|w,b), and the objective function is:

[0150]

[0151] In the formula, W and b represent the weight matrix and bias vector in each layer of the specific neural network structure, namely:

[0152]

[0153] The goal of error back propagation in the risk intelligent decision model is to minimize J(W, b: x, y), which can be achieved by updating parameters in multiple iterations using the following method:

[0154]

[0155] Among them, α is the update rate of risk intelligent decision model parameters. Constructing chain rule based on enterprise early warning model It is expressed as shown in formula 2-23:

[0156]

[0157] For the l-th layer of the neural network, by defining the error term is the error value of the enterprise early warning model, and the objective function is established about the risk characteristic value z of the first layer risk intelligent decision network (l) The partial derivative of refers to the influence of the risk characteristic value of the l-th layer neural network on the final risk alert error. (l) It also reflects the sensitivity of the final risk warning output to the risk characteristic value of the lth layer neural network and the error of the final warning model output result.

[0158] According to formula 2-19, because z (l) =W (l) ·a (l-1) +b (l) ,so:

[0159]

[0160] So we have:

[0161]

[0162] Similarly, we can get:

[0163]

[0164] The error term δ at the lth layer (l) :

[0165]

[0166] Where ⊙ is the dot product operator of the risk intelligent decision model vector, representing the multiplication of each element in the enterprise early warning model. The key to the process of formula derivation lies in the derivation of the mathematical formula conclusion of formula 2-25. There are three formulas multiplied in formula 2-25, among which the term can be defined as δ according to the enterprise early warning model. (l+1) The second formula in the early warning model is based on formula z (l+1) =W (l+1) ×a (l) +b (l) ,therefore:

[0167]

[0168] Because a (l) =f l (z (l) ), and f l (·) is the function obtained according to the calculation rules of the early warning model, so we get:

[0169]

[0170] It can be seen from formula 2-29 that within the deep neural network architecture, for any risk warning error term of the lth layer, its result will be transmitted to the l+1th layer, and then the risk warning error of the next layer will be calculated. This layer-by-layer transmission process actually reflects the reverse transmission mechanism of the error of the warning model output result. The internal specific meaning of this mechanism is: the risk indicator error term result in the deep neural network structure of the lth layer is the weighted sum of the risk intelligent decision indicator error term results of all the l+1th layer that are related to the risk indicator in the deep neural network structure, and then multiplied by the corresponding risk intelligent decision indicator, that is, the gradient corresponding to the activation function associated with the neurons in the deep neural network structure, and then calculate the error result of the risk warning output result of each layer of the neural network, which can derive the gradient value of the risk intelligent decision model parameters of each subsequent layer.

[0171] The training steps of the risk intelligent decision-making model are as follows: first, the forward propagation method is used to calculate the status and activation values ​​of various risk indicators layer by layer, so as to achieve the effective transmission of complex and multi-level risk information until it reaches the last layer of the model; second, based on the obtained risk warning results, the back propagation technology is used to quantify the error value of each layer of the neural network; finally, based on the above error values, the gradient of the parameters of each layer of the neural network is further solved, and the parameter settings of the corresponding layer are adjusted and updated accordingly. This iterative process aims to continuously optimize the warning effectiveness of the model.

[0172] 2.3 Model Training

[0173] First, construct a risk feature dataset (risk index) and a risk label dataset (risk alert level) and import them into the Wide and Deep parts as input layers. Then, calculate each part in parallel and output their own warning results. Among them, the Wide part infers and fuses the warning results of the three basic warning devices, LR, Adaboost and SVM, through evidence reasoning. For any risk intelligent warning data, the classification results of the three warning devices are equivalent to the three pieces of evidence in the evidence reasoning, and the two classifications correspond to the two evaluation levels. The warning device integration is equivalent to the fusion of the output results of the three basic warning devices using the evidence reasoning rules, and finally outputs the warning results integrating the three different types of basic warning devices. The Deep part is designed as a deep neural network (DNN) with multiple layers of neurons, which learns complex patterns from the data and outputs the warning results of the Deep part.

[0174] Finally, the warning results output by the Wide part and the Deep part are fully connected by using the Concatenate function, and the two sets of warning data are organically integrated into an output layer to predict the final extreme cash flow risk alert level of Company A. The core idea is to integrate the evidence reasoning model with interpretable ability with the deep learning model with generalization and self-learning ability, and continuously train and iterate to optimize the relevant parameters of the two models at the same time, so that the overall prediction ability of the model is optimized.

[0175] 2.4 Model Training Results

[0176] In order to closely integrate the Wide part with the Deep part and output the optimal warning result, the Dropout layer and 500 stochastic gradient descent algorithms are used in the Deep part. These strategies can optimize the iterative operation of the internal correlation parameters of the deep neural network, reduce the correlation between risk neural nodes, and thus reduce the risk of overfitting, thereby improving the warning accuracy and generalization ability of the model. After multiple training iterations and parameter adjustments, it is observed that the loss value of the training sample gradually approaches 0, which indicates that the warning accuracy of the model has reached a high level and the training of the risk warning model is basically completed. Through training optimization, the Deep part finally constructs a deep neural network structure with 3 layers. In this neural network, the number of risk warning neurons in each layer is 20, 16 and 2 respectively. Such a structural configuration can more effectively capture risk characteristics and improve warning performance.

[0177] 3.IEWD Risk Intelligent Early Warning Model - Example of Application of Extreme Cash Flow Risk Early Warning in Enterprise A

[0178] Based on the risk status and cause analysis of Company A, the construction of a risk warning indicator system and a risk intelligent warning model, the present invention will use Python software to train, test and apply the extreme cash flow risk intelligent warning model of Company A. The extreme cash flow risk data set is input, and evidence reasoning is used as the inference engine to reason and fuse the evidence of multiple base warning devices, and fully link with the deep neural network model to obtain the extreme cash flow risk intelligent warning results of my country's real estate companies. And further through the comparison of multiple groups of models, the superiority of the constructed extreme cash flow risk intelligent warning model based on Wide&Deep is verified. Finally, reasonable suggestions are given for the predicted extreme cash flow risk prevention and control of Company A to provide intelligent decision support.

[0179] 3.1 Sample selection and data sources

[0180] 3.1.1 Sample selection

[0181] Based on the design ideas of the real estate enterprise risk indicator system, this invention is designed mainly from three aspects: the requirements for real estate enterprise sample selection, the span of data research period, and the definition of enterprise extreme cash flow risk. In order to make the data acquisition more complete and accurate, the selection of real estate enterprises should fully consider:

[0182] (1) Select listed companies in the real estate and construction industries from the 2012 edition of the China Securities Regulatory Commission’s industry classification.

[0183] (2)The data period is set to 2003-2022.

[0184] (3) Eliminate enterprises marked as ST or *ST. Since the financial statements of ST enterprises are difficult to obtain and there are technical problems, the data before ST is still needed to measure the risk factors in order to study the impact of extreme cash flow risks on real estate enterprises. After sorting, we finally obtained relatively complete case company data of 256 companies and 3,340 data samples. The stock codes of the companies are shown in Table 3.

[0185] Table 3 Enterprise Stock Code List

[0186]

[0187]

[0188] 3.1.2 Data Source

[0189] The label of extreme cash flow risk of real estate enterprises is extreme cash flow risk, which needs to be classified. That is, using the historical experience distribution of the company's cash flow risk, the forecast error value ΔCF of a single enterprise at the time point [t (year), q (quarter)] and its lag of at least 20 quarters (at least 5 years) is calculated. t，q , forming an "empirical distribution"; then, the "worst case" forecast error value ΔCF of the 5% quantile tail is taken from the "empirical distribution" t，q 0.05 , which is used to measure the "cash flow shock" under extreme conditions (95% confidence interval).

[0190] Therefore, the labels of the cash flow level prediction error values ​​under extreme conditions are divided into two levels. The companies with the top 10% of the extreme cash flow risk values ​​have extreme cash flow risk, which is set to 1; the others are set to 0. Finally, all the features and labels of the extreme cash flow risk data of real estate companies are matched to form a complete extreme cash flow risk data set for real estate companies. The results are shown in Table 4.

[0191] Table 4 Dataset table

[0192]

[0193]

[0194] 3.2A Enterprise Multi-Earning Device Integration Risk Output

[0195] By constructing a multi-alarm fusion risk intelligent early warning model based on evidence reasoning, the extreme cash flow risk data set of real estate enterprises is input, and the classifier results are output, the alarm weights are determined, and evidence reasoning and fusion are performed to complete the data input and output work of the Wide component, laying the foundation for the final generation of intelligent early warning results for extreme cash flow risks of real estate enterprises.

[0196] 3.2.1 Basic warning output

[0197] First, the extreme cash flow risk dataset of real estate enterprises was input into three basic early warning devices, namely support vector machine, logistic regression and Adaboost, using Pycharm software. Then, the results of each classifier were output through the set code program, as shown in Table 5.

[0198] Table 5 Basic early warning output table

[0199]

[0200] 3.2.2 Determination of early warning device weight

[0201] Based on the extreme cash flow risk intelligent warning results output by the three basic early warning devices, SVM, LR, and Adaboost, the corresponding weights are determined according to the prediction accuracy of each basic early warning device. The Pycharm running environment equipped with Python 3.9 will output the average values ​​of four evaluation indicators, including accuracy, recall, and F1 value, after 50 runs of the three basic early warning devices, support vector machine, logistic regression, and Adaboost, as shown in Table 6, which plays an important role in the calculation of the weight of the basic early warning device. After inputting the extreme cash flow risk data set of real estate companies such as A Enterprise Co., Ltd. from 2003 to 2022, the output evaluation indicators are shown in the table, that is, the three basic early warning devices, support vector machine, logistic regression, and Adaboost, are cross-validated 5 times with 10 results, and the average values ​​of the accuracy of the 50 results are 92%, 84%, and 84%, respectively. Therefore, the corresponding weights of the three basic early warning devices are 0.36, 0.32, and 0.32, respectively.

[0202] Table 6 Statistics of three basic early warning devices

[0203]

[0204] 3.2.3 Evidence Reasoning and Fusion

[0205] After determining the weight of each basic early warning device, the importance of each piece of evidence is fixed. On this basis, the evidence reasoning theory is used to reason and fuse the results of each basic early warning device through rule activation to output the fusion early warning results of the width model. Python software is used to model and output the fusion of the early warning results of extreme cash flow risks of real estate enterprises. The output results are shown in Table 7.

[0206] Table 7 Output result table

[0207]

[0208] 3.3A Intelligent early warning of extreme cash flow risks for enterprises

[0209] The multi-warning device fusion risk intelligent early warning model based on evidence reasoning mainly performs evidence reasoning and fusion on the early warning results of multiple base early warning devices. On this basis, the evidence reasoning results are organically fully linked with the results of the deep neural network through the extreme cash flow risk intelligent early warning model based on Wide & Deep to achieve intelligent early warning. The intelligent early warning model constructed by the present invention has good memory, interpretability and generalization ability of high-frequency co-occurring feature combinations, which can produce a good support effect on the early warning results.

[0210] 3.3.1 Determination of alertness

[0211] The results of the intelligent early warning of extreme cash flow risks of real estate enterprises are fully linked to the results of evidence reasoning and deep neural network through the Wide & Deep model after the screening of key risk indicators, the output of the basic early warning results, and the integration of evidence, so as to ensure the interpretability of the intelligent decision-making results with a high accuracy. The results are shown in Table 8. The results show that the confidence levels of extreme cash flow risks of Enterprise A in 2024 are 31.43% and 68.57% respectively, indicating that Enterprise A is more likely to be exposed to extreme cash flow risks in 2024, and the management of the enterprise needs to take as many ways as possible to alleviate the cash flow pressure of the enterprise.

[0212] Table 8 Operation results

[0213]

[0214]

[0215] Presenting the extreme cash flow risk of real estate enterprises in the form of confidence distribution can enable decision makers to directly observe the level and confidence level of extreme cash flow risk, so that they can make corresponding decisions and preventive measures.

[0216] By using the intelligent early warning model (IEWD) of extreme cash flow risk of real estate enterprises based on evidence reasoning and Wide and Deep, the distribution of extreme cash flow risk of real estate enterprises in 2024 is predicted. Figure 5As shown. From the results of intelligent early warning, among the sample enterprises, 21.08% of the enterprises have extreme cash flow risks, and 78.92% of the enterprises do not have extreme cash flow risks, indicating that the extreme cash flow risks of real estate enterprises are controllable as a whole, but the scale is still large. Specifically for Company A, the probability of its extreme cash flow risk is 68.57%, which can be basically determined to be extremely risky. By analyzing the operating status of Company A, the results simulated by the intelligent early warning model constructed by the present invention are in line with the actual situation. This fully demonstrates the reliability of the intelligent early warning model for extreme cash flow risks of real estate enterprises constructed by using the base early warning device to form multi-source evidence. It also illustrates the feasibility of using the Wide&Deep model that integrates evidence reasoning and deep neural networks to intelligently warn of extreme cash flow risks of real estate enterprises.

[0217] In summary, based on the warning results of the intelligent early warning model (IEWD) for extreme cash flow risks of real estate enterprises that has been constructed, this paper selects five models, including traditional machine learning models such as random forest, support vector machine, Adaboost, separate Wide component and Deep component, for comparison. The data input of each model remains unified: first, the data missing values ​​of the real estate enterprise extreme cash flow risk warning feature data set and the risk label data set are removed; then the data is normalized, and finally the sample selection ratio of "0.75:0.25" is determined as 2505 training sample subsets and 835 test subsets. Five indicators, including accuracy, precision, recall, F1-Score and AUC, are selected to measure the effects of the five comparison models. The results of the extreme cash flow risk warning model for real estate enterprises can be divided into four categories: TP, FN, FP and TN, and the meanings are shown in Table 9.

[0218] Table 9 Meaning table

[0219]

[0220] Based on the four types of risk warning results, namely TP, FN, FP and TN, the accuracy, precision, recall and F1 value of the model are calculated.

[0221] (1) Accuracy

[0222] Accuracy refers to the ratio of the number of samples correctly classified by the classifier to all samples. The formula is:

[0223]

[0224] (2) Precision

[0225] Precision refers to the ratio of the number of correctly classified positive samples to the number of predicted positive samples. The formula is:

[0226]

[0227] (3) Recall

[0228] Recall refers to the ratio of the number of correctly classified positive samples to the actual number of positive samples. The formula is:

[0229]

[0230] (4)F1-Score

[0231] In fact, in practical applications, it is often necessary to consider Recall and Precision comprehensively to evaluate the model.

[0232] Performance, thus introducing a comprehensive evaluation index, namely F1-Score. The formula is:

[0233]

[0234] (5)AUC

[0235] AUC (Area Under the Curve) is a curve that reflects the relationship between sensitivity and specificity. It is an indicator for evaluating the performance of a classification model and is usually used to evaluate the performance of a binary classifier. The value of AUC ranges from 0.5 (equivalent to random guessing) to 1 (perfect classification). Generally speaking, the closer the AUC value is to 1, the better the classification performance of the model. This is because the higher the AUC value, the higher the probability that the model will rank positive examples first and negative examples later, that is, the stronger the model's ability to distinguish.

[0236] The prediction results of five models, including traditional machine learning models such as random forest, support vector machine, Adaboost, separate Wide component and Deep component, are shown in Table 10. As can be seen from the table, the accuracy of the intelligent early warning model (IEWD) for extreme cash flow risk of real estate enterprises constructed by the present invention is close to 0.927, and the precision, recall, F1-Score and AUC indicators are all excellent, which is significantly better than the other five comparison models. Therefore, the intelligent early warning model (IEWD) for extreme cash flow risk of real estate enterprises constructed by the present invention effectively integrates the advantages of multiple base early warning devices, which not only makes the intelligent early warning results better, but also makes the results highly interpretable due to the introduction of the evidence reasoning model, which can play a good early warning effect for extreme cash flow risks of real estate enterprises.

[0237] Table 10 Prediction results of 5 models

[0238]

[0239] Finally, the present invention also provides a terminal, such as a mobile phone, a computer, etc., which adopts the investment risk decision support method based on evidence reasoning and deep learning as described in any one of the above.

[0240] In summary, the present invention proposes an investment risk decision support method and system based on evidential reasoning and deep learning, which realizes investment risk prediction through the Wide & Deep risk warning model of evidential reasoning, and the process of rule reasoning based on the Wide & Deep risk warning model of evidential reasoning can be completely realized through a transparent and explainable rule interpretation program, avoiding the uncertainty of the "black box" and providing decision makers with a reliable and explainable decision basis for the processing process.

[0241] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0242] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0243] It will be easily understood by those skilled in the art that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An investment risk decision support method based on evidence reasoning and deep learning, characterized in that: The method includes: Obtain investment risk data and determine risk labels; Construct a Wide & Deep risk warning model based on evidence reasoning, including Wide component, Deep component, fully connected layer and output layer; The Wide component includes multiple base early warning devices, each of which outputs the confidence level of the investment risk data being classified into each risk label. Combined with the weight of each base early warning device, the basic probability distribution is calculated. Then, the basic probability distribution is combined using the evidence reasoning combination rule to obtain the first classification probability of the investment risk data being classified into each risk label. The Deep component includes a deep neural network model, which outputs the second classification probability of the investment risk data being classified into various risk labels; The fully connected layer is fully connected with the first classification probability of the Wide component and the second classification probability of the Deep component; The output layer is connected to the fully connected layer, and outputs the final classification probability of the investment risk data being classified into various risk labels; Using investment risk data and corresponding risk labels, we train a Wide & Deep risk warning model based on evidence reasoning, and use the trained Wide & Deep risk warning model based on evidence reasoning to predict investment risks.

2. The investment risk decision support method based on evidence reasoning and deep learning according to claim 1 is characterized in that: The basic probability distribution is calculated, including: Assume that the identification frame Θ = {H1, H2...H n ...H N }; Among them, H n is the confidence rating result, N is the total number of risk labels; Taking investment risk data as evidence, then: S(e i )={(H n ,β n，i )|n=1,2,...,N;i=1,2,...,L} In the formula, e i is the i-th evidence; β n,i For evidence i Rated H n The confidence level is the confidence level that the i-th investment risk data is classified as the n-th risk label by the base early warning device; L is the total number of evidences; Calculate the basic probability distribution as follows: m n,i =w i b n,i In the formula, m n,i Rated H n Evidence i The basic probability distribution of m H,i Not rated H n Evidence i The basic probability distribution of w i is the weight of the base early warning device and ∑w i =1.

3. The investment risk decision support method based on evidence reasoning and deep learning according to claim 1 or 2, characterized in that: The calculation method of the weight of the base early warning device is: In the formula, is the weight of the k-th base early warning device, accuracy k is the accuracy of the kth base early warning device, and K is the total number of base early warning devices.

4. The investment risk decision support method based on evidence reasoning and deep learning according to claim 2 is characterized in that: The basic probability distribution is combined by using the evidence reasoning combination rule to obtain the first classification probability of the investment risk data being classified into various risk labels, including: M H,i It is divided into two parts: Assume m n For evidence i and evidence j The ER rule is as follows: {H n }:m n =K[m n,i m n,j +m H,i m n,j +m n,i m H,j ] In the formula, K represents the degree of conflict between the evidences; Calculate the first classification probability of the investment risk data being classified into each risk label: In the formula, β n is the first classification probability that the investment risk data is classified as the nth risk label.

5. The investment risk decision support method based on evidence reasoning and deep learning according to claim 1 is characterized in that: The base early warning devices include LR, Adaboost and SVM.

6. The investment risk decision support method based on evidence reasoning and deep learning according to claim 1 is characterized in that: The deep neural network model is DNN deep neural network.

7. The investment risk decision support method based on evidence reasoning and deep learning according to claim 1 is characterized in that: Investment risk data refers to various indicators related to investment risks that are predetermined. Risk labels include no risk warning and risk warning.

8. An investment risk decision support system based on evidence reasoning and deep learning for implementing the investment risk decision support method based on evidence reasoning and deep learning as described in any one of claims 1 to 7, characterized in that: The system includes: a database subsystem and a model library subsystem; The database subsystem is used to store and manage investment risk data and its corresponding risk labels; The model library subsystem is used to store Wide & Deep risk warning models based on evidence reasoning; The system also has a human-computer dialogue subsystem, which includes a human-computer interaction interface; the human-computer dialogue subsystem obtains investment risk data to be decided through the human-computer interaction interface, and processes it into computer language and transmits it to the database subsystem; the database subsystem extracts risk decision data and hands it over to the model library subsystem for investment risk prediction, and finally displays the results through the human-computer interaction interface of the human-computer dialogue subsystem.

9. The investment risk decision support system based on evidence reasoning and deep learning according to claim 8 is characterized in that: The database subsystem includes the database and database management system; The database is used to store investment risk data and its corresponding risk labels, and to achieve effective data access through ETL data development and task engines. After data access, the data is cleaned and converted to ensure data quality and consistency. The database management system is used to complete the operation, maintenance and management functions of data extraction, realize the creation of data objects, query, load, add, backup, modify, delete operations of stored data, and user management and permission management of data.

10. A terminal, characterized in that: The terminal adopts the investment risk decision support method based on evidence reasoning and deep learning as described in any one of claims 1 to 7.

Images