Enterprise empty shell risk prediction method and device, computer equipment and storage medium

Abstract

The invention discloses an enterprise empty shell risk prediction method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring enterprise related data; performing feature engineering processing on the enterprise related data to obtain a processing result; inputting the processing result into a vacant shell risk prediction model to perform vacant shell risk prediction so as to obtain a prediction result; and outputting the prediction result. Wherein the empty shell risk prediction model is obtained by training a deep learning network through features obtained by performing feature engineering processing on a plurality of enterprise related data with risk level labels. By implementing the method provided by the invention, the accuracy, comprehensiveness and flexibility of enterprise vacant shell risk prediction can be remarkably improved, and finer risk management and wider application scenarios are supported at the same time; the problems that in the prior art, deep mining of multi-source data integration and complex feature relations is lacked, and consequently evaluation accuracy, comprehensiveness and flexibility are insufficient are solved.

Description

Technical Field

[0001] This invention relates to computers, and more specifically to methods, apparatus, computer equipment, and storage media for predicting the risk of shell companies. Background Technology

[0002] Shell companies typically refer to companies without actual business activities, assets, or employees. These companies may exist only legally, possessing a registered address and basic corporate structure, but do not engage in substantive commercial operations. In today's business management field, identifying potential shell company risks relies primarily on several methods. First, traditional rule-based analysis methods use a series of pre-defined, manually applied rules to determine whether a company is a shell. These rules usually rely on limited indicators such as registered capital, length of establishment, and operating status, and are judged using fixed thresholds. However, this method often only considers the company's business registration information and fails to comprehensively reflect the company's actual operations. Furthermore, evaluation systems based on a single data source rely solely on data from tax or business registration departments. For example, if a company has failed to file tax returns or undergo annual inspections for several consecutive years, it may be flagged as a suspected shell company. This approach, due to its single data source, cannot provide a complete perspective on the company's operating status.

[0003] Furthermore, some attempts have attempted to use basic machine learning algorithms such as decision trees and random forests for risk assessment. While progress has been made, feature engineering remains relatively simple, limiting the expressive power of the models and making it difficult to delve into the complex relationships between features. Current models tend to assess shell company risk and the activity level of their operations as two completely independent objectives, ignoring the inherent connection between them, which can lead to inconsistent assessment results. Such a separate risk and activity assessment method fails to fully utilize the mutual influence between the two, thus affecting the accuracy and reliability of the assessment. However, existing technologies also have several limitations. On the one hand, most existing assessment techniques rely heavily on data from official channels, such as business registration information and tax records, without effectively integrating publicly available data from the internet and social media information, resulting in an incomplete description of the company. On the other hand, the focus is too concentrated on basic business information and tax payment, neglecting other important dimensions such as intellectual property, online activity, and related companies. Most systems use simple binary classification labels, failing to reflect different levels of shell company risk and hindering the implementation of refined risk management. Moreover, rule-based methods are not easily adaptable to the characteristics of different industries and regions, lacking the necessary flexibility. Meanwhile, commonly used shallow machine learning models struggle to capture complex patterns and underlying correlations. Obtaining labeled data for training these models is difficult, limiting their practical application. Most techniques focus on optimizing a single objective, failing to consider the interaction between shell company risk and operational activity. Finally, the lack of effective mechanisms to ensure models reflect changes in the business environment and policy adjustments in a timely manner leads to delayed model updates. These limitations collectively constitute the challenges facing existing shell company risk assessment technologies.

[0004] Therefore, it is necessary to design a new method that significantly improves the accuracy, comprehensiveness, and flexibility of corporate shell risk prediction, while supporting more refined risk management and a wider range of application scenarios; in order to solve the problem that existing technologies lack in-depth mining of multi-source data integration and complex feature relationships, resulting in insufficient accuracy, comprehensiveness, and flexibility in assessment. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, computer equipment and storage medium for predicting the risk of shell companies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting the risk of shell companies, comprising:

[0007] Obtain relevant enterprise data;

[0008] The relevant enterprise data is subjected to feature engineering processing to obtain the processing result;

[0009] The processing and results are input into the shell company risk prediction model to predict the shell company risk and obtain the prediction results.

[0010] Output the prediction result;

[0011] The shell company risk prediction model is obtained by training a deep learning network with features obtained from several enterprise-related data with risk level labels through feature engineering.

[0012] The further technical solution is as follows: the feature engineering processing of the enterprise-related data to obtain the processing result includes:

[0013] The relevant data of the enterprise is cleaned to obtain cleaned data;

[0014] The cleaned data is then subjected to feature extraction and transformation to obtain the extraction results;

[0015] The extracted results are subjected to feature cross-interaction and combination to obtain a combined result;

[0016] Feature selection is performed from the combined results to obtain the processing result.

[0017] The further technical solution is as follows: the empty shell risk prediction model includes an input layer, a feature processing layer, a deep learning network layer and an output layer arranged in sequence; the deep learning network layer includes multiple fully connected layers, a batch normalization layer, an activation function and a Dropout layer.

[0018] The further technical solution is as follows: the prediction results include a shell company risk index and an operational activity index.

[0019] The further technical solution is as follows: the overall loss function used in the training of the shell risk prediction model is an index that comprehensively evaluates the prediction accuracy and generalization ability of the model by adjusting the relative weights of shell risk loss and business activity loss and adding an L2 regularization term.

[0020] The further technical solution is as follows: the overall loss function used during the training of the shell risk prediction model determines the weight value of the shell risk loss through grid search and cross-validation.

[0021] The further technical solution is as follows: when training the empty shell risk prediction model, Bayesian optimization technology is used to adjust the hyperparameters, wherein the hyperparameters include the learning rate and the batch size.

[0022] This invention also provides a device for predicting the risk of shell companies, comprising:

[0023] The data acquisition unit is used to acquire relevant enterprise data.

[0024] The feature engineering processing unit is used to perform feature engineering processing on the enterprise-related data to obtain the processing result;

[0025] The prediction unit is used to input the processing and results into the shell company risk prediction model to predict the shell company risk and obtain the prediction result.

[0026] An output unit is used to output the prediction result;

[0027] The shell company risk prediction model is obtained by training a deep learning network with features obtained from several enterprise-related data with risk level labels through feature engineering.

[0028] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0029] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0030] The advantages of this invention compared to existing technologies are as follows: This invention integrates multi-source heterogeneous data and performs systematic feature engineering processing on this data, including feature extraction, transformation, cross-fertilization, combination, and selection, to comprehensively capture the complex characteristic relationships of enterprises. Then, it utilizes deep learning networks to train enterprise-related data labeled with risk levels to establish an accurate shell company risk prediction model. This method not only significantly improves the accuracy, comprehensiveness, and flexibility of shell company risk prediction but also supports more refined risk management and adapts to a wider range of application scenarios. Compared to existing technologies, it effectively solves the problems of insufficient integration of multi-source data and failure to deeply explore complex characteristic relationships, thereby greatly enhancing the accuracy and reliability of the assessment. Finally, by outputting intuitive prediction results, it provides users with strong decision support.

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a schematic diagram illustrating an application scenario of the shell company risk prediction method provided in this embodiment of the invention.

[0034] Figure 2 A flowchart illustrating the method for predicting the risk of shell companies provided in an embodiment of the present invention;

[0035] Figure 3 A schematic diagram of a sub-process of the corporate shell risk prediction method provided in an embodiment of the present invention;

[0036] Figure 4 A schematic block diagram of a shell company risk prediction device provided in an embodiment of the present invention;

[0037] Figure 5 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0040] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0041] It should also be further understood that the terms "and / or" as used in this specification and the appended claims refer to any combination of one or more of the associated listed items and all possible combinations, and include such combinations.

[0042] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the corporate shell risk prediction method provided in this embodiment of the invention. Figure 2This is a schematic flowchart illustrating the enterprise shell risk prediction method provided in this embodiment of the invention. The method is applied to a server that interacts with a terminal. By integrating multi-source heterogeneous data and performing deep feature engineering processing, including data cleaning, feature extraction and transformation, feature cross-combination, and feature selection, it significantly improves the comprehensive characterization of enterprise operating status and the in-depth mining of complex feature relationships, thereby greatly enhancing the accuracy and comprehensiveness of shell risk prediction. The method trains a deep learning model containing an input layer, a feature processing layer, a deep learning network layer, and an output layer using data labeled with risk levels. During training, it employs an overall loss function that integrates shell risk loss and operating activity loss with L2 regularization. It combines grid search and cross-validation to determine weight values ​​and uses Bayesian optimization techniques to adjust hyperparameters, achieving an effective balance between model prediction accuracy and generalization ability. This also supports refined risk management and a wider range of application scenarios. This method solves the problem of insufficient integration of multi-source data and in-depth mining of complex feature relationships in existing technologies, enhancing the accuracy, comprehensiveness, and flexibility of the assessment.

[0043] Figure 2 This is a flowchart illustrating the method for predicting the risk of shell companies provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S140.

[0044] S110. Obtain relevant enterprise data.

[0045] In this embodiment, enterprise-related data refers to data collected from various information sources to identify and assess the risk of a shell company. This data covers not only the company's basic registration information but also its operational status, financial condition, tax records, intellectual property, internet activities, and relationships with other companies. The specific data categories and examples are as follows:

[0046] Business registration information includes:

[0047] Company Name: The official name of the company.

[0048] Uniform Social Credit Code (USC): A code used to uniquely identify a company.

[0049] Registered capital and paid-in capital: reflect the financial strength of a company.

[0050] Establishment date and duration: Displays the length of time the company has been in existence.

[0051] Business scope: Describes the types of business that the company is permitted to engage in.

[0052] Information on legal representatives and shareholders: Provides information on corporate control and ownership.

[0053] Tax information includes:

[0054] Tax credit rating: measures a company's performance in complying with tax laws.

[0055] Tax violation information: including whether the company has engaged in illegal activities such as tax evasion.

[0056] Tax filing records: Continuity and completeness are important indicators for judging the activity level of a business.

[0057] Transaction information includes:

[0058] Inter-company transaction records: including sales contracts, service agreements, etc.

[0059] Invoicing information: Invoice issuance information over the past few months or years, reflecting the company's actual operating situation.

[0060] Qualification information includes:

[0061] Patents, trademarks, and copyrights: demonstrate a company's investment and achievements in technological innovation.

[0062] Certifications: Proof that a company possesses the capabilities and qualifications in a specific field.

[0063] Internet information includes:

[0064] Social media activity: Assess a company’s online presence through interactions on its official accounts.

[0065] Job postings on recruitment platforms can indirectly reflect a company's expansion plans and health status.

[0066] E-commerce platform sales information: Shows the level of activity of enterprises in the field of e-commerce.

[0067] Related enterprise characteristics include:

[0068] Information on parent company, subsidiaries, and sister companies: Understanding the company's organizational structure and network.

[0069] Partners and competitors: Analyze the company's market positioning and competitive environment.

[0070] Characteristics of violations include:

[0071] Administrative penalty records: Whether the company has been penalized for violating laws and regulations.

[0072] List of dishonest judgment debtors: Whether the enterprise or its legal representative is included in the blacklist.

[0073] These data come from multiple sources, such as public databases, third-party business information service providers, social media platforms, and recruitment websites. Collecting comprehensive and accurate enterprise-related data is a fundamental step in assessing the risk of shell companies, helping to build a comprehensive feature set, thus providing strong support for subsequent risk model training and prediction. By comprehensively analyzing these features, companies that may pose a shell company risk can be identified more accurately, providing a scientific basis for business cooperation, market supervision, and financial risk control.

[0074] S120. Perform feature engineering processing on the enterprise-related data to obtain the processing result.

[0075] In this embodiment, the processing result refers to a high-quality feature set generated after a series of feature engineering steps (including data cleaning, feature extraction and transformation, feature cross-combination, and feature selection). These features will be used as input into a machine learning model for training and prediction to achieve an accurate assessment of the shell company risk.

[0076] In one embodiment, please refer to Figure 3 The above-mentioned step S120 may include steps S121 to S124.

[0077] S121. Perform data cleaning on the relevant enterprise data to obtain cleaned data.

[0078] In this embodiment, data cleaning is a crucial step in ensuring data quality and consistency. It includes the following aspects:

[0079] Missing value handling: For numerical features, the mean or median can be used to fill missing values ​​based on the feature distribution, or a special missing value prediction model can be built; for categorical features, the mode can be used to fill missing values, or a special "unknown" category can be set; for time series features, interpolation, forward filling, or backward filling can be used.

[0080] Outlier detection and handling: Identify and correct unreasonable or extreme data points to ensure data accuracy.

[0081] Duplicate record cleanup: Delete duplicate enterprise information records to ensure the consistency of the dataset.

[0082] Standardize the format: For example, standardize the date format, units, etc., to facilitate subsequent analysis.

[0083] S122. Perform feature extraction and transformation on the cleaned data to obtain the extraction results.

[0084] In this embodiment, the extraction result refers to the extracted features.

[0085] Specifically, in this step, useful features are extracted from the cleaned data and transformed as necessary to make it more suitable as input for machine learning models. This includes, but is not limited to:

[0086] Basic feature extraction: Directly extract basic attributes or variables from the raw data, such as company name, registered address, registered capital, etc.

[0087] Feature transformation: Applying specific mathematical transformations to certain features, such as Z-score standardization or Min-Max normalization for numerical features, and label encoding, one-hot encoding, or target encoding for categorical features.

[0088] Time feature processing: Convert timestamps into more meaningful forms, such as calculating the time interval between an event and the present, or decomposing dates into components such as year, month, and day.

[0089] S123. Perform feature crossing and combination on the extracted results to obtain the combined result.

[0090] In this embodiment, the combination result refers to the combined features.

[0091] This step aims to capture the interactions and correlations between different features by creating new features through combining existing features. It mainly includes:

[0092] Feature crossover: Combining two or more basic features to generate new features, such as the purchase frequency of users during holidays.

[0093] Feature combination: Further integrate multiple cross-features to form a higher-level feature representation. For example, using techniques such as polynomial transformation and nested functions, multiple cross-features can be integrated into a more representative feature set.

[0094] S124. Feature selection is performed from the combined results to obtain the processing result.

[0095] In this embodiment, the final step is to select the most valuable features from the combined results as the final feature set input to the machine learning model. The goal of feature selection is to reduce redundant information and improve model performance. Common methods include:

[0096] Filtering method: Selecting features based on statistical indicators (such as correlation coefficient, chi-square test, etc.).

[0097] Wrap-up method: Using the performance of the machine learning algorithm itself as the evaluation criterion, iteratively selects the optimal feature subset.

[0098] Embedding method: Automatic feature selection during model training, such as the L1 regularization term in Lasso regression.

[0099] By following the four steps above, a comprehensive and efficient feature set can be systematically constructed, providing strong support for subsequent risk assessment models and thus more accurately identifying the shell company risk.

[0100] S130. Input the processing and results into the shell company risk prediction model to predict the shell company risk and obtain the prediction results.

[0101] In this embodiment, the prediction results include a shell company risk index and an operational activity index.

[0102] The forecast results consist of two core indices:

[0103] Shell Company Risk Index: This is a comprehensive score reflecting the likelihood that a company is a "shell company." The higher the value, the greater the suspicion that the company is a shell company; conversely, the lower the value, the better the company's actual operations and the lower the risk of being a shell company.

[0104] The calculation method usually relies on the probability value output by the model and is adjusted in combination with business rules to ensure that the scoring system is both logical and easy to interpret.

[0105] Business Activity Index: This index measures the level of activity in a company's daily operations, encompassing information from multiple dimensions such as revenue, number of employees, and transaction records. A higher score indicates that the company is more active in the market and possesses stronger competitiveness and development potential.

[0106] This index helps identify companies that exist but are not active, providing a reference for market participants.

[0107] Application scenarios include:

[0108] Corporate credit assessment: Financial institutions can use these two indices to assess the creditworthiness of potential customers and decide whether to grant them loans or other financial services.

[0109] Market supervision: Regulatory authorities can use this information to strengthen monitoring of high-risk enterprises and prevent illegal activities.

[0110] Investment decision support: Investors can analyze these two indices to understand the health of target companies and make more informed investment choices.

[0111] In conclusion, by inputting carefully processed data into a specially designed deep learning model, not only can potential shell companies be effectively identified, but the operational activity of these companies can also be comprehensively assessed, providing crucial information for business cooperation, market supervision, and financial risk control. This process not only improves the accuracy of risk identification but also enhances the scientific rigor and reliability of decision-making.

[0112] The shell company risk prediction model is obtained by training a deep learning network with features obtained from several enterprise-related data with risk level labels through feature engineering.

[0113] In this embodiment, the empty shell risk prediction model includes an input layer, a feature processing layer, a deep learning network layer, and an output layer arranged in sequence; the deep learning network layer includes multiple fully connected layers, a batch normalization layer, an activation function, and a Dropout layer.

[0114] The overall loss function used in training the shell company risk prediction model is an indicator that comprehensively evaluates the model's prediction accuracy and generalization ability by adjusting the relative weights of shell company risk loss and operational activity loss, and adding an L2 regularization term.

[0115] The overall loss function used during the training of the shell risk prediction model determines the weight values ​​of the shell risk loss through grid search and cross-validation.

[0116] During the training of the empty shell risk prediction model, Bayesian optimization techniques are used to adjust the hyperparameters, which include the learning rate and batch size.

[0117] In one embodiment, a shell company refers to a company that has been registered but has not actually conducted business, or a company that is registered but has no substantive business activities. Such companies may be used for illegal activities, such as money laundering, fraud, tax evasion, etc., or as dummy companies to participate in bidding, disrupting market order.

[0118] The sample labeling employs a rule-based, multi-level judgment method, classifying enterprises into five risk levels: high, medium-high, medium, medium-low, and low.

[0119] High: Extremely low level of business activity, highly suspected of being a shell company.

[0120] Medium to high: Low business activity, high suspicion of being a shell company.

[0121] In China: Business activity is generally low, and there is a certain risk of it being a shell company.

[0122] Medium to low: High level of business activity, low suspicion of being a shell company.

[0123] Low: High level of business activity, low suspicion of being a shell company.

[0124] The labeling rules are as follows: First, does the Uniform Social Credit Code (USC) correspond to a contract? If not, it is directly classified as high risk. If it does, proceed to the next step. Next, does the company have a record of failing to file tax returns for three consecutive years or more? If so, it is directly classified as high risk. If not, proceed to the next step. Then, does the company meet any one of the high-risk indicators? If so, it is directly classified as high risk. If not, proceed to the next step. Finally, it checks how many medium-to-low risk indicators are met. If none are met, it checks if the company has had any contracts or credit records for more than seven years. If yes, proceed to the next step. If no, it is directly classified as low risk. Risk Level: Low; If 1 criterion is met, the risk level is directly determined to be Medium; If ≥2 criterions are met, the risk level is directly determined to be Low; Has there been no contract or credit record for more than 7 years? If yes: Are there multiple companies in remote areas? If yes, proceed to the next step; If no, the risk level is directly determined to be Low; If no, the risk level is directly determined to be Low; Is the site unmanned? If yes: Are there other medium-to-high-risk companies among the 5 shareholders? If yes, the risk level is directly determined to be Medium-High; If no, the risk level is directly determined to be Medium; If no, the risk level is directly determined to be Medium. All paths will eventually converge to the "Assessment Completed" node, indicating the end of the entire risk assessment process.

[0125] Detailed annotation rules are explained below:

[0126] Basic verification: The enterprise must be in operation and its unified social credit code must conform to the coding length rules (18 or 15 digits). If it does not meet these requirements, it will be marked as "high".

[0127] High-risk exclusion: Check if any of the following situations exist: abnormal operation for 2 or more consecutive years in the past 4 years; historical record of absconding or disappearance; if any of the above conditions are met, it will be directly marked as "high".

[0128] Low-risk quick identification: Check if any of the following conditions are met: rated as an A-level taxpayer in the past 2 years; meets the non-business registration coding rules (the first digit of the unified social credit code is not 9); has bidding information in the past 36 months; is a technology-based enterprise; has sales information on 1688; has recruitment information; has a trademark in the past 5 years; is a listed company. If any of the above conditions are met, it will be directly marked as "low".

[0129] Low to medium risk indicators: Check the following indicators: valid invoices issued in the past 12 months; number of insured persons > 3; paid-in capital ratio > 0.6 and registered capital ≥ 1 million; currently possessing valid qualifications; currently possessing valid patents; possessing trademarks in the past 5 years; being a large-scale enterprise; having social media; if ≥ 2 items are met, it is marked as "low", and if 1 item is met, it is marked as "medium".

[0130] Integrity Record: Check the company's compliance with contracts and creditworthiness: ≤4 years ago, it is marked as "low"; 4-7 years ago, it is marked as "low to medium".

[0131] Multiple companies at the same address: Check if the following conditions are met simultaneously: more than 2 affiliated companies registered by the same legal person; the registered addresses of the affiliated companies are basically the same; if met, mark the earliest established company as "low" and the rest as "medium-high"; Multiple companies with the same legal person: Check if the following conditions are met: more than 50 affiliated companies registered by the same legal representative; if met, mark the earliest 5 established companies or companies that are shareholders of affiliated companies as "low" and the rest as "medium-high";

[0132] Compliance Risk: Check if any of the following conditions are met: significant tax violations in the past 2 years; administrative penalties in the past 24 months; if met, it is marked as "Medium"; default mark: if none of the above are met, it is marked as "Low Medium".

[0133] The model's data sources are diverse, mainly including the following aspects:

[0134] Business registration information: Basic registration information, operating status, shareholder information, etc. of enterprises.

[0135] Tax information: taxpayer credit rating, tax violation information, etc.

[0136] Transaction information: inter-company transaction records, invoicing information, etc.

[0137] Qualification information: patents, trademarks, qualification certificates, etc.

[0138] Internet information: social media, job postings, e-commerce platforms, etc.

[0139] The feature classifications are shown in Tables 1 to 8.

[0140] Table 1. Basic Industrial and Commercial Characteristics

[0141]

[0142]

[0143] Table 2. Characteristics of Tax Credit

[0144]

[0145]

[0146] Table 3. Characteristics of Active Business Operations

[0147]

[0148] Table 4. Characteristics of Business Registration Changes

[0149]

[0150]

[0151] Table 5. Characteristics of Intellectual Property

[0152]

[0153]

[0154] Table 6. Characteristics of Internet Activity

[0155]

[0156] Table 7. Characteristics of Related Enterprises

[0157]

[0158]

[0159] Table 8. Characteristics of Illegal and Irregular Activities

[0160]

[0161] The extracted features include the following:

[0162] Numerical features: Fill in missing values ​​using the mean or median based on the feature distribution, or build a dedicated missing value prediction model.

[0163] Categorical features: Use the mode to fill in or set a special "unknown" category.

[0164] Time series features: use interpolation, forward padding, or backward padding.

[0165] For numerical features, Z-score normalization or Min-Max normalization is used; for categorical features, label encoding, one-hot encoding, or target encoding are used.

[0166] Basic features are fundamental attributes or variables directly extracted from the original data, and they typically reflect the essence of the problem most directly. Auxiliary features come from external data sources, derived calculations, or other supplementary information, and are used to enhance the model's understanding of complex relationships.

[0167] By combining basic and auxiliary features to generate new features, the interactions and correlations between different features can be captured. This allows for the discovery of complex patterns and regularities that cannot be revealed by a single feature, thereby improving the model's expressive power and prediction accuracy.

[0168] For example, if the basic features include "user age" and "purchase frequency", and the auxiliary features include "holiday markers", then cross features such as "purchase frequency of users during holidays" can be generated to reflect changes in user behavior within a specific time period.

[0169] Building upon cross-features, further feature combinations and transformations are performed to form higher-level feature representations. Techniques such as multinomial transformations, nested functions, and feature selection can be used to integrate multiple cross-features into a more representative feature set. This simplifies the feature space, reduces redundant information, and improves model training efficiency and generalization ability.

[0170] Multiple cross features (such as "purchase frequency of users on holidays" and "purchase frequency of users on weekdays") are combined into a comprehensive "user purchase activity" feature through weighted averaging or principal component analysis.

[0171] The final feature set obtained after the above feature engineering steps is used as input into the machine learning model, namely the empty shell risk prediction model, for training and prediction. It has strong interpretability and predictive power, and can comprehensively and accurately reflect the essence and key factors of the problem.

[0172] Select appropriate final features based on business needs and model type to ensure that the model achieves the best results in practical applications.

[0173] For example: Business activity index = f(invoicing activity, number of patents, recruitment activity).

[0174] Address clustering risk index = f(number of companies at the same address, company age, industry distribution).

[0175] Compliance risk is calculated as f(number of tax violations, number of administrative penalties, credit rating).

[0176] This feature engineering process systematically builds cross-features and combined features from basic and auxiliary features, ultimately forming high-quality final features that provide strong data support for machine learning models. This process not only improves model performance but also enhances the understanding and insight into the essence of the problem.

[0177] In this embodiment, the basic network architecture of the shell company risk prediction model includes an input layer, a feature processing layer, a pre-trained model transfer learning layer, a deep learning network, and an output layer. The input layer covers numerical feature input, categorical feature input, and temporal feature input; these raw data form the basis of the model analysis. The feature processing layer preprocesses the input data, including numerical feature normalization, categorical feature encoding, and temporal feature processing, and integrates them into a unified feature representation in the feature fusion module. The pre-trained model transfer learning uses a pre-trained enterprise risk model as a foundation, fine-tuning it through multiple fully connected layers to adapt to specific needs, improving model accuracy and generalization ability. The deep learning network is the core component, composed of multiple neural network layers, such as batch normalization layers, ReLU activation functions, Dropout layers, and multiple fully connected layers, ensuring the model's stability and robustness. Finally, the output layer provides two scoring results: a shell company risk index and a business activity index.

[0178] In shell company risk prediction models, introducing transfer learning can significantly improve model performance. First, a pre-trained model related to enterprise risk assessment is selected, such as an enterprise credit risk assessment model, an enterprise operational risk assessment model, or an enterprise fraud identification model. Then, transfer learning is applied when constructing the target model (shell company index). The input layer of the source model (pre-trained) receives the raw data, passes through a feature extraction layer, a first hidden layer, and a second hidden layer, and finally outputs the prediction or classification result through the output layer. The target model (shell company index) also starts from the input layer, but its data may differ from the source model. The target model directly uses the feature extraction layer, the first hidden layer, and the second hidden layer of the source model, and adjusts the output layer according to the target task. This weight transfer method can effectively utilize the learning results of the source model, especially when the target model has a small amount of data or high annotation costs.

[0179] The steps for implementing transfer learning are as follows: First, select a pre-trained model related to enterprise risk assessment. Then, adjust the target model structure to match the pre-trained model, transferring the weights of the pre-trained model to the corresponding layers of the target model. Fine-tuning strategies include freezing the first few layers of the network and training only the last few layers, fine-tuning the entire network using a small learning rate, or unfreezing the training layer by layer. To address the differences in feature distribution among shell companies, a domain adaptation layer is added. The advantages of transfer learning are improved initial model performance, reduced data requirements, enhanced generalization ability, and accelerated convergence.

[0180] The input layer processes numerical features, encoded categorical features, and time-series features of enterprises; the feature processing layer standardizes numerical features, embeds categorical features, uses RNN or attention mechanisms to process time series features, and fuses different types of features into a unified representation; the deep learning network extracts complex relationships between features through multiple fully connected layers, uses batch normalization layers to accelerate training and improve stability, introduces nonlinearity to enhance network expressive power, and uses dropout layers to prevent overfitting; the output layer provides the shell company risk index and the business activity index of enterprises, respectively.

[0181] The shell company risk prediction model employs a multi-objective optimization framework, simultaneously optimizing two key objectives: shell company risk index prediction and operational activity index prediction. The model first receives raw data as input, which is then transformed into feature representations reflecting enterprise characteristics through a shared feature extraction layer. These features are shared and utilized by subsequent task modules. The features output from the shared feature extraction layer enter the shell company risk task and the operational activity task, respectively, where each further processes and analyzes the features to complete specific predictions or assessments. The shell company risk loss and operational activity loss are weighted and combined to form the final weighted combined loss. The model is then optimized based on this, adjusting parameters to minimize the loss value and improve prediction accuracy and generalization ability. This entire process achieves efficient processing of shell company risk assessment and operational activity assessment, reducing computational resource consumption and improving the overall performance of the model.

[0182] In this embodiment, the overall loss function is defined as Loss = α·ShellRiskLoss + (1-α)·ActivityLoss; where α is a hyperparameter that controls the relative importance of the two objectives and takes values ​​in the range of [0,1].

[0183] For predicting the risk index of empty shells, a multi-class cross-entropy loss function is used: Where: N is the sample size; K is the number of risk levels (5 levels); y ij The true label (0 or 1) for sample i belonging to category j; p ij It is the probability that sample i belongs to category j as predicted by the model;

[0184] For predicting the business activity index, the mean squared error loss function is used: Among them, a i It is the true business activity index of sample i; It is a business activity index predicted by the model;

[0185] To prevent overfitting, an L2 regularization term is added: Where λ is the regularization intensity hyperparameter; It is the set of all weight parameters of the model;

[0186] Final loss function: TotalLoss=α·ShellRiskLoss+(1-α)·ActivityLoss+RegLoss.

[0187] To address the imbalanced sample problem, Focal Loss is introduced to improve the empty-shell risk loss function:

[0188] Where: γ is a modulation factor used to reduce the loss weight of easily classified samples; high-risk samples ("high" and "medium-high") are usually fewer, and their weight in training is increased by FocalLoss.

[0189] An adaptive weighting mechanism is introduced to dynamically adjust the weights based on changes in the loss of each objective during training. Where: α t α is the weight parameter in the t-th training round; α0 is the initial weight parameter; β is the rate adjustment hyperparameter; ShellRiskLoss t and ActivityLoss t These are the two loss values ​​in the t-th training round. This mechanism can automatically balance the optimization progress of the two objectives, preventing one objective from being over-optimized and neglecting the other.

[0190] Specifically, the multi-objective optimization strategy aims to balance and optimize the two key objectives of predicting the shell company risk index and the business activity index through a series of methods. The weighted balancing strategy dynamically adjusts the α value based on business needs to ensure that the optimization of both objectives reaches an ideal state. When the α value is large, the model focuses more on improving the accuracy of the shell company risk index prediction; conversely, when the α value is small, it focuses more on improving the accuracy of the business activity index prediction. To find the optimal α value, grid search and cross-validation methods are typically used.

[0191] Furthermore, a multi-stage training strategy is employed to optimize different objectives in stages, thereby achieving more accurate risk assessment and operational analysis. In the first stage, the model's primary task is to optimize shell company risk. To this end, a relatively high weighting coefficient α = 0.8 is set, allowing the model to focus on reducing losses associated with shell company risk and quickly learn relevant key features and patterns, laying the foundation for subsequent optimization. In the second stage, the model's objective shifts to balancing the optimization of shell company risk and operational activity. At this point, the weighting coefficient is adjusted to α = 0.5, indicating equal emphasis on both objectives. By optimizing both simultaneously, the model ensures that it can gradually improve its predictive ability for operational activity while maintaining effective control over shell company risk. Finally, in the third stage, the model's focus shifts to fine-tuning operational activity. By adjusting the weighting coefficient to α = 0.2, the model can further improve the predictive accuracy of operational activity while maintaining effective control over shell company risk. After these three stages of optimization, the final model version not only performs excellently in identifying and controlling shell company risk but also accurately assesses the company's operational activity, providing decision-makers with comprehensive and reliable information support. This comprehensive optimization process ensures that the model can better meet business needs in practical applications.

[0192] The training process of the shell company risk prediction model begins with batch data, which is typically a set of samples randomly selected from the entire dataset, used for a single iteration of model training. In the forward propagation phase, the model receives the batch data as input and performs calculations through the network structure to obtain the prediction results for shell company risk and business activity. The main purpose of this phase is to predict the input data using the current model parameters, providing a foundation for subsequent loss calculations. Based on the prediction results obtained from forward propagation, the shell company risk loss and business activity loss are calculated separately. These two loss values ​​reflect the model's prediction errors on the two tasks and are important indicators for evaluating model performance. Next, the shell company risk loss and business activity loss are weighted and summed to obtain the weighted total loss. The weights can be adjusted according to the importance of the tasks and the characteristics of the data to achieve optimal model training results. In the backpropagation phase, based on the calculated weighted total loss, the gradient information of the model parameters is calculated using optimization algorithms such as gradient descent. Finally, the model parameters are updated based on the gradient information obtained from backpropagation. By continuously iterating through forward propagation, loss calculation, backpropagation, and parameter updates, the model can gradually optimize its parameters and improve its predictive ability on both tasks.

[0193] The total loss comprehensively reflects the model's overall performance across all tasks. Generally, the total loss gradually decreases with increasing training epochs, indicating a continuous improvement in the model's predictive ability. However, in some cases, if the model overfits, the total loss may plateau or slightly increase after reaching a certain level. The empty shell risk loss specifically measures the model's performance on the empty shell risk prediction task. As training progresses, the empty shell risk loss also gradually decreases, indicating an increasing ability to identify empty shell risks. Similarly, the business activity loss specifically measures the model's performance on the business activity prediction task. As training progresses, the business activity loss also gradually decreases, indicating a continuous improvement in the model's ability to predict business activity. The α weight change reflects the changes in the weight coefficient α used to weight the empty shell risk loss and business activity loss across different training epochs. By dynamically adjusting the value of α, the optimization objectives between the two tasks can be flexibly balanced.

[0194] Regarding model training and evaluation, the training, validation, and test sets are first divided into three groups according to a 70%, 15%, and 15% ratio, respectively, using stratified sampling to ensure consistent distribution across categories. The entire process begins with data preparation, including collecting, cleaning, and organizing the dataset to ensure quality and usability. The feature engineering stage processes and transforms the raw data, extracting features useful for model training. After feature engineering, the pre-trained model is loaded, followed by initial model settings, including defining the model structure, initializing parameters, and setting the optimizer and loss function. The first training stage focuses on optimizing the empty shell risk task; the second stage aims to balance optimization of both tasks: empty shell risk and business activity; and the third stage focuses on fine-tuning the model to further optimize its performance on these two tasks. After three training stages, the model is evaluated to test its performance on test data, including metrics such as prediction accuracy, recall, and F1 score. During hyperparameter tuning, the optimal hyperparameter combination is found by adjusting the hyperparameters in the model. The final result is an optimized version that comprehensively considers both shell company risk and operational activity. This version effectively identifies and controls the risk of a shell company while accurately assessing its operational activity, providing decision-makers with comprehensive and reliable information support. Through gradual adjustments and optimizations, this process enables effective training of the multi-task learning model, helping it achieve better performance in complex scenarios.

[0195] The model evaluation employs a combination of indicators to comprehensively assess its performance. For the assessment of shell company risk, the macro-average F1-Score is used to evaluate the model's average performance across all risk levels; weighted precision and recall are used to assign different weights to different risk levels based on business importance; a confusion matrix is ​​used to analyze the prediction performance at each risk level in detail; and AUC-ROC is used to evaluate the model's ability to distinguish between high-risk and low-risk companies. For business activity, the root mean square error (RMSE) is used to assess the deviation between predicted and actual values, the mean absolute error (MAE) is used to assess the average degree of prediction deviation, and R... 2 The coefficient assessment model's ability to explain variations in business activity is evaluated. In the comprehensive assessment, a business-weighted score is used, designed based on the business scenario, and takes into account the business costs arising from different error types in the misclassification cost matrix.

[0196] For hyperparameter tuning, Bayesian optimization methods are used to fine-tune model hyperparameters, mainly involving key parameters such as learning rate, batch size, number and width of network layers, Dropout ratio, α weight coefficient, Focal Loss parameter γ, and regularization strength λ. The model deployment and application architecture begins with the data source, including internal enterprise data, external public data, and third-party data, ensuring its accuracy and completeness. This is followed by the feature processing service stage, which preprocesses and extracts features from the raw data to improve data quality and model learning efficiency. After feature processing, the model inference service uses the trained model to predict and analyze the input data, obtaining results such as shell company risk scores and business activity scores. The result storage service is responsible for persistently storing these prediction results, while the business application interface serves as a bridge connecting business systems and the model service, enabling various business applications such as pre-loan risk control, bidding review, market supervision, and due diligence for enterprise cooperation.

[0197] The real-time prediction process comprises five steps: data acquisition, feature processing, model inference, result output, and result application, ensuring real-time response to the latest information changes of enterprises. To maintain model effectiveness, three strategies are employed: periodic full updates, incremental learning updates, and triggered updates. This method not only effectively identifies shell company risks but also provides crucial decision support for financial risk prevention and control, regulation, and enterprise cooperation, demonstrating significant practical value and broad prospects for wider application.

[0198] This embodiment integrates heterogeneous data from multiple sources, including business registration information, tax information, intellectual property information, and internet activity information, to comprehensively depict the operational status of enterprises, significantly improving the accuracy and comprehensiveness of shell company identification. Secondly, it employs a five-level risk stratification method, which, compared to traditional binary classification methods, more accurately reflects the hierarchical differences in shell company risk, providing strong support for differentiated risk management. Furthermore, it introduces transfer learning, utilizing pre-trained enterprise risk model knowledge to significantly reduce reliance on large-scale labeled data. Simultaneously, by optimizing both the shell company risk index and the operational activity index, and introducing an adaptive weighting mechanism, it achieves synergistic improvement of both objectives. Deep learning network structures can capture complex nonlinear relationships between features, particularly excelling in identifying "pseudo-active" enterprises. Through a model update strategy combining regular full updates and incremental learning, the model can adapt promptly to changes in enterprise operational status and policy environment adjustments. This method is widely applicable in various scenarios such as financial risk control, bidding review, market supervision, and due diligence for enterprise cooperation, demonstrating strong adaptability and high application value. Finally, by introducing feature importance analysis and local interpretation techniques, the interpretability of the model's prediction results was enhanced, thereby improving the model's credibility and transparency.

[0199] A multi-level rule-based sample labeling method is proposed to achieve high-quality sample labeling through hierarchical judgment rules. A comprehensive feature engineering system covering eight major categories of features, including basic business registration, tax credit, and business activity, is constructed. An innovative enterprise risk pre-training model is introduced, and a network architecture with a feature processing layer, a deep learning network layer, and a dual-output layer is designed. A multi-objective optimization framework for simultaneously optimizing the shell risk index and the business activity index is designed, and an adaptive weighting mechanism is developed. A three-stage training strategy is proposed, and Bayesian optimization method is used to fine-tune the model hyperparameters. A model deployment architecture that supports real-time prediction is designed, and a model update strategy that combines regular full updates with incremental learning is developed.

[0200] In one embodiment, for sample labeling, an expert-based scoring card approach or unsupervised clustering method can be used; for feature engineering, graph neural networks or automatic feature generation methods can be introduced; for network structure, attention mechanisms, capsule networks, or Transformer architectures can be used; for loss function design, ranking-based loss functions or contrastive learning loss can be used; for training strategies, active learning strategies or adversarial training methods can be used; and for deployment, federated learning or edge computing modes can be used. These alternatives can be selected and combined according to factors such as actual application scenarios, data availability, and computing resources to meet the needs of different application environments.

[0201] S140. Output the prediction result.

[0202] The prediction results are output to the terminal for display.

[0203] The aforementioned method for predicting shell company risks integrates multi-source heterogeneous data and performs systematic feature engineering on this data, including feature extraction, transformation, cross-fertilization, combination, and selection, to comprehensively capture the complex characteristic relationships of enterprises. Then, it utilizes deep learning networks to train enterprise-related data labeled with risk levels, establishing an accurate shell company risk prediction model. This method not only significantly improves the accuracy, comprehensiveness, and flexibility of shell company risk prediction but also supports more refined risk management and adapts to a wider range of application scenarios. Compared to existing technologies, it effectively solves the problems of insufficient integration of multi-source data and failure to deeply explore complex characteristic relationships, thereby greatly enhancing the accuracy and reliability of the assessment. Finally, by outputting intuitive prediction results, it provides users with strong decision support.

[0204] Figure 4 This is a schematic block diagram of a shell company risk prediction device 300 provided in an embodiment of the present invention. Figure 4 As shown, corresponding to the above-described method for predicting the risk of shell companies, the present invention also provides a device 300 for predicting the risk of shell companies. This device 300 includes a unit for executing the above-described method for predicting the risk of shell companies, and the device can be configured in a server. Specifically, please refer to... Figure 4 The shell company risk prediction device 300 includes a data acquisition unit 301, a feature engineering processing unit 302, a prediction unit 303, and an output unit 304.

[0205] The data acquisition unit 301 is used to acquire enterprise-related data; the feature engineering processing unit 302 is used to perform feature engineering processing on the enterprise-related data to obtain processing results; the prediction unit 303 is used to input the processing and results into the shell company risk prediction model to perform shell company risk prediction to obtain prediction results; and the output unit 304 is used to output the prediction results.

[0206] The shell company risk prediction model is obtained by training a deep learning network with features obtained from several enterprise-related data with risk level labels through feature engineering.

[0207] In one embodiment, the feature engineering processing unit 302 includes a cleaning subunit, an extraction subunit, a cross-combination subunit, and a selection subunit.

[0208] The data cleaning subunit is used to clean the relevant data of the enterprise to obtain cleaned data; the extraction subunit is used to extract and transform features from the cleaned data to obtain extraction results; the cross-combination subunit is used to cross-combine features from the extraction results to obtain combination results; and the selection subunit is used to select features from the combination results to obtain processing results.

[0209] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned shell company risk prediction device 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0210] The aforementioned shell company risk prediction device 300 can be implemented as a computer program, which can, for example... Figure 5 It runs on the computer device shown.

[0211] Please see Figure 5 , Figure 5 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0212] See Figure 5 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0213] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a method for predicting corporate shell risks.

[0214] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0215] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for predicting the risk of shell companies.

[0216] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0217] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the above-described method for predicting the risk of shell companies.

[0218] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0219] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0220] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all the steps of the above-described method for predicting the risk of shell companies.

[0221] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0222] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0223] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0224] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0225] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0226] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the risk of shell companies, characterized in that, include: Obtain relevant enterprise data; The relevant enterprise data is subjected to feature engineering processing to obtain the processing result; The processing and results are input into the shell company risk prediction model to predict the shell company risk and obtain the prediction results. Output the prediction result; The shell company risk prediction model is obtained by training a deep learning network with features obtained from several enterprise-related data with risk level labels through feature engineering.

2. The method for predicting the risk of shell companies according to claim 1, characterized in that, The process of performing feature engineering on the enterprise-related data to obtain the processing result includes: The relevant data of the enterprise is cleaned to obtain cleaned data; The cleaned data is then subjected to feature extraction and transformation to obtain the extraction results; The extracted results are subjected to feature cross-interaction and combination to obtain a combined result; Feature selection is performed from the combined results to obtain the processing result.

3. The method for predicting the risk of shell companies according to claim 1, characterized in that, The empty shell risk prediction model includes an input layer, a feature processing layer, a deep learning network layer, and an output layer arranged in sequence; the deep learning network layer includes multiple fully connected layers, a batch normalization layer, an activation function, and a Dropout layer.

4. The method for predicting the risk of a shell company according to claim 3, characterized in that, The prediction results include a shell company risk index and an operational activity index.

5. The method for predicting the risk of a shell company according to claim 4, characterized in that, The overall loss function used in training the shell company risk prediction model is an indicator that comprehensively evaluates the model's prediction accuracy and generalization ability by adjusting the relative weights of shell company risk loss and operational activity loss, and adding an L2 regularization term.

6. The method for predicting the risk of a shell company according to claim 5, characterized in that, The overall loss function used during the training of the shell risk prediction model determines the weight values ​​of the shell risk loss through grid search and cross-validation.

7. The method for predicting the risk of a shell company according to claim 6, characterized in that, During the training of the empty shell risk prediction model, Bayesian optimization techniques are used to adjust the hyperparameters, which include the learning rate and batch size.

8. A device for predicting the risk of shell companies, characterized in that, include: The data acquisition unit is used to acquire relevant enterprise data. The feature engineering processing unit is used to perform feature engineering processing on the enterprise-related data to obtain the processing result; The prediction unit is used to input the processing and results into the shell company risk prediction model to predict the shell company risk and obtain the prediction result. An output unit is used to output the prediction result; The shell company risk prediction model is obtained by training a deep learning network with features obtained from several enterprise-related data with risk level labels through feature engineering.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images