Enterprise health degree evaluation method based on deep learning
Through the deep learning-based enterprise health evaluation method, combined with financial and non-financial indicators, a multi-level evaluation model is established, which solves the limitations of qualitative analysis in the evaluation of small and micro enterprises, realizes a comprehensive and objective assessment of enterprises, and provides More accurate evaluation results.
Patent Information
- Application Number
- CN202410815301.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-24
- Publication Date
- 2024-11-22
AI Technical Summary
The existing technology lacks diagnostic evaluation technology based on artificial intelligence in enterprise evaluation, especially self-diagnosis and government supervision tools for small and micro enterprises, and the evaluation method focuses on qualitative analysis and cannot comprehensively and objectively evaluate the health of enterprises.
Using the enterprise health evaluation method based on deep learning, by collecting financial and non-financial indicators, standardized processing and single-hot encoding, establishing evaluation models including embedding layer, feature extraction layer, long and short-term memory network layer and classification layer, and conducting multiple evaluation models Angle evaluation.
It achieves a comprehensive and objective evaluation of small and micro enterprises, provides more accurate and comprehensive evaluation results, overcomes the limitations of traditional qualitative analysis, and can better understand the overall situation of the enterprise.
Smart Images

Figure CN119005727A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of enterprise evaluation and relates to an enterprise health evaluation method based on deep learning. Background Art
[0002] At present, the technologies for enterprise evaluation and diagnosis systems at home and abroad are mainly concentrated on the level of enterprise credit evaluation, and the main application direction is also mainly investment and financing by financial institutions. Since small and micro enterprises are often established for a short time and their financial data is missing and irregular, the main feature of the evaluation model they developed focuses on qualitative analysis. Basically, enterprise diagnosis and evaluation technology based on artificial intelligence technology has not been established, and there is a lack of tools and applications for enterprises to self-diagnose and assist the government in supervising small and micro enterprises. Summary of the invention
[0003] The purpose of the present invention is to provide a method for evaluating enterprise health based on deep learning to solve the problems existing in the prior art: To this end, the present invention adopts the following technical solutions: A method for evaluating enterprise health based on deep learning, characterized by comprising the following steps: (1) Collecting basic information of the enterprise, including financial indicators and non-financial indicators, and then scoring the basic information to obtain enterprise health data, including continuous features and discrete features; (2) The continuous characteristics, business indicators and other continuous variables in health are standardized according to the following formula to obtain the standardized characteristic value: In the formula, x is the original eigenvalue, μ is the mean of the eigenvalue, and σ is the standard deviation of the eigenvalue; (3) Perform one-hot encoding on the discrete features in the enterprise health data to obtain a sparse matrix; (4) Divide the dataset into training set, validation set and test set. The dataset includes standardized feature values and sparse matrices; (5) Establishing an evaluation model, wherein the evaluation model includes an embedding layer, a feature extraction layer, a feature aggregation layer, a long short-term memory network layer, and a classification layer. Specifically: The embedding layer is used to encode the input data set and convert it into a dense embedding vector representation; The feature extraction layer can use convolution kernels of different sizes to simultaneously normalize the feature values Perform convolution operations to capture local features of different sizes; The feature aggregation layer is used to summarize and aggregate each local feature by using an average aggregation operation to obtain a comprehensive feature representation; The long short-term memory network layer is used to capture dependencies in the comprehensive feature representation; The classification layer is used to send the comprehensive feature representation processed by the long short-term memory network layer to the fully connected layer for classification and map the features to the prediction results of the final category space generation model; (6) Use the training set to train the evaluation model, then use the test set to test the evaluation model, and finally use the validation set to detect the test results.
[0004] Furthermore, the financial indicators in step (1) include debt-paying ability, operating ability and profitability, and the non-financial indicators include the quality of the enterprise's leaders, the quality of the enterprise, the policy environment and cooperative relationships.
[0005] Furthermore, in step (1), missing values and outliers in the collected basic information are removed, and outliers and erroneous values are corrected.
[0006] Furthermore, in step (4), the ratio of the training set, validation set and test set is 70:15:15.
[0007] The beneficial effects of the present invention are: According to the characteristics of my country's small and micro enterprises, and using factor analysis, financial and non-financial factors are analyzed, and then rewritten through an artificial intelligence model to evaluate and diagnose enterprises from multiple key dimensions such as debt capacity, operating capacity and profitability, quality of corporate leaders, corporate quality, policy environment and cooperative relationships. This successfully solves the problems existing in enterprise evaluation in previous technologies, especially the limitations caused by the emphasis on qualitative analysis. Compared with traditional evaluation methods that may be subject to the inherent problem of qualitative analysis bias when dealing with small and micro enterprises, this method achieves a comprehensive and objective evaluation. Based on considerations of multiple key dimensions, the model can understand the overall situation of the enterprise more comprehensively and from multiple angles, thereby providing more accurate and comprehensive evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 It is a structural schematic diagram of the evaluation model of the present invention; Figure 2 This is the accuracy graph of the model test; Figure 3 This is the loss curve of the model. DETAILED DESCRIPTION
[0009] The present invention is described in detail below in conjunction with embodiments and drawings: (1) Collecting basic information of the enterprise, including financial indicators and non-financial indicators, and then scoring the basic information to obtain enterprise health data. The health data includes continuous characteristics and discrete characteristics. In this embodiment, financial indicators include debt repayment ability, operating ability and profitability. If the above abilities are good, a higher score can be given, and if the above abilities are poor, a lower score can be given. Non-financial related information includes the quality of the enterprise leader, the enterprise quality, the policy environment and the cooperative relationship. Then, the missing values and outliers collected in the enterprise health data are removed, and the outliers and erroneous values are corrected to prevent affecting the evaluation accuracy.
[0010] (2) The continuous features in the enterprise health data are standardized according to the following formula. The continuous features specifically refer to various indicators in debt repayment ability, operating ability and profitability (such as asset-liability ratio, net profit growth rate, main business growth rate, accounts receivable turnover, etc.), and the standardized feature value is obtained: , standardized eigenvalues Calculated according to the following formula: In the formula, x is the original eigenvalue, μ is the mean of the original eigenvalues, σ is the standard deviation of the original eigenvalues, μ and σ They are all calculated using existing conventional methods. The value of each feature is subtracted from its mean and then divided by its standard deviation to achieve standardization. The standardized feature value will have the property of mean 0 and standard deviation 1, so that the value of the data feature is scaled to a standard normal distribution with mean 0 and standard deviation 1, thereby eliminating the dimensional differences between different features, making the model easier to converge, improving the model training speed, and better avoiding the problem of unstable model training caused by different feature value ranges.
[0011] (3) Perform one-hot encoding on the discrete features in the enterprise health data to obtain a sparse matrix.
[0012] (4) The data set is divided into a training set, a validation set, and a test set. The ratio of the training set, the validation set, and the test set is 70:15:15. Specifically, in this embodiment, the data set includes standardized feature values and the contents of the sparse matrix.
[0013] (5) Establish an evaluation model, which includes an embedding layer, a feature extraction layer, a feature aggregation layer, a self-attention mechanism layer, and a classification layer. Specifically: The embedding layer is used to convert the input index sequence into a dense embedding vector representation so that these vectors have a fixed dimension. This step is done by pytorch In nn.Embedding To achieve this, the input sequence data is passed through the embedding layer ( Embedding layer ) is used to encode the data into a fixed-dimensional vector representation. Specifically, Embedding The layer can be used to convert some important discrete features, such as the industry category, geographical location, business model and other information of the enterprise, into continuous dense vector representations. These vector representations can capture the semantic relationship and similarity between different features, so that the model can better understand and utilize these features.
[0014] The feature extraction layer is a multi-branch convolutional structure. The feature extraction layer can use convolution kernels of different sizes for convolution operations to capture local features of different sizes, which helps the model learn more global and local information. Each branch focuses on capturing different aspects of the input sequence. This multi-branch design can more effectively extract information from the input sequence and improve the performance of the model.
[0015] The feature aggregation layer is used to summarize and aggregate the features extracted by each branch, and fuse the features extracted by each branch of the feature extraction layer together. In this model, a simple averaging operation is used to aggregate the features extracted by each branch to form a comprehensive feature representation. The averaging operation is easy to implement and does not require additional parameters. It can evenly fuse the features extracted by each branch, so that the model can integrate the information of each branch. Compared with other complex feature aggregation methods, such as weighted summation or attention mechanism, the computational cost of the averaging operation is lower, which helps to reduce the computational burden of the model, and can accelerate the training and reasoning process of the model and improve efficiency.
[0016] The LSTM network layer can input the comprehensive feature representation into the LSTM network to model the input data and capture the dependencies in the comprehensive feature representation. The main role of the LSTM network layer in the model is to model the input data, especially to process the complex patterns and long-term dependencies in the input data. By introducing memory units and gating mechanisms, the LSTM network layer can selectively remember or forget the information in the input sequence to ensure that important features are retained throughout the sequence. This enables the model to capture the temporal dynamic changes and dependencies in the data, thereby more accurately understanding and representing the feature information in the input data. After the input feature information representation is processed by the LSTM layer, the model can classify or predict more effectively, improving overall performance and accuracy.
[0017] The classification layer is used to send the comprehensive feature representation processed by the long short-term memory network layer to the fully connected layer ( Fully connected layer ) for classification, the fully connected layer maps the features to the final category space and generates the prediction results of the model.
[0018] (6) Finally, the training set is used to train the evaluation model, and then the test set is used to test the evaluation model. Finally, the test results are tested on the validation set. After the test, the model's classification accuracy for the data has reached 98.57%. The model test accuracy and loss curve Figure 2 and Figure 3 shown.
Claims
1. A method for evaluating enterprise health based on deep learning, characterized in that: The following steps are involved: (1) Collecting basic information of the enterprise, including financial indicators and non-financial indicators, and then scoring the basic information to obtain enterprise health data, including continuous features and discrete features; (2) The continuous characteristics, business indicators and other continuous variables in health are standardized according to the following formula to obtain the standardized characteristic value: : In the formula, x is the original eigenvalue, μ is the mean of the eigenvalue, σ is the standard deviation of the eigenvalues; (3) Perform one-hot encoding on the discrete features in the enterprise health data to obtain a sparse matrix; (4) Divide the dataset into training set, validation set and test set. The dataset includes standardized feature values and sparse matrices; (5) Establishing an evaluation model, wherein the evaluation model includes an embedding layer, a feature extraction layer, a feature aggregation layer, a long short-term memory network layer, and a classification layer. Specifically: The embedding layer is used to encode the input data set and convert it into a dense embedding vector representation; The feature extraction layer can use convolution kernels of different sizes to simultaneously normalize the feature values Perform convolution operations to capture local features of different sizes; The feature aggregation layer is used to summarize and aggregate each local feature by using an average aggregation operation to obtain a comprehensive feature representation; The long short-term memory network layer is used to capture dependencies in the comprehensive feature representation; The classification layer is used to send the comprehensive feature representation processed by the long short-term memory network layer to the fully connected layer for classification and map the features to the prediction results of the final category space generation model; (6) Use the training set to train the evaluation model, then use the test set to test the evaluation model, and finally use the validation set to detect the test results.
2. According to the enterprise health evaluation method based on deep learning in claim 1, it is characterized in that: The financial indicators in step (1) include debt repayment ability, operating ability and profitability, and the non-financial indicators include the quality of enterprise leaders, enterprise quality, policy environment and cooperative relations.
3. According to the enterprise health evaluation method based on deep learning in claim 1, it is characterized in that: In the step (1), missing values and outliers in the collected basic information are removed, and outliers and erroneous values are corrected.
4. According to the enterprise health evaluation method based on deep learning in claim 1, it is characterized in that: The division ratio of the training set, validation set and test set in step (4) is 70:15:15.
