An intelligent management method for additional deductions based on machine learning algorithms

Abstract

The present invention discloses an intelligent management method for additional deductions based on a machine learning algorithm, which relates to the field of tax management technology. The system is integrated with a distributed file system through the JDBC / ODBC protocol to realize the automatic collection of R&D expense data, including structured financial data and unstructured documents. The preprocessing stage adopts the sliding window mean to fill missing values, Z-score standardization and One-Hot encoding technology. The feature engineering module calculates feature gain based on information entropy and realizes dimensionality reduction in combination with linear discriminant analysis. The classification model adopts the random forest algorithm, generates a decision tree through Bootstrap sampling, and node splitting is based on the Gini coefficient. The dynamic optimization module uses the gradient descent algorithm to update parameters, combined with L2 regularization to improve the generalization ability of the model. Through multi-stage feature optimization and dynamic parameter adjustment, the system realizes efficient and accurate judgment of the compliance of R&D expenses and supports rapid iteration of models under dynamic policy adjustment.

Description

Technical Field

[0001] The present invention belongs to the field of tax informatization technology, and specifically relates to an intelligent management system that integrates multimodal data processing, adaptive feature engineering and dynamic model optimization. Background Art

[0002] There are three major technical pain points in the implementation of the current additional deduction policy:

[0003] Data heterogeneity: R&D expense data is scattered across multiple data sources, including ERP systems (such as SAP and Oracle), project management platforms (such as Jira), and financial systems (such as Kingdee). Format differences make data integration difficult.

[0004] Feature complexity: A typical R&D expense dataset contains more than 80 feature dimensions, and traditional feature selection methods (such as the chi-square test) cannot effectively extract policy-sensitive features.

[0005] Model timeliness: Policy adjustments (such as the newly added "additional deduction for artificial intelligence R&D expenses" clause in 2024) require models to have rapid iteration capabilities, and traditional batch training methods are difficult to meet the needs.

[0006] Existing technologies have significant shortcomings: The hybrid model proposed in Rule-based Tax Compliance with Machine Learning (2022) still relies on manual rule annotation. The deep model in Deep Learning for Tax Fraud Detection (2023) is prone to overfitting in small sample scenarios. Existing systems lack effective utilization of the time series characteristics of R&D expense data. Summary of the Invention

[0007] Based on the above-mentioned deficiencies of the existing technologies, this method constructs a multi-level technical architecture to achieve the automated collection of R&D expense data (including structured database interface and unstructured document parsing), dimensionality reduction optimization of high-dimensional features (combining information theory and statistical learning methods), and dynamic parameter adjustment of the classification model (parameter update mechanism based on the gradient descent algorithm), and finally outputs the R&D expense compliance judgment results that comply with the latest tax policies.

[0008] The technical solution of the present invention is: an intelligent management method for additional deductions based on a machine learning algorithm, comprising the following steps:

[0009] Step 1: Establish an enterprise R&D activity data collection module, acquire multi-source heterogeneous data through structured database interfaces and distributed file systems, and perform data cleaning and standardization preprocessing. This module establishes a connection with a relational database through the JDBC / ODBC protocol and uses the Flume data flow framework to collect unstructured data such as R&D project documents, financial vouchers, and patent applications in real time. Regular expression matching technology is used to clean missing values ​​and outliers, and the Z-score standardization method is used to convert numerical data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0010] Step 2: Build a feature engineering module. Use an information entropy-based feature selection algorithm to screen key R&D expense indicators and perform feature space dimensionality reduction through linear discriminant analysis. This algorithm evaluates the information content of each feature by calculating the Shannon entropy formula, selects the top k features whose information gain exceeds a threshold, then constructs inter-class and intra-class scatter matrices. By solving the generalized eigenvalue problem to obtain the optimal projection direction, the original feature space is mapped to a d-dimensional subspace.

[0011] Step 3: Deploy a random forest classification model, set the number of decision trees to N, use the Gini coefficient as the node splitting criterion, and establish a rule for determining R&D expense compliance. The model generates M training subsets through bootstrap sampling. During training, each decision tree randomly selects mtry features for node splitting. The splitting effect is evaluated based on the Gini impurity formula, and the prediction results of all trees are finally integrated through a majority voting mechanism.

[0012] Step 4: Dynamically adjust the classification model parameters based on the gradient descent optimization algorithm to output a list of R&D expenses that meet the additional deduction policy and risk warning results; the algorithm defines the cross-entropy loss function, calculates the gradient through backpropagation, and uses the stochastic gradient descent update rule with momentum to dynamically adjust the learning rate and momentum parameters. The final output is a structured list containing expense item codes, policy matching scores, and red / yellow warning signs based on confidence thresholds.

[0013] Furthermore, the data preprocessing in step 1 includes:

[0014] Step a: Missing value filling

[0015] When preprocessing enterprise R&D activity data, missing values ​​in the data are filled using the sliding window mean method of the same enterprise's historical data. Specifically, with the time point where the current missing value is located as the center, a certain number of data points before and after are selected to form a sliding window. The mean of the historical data of the same enterprise within the window is calculated and used as the filling value for the missing value.

[0016] Step b: Standardize numeric fields

[0017] For numeric fields, Z-score standardization is implemented to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is:

[0018]

[0019] in:

[0020] x′ is the normalized data value;

[0021] x is the original data value;

[0022] Hong is the mean of all data in the numeric field;

[0023] σ is the standard deviation of all data in the numeric field;

[0024] Step c: Encoding categorical variables

[0025] For categorical variables, One-Hot encoding is used to generate a sparse matrix. One-Hot encoding is a method for converting categorical variables into numerical variables. Each categorical variable value is converted into a binary vector. The length of the vector is equal to the number of categorical variable values, in which only the corresponding value position is 1 and the rest of the positions are 0.

[0026] In the specific operation, for each categorical variable, all its possible values ​​are determined, and then a new binary feature is created for each value; in the original data set, for each sample, if its categorical variable takes a certain value, the corresponding new feature is 1, and the rest of the new features are 0.

[0027] Furthermore, the feature selection algorithm based on information entropy in step 2 is specifically as follows:

[0028] In the intelligent management system for additional deductions based on machine learning algorithms, for a given feature set

[0029] F={f1,f2,…,f n}

[0030] It is necessary to calculate the information gain value of each feature to evaluate the importance of the feature for classification; the calculation formula of the information gain value is:

[0031]

[0032] in:

[0033] IG(F|C): represents the information gain of feature F relative to category C;

[0034] H(C): is the entropy of category C, also known as category entropy, and its calculation formula is;

[0035]

[0036] Among them, C is the category set, that is, the set of all possible categories;

[0037] c is a specific category in set C;

[0038] p(c) is the probability that category c appears in the sample set S, which is calculated as where |s c | is the number of samples of category c, |S| is the total number of samples in the sample set S; slice (C) is the category entropy that measures the degree of uncertainty of category C in the absence of any feature information;

[0039] V: is the value set of feature F, that is, the set of all possible values ​​of feature F;

[0040] S: represents the sample set, which is the collection of all samples used to analyze and train the model;

[0041] S v : is the sample subset in the sample set S where the feature F takes the value v,

[0042] |S v | represents the number of samples in the subset;

[0043] H(C|S v ): is a subset of samples where the feature F takes the value v

[0044] S v The conditional entropy of category C in is calculated as

[0045]

[0046] in

[0047] p(c|S v ) is in the sample subset S v The probability of category c appearing in is calculated as

[0048] |S cv | is the sample subset S v The number of samples of category c;

[0049] Conditional entropy H(C|S v ) represents the degree of uncertainty of category C when the value of feature F is known to be v;

[0050] is the conditional entropy H(C|S v ) is weighted summed, and the weight is That is, the sample subset S vThe proportion of the sample set S;

[0051] After calculating the information gain value of each feature, the features that satisfy IG(F|C)≥θ are screened out. These features constitute the optimized feature subset; where θ is the preset threshold, and the setting of this threshold needs to be determined according to specific business needs and data characteristics.

[0052] Furthermore, the decision tree generation method of the random forest classification model in step 3 is:

[0053] In step 3 of the machine learning-based intelligent management system for additional deductions, the random forest classification model generates a decision tree using a specific method. Bootstrap sampling (a random sampling method with replacement) is performed on the original training set to generate N sub-training sets. The following operations are then recursively performed on each sub-training set until the node purity reaches a preset threshold or the number of samples is insufficient.

[0054] Step a: Randomly select candidate features

[0055] Randomly select from d features candidate features; in actual operation, a random number generator is used to select k features from the entire feature set as candidate features for the current node split;

[0056] Step b: Calculate the Gini coefficient of the candidate feature

[0057] For each candidate feature, calculate its Gini coefficient; the calculation formula of the Gini coefficient is:

[0058]

[0059] in:

[0060] Gini(D): represents the Gini coefficient of the dataset D, which measures the impurity of the dataset;

[0061] c: is the number of categories, that is, the total number of all possible categories in the dataset;

[0062] p i : is the sample proportion of category i, calculated as where |D i | is the number of samples of category i in the dataset, |D| is the total number of samples in dataset D;

[0063] Step c: Select features and split points for node splitting

[0064] After calculating the Gini coefficient of each candidate feature, select Gini(D A ) The smallest feature and segmentation point are used for node splitting; here DA It is a subset obtained by dividing according to a certain feature and split point. For continuous features, all possible split points will be traversed and the Gini coefficient of the subset divided at each split point will be calculated.

[0065] For discrete features, all possible value combinations will be considered for division; by selecting Gini (D A ) The minimum features and split points can make the purity of the divided subsets as high as possible, thereby building a more effective decision tree;

[0066] Through the above method, a decision tree is generated for the random forest classification model, and finally the classification prediction result is obtained by combining the results of multiple decision trees, which is used to determine the compliance of R&D expenses in the additional deduction intelligent management system.

[0067] Furthermore, the parameter update formula of the gradient descent optimization algorithm in step 4 is:

[0068] In step 4 of the intelligent management system for additional deductions based on machine learning algorithms, the model parameters are updated using a gradient descent optimization algorithm; this algorithm continuously iterates to find the model parameter values ​​that minimize the objective function;

[0069] The parameter update formula is:

[0070]

[0071] in:

[0072] W t+1 : represents the model parameters at the t+1th iteration; the model parameters w at the current iteration number t t Update the objective function based on its gradient, gradually approaching the parameter value that minimizes the objective function.

[0073] w t : is the model parameter of the tth iteration; in each iteration, it is adjusted according to the gradient information of the objective function to achieve the purpose of optimizing the model;

[0074] η: is the learning rate, which controls the step size of parameter update at each iteration;

[0075] is the objective function J(w) when w=w t The gradient at ; the gradient is a vector, its direction indicates the direction in which the objective function value increases fastest, and its opposite direction indicates the direction in which the objective function value decreases fastest;

[0076] The calculation formula of the objective function J(w) is:

[0077]

[0078] in:

[0079] m: is the number of samples, that is, the total number of samples used to train the model;

[0080] h w (x): is the model prediction value, which is the result of predicting the input x based on the current model parameter w; for different machine learning models, h w The specific form of (x) will vary;

[0081] x (i) : represents the input feature vector of the i-th sample;

[0082] y (i) : is the true label value of the i-th sample;

[0083] (h w (x (i) )-y (i) ) 2 Measures the square error between the model's predicted value and the true value for the i-th sample;

[0084] This part is the mean squared error (MSE) loss function, which calculates the average of the squared prediction errors of all samples and is used to measure how well the model fits the training data;

[0085] λ: L2 regularization coefficient, which controls the weight of the regularization term;

[0086] ||w||2: is the L2 norm of the model parameter w, calculated as in is the jth component of the parameter vector w;

[0087] By continuously iteratively updating the model parameters w, the objective function J(w) is gradually reduced, and finally a set of optimal model parameters is obtained.

[0088] Furthermore, the projection matrix of the linear discriminant analysis is calculated as:

[0089] In the intelligent management system for additional deductions based on machine learning algorithms, when using linear discriminant analysis to reduce the dimensionality of the feature space in step 2, it is necessary to calculate the projection matrix W. The goal of linear discriminant analysis is to find a projection matrix W that separates data of different categories as much as possible after projection, while clustering data of the same category as much as possible.

[0090] The projection matrix W is calculated by maximizing the following objective function:

[0091]

[0092] in:

[0093] W is the projection matrix, which projects the original high-dimensional feature space into a low-dimensional space;

[0094] S b It is the inter-class scatter matrix, which is used to measure the degree of dispersion between different categories; its calculation formula is:

[0095]

[0096] c is the number of categories, that is, the total number of all possible categories in the dataset;

[0097] n i is the number of samples in the i-th category, reflecting the size of the samples in the data set;

[0098] Hongi is the mean vector of the i-th class sample,

[0099] Hong is the global mean vector,

[0100] Inter-class divergence matrix S b The derivation of is based on the measurement of the difference between the mean values ​​of different categories. By calculating the deviation of the mean of each category from the global mean and taking into account the weight of the number of samples, a quantitative representation of the degree of dispersion between classes is obtained.

[0101] s w It is the intra-class scatter matrix, which is used to measure the degree of dispersion of data within the same category; its calculation formula is For each class of samples, calculate the deviation of each sample from the class mean, and then sum it up for all classes to get the intra-class scatter matrix; the intra-class scatter matrix reflects the closeness of samples in the same class;

[0102] Objective function The molecule |W T S b W| represents the degree of discreteness between different categories after projection, and the denominator |W T S w W| represents the degree of discreteness within the same category after projection. By maximizing this objective function, a projection matrix W can be found that maximizes the separation between different categories after projection and the aggregation within the same category, thereby achieving effective feature space dimensionality reduction.

[0103] In actual calculations, it is usually done by solving the generalized eigenvalue problem S b w=λS w w is used to obtain the column vectors of the projection matrix W, which are the eigenvectors corresponding to the largest eigenvalues.

[0104] The technical innovation of this invention is reflected in the following key modules:

[0105] Data acquisition and preprocessing subsystem

[0106] Multi-source data integration:

[0107] Structured data: Connect to a relational database via JDBC 4.2 protocol and use Sqoop 1.4.7 for incremental extraction

[0108] Unstructured data: Use Apache Tika 2.9.0 to parse PDF contracts and use regular expressions to extract key information such as "R&D personnel list" and "equipment lease term"

[0109] Real-time data stream: Build a message queue based on Kafka 2.8.1 to achieve a processing throughput of 5,000 data items per minute

[0110] Data cleaning technology:

[0111] Sliding window filling algorithm

[0112] The window size is automatically adjusted according to the data frequency (w=7 for daily data, w=3 for monthly data)

[0113] Outlier detection: Based on the IQR method, the calculation formula is:

[0114] Q1=25 th percentile, Q3=75 th

[0115] percentile IQR=Q3-Q1, Lower bound=Q1-1.5× IQR

[0116] Upper bound=Q3+1.5× IQR

[0117] Feature engineering optimization subsystem

[0118] Information entropy feature selection:

[0119] Implementation steps:

[0120] Calculate category entropy: slice (C) = -∑ c∈C p(c)log2p(c)

[0121] Calculate conditional entropy: slice (C|S v )=-∑ c∈C P(c|S ， )log2p(c|S v )

[0122] Information gain calculation:

[0123] Genetic algorithm is used to optimize the threshold θ, with a value range of [0.1, 0.3]

[0124] LDA dimensionality reduction implementation:

[0125] Between-class scatter matrix:

[0126] Within-class scatter matrix:

[0127] Generalized eigenvalue solution: S b w=γS w w takes the eigenvectors corresponding to the first d largest eigenvalues ​​to form the projection matrix

[0128] Classification model building subsystem

[0129] Random forest parameter configuration:

[0130] Boot strap sampling: Each sub-training set contains 60% of the original samples

[0131] Feature subset size: (d is the number of features after dimensionality reduction)

[0132] Decision tree depth: Dynamic adjustment strategy is used, with an initial depth of 5 and one layer added each time the split is completed until the Gini gain is less than 0.03

[0133] Node splitting criteria:

[0134] Gini coefficient calculation:

[0135]

[0136] Splitting conditions:

[0137]

[0138] Dynamic optimization subsystem

[0139] Objective function design:

[0140] The regularization term coefficient λ is determined by Bayesian optimization, and the search space is [0.001, 0.1]

[0141] Loss function:

[0142]

[0143] Parameter update algorithm:

[0144] The momentum term coefficient γ = 0.9, and the learning rate uses exponential decay: η t =η0×0.95 t / 100

[0145] Gradient calculation:

[0146]

[0147] Compared with the prior art, the present invention has the following beneficial effects:

[0148] 1. Multi-source heterogeneous data fusion capabilities improve R&D expense review efficiency

[0149] This invention achieves real-time integration of structured financial data and unstructured R&D documents by building a hybrid data collection system. Specific advantages include:

[0150] Cross-system data connectivity: Using the JDBC 4.2 protocol to connect to mainstream ERP systems (such as SAP S / 4HANA) and combined with the Apache Flume distributed log collection framework, it can simultaneously process relational databases such as MySQL and Oracle, as well as unstructured data such as R&D contract PDFs and patent applications stored in HDFS, overcoming the limitation of traditional systems that only support a single data source.

[0151] Real-time data stream processing: Building a data pipeline based on Kafka message queues achieves a throughput of 5,000 data items per minute, a 60% improvement in efficiency compared to traditional ETL tools (such as Informatica), ensuring the timeliness of R&D expense data. For example, an electronics company used this system to reduce the cycle time from expense data generation to analysis from 72 hours to 2 hours.

[0152] Multimodal data parsing: Using Apache Tika 2.9.0 to parse PDF contracts, we combined regular expressions to extract key information such as "R&D personnel social security payment certificate" and "equipment lease term." This overcomes the bottleneck of existing systems that rely on manual entry of unstructured data. In one application at a biopharmaceutical company, the utilization rate of unstructured data increased from 35% to 89%.

[0153] 2. Dynamic Model Optimization Mechanism Enhances Policy Adaptability

[0154] The innovative parameter update strategy designed in this paper enables the model to quickly respond to policy changes. The specific advantages are as follows:

[0155] Adaptive learning rate scheduling: Using an exponentially decaying learning rate (ηt = η0 × 0.95^(t / 100)) combined with a momentum term (γ = 0.9), the model can be retrained within 48 hours when policies are adjusted (such as changes in the additional deduction ratio), achieving an 85% improvement in response speed compared to traditional batch training. Following the 2024 policy update, one manufacturing company experienced only a 0.8% drop in model accuracy, compared to an average 5.2% drop for similar systems.

[0156] Regularization and gradient optimization: L2 regularization (λ = 0.01) combined with the AdamW optimizer effectively controls model complexity. In a test at a semiconductor company, the model's overfitting risk in a small sample size scenario (500 data points) was reduced by 63%, and the false positive rate dropped from the industry average of 8.6% to 3.7%.

[0157] Continuous feedback optimization: A model performance monitoring dashboard is established to track metrics such as the rate of change of the Gini coefficient and loss function value in real time. When the classification accuracy of a policy-sensitive feature (such as "AI R&D expenses") falls below a threshold for three consecutive days, the system automatically triggers incremental learning to ensure the model remains optimal. BRIEF DESCRIPTION OF THE DRAWINGS

[0158] Figure 1 This is the overall system architecture diagram

[0159] Figure 2 Data preprocessing flowchart

[0160] Figure 3 Constructing a graph for a random forest model DETAILED DESCRIPTION

[0161] like Figure 1-3 As shown, the method of the present invention comprises the following steps:

[0162] Step 1: Establish an enterprise R&D activity data collection module, acquire multi-source heterogeneous data through structured database interfaces and distributed file systems, and perform data cleaning and standardization preprocessing. This module establishes a connection with a relational database through the JDBC / ODBC protocol and uses the Flume data flow framework to collect unstructured data such as R&D project documents, financial vouchers, and patent applications in real time. Regular expression matching technology is used to clean missing values ​​and outliers, and the Z-score standardization method is used to convert numerical data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0163] Step 2: Build a feature engineering module. Use an information entropy-based feature selection algorithm to screen key R&D expense indicators and perform feature space dimensionality reduction through linear discriminant analysis. This algorithm evaluates the information content of each feature by calculating the Shannon entropy formula, selects the top k features whose information gain exceeds a threshold, then constructs inter-class and intra-class scatter matrices. By solving the generalized eigenvalue problem to obtain the optimal projection direction, the original feature space is mapped to a d-dimensional subspace.

[0164] Step 3: Deploy a random forest classification model, set the number of decision trees to N, use the Gini coefficient as the node splitting criterion, and establish a rule for determining R&D expense compliance. The model generates M training subsets through bootstrap sampling. During training, each decision tree randomly selects mtry features for node splitting. The splitting effect is evaluated based on the Gini impurity formula, and the prediction results of all trees are finally integrated through a majority voting mechanism.

[0165] Step 4: Dynamically adjust the classification model parameters based on the gradient descent optimization algorithm to output a list of R&D expenses that meet the additional deduction policy and risk warning results; the algorithm defines the cross-entropy loss function, calculates the gradient through backpropagation, and uses the stochastic gradient descent update rule with momentum to dynamically adjust the learning rate and momentum parameters. The final output is a structured list containing expense item codes, policy matching scores, and red / yellow warning signs based on confidence thresholds.

[0166] This paper takes the 2024 R&D expense review of a semiconductor company as an example:

[0167] Data collection configuration:

[0168] Structured data source: MySQL 8.0.28, including R&D project tables (1,200 entries) and expense details tables (35,000 entries)

[0169] Unstructured data source: 500 PDF contracts, parsed and extracted using Tika to find the "R&D personnel social security payment certificate" field

[0170] Data pipeline: Kafka cluster configured with 3 partitions and 3 ZooKeeper nodes

[0171] Preprocessing parameters:

[0172] Sliding window filling: Use a window of w=7 for the "R&D equipment usage time" field

[0173] Standardization: Set the mean value of the "R&D material purchase unit price" field to μ = 850 and the standard deviation to σ = 200

[0174] Encoding: Convert "patent type" (invention patent / utility model / design) into a 3D sparse vector

[0175] Feature Engineering Process:

[0176] Initial features: 42 cost-related indicators

[0177] Information gain screening: retain 18 features with IG ≥ 0.25 (such as average salary of R&D personnel and equipment depreciation rate)

[0178] LDA dimensionality reduction: Projecting 18-dimensional features into 8-dimensional space, with a cumulative contribution rate of 92.3%

[0179] Model training parameters:

[0180] Number of decision trees: 150, with a maximum depth of 12 per tree

[0181] Subsampling ratio: 70%, feature subset size k=4

[0182] Node splitting threshold: Gini gain ≥ 0.04, minimum number of leaf node samples 3

[0183] Dynamic optimization results:

[0184] Optimization algorithm: AdamW (weight decay coefficient 0.01)

[0185] Learning rate scheduling: cosine annealing strategy, initial n = 0.001, T _ max=500

[0186] Loss function convergence curve: The optimal value reached 0.028 at the 320th iteration

[0187] System performance indicators:

[0188] Accuracy: 95.2% (test set 2000 data)

[0189] Response time: Single data processing <50ms (based on NVIDIA A100 GPU)

[0190] Model update cycle: Retraining is completed within 48 hours after policy adjustment

[0191] This example verifies the feasibility of the system in real industrial scenarios. Feature optimization increases the model training speed by 40%, and the dynamic parameter adjustment mechanism ensures that the model performance does not drop by more than 1.5% after policy updates.

[0192] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent management method for additional deductions based on machine learning algorithms, characterized in that: The following steps are involved: Step 1: Establish an enterprise R&D activity data collection module, acquire multi-source heterogeneous data through structured database interfaces and distributed file systems, and perform data cleaning and standardization preprocessing. This module establishes a connection with a relational database through the JDBC / ODBC protocol and uses the Flume data flow framework to collect unstructured data such as R&D project documents, financial vouchers, and patent applications in real time. Regular expression matching technology is used to clean missing values ​​and outliers, and the Z-score standardization method is used to convert numerical data into a standard normal distribution with a mean of 0 and a standard deviation of 1. Step 2: Build a feature engineering module, use a feature selection algorithm based on information entropy to screen key R&D expense indicators, and perform feature space dimensionality reduction through linear discriminant analysis; The algorithm evaluates the information content of each feature by calculating the Shannon entropy formula, selects the top k features whose information gain is greater than the threshold, then constructs the inter-class scatter matrix and the intra-class scatter matrix, obtains the optimal projection direction by solving the generalized eigenvalue problem, and maps the original feature space to a d-dimensional subspace. Step 3: Deploy a random forest classification model, set the number of decision trees to N, use the Gini coefficient as the node splitting criterion, and establish a rule for determining R&D expense compliance. The model generates M training subsets through bootstrap sampling. During training, each decision tree randomly selects mtry features for node splitting. The splitting effect is evaluated based on the Gini impurity formula, and the prediction results of all trees are finally integrated through a majority voting mechanism. Step 4: Dynamically adjust the classification model parameters based on the gradient descent optimization algorithm to output a list of R&D expenses that meet the additional deduction policy and risk warning results; The algorithm defines the cross-entropy loss function, calculates the gradient through back-propagation, and dynamically adjusts the learning rate and momentum parameters using the stochastic gradient descent update rule with momentum. The final output is a structured list containing cost item codes, policy matching scores, and red / yellow warning signs based on confidence thresholds.

2. The intelligent management method for additional deductions based on a machine learning algorithm according to claim 1, characterized in that: The data preprocessing in step 1 includes: Step a: Missing value filling When preprocessing enterprise R&D activity data, missing values ​​in the data are filled using the sliding window mean method of the same enterprise's historical data. Specifically, with the time point where the current missing value is located as the center, a certain number of data points before and after are selected to form a sliding window. The mean of the historical data of the same enterprise within the window is calculated and used as the filling value for the missing value. Step b: Standardize numeric fields For numeric fields, Z-score standardization is implemented to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The calculation formula is: in: x′ is the normalized data value; x is the original data value; Hong is the mean of all data in the numeric field; σ is the standard deviation of all data in the numeric field; Step c: Encoding categorical variables For categorical variables, One-Hot encoding is used to generate a sparse matrix. One-Hot encoding is a method for converting categorical variables into numerical variables. Each categorical variable value is converted into a binary vector. The length of the vector is equal to the number of categorical variable values, in which only the corresponding value position is 1 and the rest of the positions are 0. In the specific operation, for each categorical variable, all its possible values ​​are determined, and then a new binary feature is created for each value; in the original data set, for each sample, if its categorical variable takes a certain value, the corresponding new feature is 1, and the rest of the new features are 0.

3. The intelligent management method for additional deductions based on a machine learning algorithm according to claim 1, characterized in that: The feature selection algorithm based on information entropy in step 2 is specifically as follows: In the intelligent management system for additional deductions based on machine learning algorithms, for a given feature set F={f1,f2,…,f n } It is necessary to calculate the information gain value of each feature to evaluate the importance of the feature for classification; the calculation formula of the information gain value is: in: IG(F|C): represents the information gain of feature F relative to category C; H(C): is the entropy of category C, also known as category entropy, and its calculation formula is; Among them, C is the category set, that is, the set of all possible categories; c is a specific category in set C; p(c) is the probability that category c appears in the sample set S, which is calculated as where |S c | is the number of samples of category c, |S| is the total number of samples in the sample set S; H(C) is the category entropy, which measures the degree of uncertainty of category C in the absence of any feature information; V: is the value set of feature F, that is, the set of all possible values ​​of feature F; S: represents the sample set, which is the collection of all samples used to analyze and train the model; S v : is the sample subset in the sample set S where the feature F takes the value v, |S v | represents the number of samples in the subset; H(C|S v ): is a subset of samples where the feature F takes the value v S v The conditional entropy of category C in is calculated as in p(c|S v ) is in the sample subset S v The probability of category c appearing in is calculated as |S cv | is the sample subset S v The number of samples of category c; Conditional entropy H(C|S v ) represents the degree of uncertainty of category C when the value of feature F is known to be v; is the conditional entropy H(C|S v ) is weighted summed, and the weight is That is, the sample subset S v The proportion of the sample set S; After calculating the information gain value of each feature, the features that satisfy IG(F|C)≥θ are screened out. These features constitute the optimized feature subset; where θ is the preset threshold, and the setting of this threshold needs to be determined according to specific business needs and data characteristics.

4. The intelligent management method for additional deductions based on a machine learning algorithm according to claim 1, characterized in that: The decision tree generation method of the random forest classification model in step 3 is: In step 3 of the machine learning-based intelligent management system for additional deductions, the random forest classification model generates a decision tree using a specific method. Bootstrap sampling (a random sampling method with replacement) is performed on the original training set to generate N sub-training sets. The following operations are then recursively performed on each sub-training set until the node purity reaches a preset threshold or the number of samples is insufficient. Step a: Randomly select candidate features Randomly select from d features candidate features; in actual operation, a random number generator is used to select k features from the entire feature set as candidate features for the current node split; Step b: Calculate the Gini coefficient of the candidate feature For each candidate feature, calculate its Gini coefficient; the calculation formula of the Gini coefficient is: in: Gini(D): represents the Gini coefficient of the dataset D, which measures the impurity of the dataset; c: is the number of categories, that is, the total number of all possible categories in the dataset; p i : is the sample proportion of category i, calculated as where |D i | is the number of samples of category i in the dataset, |D| is the total number of samples in dataset D; Step c: Select features and split points for node splitting After calculating the Gini coefficient of each candidate feature, select Gini(D A ) The smallest feature and segmentation point are used for node splitting; here D A It is a subset obtained by dividing according to a certain feature and split point. For continuous features, all possible split points will be traversed and the Gini coefficient of the subset divided at each split point will be calculated. For discrete features, all possible value combinations will be considered for division; by selecting Gini (D A ) The minimum features and split points can make the purity of the divided subsets as high as possible, thereby building a more effective decision tree; Through the above method, a decision tree is generated for the random forest classification model, and finally the classification prediction result is obtained by combining the results of multiple decision trees, which is used to determine the compliance of R&D expenses in the additional deduction intelligent management system.

5. The intelligent management method for additional deductions based on a machine learning algorithm according to claim 1, characterized in that: The parameter update formula of the gradient descent optimization algorithm in step 4 is: In step 4 of the intelligent management system for additional deductions based on machine learning algorithms, the model parameters are updated using a gradient descent optimization algorithm; this algorithm continuously iterates to find the model parameter values ​​that minimize the objective function; The parameter update formula is: in: w t+1 : represents the model parameters at the t+1th iteration; the model parameters w at the current iteration number t t Update the objective function based on its gradient, gradually approaching the parameter value that minimizes the objective function. w t : is the model parameter of the tth iteration; in each iteration, it is adjusted according to the gradient information of the objective function to achieve the purpose of optimizing the model; η: is the learning rate, which controls the step size of parameter update at each iteration; is the objective function J(w) when w=w t The gradient at ; the gradient is a vector, its direction indicates the direction in which the objective function value increases fastest, and its opposite direction indicates the direction in which the objective function value decreases fastest; The calculation formula of the objective function J(w) is: in: m: is the number of samples, that is, the total number of samples used to train the model; h w (x): is the model prediction value, which is the result of predicting the input x based on the current model parameter w; for different machine learning models, h w The specific form of (x) will vary; x (i) : represents the input feature vector of the i-th sample; y (i) : is the true label value of the i-th sample; (h w (x (i) )-y (i) ) 2 Measures the square error between the model's predicted value and the true value for the i-th sample; This part is the mean squared error (MSE) loss function, which calculates the average of the squared prediction errors of all samples and is used to measure how well the model fits the training data; λ: L2 regularization coefficient, which controls the weight of the regularization term; ||w||2: is the L2 norm of the model parameter w, calculated as in is the jth component of the parameter vector w; By continuously iteratively updating the model parameters w, the objective function J(w) is gradually reduced, and finally a set of optimal model parameters is obtained.

6. The intelligent management method for additional deductions based on a machine learning algorithm according to claim 3, characterized in that: The projection matrix of the linear discriminant analysis is calculated as: In the intelligent management system for additional deductions based on machine learning algorithms, when using linear discriminant analysis to reduce the dimension of the feature space in step 2, it is necessary to calculate the projection matrix W; The goal of linear discriminant analysis is to find a projection matrix W so that data of different categories are separated as much as possible after projection, and data of the same category are clustered as much as possible; The projection matrix W is calculated by maximizing the following objective function: in: W is the projection matrix, which projects the original high-dimensional feature space into a low-dimensional space; s b It is the inter-class scatter matrix, which is used to measure the degree of dispersion between different categories; its calculation formula is: c is the number of categories, that is, the total number of all possible categories in the dataset; n i is the number of samples in the i-th category, reflecting the size of the samples in the data set; μ i is the mean vector of the i-th class sample, μ is the global mean vector, Inter-class scatter matrix S b The derivation of is based on the measurement of the difference between the mean values ​​of different categories. By calculating the deviation of the mean of each category from the global mean and taking into account the weight of the number of samples, a quantitative representation of the degree of dispersion between classes is obtained. S w It is the intra-class scatter matrix, which is used to measure the degree of dispersion of data within the same category; its calculation formula is For each class of samples, calculate the deviation of each sample from the class mean, and then sum it up for all classes to get the intra-class scatter matrix; the intra-class scatter matrix reflects the closeness of samples in the same class; Objective function The molecule |W T S b W| represents the degree of discreteness between different categories after projection, and the denominator |W T S w W| represents the degree of discreteness within the same category after projection. By maximizing this objective function, a projection matrix W can be found that maximizes the separation between different categories after projection and the aggregation within the same category, thereby achieving effective feature space dimensionality reduction. In actual calculations, it is usually done by solving the generalized eigenvalue problem S b w=λS w w is used to obtain the column vectors of the projection matrix W, which are the eigenvectors corresponding to the largest eigenvalues.

Images