Multi-source data screening method and system based on neural network classification model

Abstract

The invention discloses a multi-source data screening method and system based on a neural network classification model, and the method comprises the steps: determining a target data screening rule, converting the target data screening rule into a probability model, obtaining a feature combination joint probability density function, constructing a neural network classification model for each type of data, and optimizing model parameters according to the joint probability density function. And combining the neural network classification models of different data types into a multi-layer neural network classification model, generating an adversarial sample through a probability model for training, finally determining feature importance based on the gradient of the last layer of the multi-layer neural network classification model, selecting feature combinations through clustering analysis, and screening multi-source data. And updating feature importance according to a screening result, and circularly optimizing the screening process. According to the method, various types of data sources can be processed and integrated, the breadth and depth of data utilization are improved, accurate screening of the data is achieved through combination of the probability model and the neural network, the decision quality is improved, and meanwhile good interpretability is achieved.

Description

Technical Field

[0001] The present invention relates to the field of data mining, and in particular to a multi-source data screening method and system based on a neural network classification model. Background Art

[0002] In today's digital age, data is growing explosively, and data sources are becoming increasingly diverse. Multi-source data contains rich information, but the diversity and complexity of data make it difficult for traditional data processing methods to effectively mine the value of data. There are huge differences in data format, data quality, data semantics, etc. between different data sources. In addition, the expansion of data scale has put forward higher requirements for the efficiency and accuracy of data screening. Quickly and accurately screening out valuable information from massive data has become a problem that needs to be solved urgently.

[0003] Traditional screening methods based on rules or simple statistical analysis often face problems such as high consumption of computing resources and inaccurate screening results when processing large-scale multi-source data. For example, in the financial field, it is necessary to screen out information valuable for investment decisions from multi-source data such as stock trading data, macroeconomic data, and industry reports; in the medical field, it is necessary to screen out key information related to disease diagnosis and treatment from multi-source data such as medical records, examination reports, imaging data, and genetic data. Traditional methods are difficult to meet the requirements for high efficiency and accuracy of data screening in these fields.

[0004] The multi-source data screening method and system based on the neural network classification model can realize efficient and accurate screening of multi-source data, can significantly improve data processing efficiency and accuracy, can effectively overcome the limitations of traditional methods, and has broad application prospects. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-source data screening method and system based on a neural network classification model.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The first aspect of the present invention provides a multi-source data screening method based on a neural network classification model, comprising:

[0008] S100 determines the screening rules of the target data, converts the conditions and constraints in the screening rules into a probability model, and obtains the joint probability density function between the feature combinations;

[0009] S200 constructs a neural network classification model for each type of data, and optimizes the parameters of each neural network classification model according to the shape and boundary area of ​​the joint probability density function;

[0010] S300 combines neural network classification models of different data types to obtain a multi-layer neural network classification model, and uses the probability model to generate adversarial sample data to train the multi-layer neural network classification model;

[0011] S400 determines the classification importance of features based on the gradient of the last layer of the multi-layer neural network model, performs cluster analysis according to the classification importance, and selects a feature combination for screening to screen the multi-source data;

[0012] S500 updates the classification importance of the features according to the screening results, and continuously optimizes the screening process.

[0013] As a further method, the method of converting the conditions and constraints in the screening rules into a probability model to obtain the joint probability density function between feature combinations includes:

[0014] Considering continuous feature data and discrete feature data, the conditions and constraints in the screening rules are converted into probability expressions. Specifically, for continuous feature data, the probability expression is expressed using Gaussian distribution, and for discrete feature data, the probability expression is expressed using multinomial distribution.

[0015] Determine the joint probability density between feature combinations, expressed as:

[0016]

[0017] Among them, f(h) represents the joint probability density function of the feature combination, h is the observed value of the feature combination, K is the total number of features in the feature combination, ∑ is the mean square error matrix of the feature combination, μ is the mean vector of the feature combination, and T represents the transposition operation.

[0018] As a further method, a neural network classification model is constructed for each type of data, wherein the structure of the neural network classification model includes an input layer, an embedding layer, a hidden layer and an output layer, specifically:

[0019] The input layer is used to receive the feature vector of the specified type of data;

[0020] The embedding layer contains 256 neurons and uses the ReLU activation function to preliminarily extract feature commonalities and differences;

[0021] There are three hidden layers, with 128, 64, and 32 neurons respectively. The corresponding activation functions are LeakyReLU, Tanh, and SELU, which are used for deep abstraction and transfer of features.

[0022] The output layer determines the number of neurons according to the classification target of the specified type of data, and the Softmax function outputs the probability of the category, which is expressed as:

[0023]

[0024] Among them, P(y=q|z) represents the probability that the output category is q when the input is z, y is the output category, e is a natural constant, and z q is the element of category q in the input z, Q is the total number of categories, that is, the total number of neurons in the output layer, z j is the element of category j in the input z.

[0025] As a further method, the method of optimizing the parameters of each neural network classification model according to the morphology and boundary area of ​​the joint probability density function includes:

[0026] Get the joint probability density function of all features of each type of data;

[0027] Calculate the gradient of the joint probability density function, and determine the data area where the gradient is greater than a preset threshold as the boundary area;

[0028] The fitness function is defined according to the joint probability density function of all features of each type of data, including the overall morphological fitting part and the boundary area fitting part. The expression is:

[0029]

[0030] Among them, θ is the parameter of the neural network model, α is the weight parameter of the overall morphological fitting part, P(x) is the joint probability density function, f(x;θ) is the output of the neural network model for the input x under the parameter θ, B is the boundary area, and w(x) is the weight function of the data points in the boundary area. k is the adjustment parameter of the growth rate, is the gradient of P(x), τ is the preset threshold of the gradient;

[0031] The model parameters are optimized by the sparrow algorithm to minimize the fitness function. The iteration termination condition is that the change rate of the fitness function in two consecutive iterations is less than 0.01. The expression for position update is:

[0032]

[0033] in, and are the position vectors of the i-th sparrow at the t-th iteration and the t+1-th iteration, respectively. is the position with the best fitness in the entire sparrow population at the tth iteration, a is the control parameter of the update step length, is the fitness function value of the i-th sparrow at the t+1th iteration, TDR is the mean value of the gradient change of the probability density in the boundary area, Up and Lo are the upper and lower bounds of the search space, rand is a random number generation function with a value range of [0,1], rand(1,D) is a random matrix generated by 1×D, D is the number of dimensions of the neural network model parameters to be optimized, N is the total number of individuals in the sparrow population, is the distance between the position of the i-th sparrow and the position of the optimal sparrow individual at the t-th iteration, is the position of a sparrow individual randomly selected from the sparrow population at the tth iteration.

[0034] As a further method, the method of combining neural network classification models of different data types to obtain a multi-layer neural network classification model includes:

[0035] According to the needs of the target data, the order of data type processing is determined, and the neural network structures of different data types are connected to obtain a multi-layer neural network classification model;

[0036] The connection weights between layers are optimized according to the slime mold algorithm, where the expression for local search optimization is:

[0037]

[0038] Among them, SMA is expressed as the optimization function of the slime mold algorithm, is the weight connecting the i-th layer and the j-th layer in the t-th iteration, is the output vector of the activation function of the i-th layer in the t-th iteration, is the output vector of the activation function of the jth layer in the t+1th iteration, and γ is the scaling factor; the expression of the update rule is:

[0039]

[0040] in, is the weight connecting the i-th layer and the j-th layer in the t+1th iteration, η is the learning rate, f is the loss function, is the bias of the i-th layer in the t+1th iteration, is the bias of the i-th layer in the t-th iteration.

[0041] As a further method, the method of using the probability model to generate adversarial sample data to train a multi-layer neural network classification model includes:

[0042] Determine the perturbation range of each feature based on the distribution of features in the probability model;

[0043] The conditional probability relationship in the probability model is used to determine the disturbance direction. If a feature change causes the probability of the target data classification to change in the wrong direction, the feature is preferentially disturbed in the direction that causes the classification error.

[0044] For numerical data, the key features are adjusted based on the gradient method within the perturbation range and the perturbation direction to generate adversarial sample data. For text data, the key text is located based on the perturbation range and perturbation direction, and the key text is replaced, inserted or deleted. Related vocabulary of key semantic features generates adversarial sample data;

[0045] The normal sample data and adversarial sample data are combined as the training set to train the model.

[0046] As a further method, the method for determining the classification importance of features based on the gradient of the last layer of the multi-layer neural network model includes:

[0047] Calculate the gradient of the feature in the last layer output with respect to the loss function, expressed as:

[0048]

[0049] in, is the gradient of the nth feature with respect to the loss function, M is the total number of neurons in the hidden layer, is the gradient of the hidden layer input Z, L n,m is the loss function of the nth feature in the mth neuron, Z n,m The nth feature in the mth neuron in the hidden layer input Z, A L,n,c,i,j Output A for the last convolutional layer L The partial derivative of the nth feature at the i-th height position and j-th width position in the c-th channel, where N is the number of features, C is the total number of channels in the last layer, H is the height of the feature map of the last layer, and W is the feature width of the last layer;

[0050] The absolute value of the gradient of each feature with respect to the loss function is normalized to obtain the classification importance of the feature.

[0051] As a further method, the method of performing cluster analysis according to classification importance and selecting a feature combination for screening to screen multi-source data includes:

[0052] Use clustering algorithm to divide the features into three different clusters of high importance, medium importance and low importance;

[0053] For each cluster, calculate the Pearson correlation coefficient of the features within the cluster. If the Pearson correlation coefficient is greater than the similarity threshold, remove the features with lower classification importance and retain the features with higher classification importance.

[0054] The same proportion of features are selected from each cluster to obtain feature combinations, and a data subset containing all features is extracted from multi-source data to screen the multi-source data.

[0055] As a further method, the method for updating the classification importance of the feature according to the screening result includes: setting a reward function according to the degree of matching between the screening result and the target data, and updating the classification importance of the feature through reinforcement learning based on the reward function value.

[0056] The second aspect of the present invention provides a multi-source data screening system based on a neural network classification model, comprising:

[0057] The rule modeling module is used to determine the screening rules of the target data, convert the conditions and constraints in the screening rules into a probability model, and obtain the joint probability density function between feature combinations;

[0058] A parameter tuning module, used to construct a neural network classification model for each type of data, and optimize the parameters of each neural network classification model according to the shape and boundary area of ​​the joint probability density function;

[0059] A model fusion module is used to combine neural network classification models of different data types to obtain a multi-layer neural network classification model, and use the probability model to generate adversarial sample data to train the multi-layer neural network classification model;

[0060] An importance evaluation module, used to determine the classification importance of features based on the gradient of the last layer of the multi-layer neural network model, perform cluster analysis according to the classification importance, and select a feature combination for screening to screen multi-source data;

[0061] The feedback update module is used to update the classification importance of the features according to the screening results and continuously optimize the screening process.

[0062] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0063] (1) The present invention converts the screening rules into a probability model and obtains the joint probability density function, and creates a neural network classification model, taking into account the distribution characteristics of the data and the relationship between feature combinations, and can learn the complex correlation between features, which helps to classify and screen data more comprehensively and accurately;

[0064] (2) The present invention uses a probability model to generate adversarial sample data to train a multi-layer neural network classification model, so that the model can better cope with noise, outliers and possible malicious attacks in the data, improve the stability and reliability of the model, and ensure the robustness of the screening process;

[0065] (3) The present invention determines the importance of feature classification based on the gradient of the last layer of the multi-layer neural network model, and performs cluster analysis to select feature combinations, thereby realizing automatic screening and optimization of features, avoiding the subjectivity and limitations of manual feature selection, improving the scientificity and effectiveness of feature selection, and facilitating the discovery of key information hidden in the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 1 is a flowchart of the steps of a multi-source data screening method based on a neural network classification model in an embodiment of the present invention.

[0067] Figure 2 for Figure 2 It is a logical schematic diagram of a multi-source data screening system based on a neural network classification model in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0069] Reference Figure 1 As shown, the present invention provides a multi-source data screening method based on a neural network classification model, comprising:

[0070] S100 determines the screening rules of the target data, converts the conditions and constraints in the screening rules into a probability model, and obtains the joint probability density function between the feature combinations;

[0071] In the actual evaluation, the task of medical data screening aims to screen out data features closely related to cardiovascular diseases from massive amounts of patient medical data. The data comes from multiple data sources, including basic patient information in the hospital information system, various test data in the laboratory information system, and medical record text records in the electronic medical record system.

[0072] In the actual evaluation, the screening rules were determined as follows: the patient's age was between 40 and 70 years old (continuous feature), the total cholesterol level was greater than 5.2 mmol / L (continuous feature), there was a history of hypertension (discrete feature, yes or no), and the patient's medical record contained a description of chest pain or dyspnea symptoms (discrete feature, yes or no) as the target data, which was converted into a probability model:

[0073] 1. For continuous feature data:

[0074] Age: 5,000 patient data were extracted for distribution fitting, and it was found that the mean age was 55 years old and the standard deviation was 8. Let age be A, and the probability expression is:

[0075]

[0076] Total cholesterol level: Statistics of 8,000 test reports show that it conforms to a Gaussian distribution with a mean of 5.8 mmol / L and a standard deviation of 1.2 mmol / L. Assuming the total cholesterol level is TC, the probability expression is:

[0077]

[0078] 2. For discrete feature data:

[0079] History of hypertension: In the patient population studied, the medical records of 12,000 patients were reviewed, and the proportion of patients with a history of hypertension was about 37%, that is, p1 = 0.37, and the probability of not having a history of hypertension was p2 = 0.63;

[0080] Symptoms of chest pain or dyspnea: An analysis of 15,000 medical records showed that about 30% of patients had related symptom descriptions in their medical records, p3 = 0.3, and the probability of not mentioning them was p4 = 0.7;

[0081] In the actual evaluation, the feature combination of age, total cholesterol level, history of hypertension, and chest pain or dyspnea symptoms was considered to construct a joint probability density function.

[0082] S200 constructs a neural network classification model for each type of data, and optimizes the parameters of each neural network classification model according to the shape and boundary area of ​​the joint probability density function;

[0083] In the actual evaluation, for the data type of basic patient information: input layer: receive feature vectors such as age, gender, height, and weight, a total of 4 features; for the data type of clinical examination: input layer: receive feature vectors of 5 test indicators, including total cholesterol, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, and blood sugar; for the data type of medical record text: input layer: receive feature vectors of medical record text after word vector conversion, and each medical record text is represented as a 512-dimensional vector (based on the pre-trained word vector model); obtain the joint probability density function of all features of each type of data, calculate the gradient of the joint probability density function, and determine the data area with a gradient greater than the preset threshold (set to 0.6) as the boundary area. After calculating and analyzing 5,000 sample data, the boundary area range is determined, the fitness function is defined, and the model parameters are optimized using the sparrow algorithm.

[0084] S300 combines neural network classification models of different data types to obtain a multi-layer neural network classification model, and uses the probability model to generate adversarial sample data to train the multi-layer neural network classification model;

[0085] In the actual evaluation, according to the focus of the target data requirements, the processing order is determined to process the patient's basic information data first, then the clinical test data, and finally the medical record text data, and connect the neural network structures of different data types to obtain a multi-layer neural network classification model; the slime mold algorithm is used to optimize the connection weights between layers. Set the scaling factor to 0.6 and the learning rate to 0.02. The connection weights are optimized through the local search optimization expression and update rule expression of the slime mold algorithm. The number of iterations is set to 150 times, and 300 sample data are verified in each iteration to adjust the weights;

[0086] In the actual evaluation, according to the distribution of features in the probability model: for the age feature, the perturbation range is set to ±6 years old. If the increase in age makes the probability of the target data classification change in the wrong direction, the age feature is preferentially perturbed in the increasing direction; for the total cholesterol level feature, the perturbation range is set to ±1.5mmol / L; for the discrete features of hypertension history and chest pain or dyspnea symptoms, the perturbation direction is determined according to the conditional probability relationship, where a history of hypertension and chest pain or dyspnea symptoms make the probability of classification as cardiovascular disease high, and the history of hypertension is set to no or the chest pain or dyspnea symptoms are set to no to generate adversarial sample data; for numerical data, for medical records mentioning the symptom of "chest pain", try to replace "chest pain" with words such as "chest discomfort" to generate adversarial samples; the normal sample data and the adversarial sample data are combined as the training set to train the model. A total of 20,000 normal sample data were collected, and 5,000 adversarial sample data were generated based on the above rules to form a training set for model training.

[0087] S400 determines the classification importance of features based on the gradient of the last layer of the multi-layer neural network model, performs cluster analysis according to the classification importance, and selects a feature combination for screening to screen the multi-source data;

[0088] In the actual evaluation, the gradient of the age feature with respect to the loss function is calculated, and the absolute value of the gradient of each feature with respect to the loss function is normalized to obtain the classification importance of the feature. The clustering algorithm is used to divide the features into three different clusters of high importance, medium importance, and low importance. The Pearson correlation coefficient of the features within the cluster is calculated. If the Pearson correlation coefficient is greater than 0.7, the features with lower classification importance are eliminated and the features with higher classification importance are retained. 40% of the features in the same proportion are selected from each cluster to obtain the feature combination, and a data subset containing all features is extracted from the multi-source data for screening of the multi-source data.

[0089] S500 updates the classification importance of the features according to the screening results, and continuously optimizes the screening process.

[0090] In the actual evaluation, the reward function is set according to the degree of match between the screening results and the target data (accurate diagnosis of patients with cardiovascular disease), and the classification importance of the features is updated through reinforcement learning based on the reward function value. After multiple iterative optimizations, it was found that features such as age, total cholesterol level, history of hypertension, and chest pain descriptions in medical records always maintain a high importance in the screening process, while some features with weaker correlation, such as the patient's blood type, are gradually reduced in importance or even eliminated.

[0091] In this embodiment, the method of converting the conditions and constraints in the screening rules into a probability model to obtain a joint probability density function between feature combinations includes:

[0092] Considering continuous feature data and discrete feature data, the conditions and constraints in the screening rules are converted into probability expressions. Specifically, for continuous feature data, the probability expression is expressed using Gaussian distribution, and for discrete feature data, the probability expression is expressed using multinomial distribution.

[0093] Determine the joint probability density between feature combinations, expressed as:

[0094]

[0095] Among them, f(h) represents the joint probability density function of the feature combination, h is the observed value of the feature combination, K is the total number of features in the feature combination, ∑ is the mean square error matrix of the feature combination, μ is the mean vector of the feature combination, and T represents the transposition operation.

[0096] In this embodiment, a neural network classification model is constructed for each type of data, wherein the structure of the neural network classification model includes an input layer, an embedding layer, a hidden layer and an output layer, specifically:

[0097] The input layer is used to receive the feature vector of the specified type of data;

[0098] The embedding layer contains 256 neurons and uses the ReLU activation function to preliminarily extract feature commonalities and differences;

[0099] There are three hidden layers, with 128, 64, and 32 neurons respectively. The corresponding activation functions are LeakyReLU, Tanh, and SELU, which are used for deep abstraction and transfer of features.

[0100] The output layer determines the number of neurons according to the classification target of the specified type of data, and the Softmax function outputs the probability of the category, which is expressed as:

[0101]

[0102] Among them, P(y=q|z) represents the probability that the output category is q when the input is z, y is the output category, e is a natural constant, and z q is the element of category q in the input z, Q is the total number of categories, that is, the total number of neurons in the output layer, z j is the element of category j in the input z.

[0103] In this embodiment, the method for optimizing the parameters of each neural network classification model according to the morphology and boundary area of ​​the joint probability density function includes:

[0104] Get the joint probability density function of all features of each type of data;

[0105] Calculate the gradient of the joint probability density function, and determine the data area where the gradient is greater than a preset threshold as the boundary area;

[0106] The fitness function is defined according to the joint probability density function of all features of each type of data, including the overall morphological fitting part and the boundary area fitting part. The expression is:

[0107]

[0108] Among them, θ is the parameter of the neural network model, α is the weight parameter of the overall morphological fitting part, P(x) is the joint probability density function, f(x;θ) is the output of the neural network model for the input x under the parameter θ, B is the boundary area, and w(x) is the weight function of the data points in the boundary area. k is the adjustment parameter of the growth rate, is the gradient of P(x), τ is the preset threshold of the gradient;

[0109] The model parameters are optimized by the sparrow algorithm to minimize the fitness function. The iteration termination condition is that the change rate of the fitness function in two consecutive iterations is less than 0.01. The expression for position update is:

[0110]

[0111] in, and are the position vectors of the i-th sparrow at the t-th iteration and the t+1-th iteration, respectively. is the position with the best fitness in the entire sparrow population at the tth iteration, a is the control parameter of the update step length, is the fitness function value of the i-th sparrow at the t+1th iteration, TDR is the mean value of the gradient change of the probability density in the boundary area, Up and Lo are the upper and lower bounds of the search space, rand is a random number generation function with a value range of [0,1], rand(1,D) is a random matrix generated by 1×D, D is the number of dimensions of the neural network model parameters to be optimized, N is the total number of individuals in the sparrow population, is the distance between the position of the i-th sparrow and the position of the optimal sparrow individual at the t-th iteration, is the position of a sparrow individual randomly selected from the sparrow population at the tth iteration.

[0112] In this embodiment, the method of combining neural network classification models of different data types to obtain a multi-layer neural network classification model includes:

[0113] According to the needs of the target data, the order of data type processing is determined, and the neural network structures of different data types are connected to obtain a multi-layer neural network classification model;

[0114] The connection weights between layers are optimized according to the slime mold algorithm, where the expression for local search optimization is:

[0115]

[0116] Among them, SMA is expressed as the optimization function of the slime mold algorithm, is the weight connecting the i-th layer and the j-th layer in the t-th iteration, is the output vector of the activation function of the i-th layer in the t-th iteration, is the output vector of the activation function of the jth layer in the t+1th iteration, and γ is the scaling factor; the expression of the update rule is:

[0117]

[0118] in, is the weight connecting the i-th layer and the j-th layer in the t+1th iteration, η is the learning rate, f is the loss function, is the bias of the i-th layer in the t+1th iteration, is the bias of the i-th layer in the t-th iteration.

[0119] In this embodiment, the method of using the probability model to generate adversarial sample data to train a multi-layer neural network classification model includes:

[0120] Determine the perturbation range of each feature based on the distribution of features in the probability model;

[0121] The conditional probability relationship in the probability model is used to determine the disturbance direction. If a feature change causes the probability of the target data classification to change in the wrong direction, the feature is preferentially disturbed in the direction that causes the classification error.

[0122] For numerical data, the key features are adjusted based on the gradient method within the perturbation range and the perturbation direction to generate adversarial sample data. For text data, the key text is located based on the perturbation range and perturbation direction, and the key text is replaced, inserted or deleted. Related vocabulary of key semantic features generates adversarial sample data;

[0123] The normal sample data and adversarial sample data are combined as the training set to train the model.

[0124] In this embodiment, the method for determining the classification importance of a feature based on the gradient of the last layer of the multi-layer neural network model includes:

[0125] Calculate the gradient of the feature in the last layer output with respect to the loss function, expressed as:

[0126]

[0127] in, is the gradient of the nth feature with respect to the loss function, M is the total number of neurons in the hidden layer, is the gradient of the hidden layer input Z, L n,m is the loss function of the nth feature in the mth neuron, Z n,m The nth feature in the mth neuron in the hidden layer input Z, A L,n,c,i,j Output A for the last convolutional layer L The partial derivative of the nth feature at the i-th height position and j-th width position in the c-th channel, where N is the number of features, C is the total number of channels in the last layer, H is the height of the feature map of the last layer, and W is the feature width of the last layer;

[0128] The absolute value of the gradient of each feature with respect to the loss function is normalized to obtain the classification importance of the feature.

[0129] In this embodiment, the method of performing cluster analysis according to classification importance and selecting a feature combination for screening to screen multi-source data includes:

[0130] Use clustering algorithm to divide the features into three different clusters of high importance, medium importance and low importance;

[0131] For each cluster, calculate the Pearson correlation coefficient of the features within the cluster. If the Pearson correlation coefficient is greater than the similarity threshold, remove the features with lower classification importance and retain the features with higher classification importance.

[0132] The same proportion of features are selected from each cluster to obtain feature combinations, and a data subset containing all features is extracted from multi-source data to screen the multi-source data.

[0133] In this embodiment, the method for updating the classification importance of the feature according to the screening result includes: setting a reward function according to the degree of matching between the screening result and the target data, and updating the classification importance of the feature through reinforcement learning based on the reward function value.

[0134] The second aspect of the present invention provides a multi-source data screening system based on a neural network classification model, comprising:

[0135] The rule modeling module is used to determine the screening rules of the target data, convert the conditions and constraints in the screening rules into a probability model, and obtain the joint probability density function between feature combinations;

[0136] A parameter tuning module, used to construct a neural network classification model for each type of data, and optimize the parameters of each neural network classification model according to the shape and boundary area of ​​the joint probability density function;

[0137] A model fusion module is used to combine neural network classification models of different data types to obtain a multi-layer neural network classification model, and use the probability model to generate adversarial sample data to train the multi-layer neural network classification model;

[0138] An importance evaluation module, used to determine the classification importance of features based on the gradient of the last layer of the multi-layer neural network model, perform cluster analysis according to the classification importance, and select a feature combination for screening to screen multi-source data;

[0139] The feedback update module is used to update the classification importance of the features according to the screening results and continuously optimize the screening process.

[0140] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A multi-source data screening method based on a neural network classification model, characterized in that: The following steps are involved: Determine the screening rules for the target data, convert the conditions and constraints in the screening rules into a probability model, and obtain the joint probability density function between feature combinations; Constructing a neural network classification model for each type of data, and optimizing the parameters of each neural network classification model according to the shape and boundary area of ​​the joint probability density function; Combining neural network classification models of different data types to obtain a multi-layer neural network classification model, and using the probability model to generate adversarial sample data to train the multi-layer neural network classification model; Determining the classification importance of features based on the gradient of the last layer of the multi-layer neural network model, performing cluster analysis according to the classification importance, and selecting a feature combination for screening to screen multi-source data; The classification importance of the features is updated based on the screening results to continuously optimize the screening process.

2. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method of converting the conditions and constraints in the screening rules into a probability model to obtain a joint probability density function between feature combinations includes: Considering continuous feature data and discrete feature data, the conditions and constraints in the screening rules are converted into probability expressions. Specifically, for continuous feature data, the probability expression is expressed using Gaussian distribution, and for discrete feature data, the probability expression is expressed using multinomial distribution. Determine the joint probability density between feature combinations, expressed as: Among them, f(h) represents the joint probability density function of the feature combination, h is the observed value of the feature combination, K is the total number of features in the feature combination, ∑ is the mean square error matrix of the feature combination, μ is the mean vector of the feature combination, and T represents the transposition operation.

3. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The neural network classification model is constructed for each type of data, wherein the structure of the neural network classification model includes an input layer, an embedding layer, a hidden layer and an output layer, specifically: The input layer is used to receive the feature vector of the specified type of data; The embedding layer contains 256 neurons and uses the ReLU activation function to preliminarily extract feature commonalities and differences; There are three hidden layers, with 128, 64, and 32 neurons respectively. The corresponding activation functions are LeakyReLU, Tanh, and SELU, which are used for deep abstraction and transfer of features. The output layer determines the number of neurons according to the classification target of the specified type of data, and the Softmax function outputs the probability of the category.

4. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method for optimizing the parameters of each neural network classification model according to the morphology and boundary area of ​​the joint probability density function comprises: Get the joint probability density function of all features of each type of data; Calculate the gradient of the joint probability density function, and determine the data area where the gradient is greater than a preset threshold as the boundary area; The fitness function is defined according to the joint probability density function of all features of each type of data, including the overall morphological fitting part and the boundary area fitting part. The expression is: Among them, θ is the parameter of the neural network model, α is the weight parameter of the overall morphological fitting part, P(x) is the joint probability density function, f(x;θ) is the output of the neural network model for the input x under the parameter θ, B is the boundary area, w(x) is the weight function of the data points in the boundary area, k is the adjustment parameter of the growth rate, is the gradient of P(x), τ is the preset threshold of the gradient; The model parameters are optimized by the sparrow algorithm to minimize the fitness function. The iteration termination condition is that the change rate of the fitness function in two consecutive iterations is less than 0.

01. The expression for position update is: in, and are the position vectors of the i-th sparrow at the t-th iteration and the t+1-th iteration, respectively. is the position with the best fitness in the entire sparrow population at the tth iteration, a is the control parameter of the update step length, is the fitness function value of the i-th sparrow at the t+1th iteration, TDR is the mean value of the gradient change of the probability density in the boundary area, Up and Lo are the upper and lower bounds of the parameter search space, rand is a random number generation function with a value range of [0,1], rand(1,D) is a random matrix generated by 1×D, D is the number of dimensions of the neural network model parameters to be optimized, N is the total number of individuals in the sparrow population, is the distance between the position of the i-th sparrow and the position of the optimal sparrow individual at the t-th iteration, is the position of a sparrow individual randomly selected from the sparrow population at the tth iteration.

5. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method of combining neural network classification models of different data types to obtain a multi-layer neural network classification model includes: According to the needs of the target data, the order of data type processing is determined, and the neural network structures of different data types are connected to obtain a multi-layer neural network classification model; The connection weights between layers are optimized according to the slime mold algorithm, where the expression for local search optimization is: Among them, SMA is expressed as the optimization function of the slime mold algorithm, is the weight connecting the i-th layer and the j-th layer in the t-th iteration, is the output vector of the activation function of the i-th layer in the t-th iteration, is the output vector of the activation function of the jth layer in the t+1th iteration, and γ is the scaling factor; the expression of the update rule is: in, is the weight connecting the i-th layer and the j-th layer in the t+1th iteration, η is the learning rate, f is the loss function, is the bias of the i-th layer in the t+1th iteration, is the bias of the i-th layer in the t-th iteration.

6. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method of using the probability model to generate adversarial sample data to train a multi-layer neural network classification model includes: Determine the perturbation range of each feature based on the distribution of features in the probability model; The conditional probability relationship in the probability model is used to determine the disturbance direction. If a feature change causes the probability of the target data classification to change in the wrong direction, the feature is preferentially disturbed in the direction that causes the classification error. For numerical data, the key features are adjusted based on the gradient method within the perturbation range and the perturbation direction to generate adversarial sample data. For text data, the key text is located based on the perturbation range and perturbation direction, and the key text is replaced, inserted or deleted. Related vocabulary of key semantic features generates adversarial sample data; The normal sample data and adversarial sample data are combined as the training set to train the model.

7. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method for determining the classification importance of features based on the gradient of the last layer of the multi-layer neural network model comprises: Calculate the gradient of the feature in the last layer output with respect to the loss function, expressed as: in, is the gradient of the nth feature with respect to the loss function, M is the total number of neurons in the hidden layer, is the gradient of the hidden layer input Z, L n,m is the loss function of the nth feature in the mth neuron, Z n,m The nth feature in the mth neuron in the hidden layer input Z, A L,n,c,i,j Output A for the last convolutional layer L The partial derivative of the nth feature at the i-th height position and j-th width position in the c-th channel, where N is the number of features, C is the total number of channels in the last layer, H is the height of the feature map of the last layer, and W is the feature width of the last layer; The absolute value of the gradient of each feature with respect to the loss function is normalized to obtain the classification importance of the feature.

8. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method of performing cluster analysis according to classification importance and selecting a feature combination for screening to screen multi-source data includes: Use clustering algorithm to divide the features into three different clusters of high importance, medium importance and low importance; For each cluster, calculate the Pearson correlation coefficient of the features within the cluster. If the Pearson correlation coefficient is greater than the similarity threshold, remove the features with lower classification importance and retain the features with higher classification importance. The same proportion of features are selected from each cluster to obtain feature combinations, and a data subset containing all features is extracted from multi-source data to screen the multi-source data.

9. The multi-source data screening method based on the neural network classification model according to claim 1 is characterized in that: The method for updating the classification importance of the feature according to the screening result includes: setting a reward function according to the matching degree between the screening result and the target data, and updating the classification importance of the feature through reinforcement learning based on the reward function value.

10. A multi-source data screening system based on a neural network classification model, used to execute the multi-source data screening method based on a neural network classification model according to any one of claims 1 to 9, characterized in that: The system comprises: The rule modeling module is used to determine the screening rules of the target data, convert the conditions and constraints in the screening rules into a probability model, and obtain the joint probability density function between feature combinations; A parameter tuning module, used to construct a neural network classification model for each type of data, and optimize the parameters of each neural network classification model according to the shape and boundary area of ​​the joint probability density function; A model fusion module is used to combine neural network classification models of different data types to obtain a multi-layer neural network classification model, and use the probability model to generate adversarial sample data to train the multi-layer neural network classification model; An importance evaluation module, used to determine the classification importance of features based on the gradient of the last layer of the multi-layer neural network model, perform cluster analysis according to the classification importance, and select a feature combination for screening to screen multi-source data; The feedback update module is used to update the classification importance of the features according to the screening results and continuously optimize the screening process.

Images