Data arrangement and filtering system and method based on AI model and storage medium
By introducing data screening, prediction, threshold adjustment and label correction modules into the AI model, the probability threshold of the binary classification model is dynamically adjusted, which solves the problem of category imbalance and significantly improves the recall of a few categories and the overall performance of the model.
Patent Information
- Application Number
- CN202510162264.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-14
- Publication Date
- 2025-03-18
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the binary classification task, the problem of category imbalance causes traditional classification models to tend to predict majority classes, which in turn makes the recognition effect of minority classes poor. Existing methods such as oversampling, undersampling and weighted loss functions have limitations and are difficult to adapt to different application scenarios.
By introducing data filtering module, prediction module, threshold adjustment module and label correction module into the AI model, the probability threshold of model output is dynamically adjusted to optimize the category imbalance problem. Specific steps include filtering the initial data set according to preset rules, using AI models to make predictions, dynamically adjusting the classification threshold, and correcting the prediction tag based on the re-evaluation results.
By dynamically adjusting the probability threshold, the recall rate of minority classes is significantly improved, the classification performance of the model on minority classes is improved, the bias of the model to majority classes is reduced, the balance between false positive rates and false negative rates is achieved, and the overall fairness of classification results is improved.
Smart Images
Figure CN119646716A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a data sorting and filtering system, method and storage medium based on an AI model. Background Art
[0002] In binary classification tasks, the problem of class imbalance is widespread, especially in application areas such as medical diagnosis and financial fraud detection. Class imbalance means that the number of samples in a certain class (usually the minority class) is much lower than that in another class (usually the majority class). This problem causes traditional classification models to tend to predict the majority class, resulting in poor recognition of the minority class. In order to alleviate the impact of class imbalance, many traditional methods (such as oversampling, undersampling, and weighted loss functions) attempt to adjust the distribution of training data or modify the loss function so that the model pays more attention to the minority class during training. However, these methods still have certain limitations, such as oversampling may lead to sample redundancy, undersampling may lose important information, and weighted loss functions rely on hyperparameter adjustment, which makes it difficult to adapt to different application scenarios.
[0003] To address this problem, dynamically adjusting the probability threshold of the model output has become an effective solution. Most binary classification models output a probability value between 0 and 1 when predicting, which is used to indicate the probability that a sample belongs to the positive class. Usually, when the probability value exceeds a fixed threshold (such as 0.5), the model determines the sample as a positive class, otherwise it is a negative class. However, in the case of class imbalance, using a fixed threshold may not be able to fully identify minority class samples. By dynamically adjusting the threshold, the decision boundary of the model can be optimized according to the distribution of the class and actual needs. Dynamically adjusting the threshold can not only improve the recall rate of the minority class, but also further improve the overall performance of the model by balancing the ratio of false positives and false negatives. Summary of the invention
[0004] The purpose of the present invention is to provide a data sorting and filtering system, method and storage medium based on an AI model, which dynamically adjusts the probability distribution threshold of the model prediction output to solve the category imbalance problem in binary classification tasks.
[0005] To achieve the above object, the present invention provides the following technical solution: a data sorting and filtering system based on an AI model, the system comprising: The data screening module is used to screen and obtain the initial data set for the binary classification task according to preset rules; The prediction module connected to the data screening module is used to use the AI model to predict the initial data set and output the probability value of each data item belonging to each category; The threshold adjustment module connected to the prediction module is used to dynamically determine the adjustment threshold according to the current category distribution status, and adjust the threshold to re-evaluate the probability value predicted by the model, including dividing the samples into the first category samples and the second category samples, counting the difference in the number of the first category samples and the second category samples, and calculating the dynamically adjusted classification threshold. The specific formula is: ; Among them, u represents the new classification threshold, a represents the sensitivity coefficient of the classification model, b represents the difference in the number of samples in the first category and the second category, and g represents the average importance score of all features in the data set; The label correction module connected to the threshold adjustment module is used to correct the predicted labels based on the re-evaluation results and optimize the category imbalance, including calculating the sample ratio in the data set, setting the weight parameters, and calculating the rate of label correction. The specific formula is: ; Among them, h represents the adjustment rate of label correction, v represents the label adjustment coefficient, C represents the proportion of the first category samples in the data set, D represents the proportion of the second category samples in the data set, m represents the weight parameter of the first category samples, and n represents the weight parameter of the second category samples; The number of labels is adjusted according to the rate of label modification to optimize the category distribution.
[0006] Preferably, the data screening module screens and obtains the initial data set for the binary classification task according to the preset rules, including counting the feature range of each data item, counting the proportion of the median of the statistical data item to all data items, and calculating the importance score of the data item. The specific formula is: ; Among them, Z represents the importance score of the data item, E represents the range of characteristic values of each data item, y represents the adjustment coefficient, and L represents the proportion of the median of the data item to all data items.
[0007] Preferably, the prediction module uses the AI model to predict the initial data set and outputs the probability value of each data item belonging to each category, including assigning feature importance weights to each data item, analyzing the range of feature changes, and calculating the predicted probability of the data item. The specific formula is: M = N - pq 2 +psw; Among them, M represents the predicted probability of the data item, p represents the feature importance weight assigned to each data item, q represents the rate of change of data features, s represents the category-related parameter, w represents the feature mean of the data item, and N represents the total energy value of all data items.
[0008] Preferably, the classification threshold adjustment module and the label correction module are linked via an interface.
[0009] Preferably, the threshold adjustment module dynamically determines the adjustment threshold according to the current category distribution status, and re-evaluates the threshold. The calculation formula of the sensitivity coefficient a of the classification model in the probability value predicted by the model is: a=ΔP / ΔQ; Among them, a represents the sensitivity coefficient of the classification model, ΔP represents the change in the prediction accuracy of the classification model before and after adjusting the threshold, and ΔQ represents the change in the distribution of the number of category samples when the threshold is adjusted.
[0010] Preferably, the threshold adjustment module dynamically determines the adjustment threshold according to the current category distribution status, and re-evaluates the threshold. The calculation formula of the difference b between the number of first category samples and the number of second category samples in the probability value predicted by the model is: b=R1-R2; Among them, b represents the difference between the number of samples in the first category and the number of samples in the second category, R1 represents the number of samples in the first category, and R2 represents the number of samples in the second category.
[0011] Preferably, the label correction module corrects the predicted label based on the re-evaluation result, and the calculation formula of the weight parameter m of the first category of samples in the category imbalance is optimized as follows: m=R1 / (R1+R2); Among them, the weight parameter of the first category of samples, R1 represents the number of the first category of samples, and R2 represents the number of the second category of samples.
[0012] Preferably, the label correction module corrects the predicted label based on the re-evaluation result, and the calculation formula for optimizing the weight parameter n of the second category of samples in the category imbalance is: n=R2 / (R1+R2); Among them, n represents the weight parameter of the second type of samples, R1 represents the number of first type of samples, and R2 represents the number of second type of samples.
[0013] A data sorting and filtering method based on an AI model, using the data sorting and filtering system based on an AI model, the method comprising: S1. Filter and obtain the initial data set for the binary classification task according to the preset rules; S2. Use the AI model to predict the initial data set and output the probability value of each data item belonging to each category; S3. Dynamically determine the adjustment threshold according to the current category distribution, and use the threshold to re-evaluate the probability value predicted by the model; S4. Modify the predicted labels based on the re-evaluation results to optimize the class imbalance.
[0014] A storage medium stores a computer program, which, when executed by a processor, implements the data sorting and filtering method based on an AI model.
[0015] It can be seen from the above technical solution that the present invention has the following beneficial effects: The data sorting and filtering system, method and storage medium based on the AI model, by screening and obtaining the initial data set for the binary classification task according to preset rules, using the AI model to predict the initial data set, and outputting the probability value of each data item belonging to each category, dynamically determining the adjustment threshold according to the current category distribution status, and re-evaluating the probability value predicted by the model with the threshold, correcting the predicted label based on the result of the re-evaluation, and optimizing the category imbalance, can optimize the decision boundary of the model for the category imbalance problem, make the minority class samples easier to be correctly identified by the model, thereby significantly improving the recall rate of the minority class, improving the classification performance of the model on the minority class, and effectively reducing the bias of the model for the majority class, thereby achieving a balance between the false positive rate and the false negative rate, improving the overall fairness of the classification results, and being able to more flexibly adapt to different application scenarios and data distribution changes, overcome the sample redundancy problem that may occur in the traditional oversampling method, and avoid the information loss that may be caused by undersampling, and can optimize the model performance without changing the data distribution. In addition, compared with the weighted loss function, the dynamic adjustment of the threshold does not require complex hyperparameter adjustment, which reduces the complexity of model tuning. It can comprehensively consider the overall classification performance and effectively improve the model's performance in multiple dimensions such as accuracy, precision, recall rate and F1 score by balancing the decision criteria of positive and negative samples. It can provide a more robust solution for classification problems in practical applications and dynamically adjust the probability distribution threshold of the model prediction output to solve the class imbalance problem in binary classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 This is a schematic diagram of the connection of the system modules of the present invention; Figure 2 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0018] like Figure 1 As shown, the present invention provides a technical solution: a data sorting and filtering system based on an AI model, the system comprising: The data screening module is used to screen and obtain the initial data set for the binary classification task according to preset rules; The prediction module connected to the data screening module is used to use the AI model to predict the initial data set and output the probability value of each data item belonging to each category; The threshold adjustment module connected to the prediction module is used to dynamically determine the adjustment threshold according to the current category distribution status, and adjust the threshold to re-evaluate the probability value predicted by the model, including dividing the samples into the first category samples and the second category samples, counting the difference in the number of the first category samples and the second category samples, and calculating the dynamically adjusted classification threshold. The specific formula is: ; Among them, u represents the new classification threshold, a represents the sensitivity coefficient of the classification model, b represents the difference in the number of samples in the first category and the second category, and g represents the average importance score of all features in the data set; The label correction module connected to the threshold adjustment module is used to correct the predicted labels based on the re-evaluation results and optimize the category imbalance, including calculating the sample ratio in the data set, setting the weight parameters, and calculating the rate of label correction. The specific formula is: ; Among them, h represents the adjustment rate of label correction, v represents the label adjustment coefficient, C represents the proportion of the first category samples in the data set, D represents the proportion of the second category samples in the data set, m represents the weight parameter of the first category samples, and n represents the weight parameter of the second category samples; The number of labels is adjusted according to the rate of label modification to optimize the category distribution.
[0019] The data sorting and filtering system based on the AI model first uses the data screening module to filter the initial data set that meets the preset rules from the original data to ensure data quality and applicability. The prediction module uses the AI model to process the initial data set and outputs the probability value of the data item belonging to each category based on the classification ability of the model. In order to improve the accuracy of classification, the system introduces a threshold adjustment module to dynamically adjust the classification threshold according to the current category distribution. The adjustment of the threshold is based on a specific mathematical formula, in which the sensitivity coefficient and the average value of the importance evaluation of the data feature work together to optimize the setting of the classification standard so that the classification task is more adapted to the actual distribution of the data set. Next, the label correction module corrects the classification results according to the new threshold and the calculated correction rate to reduce the impact of the category imbalance problem on the model performance. By setting the weight parameters and the scale factor, the label correction module can dynamically optimize the sample distribution, thereby further improving the classification accuracy and reliability of the model. Compared with the existing technology, this system solves the deviation problem that is prone to occur in traditional classification systems when processing category imbalanced data sets by dynamically adjusting the classification threshold and correcting the label mechanism, and significantly improves the accuracy and stability of classification. Specifically: the threshold adjustment module can dynamically optimize the classification criteria according to the real-time data distribution, enhancing the system's adaptability to different data scenarios; the label correction module effectively alleviates the negative impact of category imbalance on the classification effect by introducing the correction rate formula; the multi-module collaborative design of this system significantly improves the flexibility and robustness of the classification model, while achieving continuous optimization of the classification results to meet the needs of diversified data analysis.
[0020] The data screening module screens and obtains the initial data set for the binary classification task according to the preset rules, including counting the feature range of each data item, counting the proportion of the median of the data item to all data items, and calculating the importance score of the data item. The specific formula is: ; Among them, Z represents the importance score of the data item, E represents the range of characteristic values of each data item, y represents the adjustment coefficient, and L represents the proportion of the median of the data item to all data items.
[0021] In this implementation, the core function of the data screening module is to evaluate the importance of data items through statistics and calculations, thereby optimizing the selection process of the initial data set. Specifically, the module first analyzes the characteristic value range of each data item, and combines the statistical indicator of the proportion of the median in all data items to reflect the concentrated distribution characteristics of the data. Then, by introducing the adjustment coefficient y, combined with the formula Calculate the importance score Z for each data item. Among them, the eigenvalue range E is used to measure the degree of discreteness of the data, the median ratio L emphasizes the significance of the data in the overall distribution, and the adjustment coefficient y is flexibly set according to the actual application scenario to ensure the accuracy and applicability of the score. The scoring results can provide high-quality data input for subsequent classification tasks, significantly improving classification efficiency and performance. The present invention accurately evaluates the importance of the initial data set through a data screening module, and has the following advantages: First, by using the statistical analysis of the eigenvalue range E and the median ratio L, the distribution characteristics of the data can be fully captured, thereby ensuring the scientificity and reliability of data selection. Secondly, by introducing an adjustable adjustment coefficient y, the flexibility of the scoring formula is improved to adapt to different task requirements. In addition, the module automatically calculates the importance score Z, which can quickly screen out the key data items that have the greatest impact on the classification task, thereby significantly optimizing the input quality of the classification model and improving the processing efficiency and classification accuracy of the overall system.
[0022] The prediction module uses the AI model to predict the initial data set and outputs the probability value of each data item belonging to each category, including assigning feature importance weights to each data item, analyzing the range of feature changes, and calculating the prediction probability of the data item. The specific formula is: M = N - pq 2 +psw; Among them, M represents the predicted probability of the data item, p represents the feature importance weight assigned to each data item, q represents the rate of change of data features, s represents the category-related parameter, w represents the feature mean of the data item, and N represents the total energy value of all data items.
[0023] In this implementation, the prediction module uses an AI model to calculate the probability value of a data item belonging to each category through in-depth analysis of the initial data set. First, the system assigns a feature importance weight p to each data item to quantify its contribution to the classification task. Next, the module calculates the feature change rate q by analyzing the range of feature changes to measure the dynamic change trend of data features. Then, combining the category-related parameter s and the feature mean w of the data item, according to the formula M=N-pq 2+psw calculates the predicted probability M of the data item. Among them, the total energy value N is a global measurement indicator of the system for all data item features, providing a unified probability benchmark. The comprehensive calculation method of this module can dynamically adapt to the distribution changes of data, thereby improving the accuracy of classification results. The present invention realizes efficient prediction of data classification through the feature weight allocation and dynamic probability calculation mechanism of the prediction module. Compared with the traditional method, the main advantages of this module are: through the allocation of feature importance weights p, the sensitivity of the model to key features is effectively improved, and the classification results are optimized; combined with the feature change rate q and category-related parameters s, it can flexibly respond to dynamic changes in data distribution, thereby improving classification accuracy and robustness; based on the unified probability benchmark provided by the total energy value N, the system can ensure the balance and consistency of different data items during the calculation process. In addition, the module efficiently calculates the predicted probability of each data item, providing a reliable basis for the classification optimization and adjustment of subsequent modules.
[0024] The classification threshold adjustment module and the label correction module are linked through an interface. In this embodiment, the classification threshold adjustment module and the label correction module realize linkage operation through an interface to ensure the coordination of the adjustment of the classification threshold and the label correction process. The classification threshold adjustment module dynamically calculates the classification threshold according to the real-time data distribution, and transmits the adjusted threshold to the label correction module through an interface. After receiving the classification threshold, the label correction module further optimizes and corrects the classification result, including adjusting the label distribution, correcting the category imbalance, etc. Through the interface linkage mechanism, the two modules can share key parameters to avoid classification deviations that may be caused by independent operations, thereby realizing efficient collaborative processing of data classification tasks. This linkage mode can ensure the integrity of the data flow and the consistency of the processing steps, and provide highly integrated operation support for the system. Through the interface linkage between the classification threshold adjustment module and the label correction module, the present invention significantly improves the coordination and efficiency of the system. First, the linkage mechanism ensures the real-time adjustment of the classification threshold and the synchronization of the label correction, thereby improving the accuracy and stability of the classification results. Secondly, the interface simplifies the data interaction process between the modules, effectively reducing the complexity of the system design and the delay of data transmission. In addition, the linkage operation can dynamically adapt to different data distribution situations, which plays a significant role in dealing with category imbalance problems and further improves the robustness and flexibility of the system.
[0025] The threshold adjustment module dynamically determines the adjustment threshold according to the current category distribution, and re-evaluates the threshold. The calculation formula of the sensitivity coefficient a of the classification model in the probability value predicted by the model is: a=ΔP / ΔQ; Among them, a represents the sensitivity coefficient of the classification model, ΔP represents the change in the prediction accuracy of the classification model before and after adjusting the threshold, and ΔQ represents the change in the distribution of the number of category samples when the threshold is adjusted.
[0026] In this embodiment, the threshold adjustment module realizes dynamic correlation analysis between the prediction performance of the model and the change in the distribution of the number of category samples by calculating the sensitivity coefficient a of the classification model. Specifically, the module first obtains the change in the prediction accuracy of the model before and after adjusting the classification threshold, ΔP, and simultaneously counts the change in the distribution of the number of category samples caused by the threshold adjustment, ΔQ. Subsequently, the sensitivity coefficient a of the classification model is calculated according to the formula a=ΔP / ΔQ. The sensitivity coefficient a is used to measure the sensitivity of the model's prediction performance to changes in the distribution of category samples. The larger its value, the higher the effect optimization ability of the adjustment of the classification model under the current threshold. This process provides a scientific basis for dynamically adjusting the classification threshold, ensuring the adaptability and accuracy of the classification task under different category distributions. The present invention provides a quantitative evaluation method for classification threshold adjustment by calculating the sensitivity coefficient a, and has the following significant advantages over the prior art: the introduction of the sensitivity coefficient a enables the system to accurately evaluate the actual impact of threshold adjustment on the performance of the classification model, thereby realizing dynamic optimization of the classification process; based on the joint analysis of the change in prediction accuracy ΔP and the change in category sample distribution ΔQ, the system can effectively balance the relationship between classification accuracy and category distribution, which is particularly suitable for scenarios with uneven data distribution; the sensitivity coefficient calculation mechanism in the threshold adjustment process can adapt to a variety of classification tasks, thereby improving the versatility and stability of the system.
[0027] The threshold adjustment module dynamically determines the adjustment threshold according to the current category distribution status, and re-evaluates the threshold. The calculation formula of the difference b between the number of samples in the first category and the second category in the probability value predicted by the model is: b = R1-R2; Among them, b represents the difference between the number of samples in the first category and the number of samples in the second category, R1 represents the number of samples in the first category, and R2 represents the number of samples in the second category.
[0028] In this embodiment, the threshold adjustment module dynamically optimizes the setting of the classification threshold by calculating the difference b between the number of samples of the first category and the number of samples of the second category. The module firstly counts the number of samples of the first category R1 and the number of samples of the second category R2, and then calculates the difference value b between the number of samples of the two categories according to the formula b=R1-R2. The size of the difference value b directly reflects the degree of distribution imbalance of the two categories of samples in the data set, and serves as the core parameter for subsequent adjustment of the classification threshold. Through this dynamic evaluation method, the threshold adjustment module can adapt to the distribution of different categories, thereby optimizing the evaluation criteria of the model prediction probability and improving the performance of the classification task under the condition of category imbalance. The present invention provides a simple and effective quantitative indicator for threshold adjustment by calculating the difference b in the number of samples of the two categories, and has the following advantages over the prior art: by directly counting the difference in the number of samples of the two categories, it can quickly and accurately reflect the degree of imbalance of data distribution, thereby providing a reliable basis for the dynamic adjustment of the classification threshold; the introduction of the difference value b simplifies the calculation process of the threshold adjustment and effectively reduces the calculation complexity; the threshold optimization mechanism based on the difference in the number of samples can significantly improve the robustness and prediction accuracy of the classification model in the unbalanced data scenario; the system can adapt to the changes in the category distribution in real time to meet the needs of the dynamic data scenario.
[0029] The label correction module corrects the predicted label based on the re-evaluation results, and the calculation formula for optimizing the weight parameter m of the first class of samples in the class imbalance is: m = R1 / (R1 + R2); Among them, the weight parameter of the first category of samples, R1 represents the number of the first category of samples, and R2 represents the number of the second category of samples.
[0030] In this implementation, the label correction module optimizes the class imbalance problem by calculating the weight parameter m of the first class of samples. The module first counts the number of first class samples R1 and the number of second class samples R2, and then calculates the weight parameter m of the first class of samples according to the formula m=R1 / (R1+R2). The weight parameter m represents the proportion of first class samples in the total samples, which is used to measure the balance of class distribution and dynamically adjust the predicted labels. In this way, the label correction module can enhance the model's attention to the first class of samples in the class imbalance scenario, thereby improving the accuracy and fairness of the classification results. The present invention provides an effective quantitative means for label correction by introducing the weight parameter m of the first category of samples, and has the following significant advantages: the calculation of the weight parameter m is based on the relative proportion of the number of samples, which can accurately reflect the category distribution status, thereby optimizing the label correction strategy; by dynamically adjusting the weight of the first category of samples, the system significantly improves the recognition ability of minority categories in unbalanced data scenarios, and improves the bias of classification results; the introduction of the weight parameter simplifies the design of the label correction module, effectively reduces the calculation complexity, and ensures the real-time performance of the adjustment process; while optimizing the category distribution, the method can balance the classification accuracy and robustness, and provides a highly applicable solution for diversified data scenarios.
[0031] The label correction module corrects the predicted label based on the re-evaluation results, and the calculation formula for optimizing the weight parameter n of the second category of samples in the category imbalance is: n = R2 / (R1 + R2); Among them, n represents the weight parameter of the second type of samples, R1 represents the number of first type of samples, and R2 represents the number of second type of samples.
[0032] In this embodiment, the label correction module effectively optimizes the classification results in the case of class imbalance by calculating the weight parameter n of the second class samples. The module first counts the number of first class samples R1 and the number of second class samples R2, and then calculates the weight parameter n of the second class samples according to the formula n=R2 / (R1+R2). The weight parameter n represents the proportion of the second class samples in the total samples, and is used to dynamically adjust the classification model's attention to the second class samples. Through this weight allocation mechanism, the label correction module can balance the processing of different categories in the case of uneven class distribution, thereby improving the applicability and prediction accuracy of the classification model. The present invention realizes dynamic optimization of label correction by calculating the weight parameter n of the second category samples, and has the following significant advantages compared with the prior art: the weight parameter n is calculated based on the relative proportion of the number of second category samples, which can intuitively reflect the degree of category imbalance and provide a reliable basis for label correction; by dynamically adjusting the weight of the second category samples, the system enhances the classification model's recognition ability for minority categories and effectively improves the category imbalance problem; the introduction of weight parameters simplifies the design of the label correction module, reduces the calculation complexity, and ensures the real-time and efficiency of the adjustment process; while optimizing the category distribution, this method improves the robustness and flexibility of the classification task and is suitable for a variety of data scenarios.
[0033] like Figure 2 As shown, a data sorting and filtering method based on an AI model is also provided, using the data sorting and filtering system based on an AI model, the method includes: S1. Filter and obtain the initial data set for the binary classification task according to the preset rules; S2. Use the AI model to predict the initial data set and output the probability value of each data item belonging to each category; S3. Dynamically determine the adjustment threshold according to the current category distribution, and use the threshold to re-evaluate the probability value predicted by the model; S4. Modify the predicted labels based on the re-evaluation results to optimize the class imbalance.
[0034] In this embodiment, the method completes the sorting and classification optimization of data through the synergistic action of multiple steps. First, in step S1, the system selects an initial data set that meets the task requirements from the original data according to preset rules to ensure data quality and applicability. Then, in step S2, the system uses the AI model to predict the initial data set, and calculates the probability value of each data item belonging to each category, and the output result is the classification probability distribution. In step S3, the system optimizes the prediction criteria of the model by dynamically adjusting the classification threshold based on the current category distribution, thereby re-evaluating the classification probability of the data item. In the final step S4, the system corrects the predicted label based on the re-evaluated probability results, and effectively alleviates the category imbalance problem by adjusting the sample weight, optimizing the category distribution, and other means. Through this complete process, this method can significantly improve the performance and robustness of the classification model. The method of the present invention achieves comprehensive optimization of data collation and classification through systematic step design, and has the following significant advantages: in step S1, the initial data set is screened by preset rules, which effectively ensures the quality and applicability of data input; in step S2, the classification probability value is output by the AI model, which provides an accurate data basis for subsequent adjustment and correction; the dynamic threshold adjustment mechanism of step S3 enables the system to adapt to different category distributions, improving the flexibility and scientificity of the classification standard; the label correction method of step S4 significantly alleviates the category imbalance problem by optimizing sample weights and category distribution, thereby enhancing the accuracy and robustness of the classification model. In addition, the multi-module collaborative design of this method greatly improves the efficiency and reliability of data processing, and is suitable for a variety of classification task scenarios.
[0035] A storage medium is also provided, on which a computer program is stored, and when the computer program is executed by a processor, the data sorting and filtering method based on the AI model is implemented. In this embodiment, the computer program stored on the storage medium implements the data sorting and filtering method based on the AI model through the design of an instruction set. Specifically, when the program is loaded and executed by the processor, the following functions are implemented in sequence according to the established process: initial data set acquisition module: execute step S1 through program instructions, screen and obtain the initial data set for the binary classification task according to preset rules, and ensure that the data quality meets the task requirements; prediction module: according to step S2, call the AI model to predict the initial data set and output the classification probability of each data item; threshold adjustment module: according to step S3, dynamically analyze the current category distribution status, adjust the classification threshold and re-evaluate the probability value predicted by the model; label correction module: by executing step S4, correct the predicted label and optimize the category imbalance problem, such as dynamically adjusting the sample weight and distribution, to ensure further optimization of the model classification performance. The computer program uses modular design logic to enable each step to run independently and process collaboratively, and finally complete the entire data sorting and filtering method. The present invention provides an efficient and scalable technical solution for data sorting and filtering based on AI models by storing computer programs in a storage medium, and has the following advantages: through the automated execution of the program, the full-process processing of the method is fully realized, human intervention is reduced, and data processing efficiency is improved; the modular design of the program facilitates maintenance and expansion, and the specific implementation can be flexibly adjusted for different data scenarios; long-term storage in a storage medium facilitates cross-platform deployment and multi-device sharing, thereby improving the versatility of the system; the program can work in conjunction with the hardware system during execution, for example, in combination with a high-performance processor or acceleration chip, to further improve data processing speed and accuracy.
[0036] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. The data sorting and filtering system based on AI model is characterized by: The system comprises: The data screening module is used to screen and obtain the initial data set for the binary classification task according to preset rules; The prediction module connected to the data screening module is used to use the AI model to predict the initial data set and output the probability value of each data item belonging to each category; The threshold adjustment module connected to the prediction module is used to dynamically determine the adjustment threshold according to the current category distribution status, and use the adjustment threshold to re-evaluate the probability value predicted by the model, including dividing the samples into the first category samples and the second category samples, counting the difference in the number of the first category samples and the second category samples, and calculating the dynamically adjusted classification threshold. The specific formula is: ; Among them, u represents the new classification threshold, a represents the sensitivity coefficient of the classification model, b represents the difference in the number of samples in the first category and the second category, and g represents the average importance score of all features in the data set; The label correction module connected to the threshold adjustment module is used to correct the predicted labels based on the re-evaluation results and optimize the category imbalance, including calculating the sample ratio in the data set, setting the weight parameters, and calculating the rate of label correction. The specific formula is: ; Among them, h represents the adjustment rate of label correction, v represents the label adjustment coefficient, C represents the proportion of the first category samples in the data set, D represents the proportion of the second category samples in the data set, m represents the weight parameter of the first category samples, and n represents the weight parameter of the second category samples; The number of labels is adjusted according to the rate of label modification to optimize the category distribution.
2. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The data screening module screens and obtains the initial data set for the binary classification task according to the preset rules, including counting the feature range of each data item, counting the proportion of the median of the statistical data item to all data items, and calculating the importance score of the data item. The specific formula is: ; Among them, Z represents the importance score of the data item, E represents the range of characteristic values of each data item, y represents the adjustment coefficient, and L represents the proportion of the median of the data item to all data items.
3. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The prediction module uses the AI model to predict the initial data set and outputs the probability value of each data item belonging to each category, including assigning feature importance weights to each data item, analyzing the range of feature changes, and calculating the prediction probability of the data item. The specific formula is: M = N-pq 2 +psw; Among them, M represents the predicted probability of the data item, p represents the feature importance weight assigned to each data item, q represents the rate of change of data features, s represents the category-related parameter, w represents the feature mean of the data item, and N represents the total energy value of all data items.
4. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The classification threshold adjustment module and the label correction module are linked via an interface.
5. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The threshold adjustment module dynamically determines the adjustment threshold according to the current category distribution status, and calculates the sensitivity coefficient a of the classification model in the probability value predicted by the threshold re-evaluation model as follows: a=ΔP / ΔQ; Among them, a represents the sensitivity coefficient of the classification model, ΔP represents the change in the prediction accuracy of the classification model before and after adjusting the threshold, and ΔQ represents the change in the distribution of the number of category samples when the threshold is adjusted.
6. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The threshold adjustment module dynamically determines the adjustment threshold according to the current category distribution status, and re-evaluates the threshold. The calculation formula of the difference b between the number of first category samples and the number of second category samples in the probability value predicted by the model is: b=R1-R2; Among them, b represents the difference between the number of samples in the first category and the number of samples in the second category, R1 represents the number of samples in the first category, and R2 represents the number of samples in the second category.
7. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The label correction module corrects the predicted label based on the re-evaluation result, and the calculation formula of the weight parameter m of the first class of samples in the class imbalance is optimized as follows: m=R1 / (R1+R2); Among them, the weight parameter of the first category of samples, R1 represents the number of the first category of samples, and R2 represents the number of the second category of samples.
8. The data sorting and filtering system based on the AI model according to claim 1 is characterized in that: The label correction module corrects the predicted label based on the re-evaluation result, and the calculation formula of the weight parameter n of the second category of samples in the category imbalance is optimized as follows: n=R2 / (R1+R2); Among them, n represents the weight parameter of the second type of samples, R1 represents the number of first type of samples, and R2 represents the number of second type of samples.
9. A data sorting and filtering method based on an AI model, using the data sorting and filtering system based on an AI model according to any one of claims 1 to 8, characterized in that: The method comprises: S1. Filter and obtain the initial data set for the binary classification task according to the preset rules; S2. Use the AI model to predict the initial data set and output the probability value of each data item belonging to each category; S3. Dynamically determine the adjustment threshold according to the current category distribution, and use the threshold to re-evaluate the probability value predicted by the model; S4. Modify the predicted labels based on the re-evaluation results to optimize the class imbalance.
10. A storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the data sorting and filtering method based on the AI model described in claim 9 is implemented.