A machine learning based data reproducible versioning pipeline management method

Abstract

The present application relates to the technical field of machine learning data reproducibility, and more particularly to a data reproducible version pipeline management method based on machine learning, which submits target data to a versioned data processing pipeline, performs data analysis and generates a versioned execution record, collects a feature health degree parameter of the target data in the versioned execution record and analyzes a feature health degree index, determines whether the matching degree of feature extraction of the target data meets a standard based on the feature health degree index, collects a feature fidelity parameter of a target data reproduction result in the versioned execution record and analyzes a feature fidelity index in response to the matching degree meeting the standard, determines whether the reliability of the target data reproduction result is abnormal based on the feature fidelity index, calculates a reproduction abnormality index in response to the reliability being abnormal, and determines the cause of the reliability abnormality and corresponding processing measures based on the reproduction abnormality index. The present application improves the reliability of data traceability in the reproduction link.

Description

Technical Field

[0001] This invention relates to the field of machine learning data reproducibility technology, and in particular to a machine learning-based method for managing reproducible versioned pipelines. Background Technology

[0002] In data analysis scenarios with high reliability requirements, such as financial risk control and judicial evidence collection, existing data processing pipelines generally lack a full-chain version control mechanism. Execution records for stages such as data cleaning, feature extraction, and model inference are fragmented, making it impossible to fully trace the correspondence between input data, algorithm versions, dependent environments, and output results. Simultaneously, the lack of quantitative evaluation methods for feature extraction quality makes it difficult to perceive issues such as the decay of correlation between features and business labels, and statistical fluctuations across time windows in real time. Data reproduction is difficult to verify the integrity of the reproduction chain through technical means, leading to inconsistent results when the same data is processed by different personnel at different times. This makes auditing and traceability difficult and seriously affects the accuracy of decision-making and business compliance.

[0003] Chinese Patent Publication No. CN111274235A discloses a method for data cleaning and protocol field feature extraction of unknown protocols, comprising the following steps: First, capturing network data packets at network aggregation points; Second, marking the captured network data packets according to the capture time sequence; Third, obtaining cleaned unknown protocol data through data classification and aggregation, obtaining public data addresses, obtaining server addresses, data reclassification, and excluding known protocols; Fourth, extracting unknown protocol field features by calculating the mean and variance distribution of n-bit source server data packets, calculating the similarity of the mean and variance distribution of different n-bit source server data packets in the same type of data, and calculating the stability of similar field features; Fifth, confirming the unknown protocol field features, establishing a matching expression for protocol field feature recognition, and storing it in an unknown protocol field feature library. Summary of the Invention

[0004] To address this, the present invention provides a machine learning-based data reproducible versioned pipeline management method to overcome the problems in the prior art, which, due to the lack of full-link versioned management of data, lead to difficulties in perceiving the degradation of data processing feature extraction logic and tracing the broken links in the reproducibility pipeline, resulting in low data analysis credibility.

[0005] To achieve the above objectives, this invention provides a data reproducible versioned pipeline management method based on machine learning, comprising:

[0006] Submit the target data to the versioned data processing pipeline for data analysis and generate versioned execution records.

[0007] Collect feature health parameters of target data from versioned execution records within historical periods;

[0008] Analyze the characteristic health index based on the aforementioned characteristic health parameters;

[0009] The matching degree of the target data feature extraction is determined based on the comparison result between the feature health index and the predetermined feature health qualification threshold.

[0010] When the matching degree meets the standard, the feature fidelity parameter of the target data reproduction result is collected from the versioned execution record within the historical period.

[0011] The feature fidelity index is analyzed based on the aforementioned feature fidelity parameters;

[0012] The reliability of the target data reproduction result is determined based on the comparison result between the feature fidelity index and the predetermined feature fidelity qualification threshold.

[0013] When a reliability anomaly occurs, a reproducibility index is calculated.

[0014] Based on the aforementioned reproducibility anomaly index, the causes of reliability anomalies and corresponding handling measures are determined, including determining the adjustment range of the characteristic health qualification threshold and the adjustment range of the binding completion rate threshold.

[0015] The feature health parameters include label relevance and feature matching similarity, while the feature fidelity parameters include feature value overlap and RunID binding completion rate.

[0016] Furthermore, the process of analyzing the characteristic health index based on the aforementioned characteristic health parameters includes:

[0017] Extract the label relevance and feature matching similarity of target data from versioned execution records within historical periods;

[0018] The ratio of the calculated label relevance to a predetermined label relevance threshold is determined as the first health factor;

[0019] The ratio of the calculated feature matching similarity to a predetermined feature matching similarity threshold is used as the second health factor.

[0020] The sum of the first health factor and the second health factor is determined as the characteristic health index.

[0021] Furthermore, the process of determining whether the matching degree of the target data feature extraction meets the standard based on the comparison result between the feature health index and the predetermined feature health qualification threshold includes:

[0022] If the feature health index is less than or equal to the predetermined feature health qualification threshold, then the matching degree is determined to be non-compliant with the standard.

[0023] If the feature health index is greater than the predetermined feature health qualification threshold, then the matching degree is determined to meet the standard.

[0024] Furthermore, the process of analyzing the feature fidelity index based on the feature fidelity parameters includes:

[0025] Extract the feature value overlap and RunID binding completion rate of the target data reproduction results from the versioned execution records within the historical period;

[0026] The ratio of the calculated eigenvalue overlap degree to the predetermined eigenvalue overlap degree threshold is determined as the first fidelity factor;

[0027] The ratio of the binding completion rate to the predetermined binding completion rate threshold is determined as the second fidelity factor;

[0028] The sum of the first fidelity factor and the second fidelity factor is determined as the feature fidelity index.

[0029] Furthermore, the process of determining whether the reliability of the target data reproduction result is abnormal based on the comparison result between the feature fidelity index and the predetermined feature fidelity pass threshold includes:

[0030] If the feature fidelity index is less than or equal to the predetermined feature fidelity pass threshold, then the reliability is determined to be abnormal.

[0031] If the feature fidelity index is greater than the predetermined feature fidelity qualification threshold, then the reliability is determined to be normal.

[0032] Furthermore, the process of calculating the recurrence anomaly index includes:

[0033] Extract the predetermined feature fidelity pass threshold and feature fidelity index;

[0034] Calculate the difference between the predetermined feature fidelity pass threshold and the feature fidelity index;

[0035] The difference is determined as the recurrence anomaly index.

[0036] Furthermore, the process of determining the cause of reliability anomalies based on the reproducibility anomaly index includes:

[0037] If the recurrence anomaly index is less than or equal to the predetermined recurrence anomaly index threshold, it is determined that there is a slight degradation in the feature extraction logic;

[0038] If the anomaly index is greater than the predetermined anomaly index threshold, it is determined that the data reproduction link is broken.

[0039] Furthermore, the process of matching appropriate handling measures based on the cause of the reliability anomaly includes:

[0040] If the cause is a slight degradation in the feature extraction logic, then determine the adjustment range of the feature health qualification threshold;

[0041] If the cause is a broken data reproduction link, then determine the adjustment range of the binding completion rate threshold.

[0042] Furthermore, the adjustment range of the characteristic health qualification threshold is determined based on the recurrence anomaly index, wherein the characteristic health qualification threshold is positively correlated with the recurrence anomaly index.

[0043] Furthermore, the adjustment range of the binding completion rate threshold is determined based on the recurrence anomaly index, wherein the binding completion rate threshold is positively correlated with the recurrence anomaly index.

[0044] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a machine learning-based method for managing reproducible versioned data pipelines. The method involves submitting target data to a versioned data processing pipeline for data analysis and generation of versioned execution records; collecting feature health parameters of the target data from the versioned execution records within historical periods; analyzing the feature health index based on the feature health parameters; determining whether the matching degree of the target data feature extraction meets the standard based on the comparison result of the feature health index and a predetermined feature health qualification threshold; in response to the matching degree meeting the standard, collecting feature fidelity parameters of the target data reproduction results from the versioned execution records within historical periods; analyzing the feature fidelity index based on the feature fidelity parameters; determining whether the reliability of the target data reproduction results is abnormal based on the comparison result of the feature fidelity index and a predetermined feature fidelity qualification threshold; in response to reliability anomalies, calculating a reproduction anomaly index; and determining the cause of the reliability anomaly and its corresponding handling measures based on the reproduction anomaly index. This invention overcomes the problems in the prior art, which are caused by the lack of full-link version management of data, resulting in difficulties in perceiving the degradation of data processing feature extraction logic and in tracing the broken links, leading to low data analysis credibility.

[0045] In particular, this invention obtains a feature health parameter by collecting the label relevance and feature matching similarity of target data in versioned execution records within historical periods, and then quantifies and sums the feature health parameter to obtain a feature health index, which is used to characterize the accuracy of the feature extraction results of target data in versioned execution records. This allows for an objective judgment on whether the matching degree of target data feature extraction meets the standard, thereby improving the stability of the output quality of the versioned data processing pipeline during long-term operation.

[0046] In particular, this invention obtains feature fidelity parameters by further collecting the feature value overlap of the target data reproduction results and the RunID binding completion rate in the versioned execution records within historical periods, and calculates the feature fidelity index from the feature fidelity parameters to reflect the overlap between the target data reproduction results and the original results in the versioned execution records. This can objectively assess the performance degradation and breakage that occurs at any stage in the versioned data processing pipeline, thereby improving the stability of the data analysis process.

[0047] In particular, this invention determines whether the matching degree of the target data feature extraction meets the standard by comparing the feature health index with the predetermined feature health qualification threshold. If the feature health index is greater than the predetermined feature health qualification threshold, the matching degree of the target data feature extraction is determined to meet the standard, allowing it to proceed to the subsequent judgment stage. By comparing the feature fidelity index with the predetermined feature fidelity qualification threshold, when the feature fidelity index does not exceed the predetermined feature fidelity qualification threshold, it indicates that the reliability of the target data reproduction result is abnormal, thereby improving the decision-making efficiency of pipeline output quality management.

[0048] In particular, this invention, after determining that the reliability of the target data reproduction result is abnormal, calculates the difference between a predetermined feature fidelity pass threshold and a feature fidelity index as a reproduction anomaly index. The magnitude of the reproduction anomaly index characterizes the severity of the reliability anomaly problem, determining whether the anomaly is caused by a slight degradation in the feature extraction logic or a break in the data reproduction link, and matching corresponding processing measures: determining the adjustment range of the feature health pass threshold and the adjustment range of the binding completion rate threshold. This improves the accuracy of anomaly problem resolution. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the steps of the machine learning-based data reproducible versioned pipeline management method according to an embodiment of the present invention.

[0050] Figure 2 This invention provides a logical decision diagram for determining whether the matching degree of target data feature extraction conforms to the standard based on the feature health index.

[0051] Figure 3 The present invention provides a logical judgment diagram for determining whether the reliability of the target data reproduction result is abnormal based on the feature fidelity index.

[0052] Figure 4 This invention provides a logical decision diagram for determining the causes of reliability anomalies and corresponding handling measures based on the reproducibility anomaly index. Detailed Implementation

[0053] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0054] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0055] Please see Figure 1 The diagram shows the steps of a machine learning-based data reproducible versioned pipeline management method according to an embodiment of the present invention. The present invention provides a machine learning-based data reproducible versioned pipeline management method, comprising:

[0056] Step S1: Submit the target data to the versioned data processing pipeline for data analysis and generate versioned execution records;

[0057] Step S2: Collect the feature health parameters of the target data in the versioned execution records within the historical period;

[0058] Step S3: Analyze the feature health index based on the feature health parameters;

[0059] Step S4: Determine whether the matching degree of the target data feature extraction meets the standard based on the comparison result between the feature health index and the predetermined feature health qualification threshold;

[0060] Step S5: In response to the matching degree meeting the standard, collect the feature fidelity parameter of the target data reproduction result in the versioned execution record within the historical period;

[0061] Step S6: Analyze the feature fidelity index based on the feature fidelity parameters;

[0062] Step S7: Determine whether the reliability of the target data reproduction result is abnormal based on the comparison result between the feature fidelity index and the predetermined feature fidelity qualification threshold;

[0063] Step S8: In response to reliability anomalies, calculate the reproducibility index;

[0064] Step S9: Determine the cause of reliability anomalies and corresponding handling measures based on the reproducibility anomaly index, including determining the adjustment range of the characteristic health qualification threshold and the adjustment range of the binding completion rate threshold.

[0065] The feature health parameters include label relevance and feature matching similarity, while the feature fidelity parameters include feature value overlap and RunID binding completion rate.

[0066] Understandably, the specific steps of a versioned data processing pipeline include: input data cleaning, data feature extraction, data statistical analysis, visualization chart generation, and integrating the output results of the aforementioned steps to generate a versioned execution record. The versioned data processing pipeline matches a unique RunID to the input data involved in each data analysis request. When the versioned data processing pipeline starts, it collects and locks the hash value of the currently used input data, the script code submission version of each processing step, the parameter configuration file, and the runtime environment image identifier, and matches them with a globally unique RunID. The naming and storage of all intermediate files and the final report generated during the execution of all data processing steps are strongly bound to the matched RunID. This ensures that any data analysis output result can be traced back to the precise data state and code logic relied upon at the time of data input, thereby improving the transparency of the analysis process and the reliability of the conclusions.

[0067] It's understandable that label relevance refers to the standardized mutual information value between target data features and target identifiers, derived from the feature matrix and corresponding label vectors stored in the versioned execution record. A higher mutual information value indicates a stronger correlation between the target data features and the target identifier.

[0068] As can be understood, feature matching similarity refers to the Euclidean distance between corresponding points in space between two feature vectors before and after feature extraction. The larger the distance value, the greater the difference between the vectors, the lower the feature matching similarity, and the worse the reproducibility.

[0069] It is understandable that the feature value overlap refers to the proportion of features whose feature values ​​are consistent with the original feature values ​​after the target data is reproduced within a historical period to the total number of features. The higher the feature value overlap, the greater the similarity between the reproduced result and the original result, and the better the reproduction quality.

[0070] As can be understood, the RunID binding completion rate refers to the proportion of files whose output results at each step of the versioned data processing pipeline within a historical period are successfully bound to a unique RunID, out of the total number of files. A higher binding completion rate indicates a stronger link tracing completion rate.

[0071] In this embodiment, the single acquisition cycle is preset, and the preferred single acquisition cycle is 50 μs.

[0072] This invention provides a machine learning-based method for managing reproducible versioned data pipelines. The method involves submitting target data to a versioned data processing pipeline for data analysis and generation of versioned execution records. It collects feature health parameters of the target data from the versioned execution records within historical periods; analyzes a feature health index based on these parameters; determines whether the matching degree of the target data feature extraction meets a standard based on a comparison of the feature health index with a predetermined feature health qualification threshold; when the matching degree meets the standard, it collects feature fidelity parameters of the target data reproduction results from the versioned execution records within historical periods; analyzes a feature fidelity index based on these parameters; determines whether the reliability of the target data reproduction results is abnormal based on a comparison of the feature fidelity index with a predetermined feature fidelity qualification threshold; when reliability is abnormal, it calculates a reproduction anomaly index; and determines the cause of the reliability anomaly and corresponding handling measures based on the reproduction anomaly index. This invention improves the reliability of data traceability in the reproduction chain.

[0073] Specifically, the process of analyzing the characteristic health index based on the aforementioned characteristic health parameters includes:

[0074] Extract the label relevance and feature matching similarity of target data from versioned execution records within historical periods;

[0075] The ratio of the calculated label relevance to a predetermined label relevance threshold is determined as the first health factor;

[0076] The ratio of the calculated feature matching similarity to a predetermined feature matching similarity threshold is used as the second health factor.

[0077] The sum of the first health factor and the second health factor is determined as the characteristic health index.

[0078] In this embodiment, the predetermined label relevance threshold is preset. Specifically, label relevance samples within six historical collection periods are predetermined, and the predetermined label relevance threshold is determined based on the average value of the label relevance samples. The threshold is determined within the range [0.5, 0.7]. Based on the requirements of feature discrimination in the actual business scenario, an optimal value that meets 85% of the upper limit of the measured value is selected. In this embodiment, the predetermined label relevance threshold is preferably 0.6.

[0079] In this embodiment, the predetermined feature matching similarity threshold is preset. Specifically, feature matching similarity samples within 6 historical collection periods are predetermined, and the predetermined feature matching similarity threshold is determined based on the average value of the feature matching similarity samples. The threshold is determined within the range [0.2, 0.3]. Based on the tolerance of feature stability in the actual business scenario, an optimal value that meets 90% of the upper limit of the measured value is selected. In this embodiment, the predetermined feature matching similarity threshold is preferably 0.27.

[0080] This invention sets a baseline threshold based on historical data collection cycles, associates quality standards with historical operating status, and sums the feature health index by the ratio of the feature health parameter to the tag correlation threshold and the tag correlation threshold, respectively. This enables the feature health index evaluation criteria to adapt to different data sources and analysis tasks, thereby improving the accuracy of quality monitoring in versioned data processing pipelines.

[0081] Please see Figure 2 As shown, this is a logical judgment diagram for determining whether the matching degree of target data feature extraction meets the standard based on the feature health index in an embodiment of the present invention. The process of determining whether the matching degree of target data feature extraction meets the standard based on the comparison result of the feature health index and the predetermined feature health qualification threshold includes:

[0082] If the feature health index is less than or equal to the predetermined feature health qualification threshold, then the matching degree is determined to be non-compliant with the standard.

[0083] If the feature health index is greater than the predetermined feature health qualification threshold, then the matching degree is determined to meet the standard.

[0084] In this embodiment, the predetermined characteristic health qualification threshold is preset. Specifically, the average value of the characteristic health index over 10 historical collection periods is predetermined, and the predetermined characteristic health qualification threshold is determined based on the average value of the characteristic health index, or determined within the range [1.57, 2.28]. Based on the business risk tolerance level and quality assurance principle, an optimal value that meets 92% of the upper limit of the measured value is selected. In this embodiment, the predetermined characteristic health qualification threshold is preferably 2.10.

[0085] By associating the qualified threshold of feature health with historical performance, the present invention makes the judgment of the current feature extraction quality based on the statistical evaluation of recent historical performance. This can respond to the slow changes in data features and provide early warning of the performance degradation trend of feature extraction, thereby improving the stability of the quality control of the versioned data processing pipeline.

[0086] Specifically, the process of analyzing the feature fidelity index based on the feature fidelity parameters includes:

[0087] Extract the feature value overlap and RunID binding completion rate of the target data reproduction results from the versioned execution records within the historical period;

[0088] The ratio of the calculated eigenvalue overlap degree to the predetermined eigenvalue overlap degree threshold is determined as the first fidelity factor;

[0089] The ratio of the binding completion rate to the predetermined binding completion rate threshold is determined as the second fidelity factor;

[0090] The sum of the first fidelity factor and the second fidelity factor is determined as the feature fidelity index.

[0091] In this embodiment, the predetermined feature value overlap threshold is preset. Specifically, feature value overlap samples within six historical acquisition periods are predetermined, and the predetermined feature value overlap threshold is determined based on the average value of the feature value overlap samples, or determined within the range [98.2, 99.4]. Based on the accuracy requirements of reproduction and the tolerance of storage and calculation errors, an optimal value that meets the upper limit of the measured value of 99% is selected. In this embodiment, the predetermined feature value overlap threshold is preferably 98.4%.

[0092] In this embodiment, the predetermined binding completion rate threshold is preset. Specifically, binding completion rate samples within six historical collection periods are predetermined, and the predetermined binding completion rate threshold is determined based on the average value of the binding completion rate samples. The threshold is determined within the range [98.3, 99.8]. Based on the completeness requirements of audit traceability and the performance of source data management, an optimal value that meets the upper limit of the measured value of 99% is selected. In this embodiment, the predetermined binding completion rate threshold is preferably 98.8%.

[0093] This invention sets two thresholds for feature fidelity evaluation parameters and calculates the ratios of the feature fidelity evaluation parameters to the thresholds, thereby quantifying the feature fidelity parameters into a feature fidelity index. This ensures the integrity of the data chain used for traceability and enables versioned data processing pipelines to identify data reproduction deviations and link risks, thus improving the accuracy of reproduction results.

[0094] Please see Figure 3 As shown, this is a logic diagram for determining whether the reliability of the target data reproduction result is abnormal based on the feature fidelity index in an embodiment of the present invention. The process of determining whether the reliability of the target data reproduction result is abnormal based on the comparison result of the feature fidelity index and the predetermined feature fidelity pass threshold includes:

[0095] If the feature fidelity index is less than or equal to the predetermined feature fidelity pass threshold, then the reliability is determined to be abnormal.

[0096] If the feature fidelity index is greater than the predetermined feature fidelity qualification threshold, then the reliability is determined to be normal.

[0097] In this embodiment, the predetermined feature fidelity qualification threshold is preset. Specifically, the average value of the feature fidelity index over 10 historical acquisition periods is predetermined. The predetermined feature fidelity qualification threshold is determined based on the average value of the feature fidelity index, within the range [1.991, 2.020]. Based on the requirements for reproducibility and auditability, an optimal value that meets the upper limit of the measured value of 99.5% is selected. In this embodiment, the predetermined feature fidelity qualification threshold is preferably 2.010.

[0098] By setting a predetermined feature fidelity qualification threshold within an interval determined by historical performance data and selecting an optimal value based on the requirements for reproducibility and auditability, the invention enables the interpretability of the conclusions of each reproducibility audit, thereby improving the reliability and long-term quality stability of the versioned data processing pipeline output results.

[0099] Specifically, the process of calculating the recurrence anomaly index includes:

[0100] Extract the predetermined feature fidelity pass threshold and feature fidelity index;

[0101] Calculate the difference between the predetermined feature fidelity pass threshold and the feature fidelity index;

[0102] The difference is determined as the recurrence anomaly index.

[0103] Understandably, the anomaly reproducibility index is a positive value. Its magnitude reflects the degree of deviation between the actual performance of the versioned data processing pipeline and the expected standard. A higher anomaly reproducibility index indicates a more severe anomaly, while a lower index indicates a smaller deviation and a less severe anomaly.

[0104] This invention transforms the degree of reliability anomaly into a clear quantitative indicator by calculating the difference between the feature fidelity pass threshold and the actual feature fidelity index. This provides a precise input basis for subsequent root cause analysis and differentiated processing, allowing the formulation and selection of processing measures to be based on the severity of the anomaly, thereby improving the accuracy of versioned data processing pipeline operation and maintenance response.

[0105] Please see Figure 4 As shown, this is a logic diagram illustrating how the present invention determines the causes of reliability anomalies and corresponding handling measures based on a reproducibility anomaly index. The process of determining the causes of reliability anomalies based on the reproducibility anomaly index includes:

[0106] If the recurrence anomaly index is less than or equal to the predetermined recurrence anomaly index threshold, it is determined that there is a slight degradation in the feature extraction logic;

[0107] If the anomaly index is greater than the predetermined anomaly index threshold, it is determined that the data reproduction link is broken.

[0108] In this embodiment, the predetermined anomaly index threshold is preset. Specifically, the average value of the anomaly index over 16 historical collection periods is predetermined. The predetermined anomaly index threshold is determined based on the average value of the anomaly index and is set within the range [0.001, 0.019]. Based on the optimization principle of anomaly handling resource allocation and the tolerance for business continuity risks, an optimal value that meets 80% of the upper limit of the measured value is selected. In this embodiment, the predetermined anomaly index threshold is preferably 0.015.

[0109] This invention establishes classification criteria for problems of different severity levels, discretizing continuous reproducible anomaly indices into two categories: minor degradation and link breakage. The grading and judgment mechanism based on the quantitative index enables the system to take differentiated measures for problems of different severity levels, thereby improving the intelligence level and resource utilization efficiency of the entire operation and maintenance system.

[0110] Specifically, the process of matching appropriate handling measures based on the cause of the reliability anomaly includes:

[0111] If the cause is a slight degradation in the feature extraction logic, then determine the adjustment range of the feature health qualification threshold;

[0112] If the cause is a broken data reproduction link, then determine the adjustment range of the binding completion rate threshold.

[0113] Understandably, by determining the adjustment range of the feature health qualification threshold, the quality standards of the feature extraction process can be progressively optimized. When an anomaly is determined to stem from a slight degradation in feature logic, the feature health qualification threshold will be increased, prompting the versioned data processing pipeline to perform more stringent self-filtering during the feature generation stage. This will prevent low-quality features from flowing into subsequent analysis stages from the source, thereby proactively suppressing and correcting the downward trend in data product quality.

[0114] Understandably, by determining the adjustment range of the binding completion rate threshold, rapid location and recovery from severe systemic failures can be achieved. When an anomaly is determined to be caused by a break in the data reproduction link, an isolated deep verification of the complete execution environment associated with the problematic RunID will be automatically triggered, thereby improving the efficiency of troubleshooting severe failures.

[0115] This invention, through the synergistic effect of processing measures, can cover the entire spectrum of problem handling, from minor performance drift to severe system failures. Determining the adjustment range of the characteristic health qualification threshold serves as a preventative, continuously operating fine-tuning mechanism, responsible for maintaining a stable baseline of daily analysis quality. Conversely, determining the adjustment range of the binding completion rate threshold serves as an on-demand, radical remediation measure, responsible for handling destructive anomalies. This improves the reliability of the layered assurance throughout the entire lifecycle of the versioned data processing pipeline.

[0116] Specifically, the adjustment range of the characteristic health qualification threshold is determined based on the recurrence anomaly index, wherein the characteristic health qualification threshold is positively correlated with the recurrence anomaly index.

[0117] In this embodiment, the adjustment range for determining the qualified threshold of the characteristic health status is specifically as follows: ;in, The adjusted threshold for the health status of the features. The current threshold for the qualified characteristic health level is determined in the example within the range [1.57, 2.28], with a preferred threshold of 2.10. The current anomaly index is determined in the example within the range [0.001, 0.015], with a preferred anomaly index of 0.010. The ratio of label relevance to a predetermined label relevance threshold is used in this embodiment to determine the ratio within the range [0.83, 1.17]. Preferably, the ratio of label relevance to the predetermined label relevance threshold is 1.10. The ratio of feature matching similarity to a predetermined feature matching similarity threshold is defined in the example within the range [0.74, 1.11], and preferably, the ratio is 1.03. The preferred adjustment coefficient is the adjustment coefficient used in this embodiment. It is 40.

[0118] This invention, through the construction of an adjustment formula for the feature health qualification threshold associated with the reproducibility anomaly index, can feed back reliability anomalies discovered in the reproducibility process to the quality admission criteria for feature extraction. This enables the overall quality control to form a feedback loop based on objective data, thereby improving the stability of the self-maintaining output quality of the versioned data processing pipeline.

[0119] Specifically, the adjustment range of the binding completion rate threshold is determined based on the recurrence anomaly index, wherein the binding completion rate threshold is positively correlated with the recurrence anomaly index.

[0120] In this embodiment, the specific formula for adjusting the completion rate threshold is as follows: ;in, The adjusted binding completion rate threshold. The current binding completion rate threshold is determined within the range [0.99, 1.01] in this embodiment, with a preferred binding completion rate threshold of 1.0. The current anomaly index is determined within the range [0.015, 0.019] in this embodiment, with a preferred anomaly index of 0.017. The repair coefficient is the preferred repair coefficient in the embodiments. It is 20.

[0121] This invention establishes a positive correlation adjustment formula between the binding completion rate threshold and the reproducibility anomaly index. This allows for a proportional increase in the baseline for metadata binding completion rate when reliability anomalies are detected during the reproducibility phase. As the reproducibility anomaly index increases, the binding completion rate threshold rises accordingly, prompting the versioned data processing pipeline to self-reinforce itself at the metadata management level. This gradually reduces the risk of traceability failures due to missing or inaccurate records, thereby improving the transparency and auditability of historical operations within the versioned data processing pipeline.

[0122] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A data reproducible versioned pipeline management method based on machine learning, characterized in that, include: Submit the target data to the versioned data processing pipeline for data analysis and generate versioned execution records. Collect feature health parameters of target data from versioned execution records within historical periods; Analyze the characteristic health index based on the aforementioned characteristic health parameters; The matching degree of the target data feature extraction is determined based on the comparison result between the feature health index and the predetermined feature health qualification threshold. When the matching degree meets the standard, the feature fidelity parameter of the target data reproduction result is collected from the versioned execution record within the historical period. The feature fidelity index is analyzed based on the aforementioned feature fidelity parameters; The reliability of the target data reproduction result is determined based on the comparison result between the feature fidelity index and the predetermined feature fidelity qualification threshold. When a reliability anomaly occurs, a reproducibility index is calculated. Based on the aforementioned reproducibility anomaly index, the causes of reliability anomalies and corresponding handling measures are determined, including determining the adjustment range of the characteristic health qualification threshold and the adjustment range of the binding completion rate threshold. The feature health parameters include label relevance and feature matching similarity, while the feature fidelity parameters include feature value overlap and RunID binding completion rate.

2. The data reproducible versioned pipeline management method based on machine learning according to claim 1, characterized in that, The process of analyzing the characteristic health index based on the aforementioned characteristic health parameters includes: Extract the label relevance and feature matching similarity of target data from versioned execution records within historical periods; The ratio of the calculated label relevance to a predetermined label relevance threshold is determined as the first health factor; The ratio of the calculated feature matching similarity to a predetermined feature matching similarity threshold is used as the second health factor. The sum of the first health factor and the second health factor is determined as the characteristic health index.

3. The data reproducible versioned pipeline management method based on machine learning according to claim 1, characterized in that, The process of determining whether the matching degree of the target data feature extraction meets the standard based on the comparison result between the feature health index and the predetermined feature health qualification threshold includes: If the feature health index is less than or equal to the predetermined feature health qualification threshold, then the matching degree is determined to be non-compliant with the standard. If the feature health index is greater than the predetermined feature health qualification threshold, then the matching degree is determined to meet the standard.

4. The data reproducible versioned pipeline management method based on machine learning according to claim 1, characterized in that, The process of analyzing the feature fidelity index based on the aforementioned feature fidelity parameters includes: Extract the feature value overlap and RunID binding completion rate of the target data reproduction results from the versioned execution records within the historical period; The ratio of the calculated eigenvalue overlap degree to the predetermined eigenvalue overlap degree threshold is determined as the first fidelity factor; The ratio of the binding completion rate to the predetermined binding completion rate threshold is determined as the second fidelity factor; The sum of the first fidelity factor and the second fidelity factor is determined as the feature fidelity index.

5. The data reproducible versioned pipeline management method based on machine learning according to claim 1, characterized in that, The process of determining whether the reliability of the target data reproduction result is abnormal based on the comparison result between the feature fidelity index and the predetermined feature fidelity pass threshold includes: If the feature fidelity index is less than or equal to the predetermined feature fidelity pass threshold, then the reliability is determined to be abnormal. If the feature fidelity index is greater than the predetermined feature fidelity qualification threshold, then the reliability is determined to be normal.

6. The data reproducible versioned pipeline management method based on machine learning according to claim 1, characterized in that, The process of calculating the recurrence anomaly index includes: Extract the predetermined feature fidelity pass threshold and feature fidelity index; Calculate the difference between the predetermined feature fidelity pass threshold and the feature fidelity index; The difference is determined as the recurrence anomaly index.

7. The data reproducible versioned pipeline management method based on machine learning according to claim 1, characterized in that, The process of determining the cause of reliability anomalies based on the aforementioned reproducibility anomaly index includes: If the reproducibility anomaly index is less than or equal to the predetermined reproducibility anomaly index threshold, the cause of the reliability anomaly is determined to be a slight degradation in the feature extraction logic. If the reproducibility anomaly index is greater than the predetermined reproducibility anomaly index threshold, then the cause of the reliability anomaly is determined to be a data reproduction link break.

8. The data reproducible versioned pipeline management method based on machine learning according to claim 7, characterized in that, The process of matching appropriate handling measures based on the causes of the reliability anomalies includes: If the cause is a slight degradation in the feature extraction logic, then determine the adjustment range of the feature health qualification threshold; If the cause is a broken data reproduction link, then determine the adjustment range of the binding completion rate threshold.

9. The data reproducible versioned pipeline management method based on machine learning according to claim 8, characterized in that, The adjustment range of the characteristic health qualification threshold is determined based on the recurrence anomaly index, wherein the characteristic health qualification threshold is positively correlated with the recurrence anomaly index.

10. The machine learning-based data reproducible versioned pipeline management method according to claim 8, characterized in that, The adjustment range of the binding completion rate threshold is determined based on the recurrence anomaly index, wherein the binding completion rate threshold is positively correlated with the recurrence anomaly index.

Images