A project management system development method based on low-code configuration

Abstract

The application discloses a project management system development method based on low-code configuration and relates to the technical field of computer software.The application significantly improves the prediction adaptability and practicability of the project management system in new project types through the architecture of low-code configuration and meta-learning fusion;the low-code interface allows non-technical users to intuitively configure project parameters, such as the construction period or risk indicators, reduces the system customization threshold, enables enterprises to quickly respond to innovative project demands, and does not need to rely on professional development teams;the meta-learning module utilizes the general knowledge of a pre-trained model and can complete model adjustment only by using a small amount of new data, thereby overcoming the short board of weak generalization ability of traditional models in a data sparse scene;the few-sample self-adaptive mechanism is optimized through internal and external circulation, ensures that the model converges quickly while maintaining stability, and reduces the training time and resource consumption;the verification process in the deployment stage enhances the system reliability, and dynamic consistency checking can timely detect model output deviation.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, and in particular to a method for developing a project management system based on low-code configuration. Background Technology

[0002] As enterprises increase their investment in smart city, aerospace exploration, and emerging technology projects, project management systems are increasingly incorporating machine learning-based predictive models to support automated estimation and risk warning of schedules, costs, and resource requirements. Most existing systems rely on supervised training using historical data from similar projects, achieving relatively stable results under identically or approximately distributed scenarios.

[0003] However, when the characteristics of the target project differ significantly from the training data, data distribution shift (including covariate shift and concept drift) and out-of-domain (OOD) problems are likely to occur. For example, when shifting from traditional buildings to smart city development, or from routine IT operations and maintenance to aerospace payload development, changes occur in mission structure, technology maturity level (TRL), compliance constraints, and supply chain uncertainties, leading to deviations in the model's output of schedule / cost / resource and risk estimates, reduced calibration, and consequently, budget overruns and schedule delays.

[0004] To enhance cross-scenario generalization capabilities, some solutions employ transfer learning, meta-learning, or incremental / continuous learning, using multi-domain pre-training and online fine-tuning to alleviate cold start to some extent. However, in novel projects with sparse data and zero / few samples, manual feature engineering, additional annotation, or frequent model retraining are still necessary, resulting in high deployment and maintenance costs, long adaptation lags, and difficulty in meeting the real-time and robustness requirements of rapidly iterating industries.

[0005] Therefore, how to improve the cross-domain adaptability, calibration and timeliness of project forecasting and risk identification, and reduce the cost of manual intervention and retraining, under the condition of insufficient similar historical data, has become an urgent technical problem to be solved. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] This invention provides a low-code configuration-based project management system development method to solve the problems of existing project management systems in new project types, such as slow adaptation of prediction models, low accuracy, and difficulty in dealing with unknown variables due to data sparsity, and reliance on historical patterns.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] This invention provides a method for developing a project management system based on low-code configuration, comprising:

[0010] Step S1: Provide a low-code configuration interface and receive project configuration information, wherein the project configuration information includes at least a feature description of the target project type and the prediction task type;

[0011] Step S2: Based on the project configuration information, obtain initial data related to the target project type from at least one data source, and perform data cleaning and feature extraction on the initial data;

[0012] Step S3: Based on meta-learning, perform few-sample adaptation on the pre-trained prediction model to obtain the adapted prediction model.

[0013] Step S4: Deploy the adapted prediction model to the project management system to perform the prediction task, and verify the deployment results;

[0014] The pre-trained prediction model is trained based on historical project data from multiple domains.

[0015] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, the low-code configuration interface includes a drag-and-drop component library, and users define project parameters by selecting and combining components. The project parameters include at least the project duration, resource requirements, and / or risk indicators.

[0016] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, the meta-learning adopts a model-independent meta-learning algorithm, and the few-sample adaptation includes updating the model weights within a small number of gradient step updates.

[0017] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, the data source includes an external database and / or a real-time input stream;

[0018] The data cleaning includes missing value processing, outlier removal, and time alignment.

[0019] The feature extraction includes the construction of structured task decomposition features and contextual environment features.

[0020] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, the pre-trained prediction model is a neural network model, and its pre-training data covers multiple project domains.

[0021] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, the verification includes checking the consistency verification threshold and calibration index of the adapted prediction model, and triggering rollback or re-adaptation when the verification fails.

[0022] The definition methods for the consistency verification threshold include:

[0023] a) In the outer loop After the next iteration, on the validation set, sort by the nearest The prediction interval and probability output are generated in a round, and the target coverage rate is assumed to be 100%. The significance level is ;

[0024] b) Interval coverage within the statistical validation window Calculate the expected calibration error of the bins. And measured by the stability of the input distribution. Characterizes the drift intensity;

[0025] c) Consistency statistics are constructed using coverage bias, calibration error, and drift, and thresholds and pass rules are given accordingly:

[0026] , ,

[0027] ,

[0028] in, Indicates the outer loop number Consistency statistics for each iteration. This represents the empirical coverage of the prediction interval within the validation window. Indicates coverage target, express The weighting coefficients, This represents the expected calibration error obtained from binning. express The weighting coefficients, A scalar measure representing the drift of the input or label distribution. This indicates the consistency check threshold for this round. Indicates that recently On the wheel The moving average, This represents the corresponding moving standard deviation. Represents the standard normal quantile, satisfying , Indicates the outer circular index. The cumulative distribution function represents the standard normal distribution. Indicate its quantile function, Indicates the significance level;

[0029] d) Based on the current estimate To achieve the target, the index is adjusted and limited to the allowable range:

[0030] ,

[0031] in, Indicates the threshold to be used in the next round. This indicates an update to the stride coefficient. This indicates the current target threshold based on statistics. This indicates that the scalar Limited to the range The function, and These represent the lower and upper bounds of the threshold, respectively.

[0032] e) If it occurs continued Wheel, or If the value exceeds the baseline mean plus three standard deviations before deployment, a rollback will be triggered and the process will be switched to a smaller sample re-adaptation process.

[0033] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, user feedback is collected through the low-code configuration interface, the user feedback includes confirmation and / or correction instructions for the prediction results, and the user feedback is structured and stored as data for subsequent learning.

[0034] As a preferred embodiment of the low-code configuration-based project management system development method of the present invention, wherein: based on the user feedback, an incremental learning process is triggered, the incremental learning updates the adapted prediction model online while maintaining the existing task performance, and terminates the online update when the stopping condition is met.

[0035] As a preferred embodiment of the project management system development method based on low-code configuration described in this invention, the deployment includes canary release or A / B evaluation, and the error and stability indicators of the online version and the control version are monitored within the evaluation window, and the system automatically rolls back when the indicators drop below a preset threshold.

[0036] As a preferred embodiment of the low-code configuration-based project management system development method of the present invention, the method further includes a step of generating an adaptability report, wherein the adaptability report includes performance indicators, applicability recommendations, and records of model updates that have occurred for the target project type.

[0037] The beneficial effects of this invention are as follows: Through an architecture that integrates low-code configuration and meta-learning, this invention significantly improves the predictive adaptability and practicality of the project management system in new project types. The low-code interface allows non-technical users to intuitively configure project parameters, such as schedule or risk indicators, lowering the system customization threshold and enabling enterprises to quickly respond to innovative project needs without relying on a professional development team. The meta-learning module utilizes the general knowledge of pre-trained models, requiring only a small amount of new data to complete model adjustments, overcoming the weakness of traditional models' generalization ability in sparse data scenarios. This few-sample adaptive mechanism, through internal and external loop optimization, ensures that the model converges quickly while maintaining stability, reducing training time and resource consumption.

[0038] The verification process during the deployment phase of this invention enhances system reliability. Dynamic consistency verification can promptly detect model output deviations and, combined with user feedback, trigger incremental learning, forming a continuous optimization cycle. This avoids decision-making errors caused by prediction distortion, effectively controlling schedule delays and cost overruns, especially in high-risk industries such as construction or R&D. The canary release and A / B evaluation mechanisms further ensure deployment stability, while the automatic rollback function intervenes quickly when indicators are abnormal, improving system robustness.

[0039] This invention enhances the agility and scalability of project management systems. Low-code configuration enables the system to flexibly adapt to projects across multiple domains, while meta-learning endows the model with lifelong learning capabilities, adapting to environments of rapid technological iteration. Users participate in model optimization through structured feedback, enhancing system transparency and trust. This method reduces development costs while improving prediction accuracy, providing enterprises with practical tools to support their digital transformation. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0041] Figure 1 This is a flowchart illustrating the development method of a project management system based on low-code configuration in this embodiment. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0043] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0044] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0045] This application proposes a development method for a project management system based on low-code configuration, combining... Figure 1 As shown, the method includes:

[0046] Step S1: Provide a low-code configuration interface to receive project configuration information. This information includes at least a feature description of the target project type and a predicted task type. In this embodiment, the feature description of the target project type refers to structured information that business personnel can input in the interface using field-value pairs. Typical examples include task decomposition granularity, dependencies, constraints, milestones, and risk concerns. This information is obtained by the interface component receiving user input or loading from an existing project template. The predicted task type refers to the target that the system needs to output, such as duration, cost, resource usage, or risk classification. This is obtained by selecting the corresponding task card in the low-code interface. Specifically, the project template is used as the initial source by default; if no template is available, a blank configuration is started. Fields can be added or deleted in the interface. Basic fields are enabled by default and can be expanded in subsequent steps. The settings are based on domain expert rules and historical implementation experience before deployment. Furthermore, if the user does not provide certain required fields, the system automatically completes them with a minimal field set and marks them as pending confirmation, ensuring the process can continue. Optionally, if the interface interaction is abnormal or storage resources are insufficient, the system will only persist the confirmed fields and pause the addition of new fields until resources are restored.

[0047] Step S2 involves obtaining initial data related to the target project type from at least one data source based on project configuration information, and performing data cleaning and feature extraction on the initial data. Specifically, the initial data refers to historical records or real-time reported data that are comparable to the target project type in terms of time, region, or technology domain. In engineering terms, it is the minimum available subset of data used to support small-sample adaptive testing. The initial data is obtained by initiating a read from an external database or real-time input stream according to the source indication in the project configuration information. For example, by default, data within the most recent evaluation window is read, prioritizing the use of the original timestamp from the source system. The cleaning rules are: missing data is filled with the nearest neighbor from the same source or the median of the same task's history; outliers are removed using the quantile method; and time is aligned to the finest granularity of the source. The settings are based on the error-sample size curve of the validation set before deployment and recommendations from domain experts. Similarly, if there are records that cannot be parsed, they are invalidated, and the original line number is retained for traceability. Optionally, when the target project type lacks directly comparable data, the system only uses the closest subset from the same domain and retains the source identifier during the feature extraction stage for subsequent calibration. Furthermore, when the data connection fails or the real-time stream is temporarily interrupted, the system rolls back to the last successful snapshot and restricts feature extraction to the data within the snapshot, ensuring that the process is not interrupted.

[0048] Step S3: Based on meta-learning, perform few-sample adaptation on the pre-trained prediction model to obtain the adapted prediction model.

[0049] In this embodiment, few-shot adaptive engineering means completing parameter updates for the target project type within a limited annotation and configuration time. The method of acquisition is to optimize the model parameters in a finite number of steps according to the inner and outer loop relationships given in the embodiment. Specifically, by default, a combination of outer loop batch task training and inner loop finite-step updates is used. The number of steps and learning rate can be adjusted within an allowable range based on the pre-deployment validation results, set according to the joint optimality of validation set loss and calibration metrics. Further, if the support set size is insufficient, the system prioritizes supplementing with high-confidence samples of the target project type and records them as low-weight samples in the query set evaluation to reduce the risk of overfitting. When computational resources are insufficient to complete all task batches, task partitioning is used for sequential execution while maintaining input-output consistency, ensuring that the resulting adapted model meets the predetermined constraints under the same validation conditions. If gradient explosion occurs or optimization fails to converge, the system enables gradient pruning and early stopping, and rolls back to the most recently validated parameter snapshot.

[0050] Step S4 involves deploying the adapted prediction model to the project management system to execute prediction tasks and verifying the deployment results. For example, the deployment process means registering the adapted model as an online callable version and connecting it to the inference service. The verification process means performing consistency and calibration checks on the online predictions within a defined window and determining whether to maintain, roll back, or re-adapt based on this. Specifically, the default method for online testing is gray-scale or A / B testing. The evaluation window length and the ratio of control traffic are set based on the pre-deployment evaluation and operational rules and can be adjusted once during the window period. Optionally, if the online traffic is insufficient to complete the statistical test in a short time, the evaluation window is extended and the existing version is kept unchanged until the minimum sample requirement is met. Furthermore, in the event of service unavailability or response timeout, the system switches to the previous stable version and records the triggering reason and timestamp to ensure business continuity.

[0051] The pre-trained prediction model is trained based on historical project data from multiple domains. In this embodiment, historical project data from multiple domains refers to a set of labeled samples covering different project types, scales, and constraints. This data is obtained by aggregating existing project libraries and external licensed data sources and uniformly encoding them. Specifically, the default is to use the same task definition and evaluation criteria as the target domain. If historical criteria differ, criterion mapping is performed during the aggregation phase, and the mapping relationship is retained for traceability description in the specification. This setting is based on historical error stability and sample coverage. Optionally, if the data volume in one domain is much higher than in other domains, to avoid imbalance effects, the samples are reweighted or resampled during the pre-training phase without changing the input / output interface definitions.

[0052] In one embodiment, the low-code configuration interface includes a drag-and-drop component library. Users define project parameters by selecting and combining components. Project parameters include at least duration, resource requirements, and / or risk indicators. Further, the drag-and-drop component library represents a visual set of fields and validators, obtained by loading component metadata on the interface and binding input controls according to project parameters. Specifically, field integrity and value range validation are enabled by default, with the value range set based on historical statistics and domain expert constraints. If user input exceeds the range, the system replaces it with the nearest valid value and provides a prompt. Optionally, if component rendering fails, a basic input box with the same function is used to ensure uninterrupted configuration.

[0053] In one embodiment, meta-learning employs a model-independent meta-learning algorithm to achieve rapid adaptation. Few-sample adaptation includes updating model weights within a small number of gradient steps. In this embodiment, the model-independent meta-learning algorithm is an engineering meaning of a parameterized meta-learning framework that does not depend on the model structure. It is obtained by updating meta-parameters using aggregated gradients of task loss in the outer loop and updating task-specific parameters using support set loss in the inner loop. Specifically, by default, adaptation is completed within a finite number of steps, and the loss and gradient pruning ratio at each step are recorded as audit information. This is based on the validation set query loss and calibration performance not being inferior to the upper limit threshold. Optionally, if a batch of tasks is temporarily unavailable, an outer loop is completed first with a subset of existing tasks, and then completed when subsequent batches are available, without changing the final output parameter interface. If any task exhibits a numerical anomaly in the inner loop, the system skips the update of that task in that round and processes it with zero weight during aggregation in the outer loop, ensuring it does not affect overall convergence.

[0054] Methods for updating gradients in a small number of steps include:

[0055] 1) Task and Data Configuration: Set the number of task batches. The number of samples supported per task Query set sample size Upper limit of outer loop iteration Early stop observation window inner loop steps ;

[0056] 2) Learning Rate and Regularization: Define the learning rate of the inner loop as... outer loop learning rate Gradient clipping threshold Weight decay coefficient ;

[0057] 3) K-step inner loop update: In the outer loop... The next iteration initializes with meta-parameters. Starting with the task, perform K-step updates on its support set for each task:

[0058] , ,

[0059] in, Indicates the outer loop number In the next iteration, the inner loop... The parameter vector of the step, Indicates the next parameter, Indicates the first Initialize the meta-parameters of the outer loop (the right side of the above equation is used as the initial value). This represents the learning rate of the inner loop. This represents the gradient operator with respect to the parameter vector. Indicates the task loss based on the support set, superscript Indicates the outer loop counting index. and Indicates the index of the inner loop step. This indicates the total number of steps in the inner loop;

[0060] 4) Outer loop relationship and parallel summary: Each task in the first... The parameters obtained from the step are used to calculate the loss for its query set, according to After calculating the mean, with right Get an update Apply gradient to each step during update. Cut, and with Perform weight decay;

[0061] 5) Stopping condition: If the query loss of the validation set is in the most recent... The relative change within the next outer loop is below the threshold. Or the external circulation reaches Then stop if the calibration is verified to be continuous. If the condition deteriorates further, stop the cycle early and roll back to the optimal external circulation cycle.

[0062] Log records are fixed to include Query the loss, pruning ratio, and rollback point for integration with subsequent adaptability reports and verification processes;

[0063] Specifically, it is organized in a decoupled manner between inner and outer loops: the inner loop performs a finite number of steps on the support set, emphasizing the use of small steps in conjunction with gradient clipping and weight decay to suppress oscillations; the outer loop aggregates targets across tasks on the query set, thereby driving the meta-parameters to move along a general direction and avoiding overfitting to a single task.

[0064] The stopping conditions employ windowed relative changes and calibrating observations to reduce the impact of random fluctuations, and are combined with rollback to ensure the traceability of the pre-deployment state; log fields cover outer loop count, inner loop steps, loss, pruning, and rollback point, facilitating integration with subsequent validation strategies and the generation of adaptive reports; the overall structure is adapted to pipeline forms triggered by low-code interfaces, and parameter values ​​can be locally fine-tuned according to the characteristics of the project domain to meet the balance requirements between response speed and robustness for different prediction task types;

[0065] In one embodiment, the data source includes an external database and / or a real-time input stream. Optionally, the external database refers to an authorized readable structured storage system, accessed by connecting with read-only permissions and retrieving the required fields according to the project configuration information. The real-time input stream refers to a data channel that continuously generates events or metrics, accessed via subscription or polling. Specifically, by default, batch reading from the database is prioritized, with the real-time stream only used to fill gaps within the window; this setting is based on read latency and consistency requirements. When access is restricted or bandwidth is insufficient, the system reduces the read frequency and records an alarm without interrupting subsequent processes.

[0066] Data cleaning includes missing value handling, outlier removal, and time alignment. Similarly, missing value handling involves imputing data using robust statistics from the same source task; outlier removal involves identifying and removing extreme samples based on empirical quantiles or robust scales; and time alignment involves mapping timestamps from different sources to a common reference axis while maintaining event order. Specifically, by default, median or nearest neighbor data from the same source history is used for imputation; outliers are removed using quantile thresholds; and time alignment is performed with the finest available granularity. These settings are based on the error-retention rate tradeoff curve before deployment. If alignment fails, the data is sorted by source timestamp and imputed with default tags to ensure feature extraction can continue.

[0067] Feature extraction includes the construction of structured task decomposition features and contextual features. For example, structured task decomposition feature engineering refers to the generation of learnable vectors from the project's decomposition structure, dependencies, and milestones, while contextual feature engineering refers to the encoding of environmental variables such as regulations, supply constraints, or technology maturity. Specifically, by default, a combination of discrete one-hot encoding and continuous field normalization is used, and source identifiers are retained in the feature set to support subsequent validation; the setting is based on the improvement in prediction error and calibration on the validation set. Optionally, if some contextual fields are unavailable, they are used as default placeholders, and the placeholder percentage is indicated in the adaptation report.

[0068] In one embodiment, the pre-trained prediction model is a neural network model whose pre-training data covers multiple project domains to learn general patterns.

[0069] Furthermore, the neural network model engineering refers to a predictive model with end-to-end representation learning capabilities. This model is acquired by training on multi-domain data in an offline environment and outputting transferable parameter initializations. Specifically, by default, a comparison of freeze and fine-tuning strategies is performed after pre-training, and the more robust strategy on the target domain is selected; this is based on multi-domain cross-validation and pre-deployment evaluation results. If the target domain differs significantly from the pre-training domain, the system uses a smaller update step size for some early layer parameters before adaptation to reduce the risk of negative transfer.

[0070] In one embodiment, validation includes checking the consistency check threshold and calibration index of the adapted prediction model, and triggering a rollback or re-adaptation if validation fails. In this embodiment, the consistency check threshold is defined as a dynamic pass boundary constructed based on statistics within a window; the calibration index is defined as a measure reflecting the degree of matching between the probability output and the empirical frequency, obtained by statistically analyzing online samples within the validation window. Specifically, by default, the decision statistic is synthesized using coverage, calibration error, and distribution drift metrics, and the threshold is constructed based on the moving mean, standard deviation, and quantiles; the setting is based on the false positive-false negative trade-off on the validation set before deployment and domain expert recommendations. Optionally, if the samples within the window are insufficient to complete the decision, the window is extended and the existing version remains unchanged until the minimum sample condition is met. After triggering a rollback, the few-sample re-adaptation process is immediately initiated, and the input and output of this decision are recorded for subsequent reporting.

[0071] The methods for defining consistency verification thresholds include:

[0072] a) In the outer loop After the next iteration, on the validation set, sort by the nearest The prediction interval and probability output are generated in a round, and the target coverage rate is assumed to be 100%. significance level ;

[0073] b) Interval coverage within the statistical validation window Calculate the expected calibration error of the bins. And measured by the stability of the input distribution. The drift intensity is characterized by a scalar value (such as the Population Stability Index calculated using quantile migration). Specifically, the values ​​of coverage, calibration error, and drift metric are statistically determined based on the validation window set in the embodiment. The method of obtaining these values ​​is to estimate the frequency of online samples within the window according to predetermined bins and calculate the corresponding deviation or stability index. Typical value ranges are jointly provided by the validation set before going online and the historical baseline to ensure stable validation efficacy under real traffic conditions. Optionally, if the bin boundaries do not match the traffic distribution, resulting in sparse samples in a single bin, the binning is reset to a coarser granularity while maintaining the same statistical caliber.

[0074] c) Consistency statistics are constructed using coverage bias, calibration error, and drift, and thresholds and pass rules are given accordingly:

[0075] , ,

[0076] ,

[0077] in, Indicates the outer loop number Consistency statistics for each iteration. This represents the empirical coverage of the prediction interval within the validation window. Indicates coverage target, express The weighting coefficients, This represents the expected calibration error obtained from binning. express The weighting coefficients, A scalar measure representing the drift of the input or label distribution. This indicates the consistency check threshold for this round. Indicates that recently On the wheel The moving average, This represents the corresponding moving standard deviation. Represents the standard normal quantile, satisfying , Indicates the outer circular index. The cumulative distribution function represents the standard normal distribution. Indicate its quantile function, Indicates the significance level;

[0078] d) Based on the current estimate To achieve the target, the index is adjusted and limited to the allowable range:

[0079] ,

[0080] in, Indicates the threshold to be used in the next round. Indicates updating the stride coefficient (suggested) ), This indicates the current target threshold based on statistics. This indicates that the scalar Limited to the range The function, and These represent the lower and upper bounds of the threshold, respectively. Similarly, the exponential adjustment engineering means progressively updating the threshold based on a smooth estimate of historical statistics. This is achieved by performing a convex combination of the current and target thresholds at a predetermined step size in each update round, limiting them to preset upper and lower bounds. The typical value range is determined by the variance level of the validation set before deployment and the desired sensitivity, aiming to achieve a balance between detection rate and stability. If the upper and lower bound settings cause the threshold to remain near the limits for an extended period, a one-time reassessment is triggered, and regular updates resume in the next validation window.

[0081] e) If it occurs continued Wheel (recommended) ),or If the value exceeds the baseline mean plus three standard deviations before deployment, a rollback is triggered, and the process transitions to a smaller sample refit. In this embodiment, persistent violation and strong drift are two parallel triggering channels, representing stable violations and sudden abnormal events, respectively. The data is obtained by accumulating the number of occurrences of statistics within a continuous window or comparing the current drift with the baseline interval. The criteria are based on the limitations imposed on false positives and response lags before deployment. Optionally, if the above conditions are triggered again within the next window after rollback, the priority of refit is increased and the evaluation window is shortened until the condition is corrected.

[0082] Record The bin boundaries and PSI details are written into the adaptability report; weighting coefficients. Pre-deployment optimization of initial values ​​can be achieved through a validation set grid search or Bayesian optimization, followed by testing under real traffic. Gradual updates;

[0083] Specifically, the coverage term reflects the reliability of the prediction interval, the calibration term characterizes the deviation between the probability output and the empirical frequency, and the drift term monitors the distribution changes on the input or label side. These three are integrated into a single statistic, facilitating unified judgment in the pipeline. The threshold is constructed using the moving average, standard deviation, and quantiles, balancing sensitivity and robustness. It also suppresses threshold drift caused by short-term noise through exponential updates and upper and lower bound pruning. The triggering condition is set to a dual-channel approach of persistent over-threshold and strong drift, which helps reduce false alarms caused by occasional spikes and allows for rapid intervention when the actual traffic undergoes structural changes. At the implementation level, indicators and details are recorded to support offline review and report generation. The weight coefficients adopt an offline initialization plus online progressive correction approach to adapt to the differentiated performance of various project types in different fields. It seamlessly integrates with the rollback and re-adaptation mechanisms and can be implemented in a low-code interface as parameter cards and visualization components.

[0084] In one embodiment, user feedback is collected through a low-code configuration interface. This feedback includes confirmation and / or correction instructions for the prediction results, and is structured and stored as data for subsequent learning. Further, user feedback engineering refers to confirmation or correction information from real business stakeholders, obtained by confirming, rewriting, or supplementing a single prediction within the interface and writing it into structured storage. Specifically, by default, a timestamp, operator, and reason for change are recorded for sample weighting or filtering during subsequent incremental learning; this setting is based on data governance standards and audit requirements. Optionally, if there are conflicting or duplicate feedback, the system retains the most recent valid feedback and marks the rest as historical versions for traceability.

[0085] In one embodiment, an incremental learning process is triggered based on user feedback. Incremental learning updates the adapted prediction model online while maintaining the performance of the existing task, and terminates the online update when a stopping condition is met. For example, the online update process means applying small batches of parameter updates and maintaining interface stability without taking the service offline. This is achieved by batching newly arrived feedback samples and updating them according to predetermined constraints. Specifically, by default, a combination of performance non-degradation constraints and early stopping conditions is used to control the update magnitude; this is based on the tolerance range of key business indicators and historical replay test results before deployment. Optionally, if online resources are insufficient, the update is postponed, and the batches to be updated are queued for execution during the maintenance window to ensure that online latency is not affected.

[0086] In one embodiment, deployment includes canary releases or A / B evaluations, and within the evaluation window, error and stability metrics of the deployed version and the control version are monitored, with automatic rollback when the metrics drop below a preset threshold. In this embodiment, error and stability metrics are quantifiable measures related to prediction accuracy and volatility, obtained by aggregating and calculating from online logs within the evaluation window and comparing them with the control version. Specifically, by default, automatic rollback is performed using a preset threshold, and the rollback reason and time point are recorded; the setting is based on the experience bounds of historical online events and existing standards within the domain. Optionally, when the control version is unavailable, the system uses the most recent stable version as a reference and restores the original configuration after the evaluation.

[0087] In one embodiment, the method further includes generating an adaptation report, which includes performance metrics, applicability recommendations, and records of model updates that have occurred for the target project type. In this embodiment, the adaptation report is defined as a structured summary of the key processes and results of this adaptation and deployment validation. It is obtained by extracting necessary fields from training and deployment logs and generating the report according to a predefined template. Specifically, by default, it includes time series data of performance metrics and changes in key thresholds, as well as rollback and re-adaptation trigger records and explanations; this is set based on data governance and auditing needs. Optionally, if some logs are missing, a simplified report is generated using available fields and missing markers are inserted. The report is then automatically updated to a complete version after the logs are completed.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0089] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for developing a project management system based on low-code configuration, characterized in that, include: Step S1: Provide a low-code configuration interface and receive project configuration information, wherein the project configuration information includes at least a feature description of the target project type and the prediction task type; Step S2: Based on the project configuration information, obtain initial data related to the target project type from at least one data source, and perform data cleaning and feature extraction on the initial data; Step S3: Based on meta-learning, perform few-sample adaptation on the pre-trained prediction model to obtain the adapted prediction model. Step S4: Deploy the adapted prediction model to the project management system to perform the prediction task, and verify the deployment results; The pre-trained prediction model is trained based on historical project data from multiple domains. The verification includes checking the consistency verification threshold and calibration index of the adapted prediction model, and triggering rollback or re-adaptation if the verification fails. The definition methods for the consistency verification threshold include: a) In the outer loop After the next iteration, on the validation set, sort by the nearest The prediction interval and probability output are generated in a round, and the target coverage rate is assumed to be 100%. The significance level is ; b) Interval coverage within the statistical validation window Calculate the expected calibration error of the bins. And measured by the stability of the input distribution. Characterizes the drift intensity; c) Consistency statistics are constructed using coverage bias, calibration error, and drift, and thresholds and pass rules are given accordingly: ， ， ， in, Indicates the outer loop number Consistency statistics for each iteration. This represents the empirical coverage of the prediction interval within the validation window. Indicates coverage target, express The weighting coefficients, This represents the expected calibration error obtained from binning. express The weighting coefficients, A scalar measure representing the drift of the input or label distribution. This indicates the consistency check threshold for this round. Indicates that recently On the wheel The moving average, This represents the corresponding moving standard deviation. Represents the standard normal quantile, satisfying , Indicates the outer circular index. The cumulative distribution function represents the standard normal distribution. Indicate its quantile function, Indicates the significance level; d) Based on the current estimate To achieve the target, the index is adjusted and limited to the allowable range: ， in, Indicates the threshold to be used in the next round. This indicates an update to the stride coefficient. This indicates the current target threshold based on statistics. This means to use a scalar Limited to the range The function, and These represent the lower and upper bounds of the threshold, respectively. e) If it occurs continued Wheel, or If the value exceeds the baseline mean plus three standard deviations before deployment, a rollback will be triggered and the process will be switched to a smaller sample re-adaptation process.

2. The project management system development method based on low-code configuration as described in claim 1, characterized in that, The low-code configuration interface includes a drag-and-drop component library, where users can define project parameters by selecting and combining components. These project parameters include at least the project duration, resource requirements, and / or risk indicators.

3. The project management system development method based on low-code configuration as described in claim 1, characterized in that, The meta-learning employs a model-independent meta-learning algorithm, and the few-sample adaptation includes updating the model weights within a small number of gradient step updates.

4. The project management system development method based on low-code configuration as described in claim 1, characterized in that, The data source includes external databases and / or real-time input streams; The data cleaning includes missing value processing, outlier removal, and time alignment. The feature extraction includes the construction of structured task decomposition features and contextual environment features.

5. The project management system development method based on low-code configuration as described in claim 1, characterized in that, The pre-trained prediction model is a neural network model, and its pre-training data covers multiple project areas.

6. The project management system development method based on low-code configuration as described in claim 1, characterized in that, User feedback is collected through the low-code configuration interface. The user feedback includes confirmation and / or correction instructions for the prediction results. The user feedback is structured and stored as data for subsequent learning.

7. The project management system development method based on low-code configuration as described in claim 6, characterized in that, The incremental learning process is triggered based on the user feedback. The incremental learning updates the adapted prediction model online while maintaining the existing task performance, and terminates the online update when the stopping condition is met.

8. The project management system development method based on low-code configuration as described in claim 1, characterized in that, The deployment includes canary releases or A / B evaluations, and monitors the error and stability metrics of the online version and the control version within the evaluation window, and automatically rolls back when the metrics drop below a preset threshold.

9. The method for developing a project management system based on low-code configuration as described in claim 1, characterized in that, It also includes the step of generating an adaptation report, which includes performance metrics, applicability recommendations, and records of model updates that have occurred for the target project type.

Images