Dynamic prediction method for training learning completion time based on government procurement business

Abstract

The present application relates to the technical field of procurement business, and particularly relates to a dynamic prediction method for training learning completion time based on government procurement business. The global basic layer, the role adaptation layer and the individual fine-tuning layer in the hierarchical time series prediction model respectively depict the general law, the role difference and the individual difference of all trainees. The survival analysis framework outputs the completion time probability distribution to provide the uncertainty expression of the prediction. The dynamic correction of the prediction is realized by monitoring the behavior mutation events and the policy change events. The forward cumulative calculation and the reverse calculation are realized by the qualification expiration date and the examination batch date in the business constraint information to realize the constraint-aware prediction and reverse calculation. The method can improve the prediction accuracy of the training learning completion time, provide the probabilistic completion time prediction, realize the real-time dynamic correction of the prediction result, and form a supervision reminder closed loop based on the risk coefficient, thereby effectively supporting the advance supervision and business control of government procurement training.

Description

Technical Field

[0001] This invention relates to the field of procurement business technology, and specifically to a dynamic prediction method for training completion time based on government procurement business. Background Technology

[0002] Trainees in government procurement training encompass various roles, including review experts, purchasers, agents, and suppliers. These roles exhibit systematic differences in learning time distribution, session duration, and interruption patterns. Furthermore, training completion is subject to rigid constraints such as qualification expiration dates and assessment batch dates. However, existing training management systems generally only statically accumulate completed training hours or estimate completion time by dividing remaining hours by the average daily learning rate of all trainees. This approach treats all trainees as homogeneous, failing to differentiate learning rates based on roles and individuals, and lacking the ability to dynamically adjust to changes in learning behavior and policies. Consequently, completion time predictions are often inaccurate and lack probabilistic expression, making it difficult to support proactive supervision and business control of government procurement training. Summary of the Invention

[0003] In view of the above problems, the present invention provides a dynamic prediction method for training completion time based on government procurement business.

[0004] To achieve the above objectives, this invention provides a method for dynamically predicting the completion time of training and learning based on government procurement business, comprising:

[0005] Obtain students' time-series learning behavior data;

[0006] Anomaly cleaning and effective learning time quantification are performed on the time-series learning behavior data to remove invalid learning time from the time-series learning behavior data and obtain effective learning time data.

[0007] Feature extraction is performed on effective learning hour data to construct a time-series feature vector, which reflects learning activity, learning effectiveness, and learning stability.

[0008] Based on temporal feature vectors, a hierarchical temporal prediction model is constructed. The hierarchical temporal prediction model includes a global base layer, a role adaptation layer, and an individual fine-tuning layer. It adopts a long short-term memory network with an attention mechanism and obtains the learning rate prediction sequence through the hierarchical temporal prediction model.

[0009] The learning rate prediction sequence is processed using a survival analysis framework to generate a completion time probability distribution, which represents the probability that a trainee will complete the training before any future date.

[0010] Obtain business constraint information, which includes qualification expiration date and assessment batch date. Based on the probability distribution of completion time and business constraint information, perform forward cumulative calculation and backward calculation to generate estimated completion date and risk coefficient.

[0011] Monitor sudden changes in student behavior and policy changes. When such events are detected, dynamically adjust the probability distribution of completion time, estimated completion date, and risk coefficient.

[0012] Based on the revised risk coefficient, a supervision reminder message is generated, and the estimated completion date and supervision reminder message are output.

[0013] In some embodiments, the time-series learning behavior data is subjected to anomaly cleaning and effective learning time quantification to remove invalid learning time from the time-series learning behavior data, thereby obtaining effective learning time data, including:

[0014] Perform rule-based cleaning on the time-series learning behavior data to remove extremely short login records and records with abnormal timestamps.

[0015] Isolation forest is used to detect outliers in the time-series learning behavior data after rule cleaning. Samples with isolation depth below the isolation depth threshold are identified as outliers and removed.

[0016] From the time-series learning behavior data after removing outliers, the online time was obtained by counting the time between each login and logout, the time without operation was obtained by counting the time without operation, the time without progress during video playback was obtained by counting the time without progress during video playback, and the time of video skipped by dragging was obtained by counting the time of video skipped by dragging.

[0017] Online time is defined as nominal learning time; the sum of inactive time exceeding the dwell time threshold and video playback time without progress change is defined as idle invalid time; and drag-to-skip time is defined as fast skip time.

[0018] The sum of the invalid time spent idling and the time spent skipping quickly is determined as the invalid learning time.

[0019] Subtracting ineffective learning time from nominal learning time yields effective learning time;

[0020] Effective learning time is converted into effective learning hours to obtain effective learning hour data.

[0021] In some embodiments, feature extraction is performed on the effective learning time data to construct a time-series feature vector, including:

[0022] Effective learning time data is divided into recent window data, medium window data and full-cycle window data. Recent window data represents short-term learning status, medium window data represents recent learning habits, and full-cycle window data represents long-term learning stability.

[0023] Based on recent window data, mid-term window data, and full-cycle window data, the distribution entropy of effective learning hours in time periods, the proportion of days with effective learning hour records to the total number of window days, and the proportion of login times to the total number of window days are calculated to obtain learning activity characteristics.

[0024] Based on effective learning hour data and time-series learning behavior data, the proportion of effective learning hours to the total online time, the proportion of skipped time to the total online time, and the proportion of replay times to the total number of course segments are calculated to obtain learning effectiveness characteristics.

[0025] Based on the effective learning hour data, the difference in effective learning hours between adjacent days, the variance of single learning duration, the number of days with no consecutive effective learning hour records, and the number of days with consecutive effective learning hour records are calculated to obtain the learning stability characteristics.

[0026] The learning activity features, learning effectiveness features, and learning stability features are combined into a time-series feature vector.

[0027] In some embodiments, a hierarchical temporal prediction model is constructed based on temporal feature vectors. This model includes a global base layer, a role adaptation layer, and an individual fine-tuning layer. A long short-term memory network employing an attention mechanism is used to obtain a learning rate prediction sequence, including:

[0028] Based on the temporal feature vectors of all students, a long short-term memory network with an attention mechanism is trained to obtain a global base layer. The global base layer learns the general mapping relationship between the temporal feature vectors and the learning rate.

[0029] Based on the temporal feature vectors of students with the same role, the role parameter offset is trained on the basis of the global base layer to obtain the role adaptation layer. The role adaptation layer represents the rate difference of students with the same role relative to all students.

[0030] Based on temporal feature vectors, individual parameter offsets are trained on the basis of the global base layer and the role adaptation layer to obtain the individual fine-tuning layer. The individual fine-tuning layer represents the rate difference between the target student and the student with the same role.

[0031] For trainees who lack historical time-series feature vectors, the parameters of the role adaptation layer are used as the initial prediction parameters.

[0032] A hierarchical time series prediction model is constructed by combining the global base layer, the role adaptation layer, and the individual fine-tuning layer. The parameters of the hierarchical time series prediction model are the superposition of the parameters of the global base layer, the role adaptation layer, and the individual fine-tuning layer.

[0033] The time-series feature vector is input into the hierarchical time-series prediction model to obtain the learning rate prediction sequence, which is the predicted learning rate value for each future date.

[0034] In some embodiments, based on the temporal feature vectors of trainees in the same role, a role parameter offset is trained on top of the global base layer to obtain a role adaptation layer, including:

[0035] The roles of trainees are defined, including review experts, purchasers, agents, and suppliers. These roles are determined based on the trainees' professional status in government procurement activities.

[0036] The temporal feature vectors of students with the same role identifier are grouped into a role sample set, which contains the temporal feature vectors of all students with the same role identifier.

[0037] Based on the role sample set, the role parameter offset is trained on the basis of the global base layer. The role parameter offset is the adjustment amount of the parameters of the global base layer relative to the learning rate of the student of the same role. Each role identifier corresponds to its own role parameter offset.

[0038] The role parameter offset is determined by the learning time distribution, single learning duration and interruption mode of students with the same role. The learning time distribution represents the student's learning time preference, the single learning duration represents the student's single continuous learning duration, the interruption mode represents the frequency of student's learning interruption, and the role parameter offset represents the rate difference of students with the same role relative to all students.

[0039] The parameters of the global base layer are superimposed with the character parameter offsets to obtain the character adaptation layer. The parameters of the character adaptation layer are the sum of the parameters of the global base layer and the character parameter offsets.

[0040] In some embodiments, a survival analysis framework is used to process the learning rate prediction sequence to generate a completion time probability distribution, including:

[0041] The completion of training is defined as an event, and the data of trainees who have not yet completed training is defined as right-truncated data. Right-truncated data refers to trainees whose completion time has not yet been observed, and the completion time of right-truncated data exceeds the observation deadline.

[0042] Based on the learning rate prediction sequence, the effective learning hours are accumulated daily to obtain the effective learning hour accumulation sequence. The effective learning hour accumulation sequence is the sequence of accumulated effective learning hours for each date. The effective learning hour accumulation sequence is used to determine the time required for each trainee to complete the training.

[0043] Using a proportional hazards model within a survival analysis framework, right-truncation data and the cumulative effective training hours sequence are fitted as input to estimate the hazard function and survival function. The survival function represents the probability that training has not been completed on any given date, while the hazard function represents the instantaneous probability that training has been completed on any given date.

[0044] The cumulative probability of completing training before any future date is calculated based on the survival function. The cumulative probability of completion increases as the survival function decreases.

[0045] The cumulative completion probabilities for each future date are aggregated to generate a completion time probability distribution, which is the set of cumulative completion probabilities for each future date.

[0046] In some embodiments, business constraint information is obtained, including qualification expiration dates and assessment batch dates. Based on the probability distribution of completion time and the business constraint information, forward cumulative calculation and backward calculation are performed to generate estimated completion dates and risk coefficients, including:

[0047] Obtain business constraint information, and determine the qualification expiration date and assessment batch date from the business constraint information. The qualification expiration date and assessment batch date are rigid deadline constraints for training.

[0048] The remaining effective class hours are determined as the sum of the remaining effective class hours of compulsory courses and the remaining effective class hours of elective courses after conversion. The remaining effective class hours of elective courses are converted according to the elective conversion weight.

[0049] Based on the probability distribution of completion time, the date on which the cumulative completion probability reaches the preset probability is determined as the estimated completion date, which is the result of positive cumulative calculation.

[0050] Based on the remaining valid study hours, the qualification expiration date, and the assessment batch date, calculate the minimum daily valid study hour requirement. The minimum daily valid study hour requirement is the ratio of the remaining valid study hours to the number of days from the current date to the qualification expiration date.

[0051] The cumulative completion probability before the qualification expiration date is determined based on the probability distribution of completion time. The non-completion probability corresponding to the cumulative completion probability is determined as the overdue probability. The risk coefficient is calculated based on the overdue probability and the minimum daily effective learning hour requirement. The risk coefficient increases with the increase of the overdue probability and the increase of the minimum daily effective learning hour requirement.

[0052] In some embodiments, abrupt changes in student behavior and policy changes are monitored. When such events are detected, the probability distribution of completion time, the estimated completion date, and the risk coefficient are dynamically adjusted, including:

[0053] Monitor learners’ learning behavior and identify abrupt changes in learning behavior that deviate from the historical baseline as learner behavior abrupt events. Learning behavior includes login frequency, effective learning hours and course skipping rate.

[0054] Monitor revisions to government procurement regulations and identify those revisions that lead to changes in course hours as policy change events;

[0055] When a policy change event is detected, identify the course modules affected by the policy change event, invalidate the completed hours in the course modules that have expired, and add the hours of the newly added courses to the uncompleted hours;

[0056] For new courses, the learning rate prediction for new courses is initialized using historical learning data of similar courses. Similar courses are existing courses with content similar to the new courses.

[0057] Based on the latest learning behavior and the updated learning hours to be completed, the probability distribution of completion time, the estimated completion date, and the risk coefficient are regenerated.

[0058] When the prediction results are detected to continuously deviate from the actual learning behavior, the hierarchical time series prediction model is retrained online. Based on the retrained hierarchical time series prediction model, the completion time probability distribution, the estimated completion date, and the risk coefficient are regenerated.

[0059] In some embodiments, when a policy change event is detected, the course modules affected by the policy change event are identified, the expired completed hours in the course modules are invalidated, and the hours of newly added courses are counted as uncompleted hours, including:

[0060] Identify the government procurement regulations involved in the policy change event, including revised, repealed, or newly added regulations.

[0061] The course modules corresponding to the provisions of government procurement regulations are identified as those affected by policy change events;

[0062] Students who are learning about the course modules affected by the policy change event are identified as affected students;

[0063] The completed learning hours of the affected trainees in the course modules affected by the policy change event, based on the expired government procurement regulations, will be determined as expired learning hours. Expired learning hours are the part of the completed learning hours that corresponds to the expired government procurement regulations.

[0064] Invalidated training hours will be removed from the completed training hours of affected trainees, and the invalidated training hours will be counted as uncompleted training hours.

[0065] Compliance courses added due to policy changes will be classified as new courses, and the credit hours of these new courses will be counted as credit hours to be completed. Compliance courses are learning modules added based on the revised provisions of government procurement regulations.

[0066] Update the pending course hours based on the cancellation of expired course hours and the inclusion of hours for newly added courses.

[0067] In some embodiments, a supervision reminder message is generated based on the modified risk coefficient, and the estimated completion date and supervision reminder message are output, including:

[0068] Determine the business consequences of trainees failing to complete training on time. The business consequences are the business impact caused by trainees' failure to complete training on time, including restrictions on trainees' eligibility in government procurement activities.

[0069] Based on the risk factor, the business consequences are combined with the estimated completion date to form the reminder content, which also includes the effective learning hours that students need to complete each day;

[0070] Based on the risk factor, the notification channels for the alert content are determined, including in-site messages, SMS, and administrator notifications;

[0071] Based on the risk factor, the timing of the reminder is determined, and the reminder timing is the period when the probability of students responding to the reminder is high;

[0072] Based on the reminder content, reminder channel, and reminder timing, a supervisory reminder message is generated, and the supervisory reminder message is sent through the reminder channel at the reminder timing;

[0073] Record changes in trainees' learning behavior after supervisory reminders are sent, generate feedback on the effectiveness of the reminders, and adjust subsequent reminder strategies based on this feedback.

[0074] Output the estimated completion date and supervisor reminder information.

[0075] Unlike existing technologies, the above technical solution acquires students' temporal learning behavior data, performs anomaly cleaning and effective learning hour quantification on the temporal learning behavior data, removes ineffective learning time to obtain effective learning hour data, extracts features from the effective learning hour data to construct a temporal feature vector reflecting learning activity, learning effectiveness, and learning stability, constructs a hierarchical temporal prediction model based on the temporal feature vector including a global base layer, a role adaptation layer, and an individual fine-tuning layer, obtains a learning rate prediction sequence using a long short-term memory network with an attention mechanism, processes the learning rate prediction sequence using a survival analysis framework to generate a completion time probability distribution, obtains business constraint information including qualification expiration date and assessment batch date, generates estimated completion date and risk coefficient based on the completion time probability distribution and business constraint information through forward cumulative calculation and backward calculation, monitors student behavior mutation events and policy change events and dynamically corrects the completion time probability distribution, estimated completion date, and risk coefficient, generates supervisory reminder information based on the corrected risk coefficient, and outputs the estimated completion date and supervisory reminder information.

[0076] This method characterizes the general patterns, role differences, and individual differences of all trainees through a hierarchical time-series prediction model consisting of a global foundation layer, a role adaptation layer, and an individual fine-tuning layer. It outputs a probability distribution of completion time using a survival analysis framework to express the uncertainty of the prediction. Dynamic correction of the prediction is achieved by monitoring sudden behavioral events and policy change events. Constraint-aware prediction and backward scheduling are achieved by using qualification expiration dates and assessment batch dates from business constraint information for forward and backward calculations. Therefore, this method can improve the accuracy of training completion time prediction, provide probabilistic completion time prediction, achieve real-time dynamic correction of prediction results, and form a closed loop of supervision and reminders based on risk coefficients, thus effectively supporting the advance supervision and business control of government procurement training.

[0077] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0078] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0079] In the accompanying drawings of the instruction manual:

[0080] Figure 1 This is a schematic diagram illustrating steps S101 to S108 as described in the specific implementation method;

[0081] Figure 2 This is a schematic diagram illustrating steps S201 to S207 as described in the specific implementation method;

[0082] Figure 3 This is a schematic diagram illustrating steps S301 to S305 as described in the specific implementation method.

[0083] Figure 4 This is a schematic diagram illustrating steps S401 to S406 as described in the specific implementation method.

[0084] Figure 5 The following is a schematic diagram illustrating steps S501 to S505 as described in the specific implementation method. Detailed Implementation

[0085] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0086] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0087] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0088] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0089] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0090] Please see Figure 1 This embodiment provides a dynamic prediction method for training completion time based on government procurement business, including:

[0091] S101. Obtain students' time-series learning behavior data;

[0092] S102. Perform anomaly cleaning and effective learning time quantification on the time-series learning behavior data to remove invalid learning time from the time-series learning behavior data and obtain effective learning time data.

[0093] S103. Extract features from effective learning hour data and construct a time-series feature vector. The time-series feature vector reflects learning activity, learning effectiveness, and learning stability.

[0094] S104. Based on temporal feature vectors, a hierarchical temporal prediction model is constructed. The hierarchical temporal prediction model includes a global base layer, a role adaptation layer, and an individual fine-tuning layer. A long short-term memory network with an attention mechanism is used to obtain the learning rate prediction sequence through the hierarchical temporal prediction model.

[0095] S105. The learning rate prediction sequence is processed using a survival analysis framework to generate a completion time probability distribution, which represents the probability that a trainee will complete training before any future date.

[0096] S106. Obtain business constraint information, which includes qualification expiration date and assessment batch date. Based on the probability distribution of completion time and business constraint information, perform forward cumulative calculation and backward calculation to generate the estimated completion date and risk coefficient.

[0097] S107. Monitor sudden changes in trainee behavior and policy changes. When a sudden change in trainee behavior or a policy change is detected, dynamically adjust the probability distribution of completion time, the estimated completion date, and the risk coefficient.

[0098] S108. Generate supervisory reminder information based on the corrected risk coefficient, and output the estimated completion date and supervisory reminder information.

[0099] In step S101, the login, viewing and interaction operations of students on the learning platform are recorded by front-end tracking points and server-side logs. The recorded data is collected daily and stored in association with student and course identifiers, thus forming time-series learning behavior data. The time-series learning behavior data depicts the dynamic changes of students' learning behavior over time, providing a unified data foundation for subsequent steps.

[0100] In step S102, the judgment rules for anomaly cleaning and effective learning hour quantification can be pre-configured according to the historical distribution characteristics of time-series learning behavior data and adjusted as the collected data continues to accumulate, so that the judgment criteria for effective learning hour data are adapted to the actual learning situation of students. The processed effective learning hour data is different from the time-series learning behavior data, and retains the learning hours that reflect the students' actual input.

[0101] In step S103, feature extraction transforms the effective learning time data into a numerical representation that can be processed by the model. The time-series feature vector consists of numerical values ​​in multiple dimensions, with learning activity, learning effectiveness, and learning stability corresponding to different aspects of the numerical dimensions in the time-series feature vector.

[0102] In step S104, the long short-term memory network of the attention mechanism in the hierarchical temporal prediction model can give different degrees of attention to information at different time positions when processing temporal feature vectors, thereby capturing the temporal dependence of learning behavior and incorporating the temporal influence of historical learning behavior into the prediction.

[0103] In step S105, the survival analysis framework is suitable for handling events that occur over time, such as training completion, and the probability distribution of completion time allows the judgment of completion time to reflect the uncertainty of the prediction results.

[0104] In step S106, the business constraint information comes from the business system related to training management. The qualification expiration date and the assessment batch date reflect the rigid time requirements corresponding to the training. The forward cumulative calculation and backward calculation utilize the probability distribution of completion time and business constraint information from the two perspectives of forward time and backward time, so that the calculation results take into account both the prediction of training completion time and the rigid time constraints corresponding to the training.

[0105] In step S107, the sudden change event of trainee behavior comes from the real-time monitoring of trainee learning behavior, and the policy change event comes from the dynamic monitoring of the training-related policy environment. The dynamic correction enables the completion time probability distribution, estimated completion date and risk coefficient to respond in a timely manner to changes in learning behavior and policy environment, so as to avoid the prediction results becoming invalid due to environmental changes.

[0106] In step S108, the revised risk coefficient reflects the risk level of trainees completing the training. The supervisory reminder information varies with the level of risk, enabling the supervisory reminders to be presented in a differentiated manner for trainees with different risk levels, providing a basis for intervention decisions for training managers.

[0107] This embodiment transforms time-series learning behavior data into effective learning hour data, and then uses feature extraction to characterize the numerical features of learning behavior, thus achieving a numerical expression of learning behavior. The hierarchical time-series prediction model establishes time dependence based on time-series feature vectors, and the survival analysis framework transforms the learning rate prediction sequence into a completion time probability distribution, thereby presenting the training completion time in probabilistic form. Combined with forward cumulative calculation and backward calculation based on business constraint information, and with dynamic correction, the estimated completion date and risk coefficient are made to continuously approach the actual situation, thereby generating supervision reminder information, thus realizing dynamic prediction and timely supervision of training completion time.

[0108] Please see Figure 2 In some embodiments, the time-series learning behavior data undergoes anomaly cleaning and effective learning time quantification to remove invalid learning time from the time-series learning behavior data, resulting in effective learning time data, including:

[0109] S201. Perform rule cleaning on the time-series learning behavior data to remove ultra-short login records and abnormal timestamp records from the time-series learning behavior data;

[0110] S202. Use isolation forest to perform outlier detection on the time-series learning behavior data after rule cleaning, and identify samples with isolation depth below the isolation depth threshold as outlier records and remove them.

[0111] S203. From the time-series learning behavior data after removing outlier records, the online duration is obtained by counting the duration between each login and logout, the duration of no-operation stay is obtained by counting the duration of no-operation stay, the duration of no-progress change during video playback is obtained by counting the duration of video playback without progress change, and the drag-skipped duration is obtained by counting the duration of video skipped.

[0112] S204. Online time is defined as nominal learning time, the sum of inactive time exceeding the dwell time threshold and video playback time without progress change is defined as idle invalid time, and drag-to-skip time is defined as fast skip time.

[0113] S205. The sum of the invalid time spent idling and the time spent skipping quickly is determined as the invalid learning time.

[0114] S206. Subtract the ineffective learning time from the nominal learning time to obtain the effective learning time;

[0115] S207. Convert effective learning time into effective learning hours to obtain effective learning hour data.

[0116] In step S201, accidental clicks will generate login records that do not reflect real learning behavior, and network fluctuations will cause timestamp errors. Rule cleaning identifies and processes login records and timestamps according to pre-set rules, providing a pre-cleaned data source for outlier detection.

[0117] In step S202, the isolation forest identifies outlier samples by isolating them. Outlier samples are easier to isolate, i.e., they are exhibited as shallower isolation depths. The isolation depth threshold serves as the boundary between outlier samples and normal samples. Rule cleaning and isolation forest correspond to rule-level elimination and statistical-level detection, respectively, together constituting a dual purification of time-series learning behavior data.

[0118] In step S203, the online duration, inactive dwell time, video playback duration without progress change, and drag-and-skip time are all obtained by statistically analyzing the behavioral events in the time-series learning behavior data. Inactive dwell time corresponds to the time period when the student stays on the page but does not perform any operation, and video playback duration without progress change corresponds to the time period when the video is in playback but the playback progress has not advanced.

[0119] In step S204, before deducting various invalid durations, the online duration is first used as nominal learning duration in the calculation. Nominal learning duration reflects nominal learning input rather than actual input. The dwell time threshold is used to distinguish between normal short pauses and real idling. Idle invalid duration and quick skip duration correspond to the quantitative results of idle behavior and skip behavior, respectively.

[0120] In step S205, the invalid learning time reflects the total amount of time that students do not generate effective learning during their online period. It consists of the idle invalid time and the fast skip time, and is deducted from the nominal learning time for subsequent calculations.

[0121] In step S206, effective learning time represents the actual learning time invested by the student after deducting the time spent hanging up and skipping. The difference between nominal learning time and effective learning time is that nominal learning time does not deduct the invalid part, while effective learning time reflects the real learning investment.

[0122] In step S207, the conversion between effective learning time and effective learning hours is a conversion from time units to learning hour units. The effective learning hour data obtained after conversion reflects the total effective learning input of students and serves as the data source for the feature extraction step.

[0123] This embodiment purifies temporal learning behavior data through rule cleaning and isolated forest detection, and then statistically analyzes online time, inactive time, video playback time without progress changes, and drag-and-skip time. It distinguishes between nominal learning time, idle time, and fast-skip time, thereby removing invalid learning time from nominal learning time to obtain effective learning time that reflects the actual input. This effective learning time is then converted into effective learning hour data, enabling subsequent predictions to be based on data that reflects the actual learning input, thus improving the accuracy and reliability of predictions.

[0124] Please see Figure 3 In some embodiments, feature extraction is performed on the effective learning time data to construct a time-series feature vector, including:

[0125] S301. Divide the effective learning time data into recent window data, medium window data and full-cycle window data. Recent window data represents short-term learning status, medium window data represents recent learning habits, and full-cycle window data represents long-term learning stability.

[0126] S302. Based on recent window data, mid-term window data, and full-cycle window data, calculate the distribution entropy of effective learning hours over time periods, the proportion of days with effective learning hour records to the total number of window days, and the proportion of login times to the total number of window days to obtain learning activity characteristics.

[0127] S303. Based on the effective learning hour data and the time-series learning behavior data, calculate the proportion of the sum of effective learning hours to the sum of online time, the proportion of skipped time to the sum of online time, and the proportion of the number of replays to the number of course segments to obtain the learning effectiveness characteristics.

[0128] S304. Calculate the difference in effective learning hours between adjacent days, the variance of single learning duration, the number of days with no consecutive effective learning hours and the number of days with consecutive effective learning hours based on the effective learning hour data to obtain the learning stability characteristics.

[0129] S305. Combine the learning activity features, learning effectiveness features, and learning stability features into a time-series feature vector.

[0130] In step S301, the recent window data, the intermediate window data, and the full-cycle window data cover different time spans from near to far, forming a multi-scale observation of the learning state, so that subsequent feature calculations can capture learning patterns at multiple time scales.

[0131] In step S302, the learning activity feature is composed of multiple indicators. The distribution entropy of effective learning hours in the time period reflects the degree of dispersion of effective learning hours in the learning period. The proportion of days with effective learning hour records to the number of days in the window reflects the continuity of learning behavior. The proportion of login times to the number of days in the window reflects the frequency of learning behavior. Each indicator portrays the learning activity status of students from different perspectives.

[0132] In step S303, the learning effectiveness characteristic is also composed of multiple proportional indicators. The ratio of the sum of effective learning time to the sum of online time reflects the proportion of effective learning in the online learning time. The ratio of skipped time to the sum of online time reflects the degree of skipped content during the learning process. The ratio of the number of replays to the number of course segments reflects the degree of repeated learning of the content by the students. Each indicator describes the students' learning behavior from the perspective of learning quality.

[0133] In step S304, the learning stability characteristics are composed of indicators that characterize the fluctuation and persistence of learning. The difference in effective learning hours between adjacent days reflects the daily increase or decrease in effective learning hours. The variance of single learning duration reflects the fluctuation range of single learning duration. The number of consecutive days without effective learning hour records reflects the duration of learning interruption. The number of consecutive days with effective learning hour records reflects the duration of learning persistence. Each indicator characterizes the learner's learning behavior from the perspective of stability.

[0134] In step S305, the temporal feature vector integrates the learning activity feature, learning effectiveness feature, and learning stability feature into a unified vector representation, so that the learning state from different aspects can be read by the hierarchical temporal prediction model. The updating of the vector over time allows the temporal feature vector to reflect the dynamic evolution of the learning state.

[0135] This embodiment divides effective learning time data according to time proximity, forming multi-scale observations of recent window data, mid-term window data, and full-cycle window data. Then, it calculates learning activity characteristics, learning effectiveness characteristics, and learning stability characteristics, and finally integrates them into a time-series feature vector. This provides a unified vector form to carry multi-dimensional and multi-time-scale learning information, providing rich and accurate input for the generation of learning rate prediction sequences and improving the accuracy of prediction.

[0136] Please see Figure 4 In some embodiments, a hierarchical temporal prediction model is constructed based on temporal feature vectors. This model includes a global base layer, a role adaptation layer, and an individual fine-tuning layer. A long short-term memory network employing an attention mechanism is used to obtain a learning rate prediction sequence, including:

[0137] S401. Based on the temporal feature vectors of all students, a long short-term memory network with an attention mechanism is used for training to obtain a global base layer. The global base layer learns the general mapping relationship between the temporal feature vectors and the learning rate.

[0138] S402. Based on the temporal feature vectors of students with the same role, train the role parameter offset on the basis of the global base layer to obtain the role adaptation layer. The role adaptation layer represents the rate difference of students with the same role relative to all students.

[0139] S403. Based on temporal feature vectors, individual parameter offsets are trained on the basis of the global base layer and the role adaptation layer to obtain the individual fine-tuning layer. The individual fine-tuning layer represents the rate difference between the target student and the student with the same role.

[0140] S404. For trainees who lack historical time-series feature vectors, use the parameters of the role adaptation layer as the initial prediction parameters.

[0141] S405. Combine the global base layer, the role adaptation layer, and the individual fine-tuning layer to construct a hierarchical time series prediction model. The parameters of the hierarchical time series prediction model are the superposition of the parameters of the global base layer, the role adaptation layer, and the individual fine-tuning layer.

[0142] S406. Input the time series feature vector into the hierarchical time series prediction model to obtain the learning rate prediction sequence, which is the learning rate prediction value for each future date.

[0143] In step S401, the training uses the temporal feature vectors of all trainees as the data source. The global base layer carries the general mapping relationship between the temporal feature vectors and the learning rate. The general mapping relationship reflects the common learning pattern of all trainees and provides a basis for further adjustments at the role level and the individual level.

[0144] In step S402, there are systematic differences in the learning behavior of learners with different roles. Role parameter offset enables the role adaptation layer to reflect the rate difference of learners with the same role relative to all learners, avoiding the substitution of the uniform pattern of all learners for the actual learning pattern of different roles.

[0145] In step S403, based on the commonalities of roles reflected in the role adaptation layer, the individual parameter offset enables the individual fine-tuning layer to further reflect the rate difference of the target student relative to students of the same role, so that the prediction results are closer to the actual situation of the target student based on the adjustment at the role level.

[0146] In step S404, trainees lacking historical time-series feature vectors cannot learn individual parameter shifts through historical data. The parameters of the role adaptation layer can provide common estimates for trainees with the same role and serve as initial prediction parameters to support the start of prediction. Individual-level adjustments are made after trainees accumulate sufficient time-series feature vectors.

[0147] In step S405, the global base layer, role adaptation layer, and individual fine-tuning layer each correspond to the common patterns, role differences, and individual differences of all learners. The hierarchical time-series prediction model integrates the learning patterns at different levels into one, so that the learning patterns at different granularities work together to predict the learning rate.

[0148] In step S406, the learning rate prediction sequence serves as the basis for subsequent completion time prediction, enabling the completion time prediction to be based on the estimation of the learner's future learning pace and providing time-dimensional input for subsequent steps.

[0149] This embodiment trains a global base layer using the temporal feature vectors of all learners, a role adaptation layer using the temporal feature vectors of learners in the same role, and an individual fine-tuning layer using the temporal feature vectors. By incorporating common patterns, role differences, and individual differences layer by layer, it forms a hierarchical temporal prediction model, thereby obtaining a learning rate prediction sequence that reflects the future learning pace and providing a basis for predicting the probability of completion time. For learners who lack historical temporal feature vectors, the parameters of the role adaptation layer are used as initial prediction parameters to ensure that the prediction is still usable in scenarios where historical data is lacking, thereby improving the accuracy and coverage of completion time prediction.

[0150] Please see Figure 5 In some embodiments, based on the temporal feature vectors of trainees with the same role, a role parameter offset is trained on top of the global base layer to obtain a role adaptation layer, including:

[0151] S501. Determine the role identification of trainees. Role identification includes review experts, purchasers, agents, and suppliers. Role identification is determined based on the trainees' professional status in government procurement activities.

[0152] S502. The temporal feature vectors of students with the same role identifier are grouped into a role sample set, which contains the temporal feature vectors of all students with the same role identifier.

[0153] S503. Based on the role sample set, train the role parameter offset on the basis of the global base layer. The role parameter offset is the adjustment amount of the parameters of the global base layer relative to the learning rate of the student of the same role. Each role identifier corresponds to its own role parameter offset.

[0154] S504. The role parameter offset is determined by the learning time distribution, single learning duration and interruption mode of students with the same role. The learning time distribution represents the student's learning time preference, the single learning duration represents the student's single continuous learning duration, the interruption mode represents the frequency of student's learning interruption, and the role parameter offset represents the rate difference of students with the same role relative to all students.

[0155] S505. The parameters of the global base layer are superimposed with the character parameter offset to obtain the character adaptation layer. The parameters of the character adaptation layer are the sum of the parameters of the global base layer and the character parameter offset.

[0156] In step S501, the review experts, purchasers, agents and suppliers correspond to different participants in the government procurement activity. Their professional identities reflect the roles they play in the government procurement activity. Role identification enables participants from different participants to be distinguished by their roles, providing a basis for subsequent collection of participant data by role.

[0157] In step S502, the role sample set provides a data source for training the role parameter offset, so that the learning pattern at the role level can be inferred from the temporal feature vectors of students with the same role identifier.

[0158] In step S503, the adjustment amount reflected by the role parameter offset reflects the deviation that the parameters of the global basic layer need to be made when used for the same role learner, so that the learning pattern at the role level can be expressed separately from the global commonality.

[0159] In step S504, the learning time distribution, single learning duration and interruption mode characterize the learning behavior characteristics of students with the same role from three perspectives: time period, duration and continuity. The role parameter offset integrates the information from the three perspectives, so that the rate difference of students with the same role relative to all students can be reflected.

[0160] In step S505, the parameters of the role adaptation layer retain the common patterns of all students carried by the parameters of the global base layer, and also incorporate the role differences of the same student carried by the role parameter offset, so that the role adaptation layer can take into account both commonalities and role individualities.

[0161] This embodiment determines the role identifier of the learner and collects the role sample set, and then trains to obtain the role parameter offset. The parameters of the role adaptation layer are integrated with the parameters of the global base layer, thereby incorporating the role differences of learners of the same role into the common patterns of all learners. This allows the prediction of learning rate to be adjusted differently for learners of different roles, thus improving the accuracy of the prediction.

[0162] In some embodiments, a survival analysis framework is used to process the learning rate prediction sequence to generate a completion time probability distribution, including:

[0163] The completion of training is defined as an event, and the data of trainees who have not yet completed training is defined as right-truncated data. Right-truncated data refers to trainees whose completion time has not yet been observed, and the completion time of right-truncated data exceeds the observation deadline.

[0164] Based on the learning rate prediction sequence, the effective learning hours are accumulated daily to obtain the effective learning hour accumulation sequence. The effective learning hour accumulation sequence is the sequence of accumulated effective learning hours for each date. The effective learning hour accumulation sequence is used to determine the time required for each trainee to complete the training.

[0165] Using a proportional hazards model within a survival analysis framework, right-truncation data and the cumulative effective training hours sequence are fitted as input to estimate the hazard function and survival function. The survival function represents the probability that training has not been completed on any given date, while the hazard function represents the instantaneous probability that training has been completed on any given date.

[0166] The cumulative probability of completing training before any future date is calculated based on the survival function. The cumulative probability of completion increases as the survival function decreases.

[0167] The cumulative completion probabilities for each future date are aggregated to generate a completion time probability distribution, which is the set of cumulative completion probabilities for each future date.

[0168] In this embodiment, training completion is treated as an event, and cases where the completion time has not yet occurred during the observation period constitute truncation. Right-truncation data allows trainees who have not yet completed training to participate in model fitting, rather than being excluded due to the unknown completion time.

[0169] The cumulative effective training hours sequence reflects the process of cumulative effective training hours increasing over time. By observing the date when the cumulative effective training hours reach the level required to complete the training, we can determine the time span required to complete the training, providing a basis for estimating the completion time in the survival analysis framework.

[0170] The risk function and survival function characterize the probabilistic features of training completion time from two complementary perspectives, together forming a dual perspective on the distribution of completion time, so that the judgment of completion time can take into account both the tendency to complete and the cumulative situation of continuous non-completion.

[0171] The cumulative completion probability reflects the pattern that the likelihood of completing the training gradually increases over time, laying the foundation for summarizing the cumulative completion probabilities corresponding to each future date into a completion time probability distribution.

[0172] The completion time probability distribution presents the completion time as the probability corresponding to multiple future dates, rather than as a single date, providing a probabilistic basis for subsequent forward cumulative calculation and backward calculation combined with business constraint information.

[0173] This embodiment treats training completion as an event and retains information about trainees who have not completed the training using right-truncation data. It forms an effective learning time accumulation sequence from the learning rate prediction sequence, obtains a survival function through a proportional hazards model, calculates the cumulative completion probability, and summarizes it into a completion time probability distribution. This allows the training completion time to be characterized in the form of a probability distribution, providing a probabilistic basis for subsequent estimations based on business constraints and improving the reliability of completion time prediction.

[0174] In some embodiments, business constraint information is obtained, including qualification expiration dates and assessment batch dates. Based on the probability distribution of completion time and the business constraint information, forward cumulative calculation and backward calculation are performed to generate estimated completion dates and risk coefficients, including:

[0175] Obtain business constraint information, and determine the qualification expiration date and assessment batch date from the business constraint information. The qualification expiration date and assessment batch date are rigid deadline constraints for training.

[0176] The remaining effective class hours are determined as the sum of the remaining effective class hours of compulsory courses and the remaining effective class hours of elective courses after conversion. The remaining effective class hours of elective courses are converted according to the elective conversion weight.

[0177] Based on the probability distribution of completion time, the date on which the cumulative completion probability reaches the preset probability is determined as the estimated completion date, which is the result of positive cumulative calculation.

[0178] Based on the remaining valid study hours, the qualification expiration date, and the assessment batch date, calculate the minimum daily valid study hour requirement. The minimum daily valid study hour requirement is the ratio of the remaining valid study hours to the number of days from the current date to the qualification expiration date.

[0179] The cumulative completion probability before the qualification expiration date is determined based on the probability distribution of completion time. The non-completion probability corresponding to the cumulative completion probability is determined as the overdue probability. The risk coefficient is calculated based on the overdue probability and the minimum daily effective learning hour requirement. The risk coefficient increases with the increase of the overdue probability and the increase of the minimum daily effective learning hour requirement.

[0180] In this embodiment, the qualification expiration date and the assessment batch date correspond to the validity period of the qualification associated with the training and the assessment schedule, respectively. The qualification expiration date constrains the completion time of qualification renewal, and the assessment batch date constrains the completion time of the assessment, forming a hard node that cannot be surpassed in the training completion time.

[0181] The importance of required courses and elective courses in training compliance is different. The role of elective course weighting is to adjust the contribution of the remaining valid elective hours to the remaining valid hours, so that the weight of required courses and elective courses in fulfilling the requirements is distinguished.

[0182] The preset probability sets a confidence level for the completion time judgment. The positive cumulative calculation moves in the future to find the date when the cumulative completion probability reaches the preset probability, so that the estimated completion date has certainty in a probabilistic sense, rather than relying on point estimation of a single rate.

[0183] The minimum daily effective learning hours required reflect the minimum daily learning threshold needed to complete the training before the qualification expires. The reverse calculation method involves working backward from the qualification expiration date to the current date and distributing the remaining effective learning hours to each remaining day.

[0184] The probability of overdue payment is complementary to the cumulative completion probability before the qualification expiry date, reflecting the opposite of on-time completion; the risk coefficient combines the probability of overdue payment and the minimum daily effective learning hours required, so that the degree of risk can simultaneously reflect the uncertainty of completion and the level of pressure to complete completion.

[0185] This embodiment determines the qualification expiration date, assessment batch date, and remaining valid study hours, and uses forward cumulative calculation to obtain the estimated completion date. It then uses backward calculation to obtain the minimum daily valid study hour requirement. By combining the probability of overdue payment with the minimum daily valid study hour requirement, a risk coefficient is obtained. This simultaneously provides a prediction of completion time and a measurement of completion risk, providing a basis for subsequent supervision and reminders.

[0186] In some embodiments, abrupt changes in student behavior and policy changes are monitored. When such events are detected, the probability distribution of completion time, the estimated completion date, and the risk coefficient are dynamically adjusted, including:

[0187] Monitor learners’ learning behavior and identify abrupt changes in learning behavior that deviate from the historical baseline as learner behavior abrupt events. Learning behavior includes login frequency, effective learning hours and course skipping rate.

[0188] Monitor revisions to government procurement regulations and identify those revisions that lead to changes in course hours as policy change events;

[0189] When a policy change event is detected, identify the course modules affected by the policy change event, invalidate the completed hours in the course modules that have expired, and add the hours of the newly added courses to the uncompleted hours;

[0190] For new courses, the learning rate prediction for new courses is initialized using historical learning data of similar courses. Similar courses are existing courses with content similar to the new courses.

[0191] Based on the latest learning behavior and the updated learning hours to be completed, the probability distribution of completion time, the estimated completion date, and the risk coefficient are regenerated.

[0192] When the prediction results are detected to continuously deviate from the actual learning behavior, the hierarchical time series prediction model is retrained online. Based on the retrained hierarchical time series prediction model, the completion time probability distribution, the estimated completion date, and the risk coefficient are regenerated.

[0193] In this embodiment, the historical baseline is the benchmark level formed by the student's past learning behavior. Login frequency, effective learning hours and course skipping rate reflect the student's level of activity, engagement and concentration, respectively. Abrupt changes in learning behavior are identified as student behavior abrupt events, so that abnormal changes in the student's learning status can be detected in a timely manner.

[0194] Revisions to government procurement regulations can have a ripple effect on training courses, causing changes in course hours. Policy changes can also be reflected in the prediction of training completion times.

[0195] Policy change events cause completed hours to become invalid in course modules, while new courses are created. The hours to be completed are adjusted in two opposite directions, so that the number of hours that students need to complete can be updated as policies change.

[0196] New courses have not yet accumulated historical learning data. Similar courses can provide an initial basis for predicting the learning rate, allowing the prediction of the learning rate of new courses to be initiated even when they lack their own historical data.

[0197] The latest learning behaviors and updated pending learning hours serve as inputs for re-prediction, ensuring that the probability distribution of completion time, estimated completion date, and risk coefficient remain consistent with the learners' latest learning status and avoiding distortion due to outdated data.

[0198] The fact that the prediction results continue to deviate from the actual learning behavior means that the hierarchical time series prediction model is deviating from the real learning pattern. Online retraining corrects the deviation by updating the model, rather than relying solely on data updates, so that the model can adapt to changes in the learning pattern.

[0199] This embodiment monitors learning behavior and revisions to government procurement regulations to identify sudden changes in learner behavior and policy changes. It updates the learning hours to be completed and regenerates the probability distribution of completion time, the estimated completion date, and the risk coefficient. It also retrains the hierarchical time series prediction model online if deviations occur, thereby enabling the prediction results to continuously adapt to changes in learning behavior and the policy environment, maintaining the timeliness and accuracy of the predictions.

[0200] In some embodiments, when a policy change event is detected, the course modules affected by the policy change event are identified, the expired completed hours in the course modules are invalidated, and the hours of newly added courses are counted as uncompleted hours, including:

[0201] Identify the government procurement regulations involved in the policy change event, including revised, repealed, or newly added regulations.

[0202] The course modules corresponding to the provisions of government procurement regulations are identified as those affected by policy change events;

[0203] Students who are learning about the course modules affected by the policy change event are identified as affected students;

[0204] The completed learning hours of the affected trainees in the course modules affected by the policy change event, based on the expired government procurement regulations, will be determined as expired learning hours. Expired learning hours are the part of the completed learning hours that corresponds to the expired government procurement regulations.

[0205] Invalidated training hours will be removed from the completed training hours of affected trainees, and the invalidated training hours will be counted as uncompleted training hours.

[0206] Compliance courses added due to policy changes will be classified as new courses, and the credit hours of these new courses will be counted as credit hours to be completed. Compliance courses are learning modules added based on the revised provisions of government procurement regulations.

[0207] Update the pending course hours based on the cancellation of expired course hours and the inclusion of hours for newly added courses.

[0208] In this embodiment, there are three scenarios for government procurement regulations: revision, repeal, and addition. Revision changes the content of existing regulations, repeal renders existing regulations invalid, and addition introduces new compliance requirements. Identifying government procurement regulations can determine the specific scope of policy changes.

[0209] There is a correspondence between the provisions of government procurement regulations and course modules. Based on this correspondence, the course modules affected by policy changes can be identified, thus determining the training content affected by the regulatory changes.

[0210] The identification of affected students ensures that the subsequent cancellation of learning hours can be accurately implemented for each affected student, avoiding impact on students who are not affected by the policy change.

[0211] Only the portion of completed training hours based on expired government procurement regulations constitutes expired training hours, while training hours based on still valid government procurement regulations are retained, ensuring that the definition of expired training hours accurately corresponds to the scope of expiration of government procurement regulations.

[0212] After expired learning hours are removed from the completed learning hours of affected trainees, they are transferred back to the uncompleted learning hours. This means that learning completed based on expired government procurement regulations will no longer be considered completed and will need to be relearned, thus ensuring that the recognition of learning hours is consistent with the latest government procurement regulations.

[0213] The hours for compliance courses are included in the to-be-completed hours, transforming new compliance requirements into learning tasks that students need to complete, thus preventing new courses from being overlooked in the completion requirements.

[0214] The updated pending learning hours reflect the completion requirements for trainees after the policy changes, ensuring that subsequent forecasts and calculations are based on compliance with the latest regulations.

[0215] This embodiment identifies the provisions of government procurement regulations, locates the course modules and affected students affected by policy change events, determines invalidated learning hours, and returns them to the pending learning hours after invalidation. At the same time, the learning hours of newly added courses are included in the pending learning hours, thereby updating the pending learning hours. This ensures that the recognition of learning hours is consistent with the latest provisions of government procurement regulations, and guarantees that subsequent predictions are based on accurate completion requirements.

[0216] In some embodiments, a supervision reminder message is generated based on the modified risk coefficient, and the estimated completion date and supervision reminder message are output, including:

[0217] Determine the business consequences of trainees failing to complete training on time. The business consequences are the business impact caused by trainees' failure to complete training on time, including restrictions on trainees' eligibility in government procurement activities.

[0218] Based on the risk factor, the business consequences are combined with the estimated completion date to form the reminder content, which also includes the effective learning hours that students need to complete each day;

[0219] Based on the risk factor, the notification channels for the alert content are determined, including in-site messages, SMS, and administrator notifications;

[0220] Based on the risk factor, the timing of the reminder is determined, and the reminder timing is the period when the probability of students responding to the reminder is high;

[0221] Based on the reminder content, reminder channel, and reminder timing, a supervisory reminder message is generated, and the supervisory reminder message is sent through the reminder channel at the reminder timing;

[0222] Record changes in trainees' learning behavior after supervisory reminders are sent, generate feedback on the effectiveness of the reminders, and adjust subsequent reminder strategies based on this feedback.

[0223] Output the estimated completion date and supervisor reminder information.

[0224] In this embodiment, the business consequences inform trainees of the consequences of failing to complete the training on time, the estimated completion date informs trainees of the completion time corresponding to the current progress, the effective learning hours required each day inform trainees of the amount of learning they need to invest each day, and the reminders provide trainees with actionable learning guidance.

[0225] The reach of in-site messages, SMS messages, and administrator notifications increases progressively. The reminder channels are selected based on the risk level; the higher the risk level, the stronger the reminder channel used, ensuring that high-risk students receive stronger reach guarantees.

[0226] Determining the timing of reminders takes into account both the probability of students responding and the degree of disturbance. Sending reminders during times when students are more likely to view them increases the likelihood of a response while avoiding disturbance during periods when students are not active.

[0227] The supervisory reminder message integrates the reminder content, reminder channel, and reminder timing into a complete reminder action, so that the reminder content can be delivered to trainees at the appropriate time and through the appropriate channel.

[0228] Feedback on the effectiveness of reminders can evaluate the actual effectiveness of reminder strategies, provide a basis for adjusting subsequent reminder strategies, and form a closed loop in which reminders and feedback promote each other.

[0229] The estimated completion date and supervisory reminders are delivered to trainees and training managers, enabling trainees to understand the progress and risks, and enabling training managers to grasp the overall training status and address overdue risks in a timely manner.

[0230] This embodiment generates supervisory reminder information and delivers it to trainees by determining the business consequences and combining reminder content, channels, and timing. At the same time, it records feedback on the reminder effect to adjust subsequent reminder strategies, thereby forming a risk-adaptive supervisory reminder closed loop, increasing the proportion of trainees completing training on time, and reducing business risks caused by overdue payments.

[0231] The aforementioned technical solution characterizes the general patterns, role differences, and individual differences of all trainees through a hierarchical time-series prediction model, comprising a global foundation layer, a role adaptation layer, and an individual fine-tuning layer. It outputs a probability distribution of completion time using a survival analysis framework to express the uncertainty of the prediction. Dynamic correction of the prediction is achieved by monitoring sudden behavioral events and policy change events. Constraint-aware prediction and backward scheduling are achieved through forward accumulation and backward calculation using qualification expiration dates and assessment batch dates from business constraint information. Therefore, this method can improve the accuracy of training completion time prediction, provide probabilistic completion time prediction, achieve real-time dynamic correction of prediction results, and form a closed-loop supervision and reminder system based on risk coefficients, thereby effectively supporting the advance supervision and business control of government procurement training.

[0232] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A dynamic prediction method for training completion time based on government procurement business, characterized in that, include: Obtain students' time-series learning behavior data; The time-series learning behavior data is subjected to anomaly cleaning and effective learning time quantification processing to remove invalid learning time from the time-series learning behavior data and obtain effective learning time data. Feature extraction is performed on the effective learning hour data to construct a time-series feature vector, which reflects learning activity, learning effectiveness, and learning stability. Based on the aforementioned temporal feature vector, a hierarchical temporal prediction model is constructed. The hierarchical temporal prediction model includes a global base layer, a role adaptation layer, and an individual fine-tuning layer. It employs a long short-term memory network with an attention mechanism. The learning rate prediction sequence is obtained through the hierarchical temporal prediction model. The learning rate prediction sequence is processed using a survival analysis framework to generate a completion time probability distribution, which represents the probability that a trainee will complete training before any future date. Obtain business constraint information, which includes qualification expiration date and assessment batch date. Based on the completion time probability distribution and the business constraint information, perform forward cumulative calculation and backward calculation to generate the estimated completion date and risk coefficient. Monitor sudden behavioral events and policy change events of trainees. When a sudden behavioral event or a policy change event of trainees is detected, dynamically adjust the probability distribution of completion time, the estimated completion date and the risk coefficient. Based on the revised risk coefficient, a supervision reminder message is generated, and the estimated completion date and the supervision reminder message are output.

2. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, The time-series learning behavior data is subjected to anomaly cleaning and effective learning hour quantification processing to remove invalid learning time from the time-series learning behavior data, resulting in effective learning hour data, including: The time-series learning behavior data is cleaned using rules to remove extremely short login records and records with abnormal timestamps. Isolation forest is used to detect outliers in the time-series learning behavior data after rule cleaning. Samples with an isolation depth lower than the isolation depth threshold are identified as outliers and are removed. From the time-series learning behavior data after removing outlier records, the online duration is obtained by counting the duration between each login and logout, the duration of no-operation pause is obtained by counting the duration of no-operation pause, the duration of no-progress pause during video playback is obtained by counting the duration of no-progress pause during video playback, and the duration of video skipped by dragging is obtained by counting the duration of video skipped by dragging. The online duration is defined as the nominal learning duration. The sum of the inactive dwell time exceeding the dwell time threshold and the video playback duration without progress change is defined as the idle invalid duration. The drag-to-skip duration is defined as the fast skip duration. The sum of the invalid idle time and the quick skip time is determined as the invalid learning time; Subtract the invalid learning time from the nominal learning time to obtain the effective learning time; The effective learning time is converted into effective learning hours to obtain effective learning hour data.

3. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, Feature extraction is performed on the effective learning hour data to construct a time-series feature vector, including: The effective learning time data is divided into recent window data, medium window data and full-cycle window data. The recent window data represents the short-term learning status, the medium window data represents the recent learning habits, and the full-cycle window data represents the long-term learning stability. Based on the recent window data, the intermediate window data, and the full-cycle window data, the distribution entropy of effective learning hours in the time period, the proportion of days with effective learning hour records to the window days, and the proportion of login times to the window days are calculated to obtain the learning activity characteristics. Based on the effective learning hour data and the time-series learning behavior data, the proportion of the sum of effective learning hours to the sum of online time, the proportion of skipped time to the sum of online time, and the proportion of replay times to the number of course segments are calculated to obtain the learning effectiveness characteristics; Based on the effective learning hour data, the difference in effective learning hours between adjacent days, the variance of single learning duration, the number of days with no consecutive effective learning hour records, and the number of days with consecutive effective learning hour records are calculated to obtain the learning stability characteristics. The learning activity feature, the learning effectiveness feature, and the learning stability feature are combined into a time-series feature vector.

4. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, Based on the aforementioned temporal feature vectors, a hierarchical temporal prediction model is constructed. This model comprises a global base layer, a role adaptation layer, and an individual fine-tuning layer, employing a long short-term memory network with an attention mechanism. The learning rate prediction sequence is obtained through this hierarchical temporal prediction model, including: Based on the temporal feature vectors of all students, a long short-term memory network with an attention mechanism is trained to obtain a global base layer. The global base layer learns the general mapping relationship between the temporal feature vectors and the learning rate. Based on the temporal feature vectors of students with the same role, the role parameter offset is trained on the basis of the global base layer to obtain the role adaptation layer, which represents the rate difference of students with the same role relative to all students. Based on the temporal feature vector, individual parameter offsets are trained on the basis of the global base layer and the role adaptation layer to obtain an individual fine-tuning layer. The individual fine-tuning layer represents the rate difference between the target student and the student with the same role. For trainees who lack historical time-series feature vectors, the parameters of the role adaptation layer are used as initial prediction parameters. The global base layer, the role adaptation layer, and the individual fine-tuning layer are combined to construct a hierarchical time series prediction model. The parameters of the hierarchical time series prediction model are the sum of the parameters of the global base layer, the role adaptation layer, and the individual fine-tuning layer. The time-series feature vector is input into the hierarchical time-series prediction model to obtain the learning rate prediction sequence, which is the learning rate prediction value for each future date.

5. The dynamic prediction method for training completion time based on government procurement business according to claim 4, characterized in that, Based on the temporal feature vectors of students in the same role, a role parameter offset is trained on top of the global base layer to obtain a role adaptation layer, including: The role identifiers of the trainees are determined, including those of review experts, purchasers, agents, and suppliers. The role identifiers are determined based on the trainees' professional status in government procurement activities. The temporal feature vectors of students with the same role identifier are grouped into a role sample set, which contains the temporal feature vectors of all students with the same role identifier. Based on the character sample set, character parameter offsets are trained on the basis of the global base layer. The character parameter offset is the adjustment amount of the parameters of the global base layer relative to the learning rate of the same character learner. Each character identifier corresponds to its own character parameter offset. The role parameter offset is determined by the learning time distribution, single learning duration and interruption mode of students with the same role. The learning time distribution represents the student's learning time preference, the single learning duration represents the student's single continuous learning duration, the interruption mode represents the frequency of student's learning interruption, and the role parameter offset represents the rate difference of students with the same role relative to all students. The parameters of the global base layer are superimposed with the character parameter offset to obtain the character adaptation layer. The parameters of the character adaptation layer are the sum of the parameters of the global base layer and the character parameter offset.

6. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, The learning rate prediction sequence is processed using a survival analysis framework to generate a completion time probability distribution, including: The completion of training is defined as an event, and the data of trainees who have not yet completed training is defined as right-truncated data. The right-truncated data refers to the data of trainees whose completion time has not yet been observed, and the completion time of the right-truncated data exceeds the observation deadline. Based on the learning rate prediction sequence, the effective learning hours are accumulated daily to obtain the effective learning hour accumulation sequence. The effective learning hour accumulation sequence is a sequence of accumulated effective learning hours for each date. The effective learning hour accumulation sequence is used to determine the time required for each trainee to complete the training. Using a proportional hazards model within a survival analysis framework, the right-truncation data and the cumulative effective training hours sequence are fitted together to estimate the hazard function and survival function. The survival function represents the probability that training has not been completed on any given date, and the hazard function represents the instantaneous probability that training has been completed on any given date. The cumulative probability of completing training before any future date is calculated based on the survival function, and the cumulative probability of completion increases as the survival function decreases; The cumulative completion probabilities for each future date are aggregated to generate a completion time probability distribution, which is the set of cumulative completion probabilities for each future date.

7. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, Obtain business constraint information, including qualification expiration date and assessment batch date. Based on the completion time probability distribution and the business constraint information, perform forward cumulative calculation and backward calculation to generate estimated completion date and risk coefficient, including: Obtain business constraint information, and determine the qualification expiration date and assessment batch date from the business constraint information. The qualification expiration date and the assessment batch date are rigid deadline constraints for training. The remaining effective class hours are determined as the sum of the remaining effective class hours of compulsory courses and the remaining effective class hours of elective courses after conversion. The remaining effective class hours of elective courses are converted according to the elective conversion weight. Based on the completion time probability distribution, the date on which the cumulative completion probability reaches the preset probability is determined as the estimated completion date, which is the result of forward cumulative calculation; Based on the remaining valid study hours, the qualification expiration date, and the assessment batch date, calculate the minimum daily valid study hour requirement, which is the ratio of the remaining valid study hours to the number of days from the current date to the qualification expiration date. Based on the completion time probability distribution, the cumulative completion probability before the qualification expiration date is determined, and the non-completion probability corresponding to the cumulative completion probability is determined as the overdue probability. A risk coefficient is calculated based on the overdue probability and the daily minimum effective learning hour requirement. The risk coefficient increases with the increase of the overdue probability and the increase of the daily minimum effective learning hour requirement.

8. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, Monitor sudden behavioral events and policy change events among trainees. When a sudden behavioral event or a policy change event is detected, dynamically adjust the probability distribution of completion time, the estimated completion date, and the risk coefficient, including: Monitor students' learning behavior and identify sudden changes in learning behavior that deviate from the historical baseline as student behavior mutation events. The learning behavior includes login frequency, effective study hours, and course skipping rate. Monitor revisions to government procurement regulations and identify those revisions that lead to changes in course hours as policy change events; When the policy change event is detected, identify the course modules affected by the policy change event, invalidate the completed hours in the course modules that have expired, and include the hours of the newly added courses in the hours to be completed. For the newly added course, the learning rate prediction of the newly added course is initialized using historical learning data of similar courses, where similar courses are existing courses with content similar to the newly added course. Based on the latest learning behavior and the updated learning hours to be completed, the probability distribution of the completion time, the estimated completion date, and the risk coefficient are regenerated. When the prediction results are detected to continuously deviate from the actual learning behavior, the hierarchical time series prediction model is retrained online, and the completion time probability distribution, the estimated completion date and the risk coefficient are regenerated based on the retrained hierarchical time series prediction model.

9. The dynamic prediction method for training completion time based on government procurement business according to claim 8, characterized in that, When the policy change event is detected, the course modules affected by the policy change event are identified, the completed hours in the course modules that have expired are invalidated, and the hours of the newly added courses are counted as uncompleted hours, including: Identify the government procurement regulations involved in the policy change event, wherein the government procurement regulations are revised, repealed, or newly added; The course modules corresponding to the aforementioned government procurement regulations are identified as the course modules affected by the policy change event; Students who learn about the impact of the aforementioned policy change event will be identified as affected students; The completed learning hours of the affected trainees in the course modules affected by the policy change event, based on the expired provisions of the government procurement regulations, are determined as expired learning hours. The expired learning hours are the portion of the completed learning hours that corresponds to the expired provisions of the government procurement regulations. The invalidated learning hours shall be removed from the completed learning hours of the affected students, and the invalidated learning hours shall be included in the learning hours to be completed. The newly added compliance courses due to the policy change event are identified as the newly added courses, and the credit hours of the newly added courses are included in the credit hours to be completed. The compliance courses are learning modules added in accordance with the revised provisions of the government procurement regulations. The uncompleted hours are updated based on the invalidation of the expired hours and the inclusion of the hours for the newly added courses.

10. The dynamic prediction method for training completion time based on government procurement business according to claim 1, characterized in that, Based on the revised risk coefficient, a supervision reminder message is generated, and the estimated completion date and the supervision reminder message are output, including: Determine the business consequences of the trainee's failure to complete the training on time. The business consequences are the business impact caused by the trainee's failure to complete the training on time, including restrictions on the trainee's eligibility in government procurement activities. Based on the risk coefficient, the business consequences and the estimated completion date are combined into a reminder message, which also includes the effective study hours that the student needs to complete each day. Based on the risk coefficient, the notification channels for the reminder content are determined, including in-site messages, SMS, and administrator notifications; Based on the risk coefficient, the timing of the reminder is determined, and the reminder timing is a period when the student has a high probability of responding to the reminder; Based on the reminder content, the reminder channel, and the reminder timing, a supervision reminder message is generated, and the supervision reminder message is sent through the reminder channel at the reminder timing; Record the changes in the student's learning behavior after the supervisory reminder information is sent to form reminder effect feedback, and adjust subsequent reminder strategies based on the reminder effect feedback; Output the estimated completion date and the supervisory reminder information.

Images