Online course personalized recommendation method based on big data analysis

Abstract

The application discloses an online course individualized recommendation method based on big data analysis, comprising the following steps: S1, collecting learning behavior data and course structure data of an online course platform; S2, performing target stability calculation to generate target weight parameters and target constraint conditions; S3, generating ability credibility parameters and ability constraint conditions; S4, constructing an ability difference state vector to generate a gap interval label and a contradiction label set; S5, inputting into an improved Conditional Diffusion Model model, introducing an ability difference evolution trajectory generation mechanism to generate an ability difference evolution trajectory candidate set; S6, performing feasibility mapping to generate a transition course path set; and S7, generating a course recommendation result and updating the target weight parameters and the ability credibility parameters. The application improves the continuity of online course individualized recommendation and the stability of learning effect.

Description

Technical Field

[0001] This invention relates to the field of educational informatization and intelligent recommendation technology, and in particular to a method for personalized recommendation of online courses based on big data analysis. Background Technology

[0002] With the continuous expansion of online education platforms and the sustained growth of lifelong learning needs, course recommendation technology based on big data analytics is gradually becoming an important support for improving learning efficiency and experience. Existing online course platforms typically use information such as learning behavior data and course structure data to push content and rank courses for learners, alleviating to some extent the difficulty of selection caused by the surge in the number of courses. However, in practical applications, existing technical solutions still generally have the following shortcomings: On the one hand, existing course recommendation methods mostly construct learning profiles based on static or short-term statistical features, often focusing on historical course clicks, completion status, or simple similarity calculations. They lack joint modeling of the evolution of learning objectives and changes in learning abilities, making it difficult to reflect the stability of learners' objectives and the reliability of their abilities throughout continuous learning cycles. This leads to discrepancies between recommendation results and actual learning needs. On the other hand, the knowledge point coverage and prerequisite relationships in course structures are complex. Existing methods typically use rule matching or path search to generate learning paths, lacking a systematic characterization of differences in learning abilities. This makes it difficult to identify learners' ability gaps and structural disconnects across different knowledge point dimensions, easily resulting in courses with excessively large spans or unreachable paths. Furthermore, regarding the non-linear differences and dynamic evolution characteristics between learning objectives and learning abilities, existing recommendation methods often rely on linear scoring models or greedy path selection strategies. They lack the ability to model the diversity of the evolution process of ability gaps, making it difficult to generate continuous and feasible transitional course paths, thus affecting the stability and adaptability of recommendation results.

[0003] Therefore, how to provide personalized online course recommendations based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a personalized online course recommendation method based on big data analysis. This invention constructs learning goal vectors, learning ability vectors, and ability difference state vectors based on learning behavior data and course structure data from online course platforms. It introduces goal stability calculation, ability stability calculation, and an ability difference evolution trajectory generation mechanism. An improved Conditional Diffusion Model is used to generate a candidate set of ability difference evolution trajectories, and a transitional course path set is formed by combining knowledge point coverage relationships and knowledge point prerequisite relationships. Finally, the invention outputs course recommendation results and updates parameters based on learning behavior data. This invention achieves dynamic characterization of the gap between learning goals and learning abilities through collaborative modeling of goal constraints and ability constraints. The ability difference evolution trajectory generation mechanism improves the continuity and rationality of course path generation, possessing advantages such as highly targeted recommended paths, controllable learning transition processes, and high adaptive update capability of recommendation results.

[0005] The online course personalized recommendation method based on big data analysis according to an embodiment of the present invention includes the following steps: S1. Collect learning behavior data and course structure data from online course platforms, perform data standardization processing, and generate standardized input data; S2. Construct a learning target vector based on the standardized input data, perform target stability calculation, and generate target weight parameters; and generate target constraints based on the learning target vector and target weight parameters. S3. Construct a learning ability vector based on the standardized input data, generate an ability credibility parameter, and generate ability constraints based on the learning ability vector and the ability credibility parameter; S4. Construct a capability difference state vector based on the target constraints and capability constraints, generate gap interval labels, and generate a set of contradiction labels based on the transition intervals in the gap interval labels. S5. Using the capability difference state vector as the diffusion generation object, the learning target vector, target weight parameters, capability credibility parameters, gap interval labels and contradiction label set are input into the improved Conditional Diffusion Model. The improved Conditional Diffusion Model introduces a capability difference evolution trajectory generation mechanism to generate a capability difference evolution trajectory candidate set. S6. Construct a directed graph of the curriculum structure based on the curriculum structure data, and introduce the RCSP algorithm to perform feasibility mapping on the candidate set of the ability difference evolution trajectory to generate a set of transitional curriculum paths; S7. Generate course recommendation results based on the transition course path set, and update the target weight parameter and ability credibility parameter based on the user's learning behavior data generated by the course recommendation results.

[0006] Optionally, S1 specifically includes: Collect learning behavior data and course structure data from online course platforms to construct the original input dataset; Perform identifier recoding on the original input dataset to generate an encoded dataset; Perform time-series alignment processing on the learning behavior data in the encoded dataset to generate a time-aligned dataset; Outlier removal and missing value completion are performed on the time-aligned dataset and the encoded course structure data to generate a clean dataset. Numerical normalization is performed on the cleaned dataset. The numerical normalization process uses a linear normalization method with the maximum value minus the minimum value as the denominator and the value to be normalized minus the minimum value as the numerator to generate standardized input data.

[0007] Optionally, S2 specifically includes: Based on the course selection records, course completion status records, chapter completion ratio records, learning duration distribution records, and test result records in the standardized input data, a learning objective vector is constructed. Perform target stability calculation on the learning target vector. The target stability calculation includes statistically analyzing the magnitude of change of the learning target vector over consecutive time segments. The target stability value is calculated based on the change range of the learning target vector within consecutive time segments. The target stability value is obtained by summing the squared differences of the learning target vector between adjacent time segments and taking the average. Target weight parameters are generated based on the target stability value, and the target weight parameters and the target stability value have an inverse mapping relationship. Target constraints are generated based on the learning target vector and target weight parameters. The target constraints include: weighting the target achievement deviation between the learning target vector and the learning ability vector according to the target weight parameters to generate the allowable range of target ability difference; determining the evolution direction constraint of the target ability difference based on the time series change trend of the learning target vector; and limiting the upper limit of the rate of change of the target ability difference within a continuous time segment based on the target weight parameters.

[0008] Optionally, S3 specifically includes: Based on the chapter completion status records, chapter completion percentage records, test result records, records of knowledge points associated with incorrect questions, and records of learning time intervals in the standardized input data, a learning ability vector is constructed. Perform capability stability calculation on the learning capability vector. The capability stability calculation includes statistically analyzing the magnitude of change of the learning capability vector within a continuous time segment. The ability stability value is calculated based on the change range of the learning ability vector within consecutive time segments. The ability stability value is obtained by summing the absolute values ​​of the differences between the learning ability vectors in adjacent time segments and taking the average. Capability reliability parameters are generated based on capability stability values, and the capability reliability parameters and capability stability values ​​have an inverse mapping relationship. Capability constraints are generated based on the learning ability vector and the capability credibility parameter. The capability constraints include: weighting the change range of the learning ability vector within a continuous time segment according to the capability credibility parameter to generate a capability change credibility interval; limiting the evolution direction of the capability difference state vector based on the time series change trend of the learning ability vector; and limiting the upper limit of the change range of the capability difference state vector within a continuous time segment based on the capability credibility parameter.

[0009] Optionally, S4 specifically includes: The goal achievement deviation is calculated based on the learning target vector and the learning ability vector. The goal achievement deviation consists of the difference between the learning target vector and the learning ability vector in the corresponding knowledge point dimension. The deviation from the target achievement is weighted by the target weight parameters to generate a weighted target achievement deviation. A capability difference state vector is constructed based on the weighted target achievement deviation and capability credibility parameter. The capability difference state vector consists of capability differences in the knowledge point dimension, and the capability difference is obtained by the product of the weighted target achievement deviation and the capability credibility parameter. Perform gap interval partitioning on the capability difference state vector. The gap interval partitioning includes calculating the overall deviation intensity of the capability difference state vector. The overall deviation intensity is obtained by summing the squares of the capability differences and taking the square root. Based on the overall deviation intensity and the allowable range of target capability difference, gap interval labels are generated. The gap interval labels include directly learnable intervals, transition intervals, and unreachable intervals. Based on the capability difference state vector labeled as the gap interval, a set of gap knowledge points is extracted. The set of gap knowledge points consists of knowledge point labels whose capability difference is greater than a preset difference threshold. Based on the prerequisite relationships of knowledge points in the course structure data, a prerequisite consistency check is performed on the set of missing knowledge points to generate a broken chain set, which consists of knowledge point tags that do not satisfy the prerequisite relationships of knowledge points. A set of contradictory labels is generated based on the set of gap knowledge points and the set of broken links. The set of contradictory labels includes gap labels and broken link labels.

[0010] Optionally, the improved Conditional Diffusion Model includes a conditional coding module, a time-step coding module, a noise prediction module, and a trajectory constraint module. The conditional encoding module uses the learning target vector, target weight parameters, and capability credibility parameters as continuous feature vectors, maps the gap interval labels and contradictory label sets as discrete feature vectors, performs dimension alignment processing on the continuous feature vectors and discrete feature vectors, and concatenates them along the feature dimensions to generate a conditional embedding vector. The diffusion time step is input to the time step encoding module, which performs position encoding mapping on the diffusion time step and outputs the time step embedding vector. The capability difference state vector, conditional embedding vector, and time step embedding vector are input into the noise prediction module. The noise prediction module outputs a noise prediction vector with the same dimension as the capability difference state vector. The noise prediction module introduces a capability difference evolution trajectory generation mechanism driven by the diffusion time step. This mechanism generates the capability difference evolution trajectory by repeatedly denoising and updating the capability difference state vector in consecutive diffusion time steps while retaining the update results. The capability difference evolution trajectory generation mechanism includes: in each diffusion time step, performing a denoising update on the capability difference state vector based on the noise prediction vector; the denoising update is obtained by subtracting the noise prediction vector from the capability difference state vector; and arranging the updated capability difference state vectors obtained in consecutive diffusion time steps in chronological order to generate the capability difference evolution trajectory. By initializing with random noise multiple times and repeatedly executing the diffusion time step, multiple capability difference evolution trajectories are generated, forming a candidate set of capability difference evolution trajectories; The trajectory constraint module performs trajectory consistency constraint calculation on the candidate set of trajectory evolution with capability difference. The trajectory consistency constraint calculation includes performing L2 calculation on the difference vector of the updated capability difference state vector of adjacent time steps and taking the average to obtain the trajectory consistency loss value. The joint loss value is generated by weighting and summing the trajectory consistency loss value and the mean square error loss value between the noise prediction vector and the noise. The parameters of the noise prediction module are then updated based on the joint loss value.

[0011] Optionally, S5 specifically includes: The initial state vector for diffusion is generated based on the capability difference state vector. The initial state vector for diffusion is obtained by adding the capability difference state vector and the random noise vector. The random noise vector has the same dimension as the capability difference state vector. The learning target vector, target weight parameters, and ability credibility parameters are used as continuous feature vectors. The gap interval labels and contradictory label sets are mapped to discrete feature vectors. Dimension alignment processing is performed on the continuous feature vectors and discrete feature vectors, and they are concatenated along the feature dimensions to generate conditional embedding vectors. The initial state vector of diffusion, the conditional embedding vector, and the diffusion time step are input into the improved ConditionalDiffusion Model, and the noise prediction vector corresponding to each diffusion time step is output. The initial state vector of diffusion is denoised and updated based on the noise prediction vector. The denoising update is obtained by subtracting the noise prediction vector from the initial state vector of diffusion to obtain the updated state vector. Perform iterative updates on the updated state vector and the diffusion time step, performing the iterative updates a preset number of time steps to obtain the updated state vector sequence corresponding to the diffusion time step; The updated state vector sequence is arranged in the order of diffusion time steps to generate the capability difference evolution trajectory; Repeatedly perform random noise initialization and diffusion time step iterative update to generate multiple capability difference evolution trajectories, forming a capability difference evolution trajectory candidate set; The trajectory consistency constraint calculation is performed on the candidate set of poor-capability evolution trajectories to obtain the trajectory consistency loss value; The joint loss value is generated by weighted summation of the trajectory consistency loss value and the mean square error loss value between the noise prediction vector and the random noise vector, and the parameters of the improved Conditional Diffusion Model are updated based on the joint loss value.

[0012] Optionally, S6 specifically includes: Based on the course structure data, extract the course chapter tag set, knowledge point tag set, knowledge point coverage relationship set, and knowledge point prerequisite relationship set; A node set is constructed based on the course chapter tag set and the knowledge point tag set, and a directed edge set is constructed based on the knowledge point prerequisite relationship set. The node set and the directed edge set are combined to construct a directed graph of the course structure. Based on the set of knowledge point prerequisite relationships, the preceding counts corresponding to each knowledge point tag are statistically analyzed, an output set is constructed, and a sorted description of knowledge point prerequisite relationships is generated. The gap knowledge point extraction process is performed on the gap knowledge points in the candidate set of gap evolution trajectories. The gap knowledge points are composed of knowledge point labels in the gap state vector where the gap value is greater than a preset gap threshold. A set of candidate course paths is constructed based on a directed graph of the course structure. The course paths in the set of candidate course paths consist of a sequence of course chapter labels that satisfy the prerequisite relationship of knowledge points. Perform knowledge point coverage intersection calculation on the course paths in the candidate course path set, and generate coverage matching value based on the knowledge point coverage intersection; perform prerequisite consistency verification on the course paths in the candidate course path set, and generate prerequisite violation set; The RCSP algorithm input is constructed based on the directed graph of the course structure, the coverage matching value, the prerequisite violation set and the preset constraint threshold set. The RCSP algorithm is then introduced to perform resource constraint shortest path search processing on the candidate course path set and output the path result set. Path backtracking processing is performed on the path result set to generate the transition course path set.

[0013] Optionally, the RCSP algorithm is introduced to perform resource-constrained shortest path search on the candidate course path set, outputting a path result set, specifically as follows: A path state set is constructed based on the directed graph of the course structure and the set of candidate course paths. Each path state in the path state set corresponds to a candidate course path and records the course chapter label sequence, path cost, coverage matching value, size of the prerequisite violation set, and path length. The path state set is filtered based on the preset constraint threshold set. Path states that meet the coverage matching threshold requirement, the size of the prior violation set meets the prior violation threshold requirement, and the path length meets the path length threshold requirement are written into the expandable state set. Sort the scalable state set according to path cost and select the path state with the minimum path cost as the search starting state, and write it into the determined state set at the same time. Based on the sequence of course chapter labels corresponding to the initial search state, find the set of subsequent course chapter labels that satisfy the prerequisite relationship constraint of knowledge points in the directed graph of the course structure. The subsequent course chapter tags are appended one by one to the end of the course chapter tag sequence to form an extended course path, and the path cost, overwrite matching value, prerequisite violation set size and path length are updated synchronously to generate the extended path state; The extended path states are filtered based on a preset set of constraint thresholds, and the extended path states that meet the resource constraints are written into the extended state set; the resource constraints include coverage matching constraints, prior consistency constraints, path complexity constraints, and optimization objective constraints. Perform dominance removal on path states within the scalable state set; Repeatedly search for the starting state, find subsequent course chapter tags, generate extended path states, perform constraint verification and eliminate dominance relationships until the set of expandable states is empty, and obtain a set of path states that meet the resource constraints; sort the set of path states in ascending order according to the path cost and truncate a preset number to generate a set of path results.

[0014] Optionally, S7 specifically includes: Based on the transitional course path set, extract the course chapter tag sequence set, perform path sorting processing, and sort the path in ascending order by path cost to obtain the recommended path sequence; The course recommendation results are generated based on the recommendation path sequence. The course recommendation results include the course chapter tag sequence in the recommendation path sequence and the knowledge point tag set corresponding to the course chapter tag sequence. Acquire user learning behavior data based on course recommendation results and write the learning behavior data into the time segment corresponding to the standardized input data; update the learning objective vector and recalculate the objective stability value based on the written standardized input data, and update the objective weight parameters; update the learning ability vector and recalculate the ability stability value based on the written standardized input data, and update the ability reliability parameters.

[0015] The beneficial effects of this invention are: (1) By jointly modeling the learning target vector, target weight parameters and target constraints, the stability change of the learning target in continuous time segments is characterized, and the target stability is introduced into the ability difference state vector construction process to realize the dynamic constraint guidance of the learning target evolution trend on the course path generation, and improve the long-term consistency between the course recommendation results and the real learning needs. (2) By continuously characterizing the learning ability change process through learning ability vector, ability credibility parameter and ability constraint, the ability stability is directly applied to the evolution control of the ability difference state vector, so that the ability difference state maintains smooth convergence characteristics in the time dimension, effectively avoiding the interference of ability assessment fluctuations on the continuity and feasibility of the course recommendation path. (3) Introduce an improved Conditional Diffusion Model and a capability difference evolution trajectory generation mechanism to expand the capability difference state from single-point prediction to a multi-time-step continuous evolution trajectory candidate set, and filter smooth reachable trajectories through trajectory consistency constraints, so that the course path construction process has continuity, controllability and multi-solution coverage capability. (4) By constructing a set of transitional course paths in collaboration with the directed graph of course structure, the prior relationship description of knowledge points and the RCSP algorithm, the feasibility mapping of the capability gap repair path is realized under the joint control of the coverage matching value, the prior violation set and the resource constraint threshold, which significantly reduces the probability of course span mismatch and path inaccessibility problems. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a flowchart of the online course personalized recommendation method based on big data analysis proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved Conditional Diffusion Model proposed in this invention; Figure 3 This is a data flow diagram of the online course personalized recommendation method based on big data analysis proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1-3 The personalized recommendation method for online courses based on big data analysis includes the following steps: S1. Collect learning behavior data and course structure data from online course platforms, perform data standardization processing, and generate standardized input data; S2. Construct a learning target vector based on standardized input data, perform target stability calculation, and generate target weight parameters; and generate target constraints based on the learning target vector and target weight parameters. S3. Construct a learning ability vector based on standardized input data and generate an ability credibility parameter; and generate ability constraints based on the learning ability vector and the ability credibility parameter. S4. Construct a capability difference state vector based on the target constraints and capability constraints, generate gap interval labels, and generate a set of contradiction labels based on the transition intervals in the gap interval labels. S5. Using the capability difference state vector as the diffusion generation object, the learning target vector, target weight parameters, capability credibility parameters, gap interval labels and contradiction label set are input into the improved Conditional Diffusion Model. The improved Conditional Diffusion Model introduces a capability difference evolution trajectory generation mechanism to generate a candidate set of capability difference evolution trajectories. S6. Construct a directed graph of the curriculum structure based on the curriculum structure data, and introduce the RCSP algorithm to perform feasibility mapping on the candidate set of poor ability evolution trajectories to generate a set of transitional curriculum paths; S7. Generate course recommendation results based on the set of transitional course paths, and update the target weight parameters and ability credibility parameters based on the learning behavior data generated by users on the course recommendation results.

[0019] In this embodiment, S1 specifically refers to: Collect user course click records, course start records, course completion status records, chapter completion percentage records, learning time distribution records, test result records, as well as course corresponding knowledge point composition relationship data and knowledge point prerequisite relationship data from online course platforms. Construct the original input dataset according to user identifier field, course identifier field, knowledge point identifier field and time stamp field. The original input dataset uses the user dimension as the row index and the course identifier and knowledge point identifier combination dimension as the column index, and populates the corresponding positions with learning behavior numerical feature vectors. The original input dataset is subjected to identifier recoding. User identifiers are mapped to consecutive integer indices according to their appearance order, and course identifiers are mapped to consecutive integer indices according to the course release order. Knowledge point identifiers are mapped to consecutive integer indices according to the course structure hierarchy, forming a unified index matrix structure. The corresponding identifier fields in the original input dataset are replaced with index values ​​to generate an encoded dataset. The encoded dataset maintains the row and column structure unchanged, and only the identifier fields are numerically unified. Time series alignment processing is performed on the learning behavior data in the coded dataset. The time granularity is set, and the time stamp is divided by the time granularity length and the integer is taken as the time segment index. The learning behavior values ​​corresponding to the same user index, course index, knowledge point index and time segment index are summed and averaged. The aggregation results are filled into the corresponding time segment positions to form a learning behavior matrix arranged continuously according to time segments, and a time-aligned dataset is generated. Outlier removal is performed on the time-aligned dataset and the encoded course structure data. Records with learning duration greater than the mean plus 3 times the standard deviation are marked as outliers and deleted. Records with chapter completion ratios less than 0 or greater than 1 are deleted. At the same time, missing positions are filled by linear interpolation based on the values ​​of the same user index in adjacent time segments. For positions with more than 5 consecutive missing time segments, the mean of the same course index is used for filling, forming a clean dataset with continuous and complete values. Numerical normalization is performed on the cleaned dataset. The maximum and minimum values ​​of each feature column are calculated. The original feature value minus the minimum value of the corresponding column is used as the numerator, and the maximum value minus the minimum value of the corresponding column is used as the denominator. The ratio of the numerator to the denominator is used as the normalization result and filled into the corresponding position to form standardized input data with values ​​ranging from 0 to 1.

[0020] In this embodiment, S2 specifically refers to: A learning objective vector is constructed based on standardized input data including course selection records, course completion status records, chapter completion ratio records, learning time distribution records, and test result records. The learning objective vector is a numerical vector formed by arranging the knowledge point weight values ​​corresponding to course selection records, completion marker values ​​corresponding to course completion status, proportion values ​​corresponding to chapter completion ratios, time percentage values ​​corresponding to learning time distribution, and test result scores along the knowledge point dimension. The learning objective vector is organized according to the knowledge point dimension, with each knowledge point dimension corresponding to a target component. Course selection records are used to tag the frequency of user-initiated course selections within the statistical time range; course completion status records are used to tag whether course chapters have met completion conditions; chapter completion ratio records are used to characterize the degree of course chapter completion; learning time distribution records are used to characterize the intensity of time invested in course chapters corresponding to different knowledge points; and test result records are used to characterize the target orientation of knowledge point mastery results. The target stability value is calculated based on the change of the learning target vector within consecutive time segments. The target stability value is calculated by squaring the difference of the target value between adjacent time segments for each knowledge point dimension, summing all the squaring results, and finally dividing the sum by the number of time segments involved in the calculation to obtain the average value. The larger the target stability value, the more obvious the fluctuation of the learning target over time, and the smaller the target stability value, the more stable the learning target remains. Target weight parameters are generated based on the target stability value. The target weight parameters are inversely mapped to the target stability value. The target weight parameters are calculated by taking 1 as the initial weight benchmark value and dividing the initial weight benchmark value by the sum of the target stability value and the smoothing factor to obtain the target weight parameters. The smoothing factor is used to avoid the situation where the weight increases infinitely when the target stability value is close to zero. Preferably, the smoothing factor is 0.01. The generated target weight parameters are used to characterize the constraint strength of the learning target in the subsequent ability difference modeling process. The target constraints are generated based on the learning target vector and target weight parameters. The allowable range of target ability difference in the target constraints is set by weighting the difference between the learning target vector and the learning ability vector in the corresponding knowledge point dimension with the target weight parameters. The upper boundary is obtained by adding a preset tolerance range to the weighted difference, and the lower boundary is obtained by subtracting the preset tolerance range from the weighted difference. Preferably, the preset tolerance range is set to 0.2 to limit the change of the ability difference state within a reasonable target deviation range. The ability difference state vector is formed by weighting the difference between the learning target vector and the learning ability vector in the corresponding knowledge point dimension with the target weight parameter and the ability credibility parameter. The vector element is the ability difference corresponding to each knowledge point. The evolution direction constraint of the target ability difference is determined based on the time series change trend of the learning target vector. The evolution direction constraint is determined by statistically analyzing the change sign of the learning target vector in consecutive time segments. When the target value of the same knowledge point dimension shows an upward trend in multiple consecutive time segments, the change direction of the ability difference state vector in that knowledge point dimension is constrained to narrow the gap. When the target value of the same knowledge point dimension shows a downward trend in multiple consecutive time segments, the change direction of the ability difference state vector in that knowledge point dimension is constrained to maintain or slowly widen the gap. The target weight parameter limits the upper limit of the rate of change of the target capability difference within consecutive time segments. The upper limit of the rate of change is obtained by multiplying the target weight parameter by a preset maximum change ratio. Preferably, the preset maximum change ratio is set to 0.3, which is used to limit the maximum change amplitude of the capability difference state within adjacent time segments, and prevent the capability difference state from undergoing drastic changes during the evolution process, thereby forming a complete and fully disclosed target constraint condition.

[0021] In this embodiment, S3 specifically refers to: Learning ability vectors are constructed based on standardized input data, including chapter completion status records, chapter completion ratio records, test result records, records of knowledge points associated with incorrect answers, and learning time interval records. The learning ability vectors are organized according to knowledge point dimensions, with each knowledge point dimension corresponding to an ability component. Chapter completion status records are used to characterize whether the corresponding knowledge point course chapter has met the completion conditions set by the platform. Chapter completion ratio records are used to characterize the completion depth of the corresponding knowledge point course chapter. Test result records are used to characterize the correctness of answers to the corresponding knowledge points. Records of knowledge points associated with incorrect answers are used to characterize the frequency distribution of knowledge points that were incorrectly answered during the test. Learning time interval records are used to characterize the time span between two adjacent learning behaviors for the same knowledge point. The ability stability value is calculated based on the change of the learning ability vector within consecutive time segments. The calculation method is to take the absolute value of the difference in ability value between adjacent time segments for each knowledge point dimension, then sum all the absolute value results, and finally divide the summation result by the number of time segments involved in the calculation to obtain the average value. The larger the ability stability value, the more obvious the fluctuation of learning ability over time, and the smaller the ability stability value, the more stable the learning ability remains. Ability reliability parameters are generated based on ability stability values. The ability reliability parameters are inversely mapped to the ability stability values. The ability reliability parameters are calculated by taking 1 as the initial reliability benchmark value and dividing the initial reliability benchmark value by the sum of the ability stability value and the smoothing factor. The smoothing factor is used to avoid the reliability parameters from increasing infinitely when the ability stability value is close to zero. Preferably, the smoothing factor is 0.01. The generated ability reliability parameters are used to characterize the reliability of the learning ability assessment results. Based on the learning ability vector and the ability credibility parameter, the ability constraint condition is generated. The ability change credibility interval in the ability constraint condition is set by weighting the change amplitude of the learning ability vector in a continuous time segment with the ability credibility parameter and setting the upper and lower boundaries. The upper boundary is obtained by adding the preset tolerance amplitude to the weighted change amplitude, and the lower boundary is obtained by subtracting the preset tolerance amplitude from the weighted change amplitude. Preferably, the preset tolerance amplitude is set to 0.15 to limit the ability change to be within the credibility range. The evolution direction of the ability gap state vector is determined based on the time series change trend of the learning ability vector. The evolution direction is determined by statistically analyzing the change direction of the learning ability vector in continuous time segments. When the ability value of the same knowledge point dimension shows a continuous upward trend, the evolution direction of the ability gap state vector in the corresponding dimension is determined to be the direction of narrowing the ability gap. When the ability value of the same knowledge point dimension shows a continuous downward trend, the evolution direction of the ability gap state vector in the corresponding dimension is determined to be the direction of maintaining or slowly widening the ability gap. The ability credibility parameter limits the upper limit of the change range of the ability-poor state vector within consecutive time segments. The upper limit of the change range is obtained by multiplying the ability credibility parameter by a preset maximum change ratio. Preferably, the preset maximum change ratio is set to 0.25 to limit the maximum change range of the ability-poor state within adjacent time segments, so as to prevent the ability-poor state from exhibiting drastic fluctuations that do not conform to the learning behavior characteristics during the evolution process, thereby forming a fully disclosed and logically closed ability constraint condition.

[0022] Both the learning objective vector and the learning ability vector are constructed using the knowledge point labels in the course structure data as a unified index dimension.

[0023] Learning behavior indicators from different sources are merged into the corresponding knowledge point dimensions through knowledge point label mapping relationships, and target values ​​or ability values ​​under the knowledge point dimensions are generated through weighted aggregation, so that the learning target vector and the learning ability vector are consistent in terms of the number of dimensions, the meaning of dimensions, and the order of arrangement.

[0024] In this embodiment, S4 specifically refers to: Based on the learning target vector and learning ability vector aligned under the same knowledge point dimension, the target achievement deviation is calculated dimension by dimension. The target achievement deviation consists of the difference between the learning target vector and the learning ability vector in the corresponding knowledge point dimension. The deviation from the target achievement is weighted by the target weight parameters to generate a weighted target achievement deviation. A capability difference state vector is constructed based on the weighted target achievement deviation and capability credibility parameter. The capability difference state vector consists of capability differences in the knowledge point dimension, and the capability difference is obtained by the product of the weighted target achievement deviation and the capability credibility parameter. Perform gap interval partitioning on the capability difference state vector. The gap interval partitioning includes calculating the overall deviation intensity of the capability difference state vector. The overall deviation intensity is obtained by summing the squares of the capability difference values ​​in the capability difference state vector and taking the square root. Gap interval labels are generated based on the overall deviation intensity and the allowable range of the target capability difference. These labels include directly learnable intervals, transition intervals, and unreachable intervals. The overall deviation intensity is obtained by summing the squares of the capability differences in each knowledge point dimension of the capability difference state vector and taking the square root. The allowable range of the target capability difference is composed of the minimum and maximum allowable deviation thresholds defined by the target constraints. An interval comparison is performed between the overall deviation intensity and the allowable range of the target capability difference. When the overall deviation intensity is less than the minimum allowable deviation threshold, a directly learnable interval label is generated. When the overall deviation intensity is greater than or equal to the minimum allowable deviation threshold and less than or equal to the maximum allowable deviation threshold, a transition interval label is generated. When the overall deviation intensity is greater than the maximum allowable deviation threshold, an unreachable interval label is generated. Preferably, the minimum allowable deviation threshold is set between 0.15 and 0.30, and the maximum allowable deviation threshold is set between 0.60 and 0.85. Based on the capability difference state vector labeled as the gap interval, a set of gap knowledge points is extracted. The set of gap knowledge points consists of knowledge point labels whose capability difference is greater than a preset difference threshold. Based on the prerequisite relationships of knowledge points in the course structure data, a prerequisite consistency check is performed on the set of missing knowledge points to generate a broken chain set, which consists of knowledge point tags that do not satisfy the prerequisite relationships of knowledge points. A set of contradictory labels is generated based on the set of gap knowledge points and the set of broken links. The set of contradictory labels includes gap labels and broken link labels.

[0025] In this embodiment, the improved Conditional Diffusion Model includes a conditional coding module, a time-step coding module, a noise prediction module, and a trajectory constraint module, specifically: In the conditional encoding module, the learning target vector, target weight parameters, and capability credibility parameters are represented as continuous feature vectors. Each dimension of the continuous feature vector corresponds to a numerical description under the knowledge point dimension. The gap interval labels and the contradiction label set are represented as discrete feature vectors. The discrete feature vectors are expanded using one-hot encoding. The gap interval labels are mapped to discrete vectors of length 3, corresponding to the directly learnable interval, the interval requiring transition, and the unreachable interval, respectively. The contradiction label set is mapped to a binary label vector with the same number of knowledge points. The gap label position is set to 1, and the non-gap position is set to 0. The chain break label position is set to 1, and the non-chain break position is set to 0. Dimensional alignment processing is performed on the continuous feature vectors and discrete feature vectors. Dimensional alignment processing maps features from different sources to the same dimensional space through linear mapping. The preferred mapping dimension is 128. The features after dimensional alignment processing are concatenated along the feature dimensions to generate a conditional embedding vector. In the conditional embedding vector, different types of features are arranged in the order of learning target vector, target weight parameters, capability credibility parameters, gap interval labels, and contradiction label set. In the time step encoding module, the diffusion time step is represented as an integer sequence. The range of the diffusion time step is preferably set to 1 to 50. Position encoding mapping is performed on the diffusion time step. The position encoding mapping is formed by combining the sine and cosine functions of the diffusion time step to form the time step embedding vector. The dimension of the time step embedding vector is consistent with the dimension of the conditional embedding vector, which is used to characterize the evolution position of the poor ability state vector in different diffusion stages. In the noise prediction module, the capability difference state vector, conditional embedding vector, and time step embedding vector are concatenated along the feature dimension and input into the prediction network that combines multi-layer linear transformation and nonlinear activation. The prediction network outputs a noise prediction vector, in which each dimension corresponds to the noise estimate of the knowledge point dimension in the capability difference state vector. In the capability difference evolution trajectory generation mechanism, within each diffusion time step, the capability difference state vector is denoised and updated based on the noise prediction vector. The denoising and updating process obtains the updated capability difference state vector by subtracting the corresponding dimension value from the corresponding dimension value in the noise prediction vector. The updated capability difference state vectors obtained in consecutive diffusion time steps are stored in the order of diffusion time steps to form the capability difference evolution trajectory. The capability difference state vectors between adjacent time steps in the capability difference evolution trajectory reflect the gradual change of capability difference over time. By performing multiple independent initializations on the random noise vector, the dimension of the random noise vector is consistent with the capability difference state vector, and the value range of the random noise vector is preferably set to -1 to 1. Combined with the repeated execution of the denoising and update process in the diffusion time step, multiple capability difference evolution trajectories are generated, forming a candidate set of capability difference evolution trajectories. In the trajectory constraint module, a consistency evaluation is performed on each capability difference evolution trajectory in the candidate set of capability difference evolution trajectories. The trajectory consistency loss value is obtained by calculating the L2 norm of the difference vector of the capability difference state vector updated in adjacent diffusion time steps and taking the average value. The trajectory consistency loss value is used to measure the smoothness of the capability difference evolution trajectory in the time dimension. The trajectory consistency loss value and the mean square error loss value between the noise prediction vector and the random noise vector are weighted and summed according to preset weights, preferably set to 0.7 and 0.3, to generate a joint loss value. Based on the joint loss value, the parameters in the noise prediction module are updated in reverse, so that the generated capability difference evolution trajectory has continuity and stability while satisfying the condition constraints.

[0026] In this implementation, because the ability difference state vector is simultaneously influenced by learning objective constraints, ability credibility constraints, and curriculum structure constraints during its temporal evolution, the Conditional Diffusion Model only statically encodes the conditional information and guides the random generation process. This makes it difficult to characterize the evolution direction and continuity of the ability difference state over continuous time, easily leading to discontinuous changes in the generated results at adjacent diffusion stages, thus weakening the feasibility and stability of the transitional curriculum path construction. To address this, an ability difference evolution trajectory generation mechanism is introduced, extending the single diffusion result generation method to a continuous denoising and updating process over multiple diffusion time steps. By retaining the staged update results of the ability difference state vector within the diffusion time dimension, an ability difference evolution trajectory with a temporal order is formed, transforming the generation process from static condition guidance to a dynamic evolution process jointly regulated by objective and ability constraints. Simultaneously, a trajectory constraint module in the model structure applies continuity constraints to the candidate set of ability difference evolution trajectories, ensuring smooth changes in the ability difference state between adjacent diffusion time steps. This avoids abrupt generation results interfering with the curriculum path mapping, strengthening the stability and interpretability of the generated trajectory from a structural perspective. By uniformly encoding the learning target vector, target weight parameters, ability credibility parameters, gap interval labels, and contradictory label sets into a conditional embedding vector, and continuously participating in noise prediction and denoising updates during the diffusion process, a deep coupling between the ability difference state generation process and the learning target, learning ability, and curriculum structure is achieved. It has the ability to perform multi-path prediction and constraint screening at the ability difference evolution level, thus breaking through the limitation of traditional conditional diffusion models that only focus on the final state generation in terms of technical implementation.

[0027] In this embodiment, S5 specifically refers to: When constructing the initial state vector for diffusion based on the capability difference state vector and the random noise vector, the random noise vector adopts a zero-mean Gaussian distribution vector with the same dimension as the capability difference state vector, and the noise amplitude is controlled within the range of 0.5 to 1.5. The initial state vector for diffusion is formed by adding it dimension by dimension to the capability difference state vector, so that the diffusion generation process has random perturbation characteristics and forms a diversified evolution starting point. When the learning target vector, target weight parameters, and capability credibility parameters are used as continuous feature vectors, the continuous feature vectors are arranged in numerical order and scaled uniformly. The value range of the continuous feature vectors after scaled uniformity is limited to 0 to 1. When the gap interval labels and contradiction label sets are mapped to discrete feature vectors, the directly learnable interval, the transition interval, and the unreachable interval are mapped to fixed-length binary vectors through one-hot encoding. The gap label and the broken chain label are mapped to binary labels with 1 at the corresponding position. The continuous feature vectors and discrete feature vectors are aligned in dimension by padding with zeros and concatenated along the feature dimensions to generate a conditional embedding vector. The overall dimension of the conditional embedding vector is of the same order of magnitude as the dimension of the capability difference state vector. When the diffusion initial state vector, conditional embedding vector, and diffusion time step are input into the improved ConditionalDiffusion Model, the diffusion time step value is set to the range of 1 to 100. The diffusion time step is used as a discrete integer input time step encoding module. The time step encoding result and the conditional embedding vector are fused in the feature dimension and used as the conditional input of the noise prediction module, so that the noise prediction process is simultaneously affected by diffusion stage information and target constraint information. When performing denoising update on the initial state vector of diffusion based on the noise prediction vector, a denoising update is performed once in each diffusion time step. The updated state vector is obtained by subtracting the corresponding noise prediction value from the current state vector dimension by dimension. The update magnitude is jointly adjusted by the target weight parameter and the capability confidence parameter, and the change magnitude of the capability poor state vector in a single diffusion time step is limited to no more than 0.2. When performing iterative updates on the updated state vector and the diffusion time step, the number of iterations is set to 50 to 100, forming an updated state vector sequence that corresponds one-to-one with the diffusion time step. Each state vector in the updated state vector sequence retains complete capability difference state information and is arranged in ascending order according to the diffusion time step, forming the capability difference evolution trajectory. When repeatedly performing random noise initialization and diffusion time step iterative update, the number of random noise initializations is set to 10 to 30. Different random noise initializations correspond to different capability difference evolution trajectories. By summarizing, a candidate set of capability difference evolution trajectories is formed, so that the candidate set covers multiple potential capability repair paths. When performing trajectory consistency constraint calculation on the candidate set of capability difference evolution trajectories, the trajectory consistency loss value is obtained by averaging the sum of squares of the difference vectors between the updated capability difference state vectors corresponding to adjacent diffusion time steps. The trajectory consistency loss value is used to measure the continuity of capability difference states in the diffusion time dimension. When weighted summing the trajectory consistency loss value and the mean square error loss value between the noise prediction vector and the random noise vector, the weighting coefficients are set to 0.6 and 0.4 to generate a joint loss value. Based on the joint loss value, the parameters of the noise prediction module are updated so that the candidate set of poor capability evolution trajectories simultaneously meets the constraints in terms of denoising accuracy and temporal continuity.

[0028] In this embodiment, S6 specifically refers to: Based on the course structure data, we extract the following sets: course chapter tag set, knowledge point tag set, knowledge point coverage relationship set, and knowledge point prerequisite relationship set. The course chapter tag set is obtained by combining the course identifier and chapter identifier in the course structure data and removing duplicates. The knowledge point tag set is obtained by obtaining the knowledge point identifier in the course structure data and removing duplicates. The knowledge point coverage relationship set consists of the corresponding records of course chapter tags and knowledge point tags. The knowledge point prerequisite relationship set consists of the corresponding records of prerequisite knowledge point tags and subsequent knowledge point tags. A node set is constructed based on the course chapter label set and the knowledge point label set, and a directed edge set is constructed based on the knowledge point prerequisite relationship set. The node set and the directed edge set are combined to construct a directed graph of the course structure. The node set contains two types of nodes: course chapter nodes and knowledge point nodes. The directed edge set contains knowledge point prerequisite edges and retains the edge directions from prerequisite knowledge point labels to successor knowledge point labels. The directed graph of the course structure is stored in the form of an adjacency list, which registers the list of successor knowledge point labels for each knowledge point label. The knowledge point coverage relationship set is transcribed into a coverage index table, which registers the list of covered knowledge point labels for each course chapter label. Based on the set of prerequisite relationships for knowledge points, the predecessor count for each knowledge point tag is calculated, where the predecessor count equals the number of prerequisite relationships pointing to the corresponding knowledge point tag. An output set is constructed, consisting of knowledge point tags with a predecessor count of zero. Knowledge point tags are extracted from the output set and written into a sorting sequence. The set of successor knowledge point tags corresponding to the knowledge point tags is extracted from the directed graph of the course structure. For each knowledge point tag in the set of successor knowledge point tags, the predecessor count is decremented by 1, and when the result of decrementing by 1 is 0, it is written into the output set. This writing and updating is repeated until the output set is empty, generating a sorted description of knowledge point prerequisite relationships. The knowledge point prerequisite relationship constraint is that any knowledge point label in the course chapter label sequence must satisfy the order constraint that the position of the preceding knowledge point label is less than the position of the subsequent knowledge point label in the knowledge point prerequisite relationship set; The gap knowledge point extraction process is performed on the capability difference evolution trajectory candidate set. The gap knowledge point consists of the knowledge point label in the capability difference state vector whose capability difference value is greater than a preset difference threshold. The preset difference threshold is obtained by statistically analyzing the distribution of capability difference values ​​in the standardized input data. The threshold is the sum of the mean capability difference value and the standard deviation capability difference value. The mean capability difference value is equal to the sum of the capability difference values ​​of all knowledge point dimensions divided by the number of knowledge point dimensions. The standard deviation capability difference value is equal to the sum of the squares of the differences between the capability difference values ​​of all knowledge point dimensions and the mean capability difference value divided by the number of knowledge point dimensions and then taking the square root. The gap knowledge point set is stored in a deduplicated set structure and the number of gap knowledge points is recorded. A candidate course path set is constructed based on a directed graph of the course structure. Each course path in the candidate course path set consists of a sequence of course chapter labels that satisfy the prerequisite relationship constraint of knowledge points. The construction of the candidate course path set employs backtracking enumeration, which selects course chapter labels that intersect with the missing knowledge points from the coverage index table as candidate path starting points. In each expansion, the backtracking enumeration selects the next course chapter label from the coverage index table and appends it to the end of the course chapter label sequence. The course chapter label sequence is appended sequentially, with the append position at the end of the sequence. After each append, a temporary covered knowledge point set is formed, which is obtained by taking the union of all covered knowledge point labels within the sequence. Finally, a temporary knowledge point sequence is formed after each append, which is obtained by expanding the corresponding covered knowledge point labels of the course chapter labels within the sequence according to the order of the course chapter label sequence. The knowledge point coverage intersection is calculated for each course path in the candidate course path set, and a coverage matching value is generated based on the knowledge point coverage intersection. The knowledge point coverage intersection is obtained by taking the intersection of the gap knowledge point set and the temporary coverage knowledge point set. The coverage matching value is equal to the number of knowledge point coverage intersections divided by the number of gap knowledge points. The number of knowledge point coverage intersections is equal to the count value of the knowledge point labels in the intersection. The coverage matching value ranges from 0 to 1. The prerequisite consistency check is performed on the course paths in the candidate course path set to generate a prerequisite violation set. The prerequisite consistency check then searches the adjacency list for each knowledge point label in the temporary knowledge point sequence to obtain a prerequisite precondition set. This prerequisite precondition set consists of all prerequisite knowledge point labels pointing to the current knowledge point label from the knowledge point prerequisite relation set. For each prerequisite knowledge point label in the prerequisite precondition set, the check determines whether its position index in the temporary knowledge point sequence is less than the position index of the current knowledge point label. The position index is obtained from the first occurrence of the knowledge point label in the temporary knowledge point sequence. If any prerequisite knowledge point label's position index is greater than or equal to the current knowledge point label's position index, the current knowledge point label and its corresponding prerequisite knowledge point label are written into the prerequisite violation set. The prerequisite violation set is stored using a binary tuple list structure, where each tuple consists of a prerequisite knowledge point label and its successor knowledge point label. The size of the prerequisite violation set is equal to the number of binary tuples. The RCSP algorithm is constructed based on the directed graph of the course structure, the coverage matching value, the prerequisite violation set and the preset constraint threshold set. The RCSP algorithm is then used to perform resource constraint shortest path search on the candidate course path set, and the output path result set is generated. Perform path backtracking processing on the path result set to generate a transition course path set; the path backtracking processing preserves the course chapter label order for each course chapter label sequence in the path result set and generates transition course path labels; the transition course path labels are generated by combining the sequence number in the path result set with the course chapter label sequence length and deduplicating them.

[0029] In this implementation, the RCSP algorithm is introduced to perform resource-constrained shortest path search on the candidate course path set, and the output path result set is as follows: A path state set is constructed based on the directed graph of the course structure and the set of candidate course paths. Each path state in the path state set corresponds to a sequence of course chapter labels, and the path cost, coverage matching value, size of the prerequisite violation set, and path length are recorded synchronously. The path length is obtained by the number of chapters in the course chapter label sequence. The path cost is obtained by multiplying the path length by the average learning time of the chapters in the course chapter label sequence. The average learning time of the chapters is obtained by summing the learning times of each chapter in the course structure data and dividing by the number of chapters. The intersection of the knowledge point tag set covered by the course chapter tag sequence and the gap knowledge point set is statistically analyzed to form the intersection number, and the coverage matching value is obtained by dividing the intersection number by the number of knowledge points in the gap knowledge point set. Record the knowledge point label pairs that violate the knowledge point prerequisite relationship constraints in the course chapter label sequence to form a prerequisite violation set, and count the number of label pairs in the prerequisite violation set to obtain the size of the prerequisite violation set; The path state set is filtered based on the preset constraint threshold set, and the extended path state that meets the resource constraint conditions is written into the expandable state set. Specifically, the path state with a coverage matching value greater than or equal to the coverage matching threshold, a prior violation set size less than or equal to the prior violation threshold, and a path length less than or equal to the path length threshold is written into the expandable state set. Arrange the path states in the scalable state set in ascending order of path cost, and select the path state with the lowest path cost to write into the determined state set to form the search starting state. Based on the sequence of course chapter labels corresponding to the starting state of the search, the set of subsequent course chapter labels that have a directed connection relationship with the last chapter label is found in the directed graph of the course structure. The set of subsequent course chapter labels consists of chapter labels that satisfy the knowledge point prerequisite relationship constraint. The subsequent course chapter tags are appended one by one to the end of the course chapter tag sequence to form a new course chapter tag sequence. The path length is updated according to the new number of chapters, the average learning time of chapters is updated according to the new chapter learning time distribution, and the path cost is recalculated. At the same time, the coverage matching value is recalculated according to the new knowledge point coverage and the size of the prerequisite violation set is recalculated according to the new prerequisite relationship violation, forming an extended path state. Based on a preset set of constraint thresholds, the extended path states are filtered, and the extended path states that meet the coverage matching threshold, the prior violation threshold, and the path length threshold are written into the extended state set. Dominance relationship elimination is performed on path states in the scalable state set. Dominance relationship elimination includes removing path states whose path cost is greater than or equal to the path cost of any path state in the determined state set, whose coverage matching value is less than or equal to the coverage matching value of the corresponding path state, whose size of the prior violation set is greater than or equal to the size of the prior violation set of the corresponding path state, and whose path length is greater than or equal to the path length of the corresponding path state from the scalable state set. In the scalable state set, re-sort according to path cost to select the new path state with the minimum path cost and write it into the determined state set. Continue to expand the chapter labels of subsequent courses, filter the constraint thresholds and remove the dominance relationship until the scalable state set is empty. The path states in the determined state set are arranged in ascending order of path cost, and a preset number are extracted to form the path result set.

[0030] In this embodiment, S7 specifically refers to: Based on the transitional course path set, a course chapter label sequence set is extracted. Each sequence in the course chapter label sequence set is formed by connecting multiple course chapter labels in the order of learning. The connection method is that the end position of the previous course chapter label is directly connected to the start position of the next course chapter label to form a linear sequence structure. Each course chapter label sequence is associated with the corresponding path cost, coverage matching value, size of the prerequisite violation set and path length. Based on the course chapter tag sequence set, the course chapter tag sequence is sorted by path. The path sorting process includes arranging the path in ascending order of path cost to form an ordered path matrix; when the path costs are the same, arranging them in descending order of coverage matching value; when the path cost and coverage matching value are the same, arranging them in ascending order of path length to generate a recommended path sequence. The course chapter tag sequence is expanded one by one based on the recommended path sequence. The corresponding knowledge point tag set is extracted according to the knowledge point coverage relationship set of the course chapter tags in the course structure data. The course chapter tag sequence and the corresponding knowledge point tag set are associated and combined one by one to form the course recommendation result. The course recommendation result is stored in a serialized structure. After the recommendation of the previous course chapter is completed, it automatically points to the recommendation of the next course chapter. Obtain user learning behavior data in the course chapters corresponding to the course recommendation results, and write the learning behavior data into the corresponding time segment position of the standardized input data according to the time stamp. Use the time index to overwrite and supplement the original data to form updated standardized input data. The learning objective vector is reconstructed based on the updated standardized input data, including course selection records, course completion status records, chapter completion ratio records, learning duration distribution records, and test result records. The value of each dimension in the learning objective vector is obtained by weighted fusion of the cumulative value of the corresponding indicator within the time segment and the historical cumulative value. The weighted fusion method is to multiply the new time segment indicator value by the time decay coefficient and add it to the historical indicator value to form the updated indicator value. The updated target stability value is formed by summing the squared differences between adjacent time segments of the updated learning target vector and taking the average. The updated target stability value is equal to the sum of the squared differences of each dimension of the learning target vector between adjacent time segments and divided by the number of dimensions. The updated target weight parameters are generated based on the updated target stability value, and the updated target weight parameters and the updated target stability value maintain an inverse mapping relationship. The learning ability vector is reconstructed based on the updated standardized input data, including chapter completion status records, chapter completion ratio records, test result records, records of knowledge points associated with incorrect questions, and records of learning time intervals. The ability value of each dimension in the learning ability vector is obtained by weighted combination of the corresponding knowledge point mastery index and completion quality index. The weighted combination method is to multiply the mastery index by the completion ratio weight and add it to the test accuracy index to form the ability value. The update capability stability value is formed by summing the absolute values ​​of the differences between adjacent time segments of the updated learning capability vector and taking the average. The update capability stability value is equal to the sum of the absolute values ​​of the differences in each dimension of the learning capability vector between adjacent time segments and divided by the number of dimensions. The update capability confidence parameter is generated based on the update capability stability value, and the update capability confidence parameter maintains an inverse mapping relationship with the update capability stability value.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a data analysis learning scenario on an online course platform. The platform includes knowledge point coverage and prerequisite relationships in course structure data, with 620 knowledge point tags, 186 course tags, 1940 course chapter tags, and 12500 user tags. Learning behavior data covers course chapter click records, course chapter start records, course chapter completion status records, chapter completion percentage records, learning time distribution records, and test result records. The platform's existing recommendation rules focus on historical clicks and popularity ranking. For users with frequently changing learning goals and significant fluctuations in learning ability, the recommendation results are prone to issues such as excessively large course chapter spans, unmet prerequisite relationships, and broken learning paths. This manifests as a high rate of course chapter interruptions after start, large fluctuations in test results, and courses remaining incomplete for extended periods. The platform needs a personalized online course recommendation method that can simultaneously characterize learning goal vectors and learning ability vectors and form a set of transitional course paths.

[0032] During platform operation, learning behavior data and course structure data undergo data standardization to generate standardized input data, with a time granularity of 7 days forming continuous time segments. Further, a learning target vector is constructed based on the standardized input data, and target stability calculation is performed. The target weight parameter is obtained by inverse mapping from the target stability value, and the target weight parameter value range is set from 0.20 to 0.90 to limit extreme weights. Simultaneously, a learning ability vector is constructed based on the standardized input data, and ability stability calculation is performed. The ability credibility parameter is obtained by inverse mapping from the ability stability value, and the ability credibility parameter value range is set from 0.30 to 0.95 to avoid an excessively narrow low-credibility interval. Further still, an ability difference state vector is constructed based on the target constraints and ability constraints. The overall deviation intensity is used to generate gap interval labels, and the difference threshold is set to 0.18 to extract the set of gap knowledge points. The broken link set is obtained by verifying the consistency of the prior relationship of knowledge points, and the contradiction label set includes gap labels and broken link labels. Subsequently, in the recommendation generation stage, the ability difference state vector is used as the diffusion generation object input to the improved Conditional Diffusion. The model uses a diffusion time step of 50. The random noise vector follows a normal distribution with a mean of 0 and a standard deviation of 1, and has the same dimension as the capability difference state vector. The noise prediction module introduces a capability difference evolution trajectory generation mechanism driven by the diffusion time step, outputting a candidate set of capability difference evolution trajectories. The candidate set of capability difference evolution trajectories is entered into a feasibility mapping. The coverage matching value is used to eliminate paths with insufficient coverage. The coverage matching value threshold is set to 0.65. When the prior violation set is not empty, it is marked as infeasible. The paths corresponding to the feasible markers are summarized to form a transition course path set. The course recommendation results are sorted by the comprehensive score of the paths and the top 12 are output. The learning behavior data generated by the user on the course recommendation results enters the parameter update. The preset learning rate is set to 0.02, forming a rolling iteration.

[0033] For ease of comparison, the same user group on the platform was used as the control group in the three months before the platform went live, using existing popularity plus collaborative filtering recommendations, and in the three months after the platform went live, the method of this invention was used as the experimental group. The statistical results are shown in Table 1.

[0034] Table 1. Comparison of the Performance of Personalized Recommendations for Online Courses

[0035] Table 1 shows that the experimental group saw a continuous increase in the course chapter start rate, course chapter completion rate, and average chapter completion ratio, while the prerequisite violation rate and 7-day dropout rate decreased simultaneously. These trends are consistent with the constraints imposed by the target constraints and ability constraints on the ability difference state vector. The increased course chapter start rate indicates that the course recommendation results are closer to the joint constraints of the learning target vector and the learning ability vector, and users' willingness to enter the learning process is more stable. The increased course chapter completion rate and average chapter completion ratio indicate that the transitional course path set is more conducive to forming a coherent learning rhythm under the constraints of knowledge point coverage and prerequisite relationships. The increase in the average test score is accompanied by a significant decrease in the prerequisite violation rate, indicating that the screening of the prerequisite violation set during the feasibility mapping stage reduces the learning frustration caused by the broken link set. The decrease in the 7-day dropout rate is related to the update of the ability credibility parameter. The update brought about by the ability fluctuation makes the ability constraints closer to the actual ability changes, reducing short-term dropouts caused by excessively large recommended path spans.

[0036] To further analyze the conversion quality from the candidate set of poor ability evolution trajectories to the set of transitional course paths, 3,000 active users in the third month were selected, and the coverage matching value, the number of prerequisite violations, the comprehensive path score and the target offset were calculated by grouping them according to the gap interval label. The results are shown in Table 2.

[0037] Table 2 Comparison of Transitional Course Path Set Mapping Quality and Parameter Updates

[0038] In Table 2, the transition interval has the highest average coverage matching value, the lowest average number of prerequisite violations, and the highest average comprehensive path score, accounting for 66.1% of the recommended courses. This indicates that the candidate set of poor capability evolution trajectories, after feasibility mapping, mainly falls within the transition interval and forms an executable set of transitional course paths. The unreachable interval has a low average coverage matching value and a high average number of prerequisite violations, accounting for only 9.4% of the recommended courses, reflecting the screening effect of coverage consistency verification and prerequisite consistency verification on infeasible markers. The average target offset and the average capability fluctuation are lowest in the transition interval, indicating that the feedback learning target vector and the feedback learning capability vector from the recommended courses are closer to the learning target vector, and the update amplitude of the target weight parameter and capability credibility parameter is more stable. The average comprehensive path score of the directly learnable interval is lower than that of the transition interval, reflecting the role of the path repair contribution value in the ranking of course recommendation results. This prioritizes the coverage of paths with greater space for repairing the poor capability state vector, thus more directly addressing the problems of excessive span and broken paths that the platform needs to solve.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A personalized online course recommendation method based on big data analysis, characterized in that, Includes the following steps: S1. Collect learning behavior data and course structure data from online course platforms, perform data standardization processing, and generate standardized input data; S2. Construct a learning target vector based on the standardized input data, perform target stability calculation, and generate target weight parameters; and generate target constraints based on the learning target vector and target weight parameters. S3. Construct a learning capability vector based on the standardized input data and generate capability credibility parameters; And based on the learning ability vector and ability credibility parameter, ability constraints are generated; S4. Construct a capability difference state vector based on the target constraints and capability constraints, generate gap interval labels, and generate a set of contradiction labels based on the transition intervals in the gap interval labels. S5. Using the capability difference state vector as the diffusion generation object, the learning target vector, target weight parameters, capability credibility parameters, gap interval labels and contradiction label set are input into the improved Conditional Diffusion Model. The improved Conditional Diffusion Model introduces a capability difference evolution trajectory generation mechanism to generate a capability difference evolution trajectory candidate set. S6. Construct a directed graph of the curriculum structure based on the curriculum structure data, and introduce the RCSP algorithm to perform feasibility mapping on the candidate set of the ability difference evolution trajectory to generate a set of transitional curriculum paths; S7. Generate course recommendation results based on the transition course path set, and update the target weight parameter and ability credibility parameter based on the user's learning behavior data generated by the course recommendation results.

2. The method for personalized online course recommendation based on big data analysis according to claim 1, characterized in that, Specifically, S1 is: Collect learning behavior data and course structure data from online course platforms to construct the original input dataset; Perform identifier recoding on the original input dataset to generate an encoded dataset; Perform time-series alignment processing on the learning behavior data in the encoded dataset to generate a time-aligned dataset; Outlier removal and missing value completion are performed on the time-aligned dataset and the encoded course structure data to generate a clean dataset. Numerical normalization is performed on the cleaned dataset. The numerical normalization process uses a linear normalization method with the maximum value minus the minimum value as the denominator and the value to be normalized minus the minimum value as the numerator to generate standardized input data.

3. The personalized online course recommendation method based on big data analysis according to claim 1, characterized in that, Specifically, S2 is: Based on the course selection records, course completion status records, chapter completion ratio records, learning duration distribution records, and test result records in the standardized input data, a learning objective vector is constructed. Perform target stability calculation on the learning target vector. The target stability calculation includes statistically analyzing the magnitude of change of the learning target vector over consecutive time segments. The target stability value is calculated based on the change range of the learning target vector within consecutive time segments. The target stability value is obtained by summing the squared differences of the learning target vector between adjacent time segments and taking the average. Target weight parameters are generated based on the target stability value, and the target weight parameters and the target stability value have an inverse mapping relationship. Target constraints are generated based on the learning target vector and target weight parameters; The target constraints include: weighting the target achievement deviation between the learning target vector and the learning ability vector according to the target weight parameters to generate the allowable range of target ability difference; determining the evolution direction constraint of the target ability difference based on the time series change trend of the learning target vector; and limiting the upper limit of the rate of change of the target ability difference within a continuous time segment based on the target weight parameters.

4. The method for personalized online course recommendation based on big data analysis according to claim 1, characterized in that, Specifically, S3 is: Based on the chapter completion status records, chapter completion percentage records, test result records, records of knowledge points associated with incorrect questions, and records of learning time intervals in the standardized input data, a learning ability vector is constructed. Perform capability stability calculation on the learning capability vector. The capability stability calculation includes statistically analyzing the magnitude of change of the learning capability vector within a continuous time segment. The ability stability value is calculated based on the change range of the learning ability vector within consecutive time segments. The ability stability value is obtained by summing the absolute values ​​of the differences between the learning ability vectors in adjacent time segments and taking the average. Capability reliability parameters are generated based on capability stability values, and the capability reliability parameters and capability stability values ​​have an inverse mapping relationship. Generate capability constraints based on learning capability vectors and capability credibility parameters; The capability constraints include: weighting the changes in the learning capability vector over consecutive time segments according to the capability confidence parameter to generate a confidence interval for capability changes; The evolution direction of the poor ability state vector is determined by the time series change trend of the learning ability vector; the upper limit of the change amplitude of the poor ability state vector within a continuous time segment is determined by the ability reliability parameter.

5. The method for personalized online course recommendation based on big data analysis according to claim 1, characterized in that, Specifically, S4 is: The goal achievement deviation is calculated based on the learning target vector and the learning ability vector. The goal achievement deviation consists of the difference between the learning target vector and the learning ability vector in the corresponding knowledge point dimension. The deviation from the target achievement is weighted by the target weight parameters to generate a weighted target achievement deviation. A capability difference state vector is constructed based on the weighted target achievement deviation and capability credibility parameter. The capability difference state vector consists of capability differences in the knowledge point dimension, and the capability difference is obtained by the product of the weighted target achievement deviation and the capability credibility parameter. Perform gap interval partitioning on the capability difference state vector. The gap interval partitioning includes calculating the overall deviation intensity of the capability difference state vector. The overall deviation intensity is obtained by summing the squares of the capability differences and taking the square root. Based on the overall deviation intensity and the allowable range of target capability difference, gap interval labels are generated. The gap interval labels include directly learnable intervals, transition intervals, and unreachable intervals. Based on the capability difference state vector labeled as the gap interval, a set of gap knowledge points is extracted. The set of gap knowledge points consists of knowledge point labels whose capability difference is greater than a preset difference threshold. Based on the prerequisite relationships of knowledge points in the course structure data, a prerequisite consistency check is performed on the set of missing knowledge points to generate a broken chain set, which consists of knowledge point tags that do not satisfy the prerequisite relationships of knowledge points. A set of contradictory labels is generated based on the set of gap knowledge points and the set of broken links. The set of contradictory labels includes gap labels and broken link labels.

6. The personalized online course recommendation method based on big data analysis according to claim 1, characterized in that, The improved Conditional Diffusion Model includes a conditional coding module, a time-step coding module, a noise prediction module, and a trajectory constraint module. The conditional encoding module uses the learning target vector, target weight parameters, and capability credibility parameters as continuous feature vectors, maps the gap interval labels and contradictory label sets as discrete feature vectors, performs dimension alignment processing on the continuous feature vectors and discrete feature vectors, and concatenates them along the feature dimensions to generate a conditional embedding vector. The diffusion time step is input to the time step encoding module, which performs position encoding mapping on the diffusion time step and outputs the time step embedding vector. The capability difference state vector, conditional embedding vector, and time step embedding vector are input into the noise prediction module. The noise prediction module outputs a noise prediction vector with the same dimension as the capability difference state vector. The noise prediction module introduces a capability difference evolution trajectory generation mechanism driven by the diffusion time step. It generates the capability difference evolution trajectory by denoising and updating the capability difference state vector multiple times in consecutive diffusion time steps and retaining the time step update results. The capability difference evolution trajectory generation mechanism includes: in each diffusion time step, performing a denoising update on the capability difference state vector based on the noise prediction vector, the denoising update is obtained by subtracting the noise prediction vector from the capability difference state vector to obtain the updated capability difference state vector; arranging the updated capability difference state vectors obtained in consecutive diffusion time steps in chronological order to generate the capability difference evolution trajectory. By initializing with random noise multiple times and repeatedly executing the diffusion time step, multiple capability difference evolution trajectories are generated, forming a candidate set of capability difference evolution trajectories; The trajectory constraint module performs trajectory consistency constraint calculation on the candidate set of trajectory evolution with capability difference. The trajectory consistency constraint calculation includes performing L2 calculation on the difference vector of the updated capability difference state vector of adjacent time steps and taking the average to obtain the trajectory consistency loss value. The joint loss value is generated by weighting and summing the trajectory consistency loss value and the mean square error loss value between the noise prediction vector and the noise. The parameters of the noise prediction module are then updated based on the joint loss value.

7. The method for personalized online course recommendation based on big data analysis according to claim 1, characterized in that, Specifically, S5 is: The initial state vector for diffusion is generated based on the capability difference state vector. The initial state vector for diffusion is obtained by adding the capability difference state vector and the random noise vector. The random noise vector has the same dimension as the capability difference state vector. The learning target vector, target weight parameters, and ability credibility parameters are used as continuous feature vectors. The gap interval labels and contradictory label sets are mapped to discrete feature vectors. Dimension alignment processing is performed on the continuous feature vectors and discrete feature vectors, and they are concatenated along the feature dimensions to generate conditional embedding vectors. The initial state vector of diffusion, the conditional embedding vector, and the diffusion time step are input into the improved ConditionalDiffusion Model, and the noise prediction vector corresponding to each diffusion time step is output. The initial state vector of diffusion is denoised and updated based on the noise prediction vector. The denoising update is obtained by subtracting the noise prediction vector from the initial state vector of diffusion to obtain the updated state vector. Perform iterative updates on the updated state vector and the diffusion time step, performing the iterative updates a preset number of time steps to obtain the updated state vector sequence corresponding to the diffusion time step; The updated state vector sequence is arranged in the order of diffusion time steps to generate the capability difference evolution trajectory; Repeatedly perform random noise initialization and diffusion time step iterative update to generate multiple capability difference evolution trajectories, forming a capability difference evolution trajectory candidate set; The trajectory consistency constraint calculation is performed on the candidate set of poor-capability evolution trajectories to obtain the trajectory consistency loss value; The joint loss value is generated by weighted summation of the trajectory consistency loss value and the mean square error loss value between the noise prediction vector and the random noise vector, and the parameters of the improved Conditional Diffusion Model are updated based on the joint loss value.

8. The method for personalized online course recommendation based on big data analysis according to claim 1, characterized in that, Specifically, S6 is: Based on the course structure data, extract the course chapter tag set, knowledge point tag set, knowledge point coverage relationship set, and knowledge point prerequisite relationship set; A node set is constructed based on the course chapter tag set and the knowledge point tag set, and a directed edge set is constructed based on the knowledge point prerequisite relationship set. The node set and the directed edge set are combined to construct a directed graph of the course structure. Based on the set of knowledge point prerequisite relationships, the preceding counts corresponding to each knowledge point tag are statistically analyzed, an output set is constructed, and a sorted description of knowledge point prerequisite relationships is generated. Based on the evolution trajectory of capability difference, gap knowledge points are extracted. Gap knowledge points are composed of knowledge point labels in the capability difference state vector where the capability difference value is greater than a preset difference threshold. A set of candidate course paths is constructed based on a directed graph of the course structure. The course paths in the set of candidate course paths consist of a sequence of course chapter labels that satisfy the prerequisite relationship of knowledge points. Perform knowledge point coverage intersection calculation on the course paths in the candidate course path set, and generate coverage matching values ​​based on the knowledge point coverage intersection; Perform prerequisite consistency checks on the course paths in the candidate course path set to generate a prerequisite violation set; Based on the directed graph of the course structure, the coverage matching value, and the prerequisite violation set, the RCSP algorithm is introduced to perform resource-constrained shortest path search on the candidate course path set and output the path result set; the path result set is then processed by path backtracking to generate the transition course path set.

9. The method for personalized online course recommendation based on big data analysis according to claim 8, characterized in that, The introduced RCSP algorithm performs resource-constrained shortest path search processing on the candidate course path set, outputting a path result set, specifically: A path state set is constructed based on the directed graph of the course structure and the set of candidate course paths. Each path state in the path state set corresponds to a candidate course path and records the course chapter label sequence, path cost, coverage matching value, size of the prerequisite violation set, and path length. The path state set is filtered based on the preset constraint threshold set. Path states that meet the coverage matching threshold requirement, the size of the prior violation set meets the prior violation threshold requirement, and the path length meets the path length threshold requirement are written into the expandable state set. Sort the scalable state set according to path cost and select the path state with the minimum path cost as the search starting state, and write it into the determined state set at the same time. Based on the sequence of course chapter labels corresponding to the initial search state, find the set of subsequent course chapter labels that satisfy the prerequisite relationship constraint of knowledge points in the directed graph of the course structure. The subsequent course chapter tags are appended one by one to the end of the course chapter tag sequence to form an extended course path, and the path cost, overwrite matching value, prerequisite violation set size and path length are updated synchronously to generate the extended path state; The extended path states are filtered based on a preset set of constraint thresholds, and the extended path states that meet the resource constraints are written into the extended state set; the resource constraints include coverage matching constraints, prior consistency constraints, path complexity constraints, and optimization objective constraints. Perform dominance removal on path states within the scalable state set; Repeatedly search for the starting state, find subsequent course chapter tags, generate extended path states, perform constraint verification and eliminate dominance relationships until the set of expandable states is empty, and obtain a set of path states that meet the resource constraints; sort the set of path states in ascending order according to the path cost and truncate a preset number to generate a set of path results.

10. The method for personalized online course recommendation based on big data analysis according to claim 1, characterized in that, Specifically, S7 is: Based on the transitional course path set, extract the course chapter tag sequence set, perform path sorting processing, and sort the path in ascending order by path cost to obtain the recommended path sequence; The course recommendation results are generated based on the recommendation path sequence. The course recommendation results include the course chapter tag sequence in the recommendation path sequence and the knowledge point tag set corresponding to the course chapter tag sequence. Acquire user learning behavior data based on course recommendation results and write the learning behavior data into the time segment corresponding to the standardized input data; The learning target vector is updated and the target stability value is recalculated based on the standardized input data after being written, and the target weight parameters are updated accordingly; the learning ability vector is updated and the ability stability value is recalculated based on the standardized input data after being written, and the ability reliability parameter is updated accordingly.

Images