A machine learning-based personalized learning path generation method

Abstract

The application discloses a kind of personalized learning path generation methods based on machine learning, belong to machine learning technical field, including the following steps: step 1, data acquisition and pre-processing, step 2, learner portrait construction, step 3, knowledge graph construction, step 4, learning state perception, step 5, initial learning path planning, step 6, learning path optimization;Through multi-source data fusion and dynamic portrait modeling, the core features of learners can be covered, and dynamic learner portraits can be accurately described, so that the path planning basis is more in line with reality, providing core support for personalized learning;In addition, through knowledge graph quantification and multi-modal state perception, a scientific knowledge system and a multi-dimensional state determination model are constructed, breaking down data and technology module barriers, enabling data linkage and sharing across all links, improving the personalized adaptation of learning paths and enhancing the model's cross-scenario generalization capability.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a method for generating personalized learning paths based on machine learning. Background Technology

[0002] With the rapid development of educational informatization and the widespread application of online learning platforms, massive amounts of learning behavior data, knowledge system data, and learner multidimensional characteristic data are constantly accumulating. How to extract individual learning patterns from this complex data and achieve truly personalized learning path planning has become a core challenge in learning.

[0003] The prior art patent document with authorization announcement number CN121094479A discloses "an adaptive personalized learning path generation method and system". The method includes: extracting a list of target capabilities from existing training models or job competency requirements to construct a capability map; generating a user capability status vector based on the user's historical course test results and practical task completion status using a weighted scoring mechanism; prioritizing capability nodes, selecting capability nodes to form a learning path based on the priority scores, allocating learning resources in combination with user preferences, and generating a task scheduling plan; issuing tasks according to the task scheduling plan and collecting behavioral data.

[0004] The patent document with authorization announcement number CN118467838A discloses a "personalized learning recommendation method and system based on machine learning", which includes matching the real-time acquired student learning behavior features with the corresponding table of learning behavior feature mapping performance map, determining the first reference tree mapping performance map that is currently called for the student, and determining the learning materials data recommended to the student based on the basic priority call weight of different second reference tree mapping performance maps in the first reference tree mapping performance map.

[0005] While existing technologies can achieve a certain degree of personalized recommendation of learning materials and generation of learning paths, taking into account both the structural dependence of learning content and the analysis of users' historical learning behavior, they suffer from several drawbacks. These include static learner profiles, a lack of quantitative knowledge graphs and scenario adaptation, reliance on a single indicator to perceive learning status, single-objective path planning that deviates from the principles of educational psychology, independent technical components that fail to form a closed-loop collaboration, and evaluation of learning effectiveness based solely on a single test score. This results in an inability to dynamically adjust learning paths, low personalization, and an inability to truly meet the real-time learning status and needs of learners in different learning scenarios. Summary of the Invention

[0006] This invention provides a personalized learning path generation method based on machine learning, which can solve the problems mentioned in the background.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for generating personalized learning paths based on machine learning, comprising: S1. Collect multi-source data from learners, knowledge systems, learning processes, and learning scenarios. Perform preprocessing operations such as cleaning, completion, outlier removal, normalization, and feature encoding on the multi-source data to form a standardized dataset. S2. Based on the standardized dataset, construct a multi-dimensional learner profile feature, quantify and model the features, establish a real-time dynamic update mechanism for the profile, and adjust the weights and filter the features according to different learning scenarios. S3. Based on the knowledge system data in the standardized dataset, perform multi-dimensional quantitative annotation on the knowledge points in the knowledge graph, construct a weighted directed knowledge graph that integrates explicit and implicit associations, and perform layering and pruning of the knowledge graph according to different learning scenarios; S4. Based on the standardized dataset, the learner profile constructed in S2, and the quantitative knowledge graph constructed in S3, perceive the learner's learning status and make quantitative judgments, and analyze the reasons for abnormal learning status. S5. Based on the dynamic learner profile of S2, the quantitative knowledge graph of S3, and the quantitative judgment result of learning status of S4, personalized initial learning path planning is carried out, which supports adaptation to learning scenarios and manual adjustment of paths. S6. Based on the quantitative judgment result of the learning state and the analysis of the cause of state abnormality in S4, optimize the initial learning path generated in S5, verify the path optimization effect in real time, and repeat the optimization and adjustment if the expected result is not achieved.

[0008] Furthermore, in S1, the collected learner data includes static data and dynamic data. The dynamic data includes learning behavior data, subjective feedback data, and physiological perception data. The knowledge system data includes basic information about knowledge points, the relationships between knowledge points, and the corresponding learning resources. The learning process data includes the learning time of knowledge points, changes in knowledge mastery, the effectiveness of learning resource usage, and the results of periodic tests. The learning scenario data includes scenario type, scenario learning objectives, and scenario learning characteristics.

[0009] Furthermore, in S2, the multi-dimensional learner profile features include basic features, knowledge mastery features, learning ability features, learning preference features, and learning status features; the hierarchical analysis method is used to assign contextual weights to the profile features, and the improved K-means++ clustering algorithm is used to cluster the learner profiles; the real-time dynamic update mechanism of the profile includes dynamic updates of knowledge mastery and real-time prediction updates of learning dynamic features.

[0010] Furthermore, in S3, the multi-dimensional quantitative annotation dimensions for knowledge points include difficulty, correlation strength, preconditions, learning time, assessment weight, and practical adaptability; the core features of the knowledge graph are retained for different learning scenarios, and the knowledge graph is layered and selectively trimmed according to scenario features.

[0011] Furthermore, in S4, features strongly correlated with the learning state are extracted from learning behavior, physiological perception, and subjective feedback to construct a quantitative dimension of the learning state that includes learning ability fit, learning efficiency, learning interest matching degree, and learning fatigue; and an interpretability analysis algorithm is used to analyze the core reasons for abnormal learning states.

[0012] Furthermore, in S5, when planning the personalized initial learning path, the multi-objective optimization of the learning path planning is constructed by combining the laws of educational psychology. Contextualized dynamic weights are assigned to each optimization objective. An improved deep reinforcement learning algorithm is used for the initial learning path planning. The algorithm incorporates the pre-constraints of the knowledge graph and the feature weights of the learner profile. The generated initial learning path includes the learning order of knowledge points, the matching of learning resources, the planning of the learning rhythm, and the stage test nodes.

[0013] Furthermore, in S6, when the learning state exceeds the threshold or the mastery of knowledge points does not meet expectations, path optimization is triggered. Based on the analysis results of the abnormal learning state, multiple targeted path optimization strategies are formulated, and the path planning model is updated using an online incremental learning algorithm.

[0014] The beneficial effects of this invention's machine learning-based personalized learning path generation method are as follows: By integrating multi-source data and dynamic profile modeling, it can cover the core characteristics of learners, accurately depict dynamic learner profiles, and make path planning more realistic, providing core support for personalized learning; In addition, by quantifying knowledge graphs and perceiving multimodal states, it constructs a scientific knowledge system and a multi-dimensional state judgment model, breaking down data and technology module barriers, enabling data linkage and sharing at each stage, improving the personalized adaptability of the learning path, and enhancing the model's cross-scenario generalization ability; At the same time, through multi-objective reinforcement learning and dynamic optimization technology, it provides scientific path planning and real-time adjustment methods, accurately matching learner needs, and quickly optimizing paths based on real-time learning states, greatly improving learning efficiency and effectiveness. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0016] Figure 1 This is a schematic diagram of the method flow for a personalized learning path generation method based on machine learning according to the present invention. Detailed Implementation

[0017] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 like Figure 1 As shown, a technical solution is provided: a personalized learning path generation method based on machine learning, comprising: Step 1: Data Acquisition and Preprocessing Collect multi-source data from learners, knowledge systems, learning processes, and learning scenarios. Perform preprocessing operations such as cleaning, completion, outlier removal, normalization, and feature encoding on the multi-source data to form a standardized dataset. Specifically, the collected learner data includes static and dynamic data. Dynamic data includes learning behavior data, subjective feedback data, and physiological perception data. Knowledge system data includes basic information about knowledge points, relationships between knowledge points, and corresponding learning resources. Learning process data includes the duration of knowledge point learning, changes in knowledge mastery, the effectiveness of learning resource usage, and stage test scores. Learning scenario data includes scenario type, scenario learning objectives, and scenario learning characteristics.

[0019] Firstly, regarding the learner dimension, static data such as grade level, age, and learning goals are extracted through the platform backend. Dynamic data such as answer records, video viewing trajectories, and learning difficulty scores are collected through terminal behavior records and subjective feedback entry points. Physiological perception data such as attention concentration and eye movement trajectories are collected through the terminal's non-facial recognition perception module. Secondly, regarding the knowledge system dimension, the education platform collects data on the definition of knowledge points in each subject, explicit relationships, and accompanying learning resources such as videos, exercises, and practical courses. Thirdly, regarding the learning process dimension, the platform collects process data such as the learning time of knowledge points, dynamic changes in mastery, the effectiveness of learning resource usage, and stage test scores. Fourthly, regarding the learning scenario dimension, the platform collects scenario type (e.g., K-12, vocational education, adult education), learning goals for exam preparation / practical / ability improvement, and fragmented / systematic, theoretical / practical scenario learning characteristics data. Then, a full-process preprocessing operation is carried out on the collected raw data. First, redundant and duplicate data collected from cross-platform and cross-terminal sources are accurately deduplicated by data feature matching. Then, for missing data, a combination of feature mean interpolation based on learners with the same profile and time-series prediction is used to complete missing value completion. Abnormal data such as extreme values ​​of answering speed and abnormal fluctuations in physiological perception data are identified and removed by the quartile method. Subsequently, feature data of different dimensions are uniformly mapped to the [0,1] interval through linear transformation to complete data normalization. Finally, category features such as learning preferences and scene types are transformed by one-hot encoding, and time-series features such as learning duration and mastery changes are extracted by time-series encoding to achieve standardized transformation of the entire data. Finally, the preprocessed data, after deduplication, completion, elimination, normalization, and feature encoding, will be structured, integrated, and classified according to learners, knowledge systems, learning processes, and learning scenarios to form a standardized dataset with complete features, consistent data, and continuous temporal sequence. At the same time, feature indexes and data traceability labels will be established for the dataset to ensure that it can directly provide high-quality standardized input data for subsequent dynamic construction of learner profiles, quantitative modeling of knowledge graphs, and multimodal perception of learning states.

[0020] Step 2: Learner Profile Construction Based on the standardized dataset, a multi-dimensional learner profile feature is constructed, the feature is quantified and modeled, a real-time dynamic update mechanism for the profile is established, and the profile features are weighted and filtered according to different learning scenarios. Specifically, the multi-dimensional learner profile features include basic features, knowledge mastery features, learning ability features, learning preference features, and learning status features; the hierarchical analysis method is used to assign contextual weights to the profile features, and the improved K-means++ clustering algorithm is used to cluster the learner profiles; the real-time dynamic update mechanism of the profile includes dynamic updates of knowledge mastery and real-time prediction updates of learning dynamic features.

[0021] First, full data corresponding to the feature dimensions are extracted from the standardized dataset to build a learner profile feature system that includes basic features, knowledge mastery features, learning ability features, learning preference features, and learning status features. Each feature dimension is quantified in the range of 0-1. The basic feature quantifies the learning style as visual / auditory / practical and encodes the features. The knowledge mastery feature quantifies the mastery of each knowledge point based on test scores and answer accuracy and calculates the knowledge forgetting rate. The learning ability feature quantifies learning efficiency and comprehension based on the mastery of knowledge points per unit time and problem-solving speed. The learning preference feature labels the resource type and learning rhythm preference based on the frequency of use of learning resources. The learning status feature combines physiological perception data and subjective feedback to quantify real-time fatigue and concentration. This forms a complete profile feature set with standardized values. Then, the Analytic Hierarchy Process (AHP) is used in conjunction with the experience of education experts. By constructing a feature judgment matrix, calculating feature weights, and performing consistency checks, contextual weights are assigned to each feature in the feature system. Then, an improved K-means++ clustering algorithm is used to perform cluster analysis on the quantified profile feature set. The temporal change of knowledge mastery is used as a hard constraint condition for clustering to avoid the bias problem of static feature clustering. Learners with similar features are grouped into the same profile category. At the same time, based on the clustering results and quantified features, an initial learner profile model is constructed to achieve accurate modeling and classification of individual learner characteristics. Finally, a two-layer real-time dynamic update mechanism for the profile is established. Based on the Ebbinghaus forgetting curve, a knowledge mastery update model is constructed, and the knowledge mastery and forgetting rate at time t are calculated and updated in real time using the following formula: In the formula, This represents the initial level of mastery of the knowledge points. For time, For personalized forgetting coefficient, Let be a natural constant, where the individualized forgetting coefficient is . The learner's memory ability is quantified, and then the Online Time Series Learning (OLSTM) algorithm is used to predict and update dynamic features such as learning efficiency, fatigue, and interest in real time. The profile features are updated within 5 seconds after new learning process data is collected. At the same time, the profile features are dynamically weighted and core features are selected for different learning scenarios such as K-12, vocational education, and adult continuing education. The updated dynamic learner profile is synchronized in real time to subsequent modules such as knowledge graph construction and learning path planning, providing them with accurate personalized feature input.

[0022] Step 3: Knowledge Graph Construction Based on the knowledge system data in the standardized dataset, the knowledge points in the knowledge graph are quantitatively labeled in multiple dimensions, and a weighted directed knowledge graph that integrates explicit and implicit associations is constructed. The knowledge graph is then layered and pruned according to different learning scenarios. Specifically, the knowledge points are quantitatively labeled in multiple dimensions, including difficulty, correlation strength, preconditions, learning time, assessment weight, and practical adaptability. The core features of the knowledge graph are retained for different learning scenarios, and the knowledge graph is layered and tailored according to the scenario features.

[0023] First, basic information, relationships, and full data of supporting learning resources for each subject knowledge point are extracted from the knowledge system data of the standardized dataset. For each knowledge point in the knowledge graph, quantitative annotations are carried out on the difficulty coefficient, association strength, preconditions, learning time, assessment weight, and practical adaptability. Educational experts first complete the initial quantitative value annotations in combination with subject teaching standards and industry practical requirements. Then, the quantitative values ​​are optimized and corrected by machine learning algorithms on the historical learning data of the education platform. The quantitative feature values ​​of all dimensions are uniformly mapped to the [0,1] interval through linear transformation. Among them, the learning time dimension is dynamically adjusted in combination with the dynamic learner profile features constructed in step 2 to match the learning ability and pace of different learners. Then, a weighted directed knowledge graph integrating explicit and implicit associations is constructed. First, based on the arrangement logic of subject textbooks and the domain experience of education experts, the explicit associations between knowledge points (precedence / consequence, parallel, and cross-disciplinary) are identified. Directed edges are constructed for each explicit association, and the quantified association strength value is used as the weight of the edge. Then, a graph attention neural network (GAT) is used to deeply mine the massive learning process data of the education platform to identify the implicit associations between knowledge points (cross-disciplinary and cross-module). Directed edges are constructed for each implicit association, and the quantified association strength weight is calculated and assigned through an algorithm. Finally, the graphs are integrated to form a weighted directed knowledge graph. In the formula, For knowledge point nodes with quantified features, For explicit and implicit directed edges with associated strength weights, To adapt the feature set to the specific context; Finally, for different learning scenarios (such as K-12 basic education, vocational skills education, and adult continuing education), the quantitative knowledge graph is modeled in a scenario-based hierarchical manner and its features are tailored accordingly. For example, in the K-12 scenario, the core features of difficulty coefficient, pre-constraints, and assessment weights are retained and layered by subject and grade. In the vocational education scenario, the core features of practical adaptability, correlation strength, and knowledge point application ability are retained and layered by vocational skills module. In the adult education scenario, the core features of learning time, assessment weights, and fragmentation adaptability are retained and layered by learning objectives. At the same time, a real-time iterative optimization mechanism for the knowledge graph is established, which feeds back the results of subsequent learning effect evaluation to the knowledge graph. The incremental learning algorithm of graph neural network is used to update the quantitative features and correlation strength weights of knowledge points in real time. The optimized scenario-based quantitative knowledge graph is synchronized to subsequent learning state perception and path planning, providing them with accurate knowledge system support.

[0024] Step 4: Learning State Awareness Based on the standardized dataset, the learner profile constructed in step 2, and the quantitative knowledge graph constructed in step 3, the learner's learning status is perceived and quantitatively determined, and the reasons for abnormal learning status are analyzed. Specifically, features strongly correlated with learning status are extracted from learning behavior, physiological perception, and subjective feedback to construct a quantitative dimension of learning status that includes learning ability fit, learning efficiency, learning interest matching degree, and learning fatigue; and interpretable analysis algorithms are used to analyze the core reasons for abnormal learning status.

[0025] First, by combining the real-time learning process data of the standardized dataset, the dynamic learner profile features of step 2, and the quantitative knowledge graph features of step 3, core features strongly correlated with the learning state are extracted from three modalities: learning behavior, physiological perception, and subjective feedback. Among them, the learning behavior features include answer accuracy, answer speed, video viewing pause frequency, and knowledge point learning time, the physiological perception features include attention concentration, eye movement trajectory changes, and frequency of terminal operation, and the subjective feedback features include learning difficulty score, interest score, and real-time fatigue self-assessment. The extracted core features are standardized and normalized to ensure that the feature values ​​are all mapped to the [0,1] interval, forming a feature input set for learning state perception. Then, using core features as input and learning ability fit, learning efficiency, learning interest matching degree, and learning fatigue quantification dimensions as output, a learning state quantification model is constructed using a multi-feature fusion random forest (RF) algorithm. The model is trained and optimized using historical learning data from the education platform. Cross-validation is used to ensure the accuracy of the model's state quantification. The trained model is deployed on the edge computing node of the learning terminal to achieve real-time inference. Within 3 seconds of collecting new learning data, the quantification values ​​of the four dimensions are output, and state thresholds are set for each dimension (e.g., learning fatigue > 0.8, learning ability fit < 0.4, and learning efficiency < 0.3 are abnormal thresholds). When the quantification value of any dimension exceeds the threshold, the system automatically marks the learning state as abnormal and triggers the subsequent analysis process. Finally, the SHAP value interpretability analysis algorithm is used to analyze the causes of learning states marked as abnormal. The outliers of the four quantitative dimensions are used as the analysis targets to calculate the contribution of each core feature to the abnormal results and accurately locate the core causes of the abnormal state. For example, low learning ability fit is due to the difficulty coefficient of the knowledge point exceeding the learner's current ability range rather than the learner's own lack of ability, and low learning efficiency is due to excessive learning fatigue rather than the mismatch between learning content and resources. At the same time, the quantitative judgment results of the learning state, the abnormal marking information and the conclusions of the cause analysis are stored in a structured manner and synchronized to the subsequent initial learning path planning and path dynamic optimization, providing real-time and interpretable state basis for the scientific planning and precise adjustment of the path.

[0026] Step 5: Initial Learning Path Planning Based on the dynamic learner profile in step 2, the quantitative knowledge graph in step 3, and the quantitative judgment result of the learning status in step 4, a personalized initial learning path is planned, which supports adaptation to the learning scenario and manual adjustment of the path. Specifically, when planning a personalized initial learning path, the algorithm combines the principles of educational psychology to construct a multi-objective optimization of the learning path planning. It assigns scenario-based dynamic weights to each optimization objective and uses an improved deep reinforcement learning algorithm for initial learning path planning. The algorithm incorporates the pre-constraints of the knowledge graph and the feature weights of the learner profile. The generated initial learning path includes the learning order of knowledge points, the matching of learning resources, the planning of the learning pace, and the stage test nodes.

[0027] First, a Markov decision process is constructed by integrating core principles of educational psychology, as shown below: The core features of the dynamic learner profile in step 2, the node / edge features of the quantitative knowledge graph in step 3, and the quantitative values ​​of the learning state in step 4 are fused and integrated to construct a 128-dimensional state space S, which comprehensively describes the current state of the learner and the knowledge system. At the same time, an action space A covering the entire content of learning path planning is defined, including three core actions: knowledge point learning order selection, learning resource type adaptation (video / exercise / practice / document), and learning rhythm planning (single knowledge point learning duration, learning rest interval). The action dimension can be dynamically expanded according to the learning scenario. The state transition probability P is calculated and determined by the knowledge mastery update model in step 2 and the knowledge graph pre-constraints in step 3, which accurately describes the transition law of the learner's state and knowledge mastery state after performing a certain learning action. Then, four core optimization objectives are constructed: maximizing knowledge mastery, minimizing learning time, maximizing the matching degree of learning interest, and maximizing the adaptability of learning difficulty. The Analytic Hierarchy Process (AHP) is used to assign scenario-based dynamic weights to each objective, combining the experience of educational experts and the characteristics of the scenario. , , , Design a comprehensive reward function: In the formula, This is the fatigue penalty coefficient. To reduce learning fatigue, , , , The objectives are to maximize knowledge mastery, minimize learning time, maximize learning interest matching, and maximize learning difficulty suitability. The Deep Deterministic Policy Gradient (DDPG) algorithm is adopted as the core algorithm for path planning and is specifically improved. The pre-constraints of the knowledge graph in step 3 are integrated into the Actor network as hard constraints for action selection to avoid generating paths that violate the learning order of knowledge points. The feature weights of the learner profile in step 2 are integrated into the Critic network to improve the personalized accuracy of the network's evaluation of the value of learning paths. The improved DDPG algorithm is trained and optimized throughout the entire process using a standardized dataset to ensure that the model has accurate and efficient personalized learning path planning capabilities. Finally, the dynamic learner profile from step 2, the quantified knowledge graph from step 3, and the quantified learning status judgment results from step 4 are used as inputs to the trained improved DDPG model. This process calculates and generates a personalized initial learning path. The path explicitly includes the optimal learning order of knowledge points, the type of learning resources suitable for each knowledge point, the learning rhythm plan with corresponding learning time and rest intervals for each knowledge point, and the specific time nodes for phased tests. Furthermore, the generated initial learning path is adapted to different learning scenarios. For example, in vocational education scenarios, the proportion of practical course resources is automatically increased, while in adult continuing education scenarios, the proportion of fragmented learning resources is automatically increased. The path also retains an interface for manual adjustment by education experts / teachers, allowing relevant personnel to make minor, refined adjustments to the learning order and single-knowledge-point learning time based on actual teaching plans and class learning needs. This integrates the scientific nature of algorithm planning with the practicality of actual teaching, resulting in a final, implementable personalized initial learning path.

[0028] Step 6: Learning Path Optimization Based on the quantitative judgment results of the learning state and the analysis of the causes of state anomalies in S4, the initial learning path generated by S5 is optimized, the path optimization effect is verified in real time, and if the expected results are not achieved, the optimization and adjustment are repeated. Specifically, when the learning status exceeds the threshold or the mastery of knowledge points does not meet expectations, path optimization is triggered. Based on the analysis results of the abnormal learning status, multiple targeted path optimization strategies are formulated, and the path planning model is updated using an online incremental learning algorithm.

[0029] First, the system monitors the quantitative dimension values ​​of the learning status output in step 4 and the learner's actual mastery of the knowledge points. When any quantitative value of the learning status exceeds a preset threshold, or when the learner's mastery of the knowledge points is lower than the preset expected value, the system triggers the learning path optimization process: it retrieves the conclusions of the abnormal learning status analysis completed in step 4, combines the dynamic learner profile features from step 2 with the quantitative features of the quantitative knowledge graph from step 3, and formulates and executes targeted path optimization strategies. The execution parameters of each strategy are deeply bound to the profile and knowledge graph. Specifically, during difficulty adaptation optimization, the system reduces the number of subsequent knowledge points. The difficulty level is adjusted by adding a review session for prerequisite basic knowledge points. When optimizing the learning pace, the learning rest interval is increased, the learning time for a single knowledge point is shortened, and a "small amount, multiple times" learning mode is adopted. When optimizing resource adaptation, the preferred learning resource types marked in the learner profile are replaced. When optimizing the sequence, the current knowledge point is skipped, and the learner returns to review the prerequisite knowledge points that have not been mastered before continuing to learn. When optimizing the duration, inefficient learning links are eliminated and targeted practice exercises are added. When optimizing interest enhancement, extended content related to the current knowledge point and marked as interesting in the learner profile is inserted to ensure the personalization and accuracy of the optimization strategy. Finally, the DDPG algorithm for online incremental learning is adopted. The abnormal feedback data of the current learning state and the process data of the optimization strategy execution are used as incremental training data to fine-tune the path planning model trained in step 5, update the Actor / Critic network parameters of the model, and then collect the learning process data of the learner in the optimized learning path in real time. The learning state is re-detected and the knowledge mastery is calculated by the learning state quantification model in step 4 to complete the real-time verification of the optimization effect. If the verification result is that the learning state has recovered to the threshold range and the knowledge mastery reaches the preset expectation, the optimization is confirmed to be effective and the optimized learning path is synchronized to the learner's learning terminal. If the expectation is not met, the cause analysis is carried out again based on the newly collected learning state data, and the appropriate optimization strategy is formulated and executed again. The above optimization and verification process is repeated until the learner's learning state and knowledge mastery both reach the preset standard.

[0030] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for generating personalized learning paths based on machine learning, characterized in that: S1. Collect multi-source data from learners, knowledge systems, learning processes, and learning scenarios. Perform preprocessing operations such as cleaning, completion, outlier removal, normalization, and feature encoding on the multi-source data to form a standardized dataset. S2. Based on the standardized dataset, construct a multi-dimensional learner profile feature, quantify and model the features, establish a real-time dynamic update mechanism for the profile, and adjust the weights and filter the features according to different learning scenarios. S3. Based on the knowledge system data in the standardized dataset, perform multi-dimensional quantitative annotation on the knowledge points in the knowledge graph, construct a weighted directed knowledge graph that integrates explicit and implicit associations, and perform layering and pruning of the knowledge graph according to different learning scenarios; S4. Based on the standardized dataset, the learner profile constructed in S2, and the quantitative knowledge graph constructed in S3, perceive the learner's learning status and make quantitative judgments, and analyze the reasons for abnormal learning status. S5. Based on the dynamic learner profile of S2, the quantitative knowledge graph of S3, and the quantitative judgment result of learning status of S4, personalized initial learning path planning is carried out, which supports adaptation to learning scenarios and manual adjustment of paths. S6. Based on the quantitative judgment result of the learning state and the analysis of the cause of state abnormality in S4, optimize the initial learning path generated in S5, verify the path optimization effect in real time, and repeat the optimization and adjustment if the expected result is not achieved.

2. The method for generating personalized learning paths based on machine learning according to claim 1, characterized in that: In S1, the collected learner data includes static data and dynamic data. Dynamic data includes learning behavior data, subjective feedback data, and physiological perception data. Knowledge system data includes basic information about knowledge points, relationships between knowledge points, and corresponding learning resources. Learning process data includes the learning time of knowledge points, changes in knowledge mastery, the effectiveness of learning resource usage, and stage test scores. Learning scenario data includes scenario type, scenario learning objectives, and scenario learning characteristics.

3. The method for generating personalized learning paths based on machine learning according to claim 1, characterized in that: In S2, the multi-dimensional learner profile features include basic features, knowledge mastery features, learning ability features, learning preference features, and learning status features; The hierarchical analysis method is used to assign contextual weights to profile features, and an improved K-means++ clustering algorithm is used to cluster learner profiles. The real-time dynamic update mechanism of the profiles includes dynamic updates of knowledge mastery and real-time prediction updates of learning dynamic features.

4. The method for generating personalized learning paths based on machine learning according to claim 1, characterized in that: In S3, the multi-dimensional quantitative annotation dimensions for knowledge points include difficulty, correlation strength, preconditions, learning time, assessment weight, and practical adaptability. The core features of the knowledge graph are retained for different learning scenarios, and the knowledge graph is layered and selectively trimmed according to the scenario features.

5. The method for generating personalized learning paths based on machine learning according to claim 1, characterized in that: In step S4, features strongly correlated with the learning state are extracted from learning behavior, physiological perception, and subjective feedback to construct a quantitative dimension of the learning state that includes learning ability fit, learning efficiency, learning interest matching degree, and learning fatigue. An interpretable analysis algorithm is used to analyze the core reasons for abnormal learning states.

6. The method for generating personalized learning paths based on machine learning according to claim 1, characterized in that: In S5, when planning a personalized initial learning path, the learning path planning is constructed by combining the laws of educational psychology. Contextual dynamic weights are assigned to each optimization objective. An improved deep reinforcement learning algorithm is used for the initial learning path planning. The algorithm incorporates the pre-constraints of the knowledge graph and the feature weights of the learner profile. The generated initial learning path includes the learning order of knowledge points, the matching of learning resources, the planning of the learning rhythm, and the stage test nodes.

7. The method for generating personalized learning paths based on machine learning according to claim 1, characterized in that: In step S6, when the learning state exceeds the threshold or the mastery of knowledge points does not meet expectations, path optimization is triggered. Based on the analysis results of the abnormal learning state, multiple targeted path optimization strategies are formulated, and the path planning model is updated using an online incremental learning algorithm.

Images