Self-adaptive personalized learning path generation method and system

Abstract

The invention provides a self-adaptive personalized learning path generation method and system, and the method comprises the steps: extracting a target capability list from an existing training model or post competency requirements, and constructing a capability map; based on a user historical course test result and a practical operation task completion condition, generating a user capability state vector by adopting a weighted scoring mechanism; performing priority scoring on the capability nodes, screening the capability nodes according to the priority score to form a learning path, distributing learning resources in combination with user preference, and generating a task scheduling plan; issuing a task according to the task scheduling plan, and collecting behavior data; calculating a behavior effectiveness score, identifying an abnormal task of which the behavior effectiveness score is lower than a threshold value, and adding a behavior deviation label; and based on the behavior deviation label matching correction strategy of the abnormal task, adjusting a path and a scheduling plan, and outputting a corrected learning path and a scheduling table.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of online education, and particularly relates to a self-adaptive personalized learning path generation method and system. BACKGROUND

[0002] With the deepening of enterprise digital transformation and the increase in the number of knowledge-based positions, how to quickly improve the position competence and ability level of employees has become the core goal of enterprise training and education system. Traditional learning systems often focus on providing content recommendations or course lists. Although this approach can provide users with certain learning resources, it lacks path planning capabilities and cannot reflect the structural dependency between learning content, making it difficult to ensure progressive and goal-oriented learning. In practice, users often face problems such as mismatch between recommended content and position goals, fragmented learning process, and uneven learning burden. For example, although enterprise training platforms can record course completion, subsequent learning recommended by the system is often based on static matching of a single tag, without considering the user's actual mastery of certain ability points and the logical order between knowledge points. This fragmented recommendation mode not only leads to low learning efficiency, but also easily causes deviation in the achievement of ability goals. In recent years, some systems have attempted to use knowledge graphs or simple machine learning models to improve path recommendations, but these methods still have many shortcomings. First, learning paths lack dynamism and are often difficult to update based on actual user performance after initial generation. Second, there is a lack of effective monitoring at the task execution level, which cannot determine whether the user's learning behavior is truly effective, but only uses course completion rate or click behavior as a standard. Third, path correction methods are too rough, often only allowing for overall rearrangement or regeneration, making it difficult to make targeted local adjustments while maintaining path stability. These shortcomings make it difficult for existing technologies to meet the needs for learning effectiveness, path rationality, and execution controllability in high-demand scenarios such as enterprise training. SUMMARY

[0003] The present application aims to design a self-adaptive personalized learning path generation method and system that can achieve a complete link from ability target modeling to path generation, from task execution monitoring to path correction, significantly improving the rationality, execution effectiveness, and dynamic adaptability of learning paths.

[0004] To achieve the above-mentioned purpose, in a first aspect of the present application, a self-adaptive personalized learning path generation method is provided, which comprises: extract a target ability list from an existing training model or job competency requirement, construct an ability graph containing a set of ability nodes and a set of edges according to the sequential dependency relationship between abilities; collect the description text corresponding to each ability node, input the processed text into a fully connected neural network after word segmentation and word embedding, and generate a node semantic vector; based on the user's historical course test results and practical task completion, a weighted scoring mechanism is used to generate a user ability state vector, and missing data is filled with the average of all users; The ability nodes are prioritized based on a structure-aware ranking function, which takes into account the difference in user mastery of the nodes, the structural dependence reflected by the in-degree and out-degree of the nodes in the graph, and the complexity of the node semantic vector; based on the priority score, the ability nodes are filtered to form a learning path, which is checked for topological validity, and learning resources are allocated based on user preferences, and a task scheduling plan is generated based on the user's daily available learning time model; According to the task scheduling plan, tasks are issued, and four types of behavior data are collected: task dwell time, content integrity flag, evaluation score, and interruption frequency; a behavior effectiveness function with an anomaly detection regular term is used to calculate the behavior effectiveness score, and abnormal tasks with a behavior effectiveness score below a threshold are identified and labeled as behavior deviation; Based on the behavior deviation label of the abnormal task, a correction strategy is matched, a path disturbance cost function is introduced to control the correction size, and the path and scheduling plan are adjusted under the constraint of total disturbance budget, and the corrected learning path and scheduling table are output.

[0005] Further, the sequential dependency relationship between abilities is determined by combining the hierarchical relationship annotated by domain experts and the sequence pattern mining results of historical user learning paths, where the sequence pattern mining uses a sliding window algorithm to extract high-frequency continuous learning node combinations.

[0006] Further, the node semantic vector is obtained by processing the description text through a pre-trained 300-dimensional word embedding model to obtain a word vector matrix, and then outputting through a two-layer fully connected neural network with ReLU activation function, with a vector dimension of 64.

[0007] Further, the specific method steps of the weighted scoring mechanism include: dividing the user ability evaluation dimensions into theoretical learning performance and practical task completion, and pre-setting the weights of each dimension according to the ability type; extracting the user's test scores and answer accuracy data in theoretical learning, as well as the completion quality scores and step compliance data in practical tasks, and mapping the two types of data to obtain normalized scores in the theoretical and practical dimensions respectively; multiply the normalized score in the theoretical dimension by the corresponding weight, add the result after multiplying the normalized score in the practical dimension by the corresponding weight, and obtain the comprehensive score of the ability node.

[0008] Further, the specific steps of the structure-aware ordering function include: calculating a mastery difference score of the user for the node, which is the difference between the score of the corresponding node in the current ability state vector and the target mastery threshold; calculating a structure dependence score, which is the weighted sum of the in-degree value and the out-degree value of the node in the ability graph, with the in-degree value being given a higher weight by a preset adjustment factor; calculating a complexity score of the node semantic vector, which is the information entropy value calculated based on the dispersion degree of the vector elements; and fusing the mastery difference score, the structure dependence score and the complexity score by a preset proportion to generate a priority score of each node.

[0009] Further, the specific steps of allocating learning resources based on user preferences include: constructing a user preference vector based on the selection frequency and completion quality of different resource types in the user's historical learning records, with the vector elements corresponding to the preference weights of each type of resource; extracting the resource demand label of the current ability node and matching the resource type with the highest weight in the preference vector; if the number of the matched resource type in the resource library is less than a preset threshold, starting a supplementary process: first checking the content keyword matching degree of the graphic-text resource with the current node, and selecting the type when the matching degree exceeds the threshold; if the graphic-text resource does not meet the requirements, checking the knowledge point coverage integrity of the PPT resource and the scene adaptation degree of the mixed case resource in turn until a resource type that meets the requirements is selected; and recording the resource allocation result for updating the weight of the user preference vector.

[0010] Further, the abnormality detection regular term is based on the group historical average evaluation score of the current ability node, and the deviation of the user score from the group average is regulated by a penalty coefficient.

[0011] Further, the path disturbance cost function calculates the disturbance cost by integrating the position offset before and after task correction, the time length variation ratio and the depth of the node in the graph.

[0012] Further, the behavior deviation label is obtained based on the following steps: When the evaluation score is lower than a preset evaluation score lower limit, the stay time is not less than a preset stay time lower limit, and the content is complete, the behavior deviation label is the ability deficiency type, and the correction strategy is intensive training; When the stay time is lower than a preset stay time lower limit, the number of interruptions exceeds a preset number of interruptions upper limit, and the evaluation score is not lower than a preset evaluation score lower limit, the behavior deviation label is the attention dispersion type, and the correction strategy is rhythm adjustment; If the content is incomplete and the stay time is lower than a preset stay time lower limit, the behavior deviation label is the incomplete execution type, and the correction strategy is process simplification; If the evaluation score is lower than the preset evaluation score lower limit and the number of interruptions exceeds the preset upper limit of the number of interruptions, the behavior deviation label is comprehensive disorder type, and the correction strategy is combined intervention.

[0013] In a second aspect of the application, a self-adaptive personalized learning path generation system is provided, comprising: A state construction module is configured to extract a target ability list from an existing training model or job competency requirement, construct an ability graph containing an ability node set and an edge set according to the sequential dependency relationship between abilities, collect the description text corresponding to each ability node, input the processed text into a full connection neural network after word segmentation and word embedding, and generate a node semantic vector; based on the historical course test results and practical operation task completion of the user, a user ability state vector is generated using a weighted scoring mechanism, and missing data is filled with the average value of all users; A path generation module is configured to score the ability nodes based on a structure-aware ranking function, which comprehensively considers the difference in the user's mastery of the nodes, the structural dependency reflected by the in-degree and out-degree of the nodes in the graph, and the complexity of the node semantic vector; the learning path is formed by screening the ability nodes according to the priority score, and after topological legality check, the learning resources are allocated in combination with the user preferences, and the task scheduling plan is generated according to the user's daily available learning time model; A behavior evaluation module is configured to issue tasks according to the task scheduling plan, collect four types of behavior data, including task stay time, content integrity flag, evaluation score, and number of interruptions; calculate the behavior effectiveness score using a behavior effectiveness function containing an abnormality detection regular term, identify abnormal tasks with a behavior effectiveness score lower than a threshold, and add a behavior deviation label; A path deployment module is configured to match the correction strategy based on the behavior deviation label of the abnormal task, introduce a path disturbance cost function to control the correction scale, adjust the path and scheduling plan under the constraint of total disturbance budget, and output the corrected learning path and scheduling table.

[0014] The beneficial technical effects of the present application are at least the following: To solve the above problems, the application provides a kind of self-adapting personalized learning path generation method and system, by constructing ability atlas and user ability state vector, learning goal is converted into directed graph structure, and the semantic representation of node is formed by combining the historical learning behavior and task performance of user, to ensure the goal-oriented nature and structural integrity of path generation.On this basis, the system introduces the ranking function of graph structure perception, considers the user ability difference, the structural position of node in atlas and its semantic complexity, and gives priority to the priority score of task node in path, generates the learning path sequence that meets the logical dependence and individual demand, and generates task scheduling plan in combination with user preference and learning time model.The system uses front-end burying point technology to collect the residence time, completion, test results and behavior stability of task in the execution phase, and through the design of behavior effectiveness function with abnormal detection regular term, it can ensure that false learning conditions such as "hang up completion", "abnormally high score" or "unstable learning" can be identified.In the path correction phase, the system selects replacement, insertion, splitting or postponement and other correction methods according to the abnormal label of task, and introduces path disturbance cost function to control the correction scale, to realize fine adjustment of local task while maintaining the overall stability of path.Through the above design, the application realizes the whole-process closed-loop management of learning path, that is, the complete link from ability target modeling to path generation, from task execution monitoring to path correction, which significantly improves the rationality, execution effectiveness and dynamic adaptive ability of learning path, and is especially suitable for enterprise training and high-demand capacity training scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0015] The application is further described by means of the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the following drawings.

[0016] Figure 1 A flow chart of the self-adapting personalized learning path generation method of the application.

[0017] Figure 2 A system framework diagram of the self-adapting personalized learning path generation system of the application. DETAILED DESCRIPTION

[0018] The embodiments of the application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are only used to explain the application, and cannot be understood as a limitation on the application.

[0019] In one or more embodiments, as Figure 1As shown, a self-adaptive personalized learning path generation method is disclosed, which comprises the following steps: S1: Extract the target ability list from the existing training model or job competence requirement, construct the ability graph containing the ability node set and the edge set according to the dependence relationship between the abilities; collect the description text corresponding to each ability node, input the full connection neural network after word segmentation and word embedding processing, and generate the node semantic vector; based on the user's historical course test results and practical operation task completion, a weighted scoring mechanism is used to generate a user ability state vector, and the missing data is filled with the average value of all users; Specifically, this step aims to provide structured input for path planning, including two core components: one is the ability graph structure representing the learning goal, and the other is the ability state vector reflecting the user's current ability mastery. This step is the basis of the entire path system, and its construction results will be used as direct input for subsequent path generation and execution scheduling, so it needs to have complete structural expression, accuracy and actual operability.

[0020] Firstly, the system extracts the target ability list from the existing training model or job competence requirement of the enterprise or platform. For example, the ability requirements set by a certain enterprise for "construction project manager" may include "construction progress control", "contract risk identification", "cost control", "team communication and cooperation", etc. The system administrator inputs these ability items and their mutual dependence relationship in a structured form in the background, such as "construction progress control" needs "engineering plan preparation" as a prerequisite, forming a directed dependence relationship between nodes. The system constructs an ability graph wherein is the set of ability nodes, is the edge set representing the dependence relationship. The graph structure is stored in the form of adjacency matrix or adjacency list for subsequent graph neural network module calls.

[0021] Each ability node needs to be mapped to a structured learnable representation, so we collect the descriptive text of its corresponding course content, task description, evaluation standard, etc. for each node. These texts can be automatically extracted from enterprise learning platforms (such as LMS), for example, the system extracts the description text of the course "construction drawing design review" as "master the standard of building construction drawing reading, and can identify and propose modification suggestions for design errors in construction drawings." All descriptions will be processed by standard word segmentation and mapped to word vectors, using a pre-trained 300-dimensional word embedding model (such as GloVe or Word2Vec), to form a word vector matrix, which is input into a two-layer fully connected neural network (64 dimensions per layer, ReLU activation) to output a fixed-length ability node semantic vector . This embedding vector will be used as the initial representation of the node in the subsequent graph neural network for path node priority learning.

[0022] At the same time of building the ability graph, the system needs to obtain the current mastery degree of each ability point of the user, i.e. the ability state vector . The generation of this vector is based on the historical learning behavior data of the user, mainly including two parts: one is the course record and test result completed by the user on the learning platform, and the other is the completion of the practical operation task or project participated by the user. These data are usually provided by the API interface of the learning platform, such as the "learning completion", "answering score" and other record fields under the SCORM or xAPI protocol, or can be exported from the structured data table of the performance system or training system. Taking the "construction progress control" ability as an example, if the user has completed the related course and passed the class test (such as scoring 85 points), and at the same time records in the actual engineering task management system that he has led a project plan preparation task, the system will record his theoretical score and practical completion status respectively.

[0023] To unify the indicators from different sources, the system adopts a weighted scoring mechanism to form a comprehensive score for each ability point : ; Wherein: represents the mastery degree of the user to the ability node , with a value range of ; represents the average test score of the user in the course corresponding to the node, from the record in the LMS, normalized to ; represents whether the user has completed the practical operation task corresponding to the node, 1 for completion and 0 for non-completion, from the project system; respectively represent the weight coefficients of theoretical learning and practical operation completion, which are preset by the system according to the ability type, for example , ; Since and are dimensionless (normalized), the weight is a real number, and the dimensions of the numerator and denominator are consistent, so the result of is dimensionless, meeting the requirement of dimensional consistency.

[0024] For missing data items, such as the user never participates in a certain course or has no related practical operation record, the system uses the mean value of the node among all users to fill in, avoiding the occurrence of untrainable items, and reducing the priority of the node in path generation.

[0025] Finally, the semantic embedding representation of each node in the ability graph needs to be built, which is used for subsequent graph calculation. The calculation process is as follows: ; Wherein: is the ability node is the corresponding text embedding matrix, is the number of words in the text description of the node; 、 is the weight matrix of the two-layer fully connected network using ReLU activation; is the output node semantic vector, representing the semantic features of the ability node; Since the network input and output are both uniform vector dimensions, the embedding process does not involve physical quantity units, and the dimensional consistency is guaranteed in the network design.

[0026] This structural semantic representation method, combined with the graph structure, provides a structural and semantic unified node representation form for subsequent graph calculation.

[0027] S2: Priority scoring of ability nodes based on a structure-aware ranking function, which integrates the difference in user's mastery of nodes, the structural dependence reflected by the in-degree and out-degree of nodes in the graph, and the complexity of the node semantic vector; filtering the ability nodes to form a learning path according to the priority score, and after topological legality check, combining user preference to allocate learning resources, and generating a task scheduling plan according to the user's daily available learning time model; Specifically, this step is based on the ability graph structure constructed in the previous stage and the user ability state vector , combined with the ability node semantic representation vector , to generate a learning path sequence for the current user , and generate a task scheduling plan for the path . The generation of this path is not only the selection and sorting of unmastered content, but also based on the structural dependence in the task graph, combined with the difference in user's current mastery, the semantic expression of the ability node, the estimation of task learning load and user preference, to perform overall sorting and scheduling. The goal is to generate an executable path that meets the structural logic, ability difference, user preference, and adaptive time rhythm, as the main control sequence of the learning system.

[0028] This step completely depends on the output variables of the previous stage: the ability graph provides the learning structural dependence, the user ability state vector provides the mastery score of each node, and the node semantic vector provides the structured semantic input features. All inputs are no longer preprocessed, and only these values are used in the graph structure for path sorting and time matching.

[0029] The core task of path ranking is to determine the "learning urgency" of each competency node, i.e. the criticality of a competency node to the current goal achievement path of the user. Traditional methods often only use the inverse value of the mastery degree (e.g. ) as the ranking indicator, ignoring the influence of the propagation effect and semantic complexity in the graph structure on the learning path. In this embodiment, there is a clear "stage-module-subtask" structure among competencies, such as "construction phase management" containing "progress", "quality", and "safety" sub-modules, and these sub-modules are further refined into multiple parallel competency nodes. If the propagation of the graph structure is not considered, the path ranking will lack learning progression and task concentration.

[0030] Therefore, this step introduces a structure-aware ranking function that integrates the following three types of features: competency difference (reflected by , the higher the mastery, the higher the weight); graph structure dependence (the higher the propagation strength of the node as a multi-path predecessor / successor node, the higher the weight); node semantic complexity (measured by the complexity of the semantic vector represented by ).

[0031] Based on the following priority scoring formula: ; Where: represents the learning urgency score of the competency node for the current user, serving as the path ranking indicator; represents the user's mastery degree of the competency node calculated in the previous stage, with a value range of ; is the in-degree of node , representing the number of other nodes it depends on, reflecting its learning depth; is the out-degree of node , representing the influence range of the node as a basic competency on other tasks; is the semantic vector of the competency node, generated by the network in the previous step, represents its norm, serving as a semantic complexity indicator; , are adjustment factors for the graph structure (e.g. , , indicating emphasis on predecessor-dependent competencies); is the semantic complexity adjustment term, used to control the weight of "language expression complex competencies" in the ranking; each part of this expression is a dimensionless value or vector norm, so the final remains dimensionless, and the dimensional consistency is established.

[0032] The scoring function simultaneously introduces a structural topology term and a semantic complexity term as bias enhancement factors. For example, although a user has a poor mastery of "safety scheme communication", the out-degree of this ability node is 0 and the semantic complexity is low, indicating that the user has weak influence on subsequent ability propagation and the meaning is simple, so the priority of the user is appropriately reduced; while another user also has a poor mastery of "construction organization design", but this ability is the starting point of multiple modules and has high semantic complexity, so the sorting score is high and the user is given priority to learn.

[0033] After sorting, the system selects task nodes from high to low as the learning path of this round, and checks the topology legality to ensure that all prerequisite nodes have been included or mastered (i.e. , the threshold is generally ). This check ensures that the path is structurally executable.

[0034] After the path is sorted, the system assigns a specific content resource set to each node in combination with the task resource library , and estimates the task execution time . The resource selection is prioritized according to the user preference vector , for example, if the user prefers video learning, the system will prioritize video content; if this type of content is missing, a recommended priority fill-in strategy will be used (e.g. image>PPT>mixed cases).

[0035] Next, the scheduling module assigns time periods to each task in a calendar manner. The system has a user learning time model , which represents the user's available learning time per day (the unit is normalized to a proportion of time period ). The system fills tasks into consecutive dates in order according to the estimated task duration , if the current date is saturated, it will be postponed to the next day. This scheduling method does not require dynamic programming algorithm, but is realized through sliding window allocation, which takes into account efficiency and rhythm matching.

[0036] S3: According to the task scheduling plan, issue tasks, collect four types of behavior data: task stay time, content integrity flag, evaluation score, and interruption frequency; use a behavior effectiveness function containing an abnormality detection regular term to calculate the behavior effectiveness score, identify abnormal tasks with a behavior effectiveness score below a threshold, and add a behavior deviation label; Specifically, this step is based on the learning path and the scheduling tableOn this basis, the learning task is actually issued to the terminal, and key behavior data is collected in the process of user executing the task. The core purpose is to judge whether the learning task is truly and effectively completed through system automatic collection and modeling without relying on user subjective feedback, and to provide clear quantitative basis for subsequent path correction and ability state update.

[0037] In actual deployment, the system pushes the task to the user terminal according to the schedule , and starts to monitor the behavior. The behavior data collection is realized through automatic burying of the client, without the need for user active operation, and has non-invasiveness and stability. However, due to the characteristics of “task-driven, strong standardization, and limited time resources” of enterprise training or pre-job learning scene, the system cannot collect a large amount of dimensional data like the Internet scene, so the behavior data collection is specially structured and simplified in this step, and only 4 types of behavior indicators that can directly map learning effectiveness are retained: : the actual residence time on the task , which is automatically timed by the terminal front end; : content integrity completion flag, such as whether the video is played to more than 95%, whether the PDF is browsed to the end, and the value is 0 or 1; : if the task contains an evaluation, the score (such as multiple-choice questions, true or false questions, quizzes, etc.) is collected and normalized to ; : the number of interruptions or switching times of the user in the task, which reflects the learning stability, and more than 1 times is abnormal.

[0038] Considering that users in actual application scenarios may have behaviors such as “hanging up to watch videos” and “quickly skimming”, directly using the above indicators for weighted summation will cause serious score distortion. Therefore, this step introduces a structure-enhanced behavior effectiveness score function to comprehensively evaluate the execution effect of each task , and introduces a structure with an explanatory weight regularization term to enhance its robustness and discriminability for enterprise training scenarios.

[0039] ; wherein: is the behavior effectiveness score of the task , the value range is not limited, and the higher the value, the more sufficient the task completion is; is the actual residence time (collected by the system front end); is the recommended learning time (calculated by the path generation module), which is used for normalization; Content completion flag, 0 or 1, determined by the system through the scroll bar position or video playback progress; Task evaluation score, normalized to , or null if no test; Number of jumps or number of interruption behaviors (e.g., switching windows, closing and re-entering); Current ability node Historical average score of other users in the same post group, used for behavior anomaly detection; to Score item weight coefficient, recommended configuration , , , ; Regularization factor for anomaly detection items, recommended setting is 0.2, indicating that when the user score deviates from the group mean, a reasonable penalty is given; all items are dimensionless data, normalized or designed to make the dimensions consistent.

[0040] The formula introduces , which is used to identify learning "false completion". For example, if a user's score in the "construction plan progress control" task is much higher than the historical average, but the behavior trajectory shows that the stay time is very short and the interruption frequency is high, the system will mark such behavior as "possibly completed with external assistance" or "not truly mastered", and this regularization term will automatically lower its value, effectively enhancing the robustness and credibility of behavior scoring. This discrimination mechanism has not appeared in existing path recommendation systems, and has obvious differences and practicality.

[0041] After calculating the of all task nodes, the system generates the score sequence . Then, according to the system's set score lower limit (such as 0.6), the behavior abnormal tasks are identified to form the abnormal task set , and further add behavior deviation labels such as "low duration", "frequent interruption", "abnormal score", etc. to each abnormal task, facilitating classification processing in the next path correction stage.

[0042] S4: Based on the behavior deviation label matching correction strategy of abnormal tasks, introduce path disturbance cost function to control the correction scale, adjust the path and scheduling plan under the total disturbance budget constraint, output the corrected learning path and scheduling table.

[0043] Specifically, this step aims to correct the original learning path and scheduling table based on the behavior score sequence output by the previous stage and the abnormal task set . Targeted corrections and reconstructions are performed. Unlike traditional recommendation systems that primarily rely on rearrangement or full replacement, this scheme's correction mechanism emphasizes the precise identification of task execution deviations and the protection of structural continuity. While ensuring the coverage of learning objectives, it achieves local fine-tuning of the path structure, ultimately outputting the corrected path. and scheduling table This step not only completes the closed-loop update of the path, but also provides a crucial behavioral adjustment channel for the adaptive learning system, making it a core module for realizing intelligent path evolution and continuous optimization.

[0044] This step requires additional input of the original path. Used to identify the contextual order and knowledge structure location of affected tasks; original scheduling table. Used to identify the time schedule, learning pace, and potential scheduling conflicts for each task.

[0045] The system first processes each abnormal task. A classification determination is performed. This determination does not rely on a deep learning model, but instead uses a rule-based model to construct an anomaly label set. Each label represents a specific anomaly pattern, such as "low duration, high score," "high bounce rate, low score," or "score deviating from the group mean." The judgment is based entirely on the behavioral data collected in step three, for example: like and Then it can be marked as ; like and Then it is marked as ; like and This indicates that the mastery level of that skill point does not match the result, and is marked as... ; The above-mentioned anomaly type identification logic enables the system to distinguish between two root causes: "content resource mismatch" and "user not aware of the problem," and to perform targeted corrective operations.

[0046] For each type of abnormal task, the system designs different correction strategies, forming a task correction operation set. And execute it in conjunction with the path context. Taking "resource mismatch" as an example, the system uses the capability graph... Find nodes resource pool And based on user preference vectors (Defined in step one) Reselect content resources If marked as "Learning Failure", then in Insert preceding capability node (Depend on Supplementary tasks consisting of edges determined in the middle) If the learning curve of a task fluctuates drastically, the task can be broken down into subtasks, such as splitting "construction progress coordination" into two microtasks: "node control principle" and "schedule Gantt chart drawing".

[0047] To control the correction intensity and maintain the path structure, a path perturbation cost function was designed. This is used to measure the offset of each correction operation from the path structure. ; in: For the task The cost of correcting disturbances; and These represent the index positions (i.e., sequence numbers) in the path before and after the task correction, used to measure the magnitude of the sorting change; The change in total duration after task adjustment, i.e. ; The depth of the capability node in the graph (measured from the root node) indicates the range of the correction effect; The structural penalty coefficient controls the weights of each disturbance term; it is typically set to... , , ; The system is based on Path correction is performed to meet constraints. The total disturbance budget is configurable and can be set according to the flexibility of the company's training plan, such as... The original path length cost. Correction tasks that satisfy the constraints will update the path one by one. Replace with correction task And adjust the scheduling table simultaneously. Consider the recommended time for new tasks And for scheduling conflict windows, a deferred insertion mechanism is used.

[0048] In one or more embodiments, such as Figure 2 As shown, an adaptive personalized learning path generation system is disclosed, the system comprising: The state construction module is used to extract a list of target capabilities from existing training models or job competency requirements, and construct a capability graph containing a set of capability nodes and a set of edges based on the sequential dependencies between capabilities; it collects the descriptive text corresponding to each capability node, processes it through word segmentation and word embedding, and then inputs it into a fully connected neural network to generate node semantic vectors; based on the user's historical course test results and practical task completion status, it uses a weighted scoring mechanism to generate user capability state vectors, and fills missing data with the average value of all users; The path generation module is configured to prioritize the ability nodes based on a structure-aware ranking function, wherein the structure-aware ranking function integrates the difference in the mastery of the nodes by users, the structural dependence reflected by the in-degree and out-degree of the nodes in the graph, and the complexity of the semantic vectors of the nodes; the ability nodes are filtered to form a learning path according to the priority score, and after a topological validity check, learning resources are allocated in combination with user preferences, and a task scheduling plan is generated according to a user daily available learning time model; The behavior evaluation module is configured to issue tasks according to the task scheduling plan, collect four types of behavior data including task stay time, content integrity flag, evaluation score, and interruption times, and calculate a behavior effectiveness score by using a behavior effectiveness function containing an abnormality detection regular term, and identify abnormal tasks with a behavior effectiveness score lower than a threshold and add a behavior deviation label. The path deployment module is configured to match a correction strategy based on the behavior deviation label of the abnormal task, introduce a path disturbance cost function to control the correction scale, adjust the path and the scheduling plan under the constraint of a total disturbance budget, and output the corrected learning path and scheduling table.

[0049] It should be noted that the specific working process of the adaptive personalized learning path generation system provided in the embodiments of the present application is the same as the process of the adaptive personalized learning path generation method described in the above embodiments, and will not be repeated here.

[0050] The embodiments of the present application also provide an adaptive personalized learning path generation device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the steps in the above adaptive personalized learning path generation method embodiments when executing the computer program, such as steps S1-S4 described in the above embodiments. Figure 1 Alternatively, the processor implements the functions of the modules in the above system embodiments when executing the computer program.

[0051] For example, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the adaptive personalized learning path generation device.

[0052] The adaptive personalized learning path generation device can be a desktop computer, a notebook, a palm computer, a cloud server, and the like. The adaptive personalized learning path generation device can include, but is not limited to, a processor, a memory. Those skilled in the art can understand that the adaptive personalized learning path generation device can also include an input / output device, a network access device, a bus, and the like.

[0053] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASAC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or can also be any conventional processor, and the like. The processor is the control center of the adaptive personalized learning path generation device, and connects various parts of the adaptive personalized learning path generation device through various interfaces and lines.

[0054] The memory can be used to store the computer programs and / or modules, and the processor realizes various functions of the adaptive personalized learning path generation device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, and the like; the data storage area can store data created according to the running of the air conditioner controller, and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0055] If the modules integrated in the adaptive personalized learning path generation device are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0056] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned various method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0057] The above is the preferred embodiment of the present application. It should be noted that those skilled in the art can make several improvements and refinements without departing from the principles of the present application. These improvements and refinements are also considered within the scope of protection of the present application.

Claims

1. An adaptive personalized learning path generation method, characterized in that, The method includes: Extract a list of target capabilities from existing training models or job competency requirements, and construct a capability graph containing a set of capability nodes and a set of edges based on the sequential dependencies between capabilities; collect the descriptive text corresponding to each capability node, process it through word segmentation and word embedding, and then input it into a fully connected neural network to generate node semantic vectors; based on the user's historical course test results and practical task completion status, use a weighted scoring mechanism to generate a user capability status vector, and fill missing data with the average value of all users. The system prioritizes capability nodes based on a structure-aware ranking function, which integrates the differences in users' mastery of nodes, the structural dependencies reflected by the in-degree and out-degree of nodes in the graph, and the complexity of the node's semantic vector. Based on the priority scores, capability nodes are selected to form learning paths. After topological validity checks, learning resources are allocated in combination with user preferences, and a task scheduling plan is generated based on the user's daily available learning time model. Tasks are issued according to the task scheduling plan, and four types of behavioral data are collected: task dwell time, content integrity flag, evaluation score, and number of interruptions. A behavioral validity score is calculated using a behavioral validity function with an anomaly detection regularization term. Abnormal tasks with behavioral validity scores below the threshold are identified and behavioral deviation labels are added. Based on the behavior deviation label matching correction strategy for abnormal tasks, a path perturbation cost function is introduced to control the correction scale. Under the constraint of total perturbation budget, the path and scheduling plan are adjusted, and the corrected learning path and scheduling table are output.

2. The adaptive personalized learning path generation method according to claim 1, characterized in that, The sequential dependencies between the capabilities are determined by fusing the hierarchical relationships annotated by domain experts with the results of sequence pattern mining of historical user learning paths. The sequence pattern mining uses a sliding window algorithm to extract combinations of high-frequency continuous learning nodes.

3. The adaptive personalized learning path generation method according to claim 1, characterized in that, The node semantic vectors are processed by a pre-trained 300-dimensional word embedding model to obtain a word vector matrix from the descriptive text, and then output through two layers of fully connected neural networks with ReLU activation functions, resulting in a vector dimension of 64 dimensions.

4. The adaptive personalized learning path generation method according to claim 1, characterized in that, The specific steps of the weighted scoring mechanism include: dividing the user's ability evaluation dimensions into theoretical learning performance and practical task completion, and pre-setting weights for each dimension according to the ability type; extracting the user's test scores and answer accuracy data in theoretical learning, and completion quality scores and step compliance data in practical tasks, and mapping the two types of data to obtain normalized scores for theoretical dimensions and normalized scores for practical dimensions; multiplying the normalized score of theoretical dimensions by the corresponding weight, and adding the result of multiplying the normalized score of practical dimensions by the corresponding weight to obtain the comprehensive score for that ability node.

5. The adaptive personalized learning path generation method according to claim 1, characterized in that, The specific steps of the structure-aware ranking function include: calculating the user's mastery difference score for nodes, where the mastery difference score is the difference between the score of the corresponding node in the current capability state vector and the target mastery threshold; calculating the structure dependency score, statistically analyzing the in-degree and out-degree values ​​of nodes in the capability graph, assigning higher weight to the in-degree values ​​through a preset adjustment factor, and then weighted summing them with the out-degree values; calculating the complexity score of the node's semantic vector, and calculating the information entropy value based on the discreteness of the vector elements; and weighting and fusing the mastery difference score, structure dependency score, and complexity score according to a preset ratio to generate a priority score for each node.

6. The adaptive personalized learning path generation method according to claim 1, characterized in that, The specific steps for allocating learning resources based on user preferences include: constructing a user preference vector based on the frequency and quality of different resource types selected in the user's historical learning records, with each vector element corresponding to the preference weight of each type of resource; extracting resource requirement tags from the current capability node and matching the resource type with the highest weight in the preference vector; if the number of matched resource types in the resource library is lower than a preset threshold, initiating a supplementary process: first checking the matching degree between text and image resources and the content keywords of the current node, selecting the type when the matching degree exceeds the threshold; if text and image resources do not meet the requirements, then sequentially checking the completeness of knowledge point coverage for PPT resources and the scenario adaptability for mixed case resources, until a suitable resource type is selected; recording the resource allocation result for updating the weights of the user preference vector.

7. The adaptive personalized learning path generation method according to claim 1, characterized in that, The anomaly detection regularization term is based on the group's historical average evaluation score of the current capability node, and the deviation of the user's score from the group mean is adjusted by a penalty coefficient.

8. The adaptive personalized learning path generation method according to claim 1, characterized in that, The path disturbance cost function calculates the disturbance cost by taking into account the position offset before and after task correction, the proportion of time change, and the depth of the node in the graph. The structural penalty coefficient controls the weight of each disturbance term.

9. The adaptive personalized learning path generation method according to claim 1, characterized in that, The behavioral deviation labels are obtained based on the following steps: When the assessment score is lower than the preset lower limit of assessment score, the dwell time is not lower than the preset lower limit of dwell time, and the content is complete, the behavior deviation label is insufficient ability type, and the correction strategy is to strengthen training; If the dwell time is lower than the preset lower limit, the number of interruptions exceeds the preset upper limit, and the assessment score is not lower than the preset lower limit, the behavior deviation label is distracted, and the correction strategy is rhythm adjustment. If the content is incomplete and the dwell time is lower than the preset dwell time limit, the behavior deviation label is incomplete execution, and the correction strategy is process simplification. If both the assessment score is below the preset lower limit and the number of interruptions exceeds the preset upper limit, the behavioral deviation is labeled as a comprehensive disorder, and the correction strategy is a combination intervention.

10. An adaptive personalized learning path generation system, characterized in that, The system includes: The state construction module is used to extract a list of target capabilities from existing training models or job competency requirements, and construct a capability graph containing a set of capability nodes and a set of edges based on the sequential dependencies between capabilities; it collects the descriptive text corresponding to each capability node, processes it through word segmentation and word embedding, and then inputs it into a fully connected neural network to generate node semantic vectors; based on the user's historical course test results and practical task completion status, it uses a weighted scoring mechanism to generate user capability state vectors, and fills missing data with the average value of all users; The path generation module is used to prioritize capability nodes based on a structure-aware ranking function. The structure-aware ranking function integrates the differences in users' mastery of nodes, the structural dependencies reflected by the in-degree and out-degree of nodes in the graph, and the complexity of the node's semantic vector. Based on the priority scores, capability nodes are selected to form learning paths. After topological validity checks, learning resources are allocated in combination with user preferences, and a task scheduling plan is generated based on the user's daily available learning time model. The behavior evaluation module is used to issue tasks according to the task scheduling plan, collect four types of behavior data: task dwell time, content integrity flag, evaluation score, and number of interruptions; calculate the behavior effectiveness score using a behavior effectiveness function with an anomaly detection regularization term, identify abnormal tasks with behavior effectiveness scores below a threshold and add behavior deviation labels; The path deployment module is used to correct the behavior deviation label matching strategy based on abnormal tasks. It introduces a path perturbation cost function to control the correction scale, adjusts the path and scheduling plan under the total perturbation budget constraint, and outputs the corrected learning path and scheduling table.

Images