A method and system for dynamically generating personalized learning paths based on a smart learning platform

Abstract

This application provides a method and system for dynamically generating personalized learning paths based on a smart learning platform. The method includes: segmenting interactive behavior data, content dwell time sequence data, and physiological feedback data according to a preset window sliding rule to generate a timestamped learning state vector; having multiple distributed path planning agents generate candidate learning path fragments in parallel on the smart learning platform; adjusting the difficulty level and content order of the candidate learning path fragments using an inner-loop optimization mechanism, and reconstructing the knowledge association structure using an outer-loop optimization mechanism to obtain optimized candidate learning path fragments; and then combining the optimized candidate learning path fragments in a personalized manner to obtain a personalized learning path corresponding to the learner. The technical solution provided in this application, through multimodal data fusion and distributed agent collaborative optimization, achieves dynamic personalized learning path generation based on learner state and effect, significantly improving path adaptation accuracy and generation efficiency.

Description

Technical Field

[0001] This application relates to the field of educational informatization technology, and in particular to a method and system for dynamically generating personalized learning paths based on a smart learning platform. Background Technology

[0002] With the development of smart education, online learning platforms need to provide highly personalized learning paths for learners with different cognitive levels, learning habits, and interests. Traditional static course arrangement is difficult to meet the needs of real-time adjustments, especially when facing large-scale learning groups. A key technical challenge is how to dynamically integrate multi-dimensional learning data and quickly generate paths that conform to knowledge logic and adapt to individual learning states. This requires the system to have fine-grained learning state awareness capabilities, real-time path optimization algorithms, and a scalable distributed computing architecture.

[0003] Existing solutions employ a single-agent path planning method based on reinforcement learning. This method generates learning paths by constructing a Markov decision process model by analyzing learners' historical answer records and clickstream data. It uses a value function to evaluate the rewards of different content nodes and dynamically adjusts the recommendation strategy based on the learner's real-time performance. While this method achieves a certain degree of personalization, it relies on a single data source for decision-making, lacks comprehensive awareness of the multi-dimensional states of the learning process, and suffers from computational efficiency bottlenecks when dealing with complex knowledge graphs and large-scale user concurrency.

[0004] The shortcomings of existing solutions are mainly reflected in two aspects: first, the data dimension is too narrow to capture physiological feedback and temporal behavioral patterns, resulting in a lack of precision in path adjustment; second, centralized decision-making mechanisms are difficult to efficiently handle multi-objective optimization (such as knowledge correlation, difficulty adaptation, and learning pace), and are prone to getting trapped in local optima. Furthermore, single-agent architectures cannot support large-scale parallel path generation across knowledge points, making it difficult to meet real-time requirements. These limitations are precisely the core problems that this application aims to solve. Summary of the Invention

[0005] This application provides a method and system for dynamically generating personalized learning paths based on an intelligent learning platform, which solves the problems in the prior art such as insufficient precision of path adjustment due to the single data dimension, low efficiency of centralized decision-making mechanisms in multi-objective optimization, and the inability of single-agent architecture to support large-scale parallel path generation.

[0006] Firstly, this application provides a method for dynamically generating personalized learning paths based on a smart learning platform, including:

[0007] Acquire learners' interactive behavior data, content dwell time data, and physiological feedback data;

[0008] The interactive behavior data, the content dwell time sequence data, and the physiological feedback data are segmented according to a preset window sliding rule to generate a learning state vector with timestamps.

[0009] Based on the learning state vector, multiple distributed path planning agents generate candidate learning path segments in parallel on the intelligent learning platform. The candidate learning path segments include learning content identifiers, difficulty levels, and knowledge association structures. The multiple distributed path planning agents coordinate to adjust by exchanging path optimization parameters that include weight scores and conflict markers.

[0010] An inner-loop optimization mechanism is used to adjust the difficulty level and content order of the candidate learning path segments, and an outer-loop optimization mechanism is used to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data to obtain optimized candidate learning path segments.

[0011] The optimized candidate learning path fragments are then combined in a personalized manner to obtain a personalized learning path for each learner.

[0012] Optionally, the step of generating candidate learning path segments in parallel on a smart learning platform by multiple distributed path planning agents based on the learning state vector includes:

[0013] Each of the distributed path planning agents extracts the set of learning content to be assigned for the corresponding time window from the content library of the smart learning platform based on the timestamp marker in the learning state vector.

[0014] Based on the set of learning content to be assigned, an initial path unit sequence is constructed according to the prerequisite relationships in the preset knowledge graph. Each unit in the initial path unit sequence contains a minimum learning time and a target association strength value. The target association strength value is the semantic association metric result between the unit and its neighboring units in the preset knowledge graph.

[0015] Based on the interaction density and physiological response amplitude in the learning state vector, the initial path unit sequence is dynamically pruned to disconnect the unit connections where the target association strength value is lower than the dynamic threshold, thus forming a primary path segment.

[0016] Through the collaborative channel of the intelligent learning platform, path optimization parameters are exchanged among multiple distributed path planning agents based on distributed message middleware. The path optimization parameters include the following: weight score and conflict marker. The weight score is a path quality quantification value calculated based on the unit association strength value and learning time allocation of the primary path segment. The conflict marker is a label used to identify content overlap or logical conflict in the primary path segments generated by different agents.

[0017] Each of the distributed path planning agents performs collaborative adjustments based on the received path optimization parameters, specifically including: if a conflict marker is detected, prioritizing conflicting path segments based on weight scores and retaining path segments with higher weight scores; if no conflict marker is detected, adjusting the unit arrangement order of the primary path segments and the minimum learning time allocation based on weight scores; when content overlap is detected between the primary path segments of different distributed path planning agents, deduplication is performed according to the hierarchical relationship in the knowledge graph, where the hierarchical relationship is defined as: the upper-level node is the core knowledge point, and the lower-level node is the derived knowledge point, and path segments containing the upper-level node are retained first during deduplication; finally, the candidate learning path segments are output.

[0018] Optionally, the step of dynamically pruning the initial path unit sequence based on the interaction density and physiological response amplitude in the learning state vector to disconnect unit connections where the target association strength value is lower than a dynamic threshold, forming a primary path segment, includes:

[0019] Based on the interaction density in the learning state vector, the sliding average of the window is calculated, and the sliding average of the window is used as the baseline density. Based on the fluctuation range of the physiological response amplitude sequence, the dynamic adjustment coefficient is calculated.

[0020] By combining the baseline density and the dynamic adjustment coefficient, a dynamic threshold is generated;

[0021] Traverse the initial path unit sequence, remove unit connections whose target association strength value is lower than the dynamic threshold, mark recombinable nodes, and retain valid unit connections whose association strength value is not lower than the dynamic threshold;

[0022] The continuity of the effective unit connections is rechecked according to the prerequisite relationship;

[0023] If the continuity check result indicates that the recombinable node cannot be connected to the effective unit connection, then based on the constraint of the minimum learning time, a transition unit is inserted or the connection order is adjusted, and the primary path segment is finally output.

[0024] Optionally, adjusting the difficulty level and content order of the candidate learning path segments using an inner-loop optimization mechanism includes:

[0025] By combining learners' ability assessment data, difficulty analysis is performed on the candidate learning path segments to obtain difficulty levels;

[0026] Gradient optimization is performed on the difficulty level to adjust the difficulty level of the candidate learning path segment, resulting in a difficulty distribution sequence;

[0027] The prerequisite constraint relationship between candidate learning path segments is analyzed, and based on the prerequisite constraint relationship, an initial content order that conforms to the knowledge progression logic is generated;

[0028] By combining the learners' historical behavioral preferences, the initial content order is adjusted to obtain an optimized content arrangement order;

[0029] The difficulty distribution sequence is fused and verified with the optimized content arrangement order to eliminate potential conflicts between the two, and the optimized candidate learning path fragments are output.

[0030] Optionally, the outer-loop optimization mechanism reconstructs the knowledge association structure based on historical learning effect data and the physiological feedback data to obtain optimized candidate learning path segments, including:

[0031] Collect historical learning effect data recorded during the historical learning process, normalize the historical learning effect data, and generate a normalized learning effect feature vector;

[0032] The learning performance feature vector is subjected to time decay weighting to obtain a weighted learning performance index.

[0033] A time-frequency domain physiological feature matrix is ​​constructed based on physiological feedback data within the historical learning period. The time-frequency domain physiological feature matrix is ​​then subjected to data alignment processing to obtain a physiological feature vector.

[0034] The association strength between the weighted learning performance index and the physiological feature vector is calculated to generate an association strength correction factor between knowledge nodes.

[0035] The association strength correction factor is input into a preset knowledge graph to obtain the reconstructed knowledge association structure;

[0036] The reconstructed knowledge association structure is then subjected to path search to generate candidate learning path fragments.

[0037] Optionally, the step of inputting the association strength correction factor into a preset knowledge graph to obtain the reconstructed knowledge association structure includes:

[0038] The association strength correction factor is input into the association strength matrix corresponding to the preset knowledge graph to generate an updated association strength matrix. Based on the updated association strength matrix and the dynamic threshold, a set of edges to be removed is generated.

[0039] The set of edges to be removed is topologically adjusted. During the adjustment process, edge removal operations are performed and a node marker graph to be reconstructed is generated. The node marker graph to be reconstructed is constrained and verified through prerequisites to generate a preliminary reconstructed graph that conforms to knowledge logic.

[0040] Based on the preliminary reconstructed graph, the association matching degree between the node to be reconstructed and the candidate node is calculated across modalities, and a set of candidate association edges is generated;

[0041] Based on the candidate set of related edges, a new set of related edges is generated through knowledge coherence evaluation;

[0042] The newly added set of associated edges is input into a preset knowledge graph for topological integrity verification. After the verification is successful, the reconstructed knowledge association structure is generated.

[0043] Optionally, after obtaining the optimized candidate learning path fragment, the method further includes:

[0044] When the change in the learning state vector exceeds the adaptive threshold calculated based on the historical fluctuation range, the number of conflict markers in the path optimization parameters exceeds the preset threshold, or the external knowledge base bound to the knowledge association structure is updated, the process of regenerating and optimizing the candidate learning path segment is triggered.

[0045] Secondly, this application provides a personalized learning path dynamic generation system based on an intelligent learning platform, comprising:

[0046] The acquisition module is used to acquire learners' interactive behavior data, content dwell time sequence data, and physiological feedback data;

[0047] The segmentation module is used to segment the interactive behavior data, the content dwell time sequence data, and the physiological feedback data according to a preset window sliding rule, and generate a learning state vector with a timestamp.

[0048] The generation module is used to generate candidate learning path segments in parallel on the intelligent learning platform by multiple distributed path planning agents based on the learning state vector. The candidate learning path segments include learning content identifiers, difficulty levels, and knowledge association structures. The multiple distributed path planning agents coordinate to adjust by exchanging path optimization parameters that include weight scores and conflict markers.

[0049] The optimization module is used to adjust the difficulty level and content order of the candidate learning path segments using an inner-loop optimization mechanism, and to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data using an outer-loop optimization mechanism, so as to obtain optimized candidate learning path segments.

[0050] The combination module is used to personalize the optimized candidate learning path fragments to obtain a personalized learning path for the learner.

[0051] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a method for dynamically generating personalized learning paths based on a smart learning platform as described in any of the first aspects.

[0052] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a method for dynamically generating personalized learning paths based on a smart learning platform as described in any of the first aspects.

[0053] This application provides a method for dynamically generating personalized learning paths based on a smart learning platform. The method includes: acquiring learner interaction behavior data, content dwell time sequence data, and physiological feedback data; segmenting the interaction behavior data, content dwell time sequence data, and physiological feedback data according to a preset window sliding rule to generate a timestamped learning state vector; based on the learning state vector, multiple distributed path planning agents generate candidate learning path segments in parallel on the smart learning platform. Each candidate learning path segment includes a learning content identifier, difficulty level, and knowledge association structure. The multiple distributed path planning agents coordinate by exchanging path optimization parameters containing weighted scores and conflict markers. An inner-loop optimization mechanism adjusts the difficulty level and content order of the candidate learning path segments, while an outer-loop optimization mechanism reconstructs the knowledge association structure based on historical learning effect data and physiological feedback data to obtain optimized candidate learning path segments. Finally, the optimized candidate learning path segments are personalized combinations to obtain the learner's corresponding personalized learning path.

[0054] This application constructs a dynamic learning state vector through multimodal data fusion, utilizes distributed intelligent agents to generate and collaboratively optimize candidate path fragments in parallel, and combines inner-loop optimization and outer-loop optimization to achieve two-level path tuning, ultimately outputting a highly personalized learning path. Its technical effects are as follows: while ensuring the logical coherence of knowledge, it accurately adapts to the learner's cognitive load, interest preferences and long-term knowledge construction needs, and improves the efficiency and real-time performance of large-scale personalized path generation through a distributed computing framework.

[0055] Furthermore, the distributed path planning agent extracts learning content to be assigned from the content library based on the timestamp of the learning state vector, and constructs an initial path unit sequence based on the prerequisite relationships of the knowledge graph. The initial sequence is dynamically pruned by interaction density and physiological response amplitude, breaking low-association connections to form primary path segments. Each agent exchanges path optimization parameters through a distributed message middleware, prioritizing the retention of high-weight segments or deduplicating based on the hierarchical relationships of the knowledge graph during collaborative adjustments, ultimately outputting candidate path segments. During dynamic pruning, a dynamic threshold is calculated by combining the moving average of interaction density and physiological response fluctuations, removing weakly associated units and reorganizing continuity, inserting transition units when necessary. Through dynamic threshold pruning and collaborative optimization, the learning path segments are ensured to simultaneously satisfy the logical coherence of knowledge and the adaptability of the learner's real-time state, avoiding content redundancy or cognitive conflict. Parallel computation and collaborative deduplication of the distributed agents improve path generation efficiency, while the dynamic threshold mechanism and the insertion of transition units ensure the flexibility and executability of the path, ultimately outputting highly personalized optimized path segments.

[0056] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart illustrating a method for dynamically generating personalized learning paths based on a smart learning platform, as provided in this application embodiment;

[0059] Figure 2 A schematic diagram of the structure of a personalized learning path dynamic generation system based on a smart learning platform provided in an embodiment of this application;

[0060] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of this application. Detailed Implementation

[0061] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0062] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 11, 12, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] Figure 1 A flowchart illustrating a method for dynamically generating personalized learning paths based on a smart learning platform, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes:

[0065] S11. Obtain learners' interactive behavior data, content dwell time data, and physiological feedback data.

[0066] Interactive behavior data consists of learners' actions on the platform, such as clicking, answering questions, and turning pages. Content dwell time series data records the time sequence of learners' stay on each content page. Physiological feedback data includes learners' physiological indicators collected through biosensors, such as heart rate, eye movements, and electroencephalograms (EEGs).

[0067] For example, in an online math course, the system records students' learning behavior in real time: when a student clicks on the "Quadratic Function" course video, the action is recorded; during video playback, the system continuously tracks the time students spend on each knowledge point, for example, spending 8 minutes on the "Vertex Formula Derivation" section and 5 minutes on the "Graph Drawing" section; simultaneously, the smart bracelet worn by the student detects a significant increase in their heart rate during the "Application Problem Analysis" stage, and the EEG headset detects a decrease in attention concentration at this time. This data is transmitted to the smart learning platform in real time, forming the raw learning dataset, providing a basis for subsequent analysis. Optionally, the EEG signals in the physiological feedback data are extracted using Fourier transform to determine the energy ratio of alpha waves (8-12Hz) and beta waves (12-30Hz), and the concentration index is calculated using the formula: Concentration = Beta wave energy / (alpha wave energy + beta wave energy) × 100%.

[0068] S12. The interactive behavior data, content dwell time sequence data, and physiological feedback data are segmented according to the preset window sliding rules to generate a learning state vector with timestamps.

[0069] The learning state vector is a feature vector calculated from interaction behavior, dwell time sequence, and physiological data, used to quantify the learning state. The window sliding rule can be a time-segmented strategy with a window size of 5 minutes and a sliding step size of 1 minute.

[0070] For example, on an online programming learning platform, the system analyzes student behavior data in 10-minute time windows. Between 9:00 and 9:10, the system recorded that the student completed three code debugging operations, spent 6 minutes on the "loop structure" knowledge point, and 4 minutes on the "function definition" section. Simultaneously, the system monitored that the student's average heart rate reached 85 beats per minute while debugging, and the focus index remained at 0.7. The system converted this data into structured features: debugging frequency 0.3 times / minute, loop structure time point percentage 60%, average heart rate fluctuation 15%, focus standard deviation 0.12, and appended the 9:00-9:10 time stamp, forming a learning state feature vector for that time period.

[0071] S13. Based on the learning state vector, multiple distributed path planning agents generate candidate learning path segments in parallel on the intelligent learning platform. The candidate learning path segments include learning content identifiers, difficulty levels, and knowledge association structures. Multiple distributed path planning agents coordinate to adjust by exchanging path optimization parameters that include weight scores and conflict markers.

[0072] The distributed path planning agent, operating in parallel, is responsible for generating and optimizing learning path segments. Candidate learning path segments are sequences of learning content generated by the distributed path planning agent and are intermediate building blocks of the final personalized learning path. Weighted scores are quantitative values ​​evaluating the quality of path segments, calculated based on association strength and learning effectiveness. Conflict markers identify logical or content conflicts between path segments generated by different agents. Learning content identifiers are unique symbols or codes that identify learning resources. Difficulty levels are quantitative values ​​of the cognitive challenge of the learning content for the current learner, and are typically dynamically adjusted. The knowledge association structure describes the logical relationship network between content within the path segments, constructed based on a pre-defined knowledge graph. Path optimization parameters are core variables controlling path generation and are dynamically adjusted through reinforcement learning strategies.

[0073] For example, in a smart English learning platform, three distributed agents simultaneously process a student's status data. Agent A, based on recent high-frequency error records, suggests prioritizing the review of the "present perfect tense" grammar point, marking it as intermediate difficulty; Agent B, combining the student's reading speed data, recommends a training module containing five short business English articles, setting the difficulty as advanced; Agent C, based on eye-tracking data, proposes inserting two video explanation units. The three agents exchange evaluation results through the platform middleware: Agent A's grammar review plan receives a weight of 85 points, but it has a time conflict with Agent C's video content. After collaborative adjustments, Agent A's grammar review and Agent B's reading training are retained, and the video explanations are adjusted to a concise version, ultimately forming an optimized learning segment containing grammar points, four reading materials, and one video.

[0074] Optionally, multiple distributed path planning agents exchange path optimization parameters through a distributed message middleware and make coordinated adjustments based on priority ranking rules and knowledge graph hierarchy deduplication strategies. Specifically, when a conflict marker is detected, conflicting path segments are prioritized according to weight scores, and path segments with high weight scores are retained. When no conflict marker is detected, the unit arrangement order and learning time allocation are adjusted according to weight scores.

[0075] S14. An inner-loop optimization mechanism is used to adjust the difficulty level and content order of candidate learning path segments, and an outer-loop optimization mechanism is used to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data to obtain optimized candidate learning path segments.

[0076] The inner-loop optimization mechanism is an algorithm module that adjusts the difficulty and order of paths in real time. The outer-loop optimization mechanism is an algorithm module that optimizes the knowledge association structure over the long term. For example, in an artificial intelligence course learning system, the inner-loop optimization mechanism detects that the difficulty rating of the "neural network backpropagation" content in the student's current path segment is too high, and automatically downgrades it from "high-level" to "medium-level" based on recent homework error rate data. Simultaneously, the accompanying "gradient descent visualization demonstration" unit is moved before this knowledge point. The outer-loop optimization mechanism analyzes the student's learning records and EEG data over the past three months, discovering a cognitive gap between "linear algebra basics" and "deep learning principles." Therefore, a "matrix calculus transition module" is added to the knowledge network, and after recalculating the association strength of each knowledge point, an optimized path segment is generated, including the adjusted difficulty order and the newly added transition content. Optionally, the inner-loop and outer-loop optimization mechanisms can be linked by sharing optimization weights.

[0077] S15. The optimized candidate learning path fragments are combined in a personalized way to obtain the learner's personalized learning path.

[0078] Personalized learning paths refer to the optimal sequence of learning content dynamically generated by intelligent algorithms based on learners' multidimensional data (behavior, cognition, physiological state) and knowledge system logic.

[0079] For example, a physics learning platform generates a final learning path for a learner. The system first selects the "Mechanics Fundamentals Review" segment as the starting point. After detecting that the learner has a good grasp of the work-energy theorem, it skips repetitive exercises and directly connects to the "Momentum Conservation Advanced Training." When the learning progress reaches the "Angular Momentum Experiment Simulation," real-time eye-tracking data indicates difficulty in understanding, and an automatic "Interactive Demonstration of Moment of Inertia" transition segment is inserted. Finally, considering the learner's long-term goals, the "Mechanical Energy Conservation Competition Problems" are adjusted to an optional module, forming a complete path from "Mechanics Fundamentals, Momentum Conservation, Moment of Inertia Demonstration, Angular Momentum Experiment to Final Exam Focus," with a total time 20% shorter than a standard course, but more precise coverage of knowledge points.

[0080] Here is a specific example:

[0081] In the intelligent mathematics learning platform, the system collects students' answer records, dwell time and concentration on each knowledge point in real time; it divides the data into 5-minute windows and calculates the average answering speed, dwell time and concentration fluctuation to generate a learning state vector; multiple agents recommend candidate path segments in parallel and optimize them collaboratively through weighted scoring and conflict marking; the inner loop mechanism adjusts the difficulty and the outer loop mechanism reconstructs the knowledge association; finally, they are combined into personalized paths and pushed to students.

[0082] By executing S11~S15, this embodiment of the application constructs a dynamic learning state representation through multimodal data fusion, utilizes distributed intelligent agents to generate and collaboratively optimize path segments, and combines inner and outer loop dual-level optimization to finally output a highly personalized learning path. While ensuring the logical coherence of knowledge, it accurately adapts to the learner's cognitive load, interest preferences and long-term learning goals, and improves the efficiency and real-time performance of large-scale path generation through a distributed computing architecture.

[0083] In one possible embodiment, S13, based on the learning state vector, multiple distributed path planning agents generate candidate learning path segments in parallel on the intelligent learning platform, including:

[0084] Step 131: Each distributed path planning agent extracts the set of learning content to be assigned for the corresponding time window from the content library of the smart learning platform based on the timestamp in the learning state vector.

[0085] The timestamp is a digital label that identifies the time point of data collection. The set of learning content to be assigned is a pool of selectable teaching content that conforms to the characteristics of a time window.

[0086] In this embodiment, the distributed path planning agent first parses the timestamp markers in the learning state vector to determine the current learning time window that needs to be planned. Each agent independently accesses the content library of the smart learning platform and filters matching learning content based on time period characteristics. The filtering process comprehensively considers metadata such as the content's publication time, popularity index, and age range to form a set of learning content to be allocated. For example, for the morning session from 9:00 to 10:00, theoretical explanation content may be prioritized; while for the evening session from 20:00 to 21:00, interactive practice content is preferred.

[0087] Step 132: Based on the set of learning content to be assigned, construct an initial path unit sequence according to the prerequisite relationships in the preset knowledge graph. Each unit in the initial path unit sequence contains a minimum learning time and a target association strength value. The target association strength value is the semantic association metric result between the unit and its neighboring units in the preset knowledge graph.

[0088] The minimum learning time is the shortest estimated time required to master the content of a unit. The initial path unit sequence is an ordered learning chain formed by initially arranging several learning units from the set of learning content to be assigned according to the prerequisite relationships of the knowledge graph.

[0089] In this embodiment of the invention, the system processes the set of learning content to be assigned based on a topological sorting algorithm of a knowledge graph. First, a directed graph of prerequisite relationships between content units is established. Then, the semantic correlation between every two adjacent units in the knowledge graph is calculated (using the cosine similarity value output by a graph embedding model). A minimum learning time is set for each unit (15 minutes for basic content, 25 minutes for advanced content), and the target correlation strength value (range 0-1) with the preceding and following units is labeled. For example, the correlation strength between the "function definition" unit and the "function properties" unit is set to 0.9, while the correlation strength with the "calculus basics" unit is set to 0.6, forming an initial path unit sequence with quantifiable relationship indicators.

[0090] Step 133: Based on the interaction density and physiological response amplitude in the learning state vector, dynamically prune the initial path unit sequence to disconnect the unit connections where the target association strength value is lower than the dynamic threshold, forming primary path segments.

[0091] Among them, the target association strength value refers to the quantitative indicator of the logical relevance between knowledge points. The primary path segment is a partially optimized local learning path. Interaction density is a quantitative value reflecting the learner's operational activity level. The formula for calculating interaction density is: Interaction Density = Number of valid operations within the current window ÷ Historical average number of operations. The historical average time range is, for example, the average of the past 7 days or the current day. Physiological response amplitude is the degree of fluctuation of physiological indicators from the baseline value. The formula for calculating physiological response amplitude is: Physiological Response Amplitude = |Current Physiological Value - Baseline Value| ÷ Baseline Standard Deviation. The dynamic threshold is calculated based on the product of 1.2 times the average interaction density and the physiological response fluctuation coefficient.

[0092] In this embodiment of the invention, the system monitors the interaction density and physiological response amplitude in the learning state vector in real time. When the interaction density is below 0.7 or the physiological response amplitude exceeds 2 standard deviations, the dynamic threshold is automatically increased. The initial path unit sequence is traversed, connections with association strength below the dynamic threshold are disconnected, and transitional content or rerouting is inserted at the breakpoints. The final retained unit connections must simultaneously meet the following conditions: association strength ≥ threshold, learning duration fit ≥ 80%, and cognitive load fluctuation ≤ 15%, forming a primary path segment that conforms to the real-time learning state. Dynamic pruning is an optimization process based on real-time data to cut off unqualified unit connections in the initial sequence and reorganize the path.

[0093] Step 134: Through the collaborative channel of the intelligent learning platform, path optimization parameters are exchanged among multiple distributed path planning agents based on the distributed message middleware. The path optimization parameters include the following: weight score and conflict marker. The weight score is the path quality quantification value calculated based on the unit association strength value of the primary path segment and the learning time allocation. The conflict marker is a label used to identify content overlap or logical conflict in the primary path segments generated by different agents.

[0094] Among them, the path quality quantification value is a numerical index calculated by comprehensively analyzing the unit association strength value and learning time allocation of the primary path segment, and is used to evaluate the rationality of the path segment.

[0095] In this embodiment, each agent exchanges path optimization parameters via a message queue. The weight score is calculated by multiplying the average association strength by the duration adaptation coefficient, ranging from 50 to 100 points. Conflict flags employ multidimensional encoding; for example, when two agents recommend path segments containing 30% of the same example questions, an overlap flag is generated; when a logical reversal occurs ("learning integrals before derivatives"), a level 5 conflict code is triggered. These parameters are synchronized in real-time within the agent cluster using a publish-subscribe model.

[0096] Step 135: Each distributed path planning agent performs collaborative adjustments based on the received path optimization parameters. Specifically, if a conflict marker is detected, conflicting path segments are prioritized based on weight scores, retaining those with higher weight scores. If no conflict marker is detected, the unit arrangement order and minimum learning time allocation of primary path segments are adjusted according to weight scores. When content overlap is detected between primary path segments of different distributed path planning agents, content deduplication is performed according to the hierarchical relationship in the knowledge graph. The hierarchical relationship is defined as: the upper-level node is the core knowledge point, and the lower-level node is the derived knowledge point. Path segments containing the upper-level node are retained first during deduplication. Finally, candidate learning path segments are output.

[0097] In this embodiment of the invention, the agent employs a multi-attribute decision-making algorithm to process and optimize parameters. For segments marked with conflict, they are sorted in descending order of weight score based on association strength or learning time efficiency ratio, retaining the top three solutions. When there is no conflict, the unit order (higher scores first) and time allocation (higher scores receive more time) are adjusted according to the score. In the content deduplication stage, the hierarchical relationship of conflicting knowledge points is extracted first (e.g., "polynomial differentiation" is the superior node of "composite function differentiation"), segments containing superior nodes are retained, and duplicate parts of derived knowledge points are removed. The final output candidate learning path segments must meet the following requirements: no content redundancy, logical coherence ≥90%, comprehensive score ≥80 points, and total time error controlled within ±10%.

[0098] Here is a specific example:

[0099] In a certain intelligent programming learning platform, at 9:00 AM, the system detected that student A's interaction intensity was 0.8 and selected 12 content items to be assigned, including programming language basics and algorithm introduction. When constructing the initial path, the association strengths of "variable types, loop structures to function definitions" were concatenated, with strengths of 0.85 and 0.78 respectively. At 10:15 AM, it was discovered that the student's heart rate suddenly increased in the "recursive algorithm" unit. The dynamic threshold was immediately adjusted from 0.65 to 0.8, and the connection with "tree structures" was broken; the original strength was 0.72. Three agents proposed adjustment schemes, and after weighted scoring (88, 92, and 85) and conflict detection, the high-scoring segment containing the core knowledge point "pointer basics" was retained, and duplicate "array exercises" were removed. Finally, an optimized path of "variables, loops, functions, pointers, and file operations" was generated, reducing the total learning time from 150 minutes to 135 minutes.

[0100] By executing steps 131-135, this embodiment of the application achieves the following through a collaborative decision-making mechanism of distributed intelligent agents: 1) dynamically generating learning paths based on multi-dimensional real-time data; 2) maintaining the integrity and logic of the knowledge system; 3) automatically eliminating content redundancy and logical conflicts; and 4) adapting to individual cognitive characteristics and state fluctuations. Compared with traditional methods, the path generation efficiency is improved by 3-5 times, and the accuracy of personalized matching is improved by more than 40%.

[0101] In one possible embodiment, step 133, dynamically pruning the initial path unit sequence based on the interaction density and physiological response amplitude in the learning state vector to disconnect unit connections where the target association strength value is lower than a dynamic threshold, forming a primary path segment, includes:

[0102] Step a1: Calculate the window sliding average based on the interaction density in the learning state vector, and use the window sliding average as the baseline density. Calculate the dynamic adjustment coefficient based on the fluctuation range of the physiological response amplitude sequence.

[0103] Interaction density refers to the frequency of learners' actions per unit of time. The sliding window average is the average number of interactions within a time window, used to smooth data fluctuations. The physiological response amplitude sequence is a numerical sequence of physiological signals (such as heart rate and eye movements) changing over time. The fluctuation range is measured by the standard deviation, which indicates the degree of dispersion of the physiological data. Physiological response amplitude = |current physiological value - baseline value| ÷ baseline standard deviation, using personalized baseline calibration, for example, collecting one week of resting physiological data as an individualized baseline during the learner's initial phase.

[0104] Optionally, the fluctuation range of the physiological response amplitude sequence is represented by the standard deviation of the physiological response amplitude sequence. The standard deviation of the physiological response amplitude sequence ÷ the historical standard deviation benchmark = dynamic adjustment factor.

[0105] Step a2: Combine the baseline density and the dynamic adjustment coefficient to generate a dynamic threshold.

[0106] The dynamic threshold is a filtering criterion for association strength that is adjusted based on the real-time learning status. The dynamic threshold is generated using a linear weighting formula: Dynamic Threshold = α × Mean Baseline Density + β × Dynamic Adjustment Coefficient, where α and β are weighting parameters.

[0107] Step a3: Traverse the initial path unit sequence, remove unit connections with target association strength values ​​lower than the dynamic threshold, mark recombinable nodes, and retain valid unit connections with association strength values ​​not lower than the dynamic threshold. Recombinable nodes are path nodes that need to be replanned due to insufficient association strength. For example, assume the initial path sequence is A, B, C to D, where the association strength between B and C is 0.5 and the dynamic threshold is 0.6. Disconnect the connection between B and C; nodes B and C are marked as recombinable nodes. Based on the hierarchical relationship of the knowledge graph, the system retrieves that the subordinate node of B is E and the superior node of C is F. It then attempts to insert transition unit E or F to form a new path A, B, E, F, C to D.

[0108] Step a4: Recheck the continuity of valid unit connections according to prerequisite relationships. Prerequisite relationships are the logical dependencies between knowledge nodes, such as the requirement to learn addition before learning multiplication.

[0109] Step a5: If the continuity check result indicates that a recombinable node cannot be connected to a valid unit connection, then based on the minimum learning time constraint, a transition unit is inserted or the connection order is adjusted, and the initial path fragment is finally output. The transition unit is auxiliary learning content used to fill in path breaks, such as examples or review materials.

[0110] Here is a specific example:

[0111] When learning "trigonometric functions," learners experience decreased interaction density and increased fluctuations in physiological response due to distraction. First, the average interaction density of the current window is calculated to be 3 times / minute, with a physiological response standard deviation of 15, generating a dynamic adjustment coefficient of 1.5. Combining the baseline density and the adjustment coefficient of 1.5, a dynamic threshold of 2.7 is generated. Units with an association strength below 2.7, such as "inverse trigonometric functions," are removed and marked as recombinable nodes. The continuity of remaining units, such as "sine function" and "cosine function," is checked, revealing logical breaks. A transitional unit, "Basic Examples of Trigonometric Functions," is inserted, and the path order is adjusted before outputting a basic path fragment.

[0112] By executing steps a1-a5, this embodiment of the application adjusts the path selection criteria based on real-time interaction and physiological data through dynamic threshold screening, knowledge logic verification, and heuristic adjustment to adapt to changes in learner status. Topological sorting ensures the path conforms to the knowledge progression relationship. Heuristic algorithms are used to quickly repair path breaks, reducing computational overhead.

[0113] In one possible embodiment, S14, adjusting the difficulty level and content order of candidate learning path segments using an inner-loop optimization mechanism, includes:

[0114] Step b1: Combine learners' ability assessment data to perform difficulty analysis on candidate learning path segments and obtain difficulty levels.

[0115] The ability assessment data includes quantitative indicators such as test accuracy, learning time efficiency ratio, and emotional motivation intensity. The difficulty level is a classification label based on the strength of the knowledge graph association and the degree of matching with the learner's ability, for example, low / medium / high.

[0116] Step b2: Perform gradient optimization on the difficulty level to adjust the difficulty level of the candidate learning path segments and obtain the difficulty distribution sequence.

[0117] Gradient optimization involves using algorithms to adjust the step size and pace of difficulty changes to avoid sudden changes in cognitive load. Difficulty distribution sequence refers to a wave-like or step-like difficulty arrangement structure formed by dynamically adjusting the difficulty levels of various knowledge points in the learning path.

[0118] Step b3: Analyze the prerequisite constraints between candidate learning path segments, and generate an initial content order that conforms to the logic of knowledge progression based on the prerequisite constraints.

[0119] Among them, the knowledge progression logic refers to a structured logical system in learning path planning that arranges learning content in a sequence from shallow to deep, from basic to complex, and from known to unknown, based on the prerequisite relationships between knowledge points and the laws of cognitive development.

[0120] Step b4: Based on the learners' historical behavioral preferences, adjust the initial content order to obtain an optimized content arrangement.

[0121] Among them, historical behavioral preference characteristics are based on learners' tendencies toward resource types, learning periods, and interaction patterns.

[0122] Step b5: Merge and verify the difficulty distribution sequence with the optimized content arrangement order to eliminate potential conflicts between the two and output the optimized candidate learning path fragments.

[0123] Among them, fusion verification checks the compatibility between the difficulty gradient and the knowledge logic, and resolves ranking conflicts.

[0124] Here is a specific example:

[0125] In a programming learning platform, the system detected a learner's basic programming language accuracy rate at 72%, and adjusted the difficulty of the "function definition" segment from level 3 to level 2. During gradient optimization, it was found that the level difference between "function to decorator" was 2, so a "higher-order function" transition segment was inserted to form a "2, 3, 4" sequence. The knowledge graph requires "class inheritance" to precede "polymorphism," but historical data showed that the student preferred case-based learning, so the "class inheritance case" was moved to the beginning while maintaining the theoretical order. During the final integration, the originally planned morning "design patterns" advanced content was moved to the afternoon, a more efficient time, and accompanied by "intermediate application cases" as a transition, outputting an optimized path that maintained both gradual difficulty and met the learner's preferences.

[0126] By executing steps b1 to b5, this embodiment of the application, through a hierarchical optimization mechanism, first independently processes the two dimensions of difficulty gradient and content order, and then intelligently integrates and eliminates conflicts, achieving: 1) the learning path conforms to the laws of cognitive development (gradual difficulty + knowledge coherence); 2) it adapts to individual learning preferences to the greatest extent; 3) it dynamically maintains the balance of each optimization goal and avoids the overall imbalance problem caused by single-dimensional optimization.

[0127] In one possible embodiment, S14, the knowledge association structure is reconstructed based on historical learning effect data and physiological feedback data using an outer-loop optimization mechanism to obtain optimized candidate learning path segments, including:

[0128] Step c1: Collect historical learning performance data recorded during the learning process, and normalize the historical learning performance data to generate a normalized learning performance feature vector. The learning performance feature vector is a structured numerical representation generated after multi-dimensional normalization of the historical learning performance data. Normalization is a common operation and will not be elaborated upon here.

[0129] Step c2: Apply time-decay weighting to the learning performance feature vectors to obtain a weighted learning performance index. Time-decay weighting dynamically adjusts data weights based on time, with more recent data having a greater impact.

[0130] Step c3: Construct a time-frequency domain physiological feature matrix based on physiological feedback data within the historical learning period, and perform data alignment processing on the time-frequency domain physiological feature matrix to obtain physiological feature vectors.

[0131] The time-frequency domain physiological feature matrix is ​​a matrix that decomposes physiological signals into time-domain and frequency-domain features. The physiological feature vector is a standardized multidimensional feature set generated by performing time-frequency domain analysis and time-series alignment on physiological feedback data.

[0132] Step c4: Calculate the correlation strength between the weighted learning performance index and the physiological feature vector to generate a correlation strength correction factor between knowledge nodes. The correlation strength correction factor is a quantitative coefficient reflecting the dynamic correlation between learning performance and physiological state, with a value ranging from 0 to 1.

[0133] Step c5: Input the association strength correction factor into the preset knowledge graph to obtain the reconstructed knowledge association structure.

[0134] The reconstructed knowledge association structure refers to the optimized knowledge network formed by dynamically adjusting the association weights and topological relationships between entities in the knowledge graph. It is achieved by iteratively updating the preset knowledge graph through association strength correction factors, so that the logical relationships between knowledge nodes are more in line with the learner's real-time cognitive state.

[0135] Step c6: Perform path search on the reconstructed knowledge association structure to generate candidate learning path fragments.

[0136] The path search employs an improved A* algorithm, combining the cumulative association strength of nodes in the knowledge graph with the learning time efficiency ratio to form a path scoring function. The formula for the path scoring function is: Path Score = Association Strength × Learning Time Efficiency Ratio. During the search process, if the score of a path segment is lower than 80% of the current optimal solution, the expansion of that branch is terminated, prioritizing paths with high association strength and high learning efficiency. The knowledge graph is traversed from the starting point to the ending point, outputting the path segment with the highest score to ensure that the path conforms to the knowledge progression logic.

[0137] Here is a specific example:

[0138] A language learning platform analyzed a user's three-month historical data, assigning a 0.82 decay weight to test scores from two weeks prior. It also discovered a strong correlation between lip electromyography (EMG) signals during pronunciation practice and listening comprehension scores. Based on this, the system increased the correlation strength between the "phonetics-listening" node, ultimately prioritizing pronunciation training before listening comprehension, resulting in a 27% improvement in the user's pronunciation accuracy.

[0139] By executing steps c1 to c6, this embodiment of the application dynamically optimizes the knowledge association model by integrating long-term learning effects with real-time physiological data, so that the generated path not only conforms to the inherent logic of the subject, but also adapts to the physiological cognitive patterns of learners. This solves the problem of insufficient personalization caused by static modeling of traditional knowledge graphs, and improves the physiological adaptability and long-term effectiveness of path recommendation.

[0140] In one possible embodiment, step c5, inputting the association strength correction factor into a preset knowledge graph to obtain the reconstructed knowledge association structure, includes:

[0141] Step c51: Input the association strength correction factor into the association strength matrix corresponding to the preset knowledge graph to generate an updated association strength matrix. Based on the updated association strength matrix and the dynamic threshold, generate a set of edges to be removed.

[0142] The association strength matrix, dynamically adjusted by an association strength correction factor, reflects the integration of real-time learning status and knowledge logic. For example, if a learner performs poorly on a "limit" knowledge point (correction factor of 0.5), the original edge weight decreases from 0.8 to 0.4, resulting in an updated matrix. The set of edges to be removed is a set of low-weight edges selected based on the updated association strength matrix and a dynamic threshold.

[0143] Step c52: Perform topological adjustment on the set of edges to be removed. During the adjustment process, perform edge removal operations and generate a node marker graph to be reconstructed. Perform constraint verification on the node marker graph to be reconstructed based on the prerequisites to generate a preliminary reconstructed graph that conforms to the knowledge logic.

[0144] The node-marked graph to be reconstructed is an intermediate graph structure containing isolated nodes or logical breaks, formed after removing low-weight edges. The preliminary reconstructed graph is an intermediate knowledge graph state formed through topology adjustments and constraint verification.

[0145] Step c53: Based on the preliminary reconstructed graph, calculate the association matching degree between the node to be reconstructed and the candidate nodes across modalities, and generate a set of candidate association edges.

[0146] Among them, cross-modal association matching degree is a semantic similarity score between nodes calculated by integrating multimodal features such as text and images, using a graph neural network.

[0147] Step c54: Based on the candidate set of related edges, generate a new set of related edges through knowledge coherence evaluation.

[0148] Among them, knowledge coherence assessment is a screening process that verifies whether newly added edges conform to the logic of knowledge progression through rule base and statistical methods.

[0149] Step c55: Input the newly added set of related edges into the preset knowledge graph for topological integrity verification. After the verification is passed, the reconstructed knowledge association structure is generated.

[0150] Among them, topological integrity verification is a core step in the knowledge graph reconstruction process, used to ensure that the global structure of the graph after adding new related edges conforms to logical rules and domain constraints.

[0151] Here is a specific example:

[0152] In the reconstruction of the mathematical knowledge graph, the system detected that the edge strength of "trigonometric functions-calculus" was 0.58, which was below the threshold of 0.6. After removal, the affected nodes were marked. Calculations showed that the cross-modal matching degree of "trigonometric functions-vectors" was 0.62, and its coherence score was evaluated as 0.5. Finally, this edge was added and passed the cycle detection, making the original rigid knowledge structure adaptable to students' cognitive characteristics of combining geometry and algebra. The accuracy of the optimized path segments improved by 15%.

[0153] By executing steps c51 to c55, this embodiment of the application achieves the following through a knowledge graph reconstruction mechanism with dynamic threshold control, while maintaining the rigor of subject logic: 1) eliminating inefficient knowledge associations; 2) intelligently establishing new connections that conform to the learner's cognitive pattern; 3) ensuring that the reconstruction process does not destroy the integrity of the original knowledge system, so that the generated path has both scientific and personalized adaptability.

[0154] In one possible embodiment, S14, after obtaining the optimized candidate learning path fragments, the method further includes:

[0155] Step 141: When the change in the learning state vector exceeds the adaptive threshold calculated based on the historical fluctuation range, the number of conflict markers in the path optimization parameters exceeds the preset threshold, or the external knowledge base bound to the knowledge association structure is updated, the process of regenerating and optimizing candidate learning path segments is triggered.

[0156] The adaptive threshold is a trigger boundary that is dynamically adjusted based on historical fluctuation ranges. The external knowledge base is a third-party data source (such as academic paper databases or industry standard documents) that is bound to the knowledge graph and is updated in real time through an application programming interface (API).

[0157] Here is a specific example:

[0158] In a math learning platform, a learner experienced a disconnect in the "Calculus Applications" node due to an update to an external textbook. The accuracy rate of the "Derivatives" knowledge point in the learning state vector suddenly dropped by 40%, five conflict markers were detected in the path optimization parameters, and a new standard solution for "Differential Equations" was added to the external knowledge base. The solution involved calculating the fluctuation amplitude using a sliding window, triggering path reconstruction; using dynamic programming to prune conflicting edges, and invoking a semantic alignment model to associate the new solution with the "Derivatives Applications" node; after generating candidate paths, topology verification was used to ensure the core link remained connected.

[0159] By executing step 141, this embodiment of the application achieves precise dynamic path adaptation, logical integrity assurance, and enhanced real-time knowledge through a triple triggering mechanism of dynamic threshold, conflict detection, and external data synchronization. Experiments show that this solution can improve the learning path matching accuracy to 92% in online education scenarios.

[0160] Figure 2 A schematic diagram of a personalized learning path dynamic generation system based on a smart learning platform, provided in an embodiment of this application, is shown below. Figure 2 As shown, the system includes:

[0161] The acquisition module 21 is used to acquire learners' interactive behavior data, content dwell time sequence data, and physiological feedback data.

[0162] The segmentation module 22 is used to segment interactive behavior data, content dwell time sequence data and physiological feedback data according to preset window sliding rules to generate a learning state vector with timestamps.

[0163] The generation module 23 is used to generate candidate learning path segments in parallel on the intelligent learning platform based on the learning state vector by multiple distributed path planning agents. The candidate learning path segments include learning content identifiers, difficulty levels and knowledge association structures. Multiple distributed path planning agents coordinate to adjust by exchanging path optimization parameters that include weight scores and conflict markers.

[0164] The optimization module 24 is used to adjust the difficulty level and content order of candidate learning path segments using an inner-loop optimization mechanism, and to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data using an outer-loop optimization mechanism, so as to obtain optimized candidate learning path segments.

[0165] The combination module 25 is used to personalize the optimized candidate learning path fragments to obtain the learner's personalized learning path.

[0166] Figure 2 The aforementioned personalized learning path dynamic generation system based on a smart learning platform can execute... Figure 1 The implementation principle and technical effects of the personalized learning path dynamic generation method based on a smart learning platform described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit performs operations in the personalized learning path dynamic generation system based on a smart learning platform in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0167] In one possible design, Figure 2 The personalized learning path dynamic generation system based on a smart learning platform, as shown in the embodiment, can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32.

[0168] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0169] The processing component 32 is used to: acquire learner interaction behavior data, content dwell time sequence data, and physiological feedback data; segment the interaction behavior data, content dwell time sequence data, and physiological feedback data according to a preset window sliding rule to generate a learning state vector with timestamps; based on the learning state vector, multiple distributed path planning agents generate candidate learning path segments in parallel on the smart learning platform, the candidate learning path segments containing learning content identifiers, difficulty levels, and knowledge association structures, the multiple distributed path planning agents coordinating adjustments by exchanging path optimization parameters containing weight scores and conflict markers; use an inner-loop optimization mechanism to adjust the difficulty level and content order of the candidate learning path segments, and use an outer-loop optimization mechanism to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data to obtain optimized candidate learning path segments; and personalize the optimized candidate learning path segments to obtain a personalized learning path corresponding to the learner.

[0170] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0171] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Random Access Memory (RAM), Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0172] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0173] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0174] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0175] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0176] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown is a method for dynamically generating personalized learning paths based on a smart learning platform.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0178] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0179] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for dynamically generating personalized learning paths based on an intelligent learning platform, characterized in that, include: Acquire learners' interactive behavior data, content dwell time data, and physiological feedback data; The interactive behavior data, the content dwell time sequence data, and the physiological feedback data are segmented according to a preset window sliding rule to generate a learning state vector with timestamps. Based on the learning state vector, multiple distributed path planning agents generate candidate learning path segments in parallel on the intelligent learning platform. The candidate learning path segments include learning content identifiers, difficulty levels, and knowledge association structures. The multiple distributed path planning agents coordinate to adjust by exchanging path optimization parameters that include weight scores and conflict markers. The weight score is a path quality quantification value calculated based on the unit association strength value of the primary path segment and the learning time allocation. An inner-loop optimization mechanism is used to adjust the difficulty level and content order of the candidate learning path segments, and an outer-loop optimization mechanism is used to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data to obtain optimized candidate learning path segments. The optimized candidate learning path fragments are combined in a personalized way to obtain the learner's personalized learning path. The difficulty level and content order of the candidate learning path segments are adjusted using an inner-loop optimization mechanism, including: By combining learners' ability assessment data, difficulty analysis is performed on the candidate learning path segments to obtain difficulty levels; Gradient optimization is performed on the difficulty level to adjust the difficulty level of the candidate learning path segment, resulting in a difficulty distribution sequence; The prerequisite constraint relationships between candidate learning path segments are analyzed, and an initial content order that conforms to the knowledge progression logic is generated based on the prerequisite constraint relationships. By combining the learners' historical behavioral preferences, the initial content order is adjusted to obtain an optimized content arrangement order; The difficulty distribution sequence is fused and verified with the optimized content arrangement order to eliminate potential conflicts between the two and output the optimized candidate learning path fragments. The outer-loop optimization mechanism reconstructs the knowledge association structure based on historical learning effect data and the physiological feedback data to obtain optimized candidate learning path segments, including: Collect historical learning effect data recorded during the historical learning process, normalize the historical learning effect data, and generate a normalized learning effect feature vector; The learning performance feature vector is subjected to time decay weighting to obtain a weighted learning performance index. A time-frequency domain physiological feature matrix is ​​constructed based on physiological feedback data within the historical learning period. The time-frequency domain physiological feature matrix is ​​then subjected to data alignment processing to obtain a physiological feature vector. The association strength between the weighted learning performance index and the physiological feature vector is calculated to generate an association strength correction factor between knowledge nodes. The association strength correction factor is input into a preset knowledge graph to obtain the reconstructed knowledge association structure; The reconstructed knowledge association structure is then subjected to path search to generate candidate learning path fragments.

2. The method according to claim 1, characterized in that, Based on the learning state vector, multiple distributed path planning agents generate candidate learning path fragments in parallel on the intelligent learning platform, including: Each of the distributed path planning agents extracts the set of learning content to be assigned for the corresponding time window from the content library of the smart learning platform based on the timestamp marker in the learning state vector. Based on the set of learning content to be assigned, an initial path unit sequence is constructed according to the prerequisite relationships in the preset knowledge graph. Each unit in the initial path unit sequence contains a minimum learning time and a target association strength value. The target association strength value is the semantic association metric result between the unit and its neighboring units in the preset knowledge graph. Based on the interaction density and physiological response amplitude in the learning state vector, the initial path unit sequence is dynamically pruned to disconnect the unit connections where the target association strength value is lower than the dynamic threshold, thus forming a primary path segment. Through the collaborative channel of the intelligent learning platform, path optimization parameters are exchanged among multiple distributed path planning agents based on distributed message middleware. The path optimization parameters include the following: weight score and conflict marker. The weight score is a path quality quantification value calculated based on the unit association strength value and learning time allocation of the primary path segment. The conflict marker is a label used to identify content overlap or logical conflict in the primary path segments generated by different agents. Each of the distributed path planning agents performs collaborative adjustments based on the received path optimization parameters, specifically including: if a conflict marker is detected, prioritizing conflicting path segments based on weight scores and retaining path segments with higher weight scores; if no conflict marker is detected, adjusting the unit arrangement order of the primary path segments and the minimum learning time allocation based on weight scores; when content overlap is detected between the primary path segments of different distributed path planning agents, deduplication is performed according to the hierarchical relationship in the knowledge graph, where the hierarchical relationship is defined as: the upper-level node is the core knowledge point, and the lower-level node is the derived knowledge point, and path segments containing the upper-level node are retained first during deduplication; finally, the candidate learning path segments are output.

3. The method according to claim 2, characterized in that, The step of dynamically pruning the initial path unit sequence based on the interaction density and physiological response amplitude in the learning state vector to disconnect unit connections where the target association strength value is lower than a dynamic threshold, forming a primary path segment, includes: Based on the interaction density in the learning state vector, the sliding average of the window is calculated, and the sliding average of the window is used as the baseline density. Based on the fluctuation range of the physiological response amplitude sequence, the dynamic adjustment coefficient is calculated. By combining the baseline density and the dynamic adjustment coefficient, a dynamic threshold is generated; Traverse the initial path unit sequence, remove unit connections whose target association strength value is lower than the dynamic threshold, mark recombinable nodes, and retain valid unit connections whose association strength value is not lower than the dynamic threshold; The continuity of the effective unit connections is rechecked according to the prerequisite relationship; If the continuity check result indicates that the recombinable node cannot be connected to the effective unit connection, then based on the constraint of the minimum learning time, a transition unit is inserted or the connection order is adjusted, and the primary path segment is finally output.

4. The method according to claim 1, characterized in that, The step of inputting the association strength correction factor into a preset knowledge graph to obtain the reconstructed knowledge association structure includes: The association strength correction factor is input into the association strength matrix corresponding to the preset knowledge graph to generate an updated association strength matrix. Based on the updated association strength matrix and the dynamic threshold, a set of edges to be removed is generated. The set of edges to be removed is topologically adjusted. During the adjustment process, edge removal operations are performed and a node marker graph to be reconstructed is generated. The node marker graph to be reconstructed is constrained and verified through prerequisites to generate a preliminary reconstructed graph that conforms to knowledge logic. Based on the preliminary reconstructed graph, the association matching degree between the node to be reconstructed and the candidate node is calculated across modalities, and a set of candidate association edges is generated; Based on the candidate set of related edges, a new set of related edges is generated through knowledge coherence evaluation; The newly added set of associated edges is input into a preset knowledge graph for topological integrity verification. After the verification is successful, the reconstructed knowledge association structure is generated.

5. The method according to claim 1, characterized in that, After obtaining the optimized candidate learning path fragments, the method further includes: When the change in the learning state vector exceeds the adaptive threshold calculated based on the historical fluctuation range, the number of conflict markers in the path optimization parameters exceeds the preset threshold, or the external knowledge base bound to the knowledge association structure is updated, the process of regenerating and optimizing the candidate learning path segment is triggered.

6. A personalized learning path dynamic generation system based on an intelligent learning platform, characterized in that, A method for dynamically generating personalized learning paths based on a smart learning platform as described in any one of claims 1 to 5, comprising: The acquisition module is used to acquire learners' interactive behavior data, content dwell time sequence data, and physiological feedback data; The segmentation module is used to segment the interactive behavior data, the content dwell time sequence data, and the physiological feedback data according to a preset window sliding rule, and generate a learning state vector with a timestamp. The generation module is used to generate candidate learning path segments in parallel on the intelligent learning platform by multiple distributed path planning agents based on the learning state vector. The candidate learning path segments include learning content identifiers, difficulty levels, and knowledge association structures. The multiple distributed path planning agents coordinate to adjust by exchanging path optimization parameters that include weight scores and conflict markers. The optimization module is used to adjust the difficulty level and content order of the candidate learning path segments using an inner-loop optimization mechanism, and to reconstruct the knowledge association structure based on historical learning effect data and physiological feedback data using an outer-loop optimization mechanism, so as to obtain optimized candidate learning path segments. The combination module is used to personalize the optimized candidate learning path fragments to obtain a personalized learning path for the learner.

7. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a method for dynamically generating personalized learning paths based on a smart learning platform as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements a method for dynamically generating personalized learning paths based on a smart learning platform as described in any one of claims 1 to 5.

Images