Learning path intelligent recommendation system based on user behavior big data analysis

Abstract

The invention relates to the technical field of learning path recommendation, and discloses a learning path intelligent recommendation system based on user behavior big data analysis, and the system comprises a behavior data collection module which is used for collecting behavior data of a learner on a learning platform; the behavior inertia modeling module is used for extracting behavior inertia parameters of the learner based on the behavior data; the task capability graph construction module is used for constructing a task capability graph; calculating a grasp score according to the test score, and extracting a sub-graph map; and the recommendation path generation module is used for generating an optimal learning recommendation path according to the path recommendation engine, and selecting the optimal learning recommendation path as a recommendation result based on the path adaptation degree score. According to the method, personalized, rhythm-adaptive and resource-optimized learning path recommendation is dynamically generated on the basis of user behavior big data analysis in combination with post capability requirements.

Description

Technical Field

[0001] The present invention belongs to the technical field of learning path recommendation, and particularly relates to an intelligent learning path recommendation system based on big data analysis of user behavior. Background Art

[0002] With the rapid development of vocational education, vocational colleges increasingly attach importance to providing personalized learning path design according to different professional directions and job ability requirements. The traditional teaching mode often adopts a unified curriculum arrangement and a fixed teaching progress, which is difficult to fully consider the differences of learners in terms of learning rhythm, cognitive ability, knowledge mastery, and resource usage preferences, resulting in unsatisfactory learning effects and difficulty in accurately achieving the job ability training objectives. Some existing path recommendation methods mainly sort simply according to the curriculum plan or basic prerequisite relationships, lacking in-depth mining and dynamic analysis of learners' learning behavior data, and unable to adjust the recommended path in a timely manner according to the actual learning status of learners. At the same time, current intelligent applications in vocational education mostly focus on resource integration and information management, and rarely realize the dynamic modeling and application of the systematic association between job tasks, ability indicators, and course knowledge points. Summary of the Invention

[0003] The present invention provides an intelligent learning path recommendation system based on big data analysis of user behavior, which solves the technical problems in related technologies that the learning path cannot be dynamically adjusted according to the actual behavior of learners and lacks job ability orientation and personalized rhythm matching.

[0004] The present invention provides an intelligent learning path recommendation system based on big data analysis of user behavior, including:

[0005] A behavior data collection module, configured to collect the behavior data of learners on the learning platform, wherein the behavior data includes: click behavior, learning duration, learning completion rate, and test scores;

[0006] A behavior inertia modeling module, configured to extract the behavior inertia parameters of learners based on the behavior data, and the behavior inertia parameters include: learning rhythm stability coefficient, cognitive load fluctuation coefficient, and resource type efficiency characteristics;

[0007] A task ability map construction module, configured to construct a task ability map according to the association relationships between job tasks and ability indicators, ability indicators and course modules, and course modules and knowledge points predefined by the learning platform, where the task ability map includes job task nodes, ability indicator nodes, course module nodes, and knowledge point nodes, as well as the associated edges between the nodes;

[0008] Calculate the mastery score of each ability indicator node according to the test scores;

[0009] Extract a sub-graph map from the task ability map according to the mastery score and the behavior inertia parameter;

[0010] A recommended path generation module, which is used to generate an optimal learning recommendation path based on the behavior inertia parameter and the sub-graph map according to the path recommendation engine. Among them, the path recommendation engine is used to generate Q candidate learning recommendation paths and select the optimal learning recommendation path as the recommendation result based on the path fitness score. The candidate learning recommendation path is a sequence of course module nodes and knowledge point nodes arranged in order, and Q is a positive integer.

[0011] Further, construct a time rhythm dispersion coefficient through the ratio of the standard deviation of the learning peak time point to the mean of the learning peak time point. Calculate the Pearson coefficient between the learning duration and the test score of the learner within the first preset time period, and perform normalization processing through the sigmoid function to obtain the first intermediate feature. Combine the time rhythm dispersion coefficient and the first intermediate feature to obtain the learning rhythm stability coefficient;

[0012] Construct an error rate deviation coefficient based on the ratio of the difference between the number of test wrong questions of the learner and the mean of the number of test wrong questions to the standard deviation of the number of test wrong questions. Construct a behavior operation anomaly coefficient based on the learner's page stay time, the number of page rollbacks in a single learning, and the average number of page rollbacks of other learners on the platform in a single learning. Combine the error rate deviation coefficient and the operation anomaly coefficient to obtain the cognitive load fluctuation coefficient;

[0013] Obtain the second intermediate feature through the ratio of the learning duration of the learner on one type of resource to the total learning duration on all resource types, and obtain the third intermediate feature through the ratio of the average test score of the learner to the full score of the test. Combine the second intermediate feature and the third intermediate feature to obtain the resource type efficiency feature.

[0014] Further, the association relationships in the task ability map include:

[0015] There is a one-to-many association between the job task node and the ability index node;

[0016] There is a one-to-many association between the ability index node and the course module node;

[0017] There is a one-to-many association between the course module node and the knowledge point node;

[0018] Directed edges are established between knowledge point nodes according to the front-to-back dependency relationship of the knowledge system;

[0019] And generate directed edges between different types of nodes to construct a nested map structure of job task - ability index - course module - knowledge point.

[0020] Further, the mastery score of the ability index node is calculated based on the test scores associated with the ability index node. Specifically, the test scores of the ability index node are normalized and then weighted and summed to obtain the mastery score.

[0021] Further, according to the mastery score and the behavior inertia parameter, a sub-graph map is extracted from the task ability map. The specific steps include:

[0022] S201, screening the ability index nodes with a mastery score greater than the first preset threshold and a learning rhythm stability coefficient greater than the second preset threshold;

[0023] S202, extracting the job task nodes and course module nodes connected to the screened ability index nodes, as well as the knowledge point nodes connected to the course module nodes;

[0024] S203, obtaining the sub-graph map according to the screened ability index nodes, the extracted job task nodes, course module nodes and knowledge point nodes, and the edges between the nodes.

[0025] Further, the recommended path generation module imposes the following internal constraint conditions during the candidate learning recommended path generation process executed by the path recommendation engine:

[0026] The order of the course module nodes and knowledge point nodes is the sequential learning order between the course module nodes and knowledge point nodes in the sub-graph map;

[0027] There is an edge between adjacent knowledge point nodes in the candidate learning recommended path of the learner;

[0028] The estimated total learning time of the candidate learning recommended path of the learner does not exceed the fourth preset threshold.

[0029] Further, the recommended path generation module further includes: receiving intervention information input from an external port as an external constraint condition imposed during the candidate learning recommended path generation process executed by the path recommendation engine. The intervention information includes:

[0030] The forced priority learning order of the specified course module nodes;

[0031] Specifying that the candidate learning recommended path includes specific knowledge point nodes;

[0032] Specifying that the estimated total learning time of the candidate learning recommended path is not less than the fifth preset threshold.

[0033] Further, the recommended path generation module performs a structured encoding on the candidate learning recommended path. The specific steps include:

[0034] S301. Hierarchically encode the course module nodes and knowledge point nodes in the candidate learning recommendation path according to the dependency relationship in the sub-graph atlas to generate a node sequence with topological serial numbers.

[0035] S302. Map the resource type efficiency characteristics in the behavior inertia parameters to node weight labels, and form an encoded vector with the node sequence generated in S301.

[0036] Furthermore, the calculation process of the path fitness score includes:

[0037] S401. Detect the nodes in the node sequence that violate the dependency relationship in the sub-graph atlas as conflict nodes, assign a penalty value to the conflict nodes, and generate an order rationality score in combination with the hierarchical depth of the conflict nodes.

[0038] S402. Set the difficulty levels of the course module nodes and knowledge point nodes, count the node pairs in the candidate learning recommendation path where the difference in difficulty levels between adjacent nodes exceeds 2 levels, and calculate the difficulty gradient smoothness score through exponential decay.

[0039] S403. Perform weighted addition on the order rationality score and the difficulty gradient smoothness score to obtain the path fitness score.

[0040] Furthermore, the steps for the recommendation path generation module to select the optimal learning recommendation path based on the path fitness score include:

[0041] S501. Sort the Q candidate learning recommendation paths in descending order of the path fitness score, and retain a preset number of candidate learning recommendation paths.

[0042] S502. Perform a first mutation operation on the retained candidate learning recommendation paths with a first preset probability. The first mutation operation includes:

[0043] Randomly select a course module node in the candidate learning recommendation path.

[0044] Retrieve the knowledge point nodes associated with the course module node in the sub-graph atlas but not in the candidate learning recommendation path, and filter out the knowledge point nodes whose resource type efficiency characteristics are higher than the sixth preset threshold.

[0045] Replace the knowledge point nodes connected to the course module node in the node sequence with the filtered knowledge point nodes.

[0046] S503. Recalculate the path fitness score for the candidate learning recommendation path after the first mutation operation.

[0047] S504. Repeat the execution of S501 to S503 until the improvement of the highest path fitness score in the candidate learning path does not exceed the seventh preset threshold in consecutive U rounds of iteration, and use the candidate learning recommendation path with the highest path fitness score in the U-th round as the optimal learning recommendation path, where U is a positive integer.

[0048] The beneficial effects of the present invention are as follows: By collecting multi-dimensional behavioral data of learners on the learning platform, extracting behavioral inertia parameters, combining with the multi-layer nested task ability map constructed by job tasks, ability indicators, course modules and knowledge points, dynamically extracting personalized sub-maps, and based on the path fitness score and mutation optimization mechanism, generating the optimal learning recommendation path that conforms to the personality characteristics of learners and job requirements; By introducing external intervention constraints, realizing the coordinated regulation of path recommendation, teaching objectives and job requirements. The present invention can effectively improve the personalized matching degree of the learning path, optimize the learning rhythm and cognitive load distribution, improve the resource utilization efficiency, and enhance the job competency and learning completion quality of vocational college students. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a schematic diagram of the modules of the learning path intelligent recommendation system based on user behavior big data analysis of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Now, the subject matter described herein will be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed can be changed without departing from the protection scope of the content of this specification. Each example can omit, substitute or add various processes or components as needed. Additionally, the features described relative to some examples can also be combined in other examples.

[0051] As Figure 1 shown, the learning path intelligent recommendation system based on user behavior big data analysis includes:

[0052] A behavior data collection module 101 for collecting the behavior data of learners on the learning platform, where the behavior data includes: click behavior, learning duration, learning completion rate, and test scores;

[0053] A behavior inertia modeling module 102 for extracting the behavior inertia parameters of learners based on the behavior data, and the behavior inertia parameters include: learning rhythm stability coefficient, cognitive load fluctuation coefficient, and resource type efficiency characteristics;

[0054] The task ability graph construction module 103 is used to construct a task ability graph according to the association relationships between the job tasks and ability indicators predefined by the learning platform, between the ability indicators and course modules, and between the course modules and knowledge points. The task ability graph includes job task nodes, ability indicator nodes, course module nodes, and knowledge point nodes, as well as the association edges between the nodes;

[0055] Calculate the mastery score of each ability indicator node according to the test scores;

[0056] Extract a sub-graph from the task ability graph according to the mastery score and the behavior inertia parameter;

[0057] The recommended path generation module 104 is used to generate an optimal learning recommendation path based on the behavior inertia parameter and the sub-graph according to the path recommendation engine. Among them, the path recommendation engine is used to generate Q candidate learning recommendation paths and select the optimal learning recommendation path as the recommendation result based on the path fitness score. The candidate learning recommendation path is a sequence of course module nodes and knowledge point nodes arranged in order, and Q is a positive integer.

[0058] In an embodiment of the present invention, by embedding points on the front-end page, click behaviors are obtained and recorded. The click behaviors include: resource access clicks, page jump clicks, control operation clicks, and page back operations; the learning duration is obtained by recording the time when the learner enters and leaves the page; the learning completion rate is calculated by reading the video playback progress, the number of completed questions, etc.; the test scores of the learner are obtained through the back-end interface; since the above data have different sources and inconsistent dimensions, in order to ensure the accuracy of subsequent data fusion and calculation, they are normalized.

[0059] In an embodiment of the present invention, the calculation formula of the learning rhythm stability coefficient is: , where represents the learning rhythm stability coefficient, which is used to measure the overall rhythm stability. The higher the learning rhythm stability, the more stable it is. represents the standard deviation of the learning peak time points, which is used to measure the volatility. represents the mean of the learning peak time points, represents the regular learning time points, and sigmoid represents the sigmoid function. represents the sequence composed of the learning durations of N days. represents the sequence composed of the corresponding test scores of the next day. represents and the Pearson correlation coefficient between them, which is used to measure the coherence between learning behavior and learning effect;

[0060] The calculation formula of the cognitive load fluctuation coefficient is: , where denotes the cognitive load fluctuation coefficient, which measures the degree of cognitive fluctuation of learners during the learning process by recognizing fluctuations from two aspects: the degree of fluctuation of error performance and the abnormal intensity of behavioral operations. denotes the tanh function. denotes the number of wrong questions in the t-th test. denotes the average value of the number of wrong questions in the most recent 10 tests. denotes the standard deviation of the number of wrong questions in the most recent 10 tests. denotes the number of page backward operations during a single learning session, and d denotes the residence time on the current page. denotes the average residence time of similar resources. denotes the average value of the number of page backward operations during a single learning session of other learners on the platform.

[0061] The calculation formula for the efficacy characteristics of the resource type is as follows: , where denotes the efficacy characteristics of the resource type, which is used to measure the learning promotion effect of a resource type on the current user. denotes the learning duration of the learner on the resource type r. denotes the learning duration of the learner on all resource types. Specifically, the resource types include: course videos, course documents, course test questions, and external link resources. denotes the average value of the test scores of the learner. denotes the full score value of the test.

[0062] In an embodiment of the present invention, the task ability map is used to represent the multi-level association relationship among job tasks, ability indicators, course modules, and knowledge points in the learning platform;

[0063] Among them, the task ability map includes: job task nodes, ability indicator nodes, course module nodes, and knowledge point nodes; the job task nodes represent the task objectives to be completed under the actual job; the ability indicator nodes: represent the specific ability items required to complete the job tasks; the course module nodes represent the modular teaching units for ability cultivation in the course system provided by the platform; the knowledge point nodes represent the specific knowledge content units subdivided in each course module.

[0064] In an embodiment of the present invention, the association relationships in the task ability map include:

[0065] There is a one-to-many association between the job task nodes and the ability indicator nodes;

[0066] There is a one-to-many association between the ability indicator nodes and the course module nodes;

[0067] There is a one-to-many association between the course module nodes and the knowledge point nodes;

[0068] Directed edges are established between knowledge point nodes according to the front - to - back dependency relationship of the knowledge system;

[0069] And directed edges are generated between different types of nodes to construct a nested graph structure of job tasks - competency indicators - course modules - knowledge points. Among them, the directed edges between nodes point from the job task nodes to the competency indicator nodes, from the competency indicator nodes to the course module nodes, and from the course module nodes to the knowledge point nodes. The directed edges between knowledge points are determined according to the order of knowledge mastery.

[0070] By constructing the nested graph structure of job tasks - competency indicators - course modules - knowledge points, a full - chain mapping from job requirements to the knowledge - mastery path can be realized, making learning recommendations more accurate and systematic.

[0071] In an embodiment of the present invention, the mastery score of the competency indicator node is calculated based on the test scores associated with this competency indicator node. Among them, the test scores of this competency indicator node are normalized and weighted and summed to obtain the mastery score. The calculation formula for the mastery score is: , represents the mastery degree of the competency indicator node c, c represents the index of the competency indicator node, M represents the number of tests, j represents the index of the test, represents the test score of the learner in the j - th test of the competency indicator node c, represents the full score of the j - th test.

[0072] In an embodiment of the present invention, according to the mastery score and the behavior inertia parameter, a sub - graph is extracted from the task - competency graph. The specific steps include:

[0073] S201, screen out the competency indicator nodes with a mastery score greater than the first preset threshold and a learning rhythm stability coefficient greater than the second preset threshold;

[0074] S202, extract the job task nodes and course module nodes connected to them according to the screened - out competency indicator nodes, and the knowledge point nodes connected to the course module nodes;

[0075] S203, obtain the sub - graph according to the screened - out competency indicator nodes, the extracted job task nodes, course module nodes and knowledge point nodes, and the edges between the nodes.

[0076] In an embodiment of the present invention, the recommended path generation module imposes the following internal constraint conditions during the candidate learning recommended path generation process executed by the path recommendation engine:

[0077] The order of the course module nodes and knowledge point nodes is the sequential learning order between the course module nodes and knowledge point nodes in the sub - graph.

[0078] There is an edge between adjacent knowledge point nodes in the candidate learning recommendation path of the learner to ensure the coherence of the knowledge points in the learning path;

[0079] The estimated total learning time of the candidate learning recommendation path of the learner does not exceed the fourth preset threshold, where the estimated total learning time is estimated based on the learning duration of the learner in the historical learning process.

[0080] In an embodiment of the present invention, the recommendation path generation module further includes: a module for receiving intervention information input from an external port as an external constraint condition imposed during the execution of the candidate learning recommendation path generation process by the path recommendation engine. The external port includes: a teacher management terminal, an enterprise learning management platform, and a third-party teaching interface. The intervention information includes:

[0081] Specify the forced priority learning order of the specified course module nodes to ensure that key courses are presented first;

[0082] Specify that the candidate learning recommendation path includes specific knowledge point nodes to prevent the learning path from missing the core knowledge content of the post;

[0083] Specify that the estimated total learning time of the candidate learning recommendation path is not less than the fifth preset threshold to avoid the path length being too short and resulting in insufficient learning depth.

[0084] In an embodiment of the present invention, the steps for the path recommendation engine to generate Q candidate learning recommendation paths include:

[0085] According to the sequence dependency relationship between the course module nodes and the knowledge point nodes in the subgraph atlas, traverse the subgraph to construct a node sequence that meets the internal constraint conditions and the external constraint conditions;

[0086] Randomly insert node switching and backtracking operations during the generation process to avoid all paths being overly concentrated at the beginning or end;

[0087] When the number of generated candidate learning recommendation paths reaches Q, stop generating.

[0088] In an embodiment of the present invention, the recommendation path generation module performs a structured encoding on the candidate learning recommendation path. The specific steps include:

[0089] S301, hierarchically encode the course module nodes and the knowledge point nodes in the candidate learning recommendation path according to the dependency relationship in the subgraph atlas to generate a node sequence with topological serial numbers. Specifically, the topological serial number represents the serial number of the node, and the node sequence is expressed as: ;

[0090] S302, mapping the resource type efficiency feature in the behavioral inertia parameter into a node weight label, and forming a coding vector with the node sequence generated in S301. Specifically, the node weight label is a characteristic value of the resource type assigned to each node in the candidate learning recommendation path. For example, the node sequence is: , the encoding vector is:

[0091] In one embodiment of the present invention, the calculation process of the path fitness score includes:

[0092] S401, detect nodes in the node sequence that violate the dependency relationship in the subgraph graph as conflict nodes, assign penalty values ​​to the conflict nodes, and generate a sequence rationality score based on the hierarchical depth of the conflict nodes. The calculation formula for the sequence rationality score is: ,in, It represents the order rationality score, which is used to measure the degree to which the arrangement order of course module nodes and knowledge point nodes in the candidate learning recommendation path complies with the knowledge system dependency and teaching logic, ensuring the coherence of the learning process and the rationality of the knowledge structure. P represents the number of conflicting nodes, and o represents the index of the conflicting node. Indicates the level depth of the conflicting node in the subgraph. For example, the level depth of the course module node is 1, and the level depth of the knowledge point node is 2. represents the first weight coefficient;

[0093] S402, setting the difficulty level of the course module nodes and knowledge point nodes, counting the node pairs whose difficulty level difference between adjacent node pairs in the candidate learning recommendation path exceeds 2 levels, and obtaining the difficulty gradient smoothness score through exponential decay calculation. The calculation formula of the difficulty gradient smoothness score is: ,in, represents the difficulty gradient smoothness score, which is used to evaluate the smoothness of the change in difficulty level of adjacent content in the candidate learning recommendation path. e represents a natural constant. represents the attenuation intensity coefficient, represents the tolerance factor, Indicates the number of node pairs whose difficulty levels differ by more than 2 levels;

[0094] S403, weighted addition of the order rationality score and the difficulty gradient smoothness score is performed to obtain a path fitness score. The calculation formula of the path fitness score is: ,in, It represents the path fitness score, which is used to comprehensively evaluate the overall optimization degree of the candidate learning recommendation path in terms of knowledge coherence and difficulty transition smoothness, quantify the matching level between the path and the user behavior characteristics and knowledge system requirements, and guide the present invention to screen out the optimal learning recommendation path. and respectively represent the second weight coefficient and the third weight coefficient.

[0095] In an embodiment of the present invention, the steps of the recommended path generation module for selecting the optimal learning recommended path based on the path fitness score include:

[0096] S501, arrange the Q candidate learning recommended paths in descending order of the path fitness score, and retain a preset number of candidate learning recommended paths;

[0097] S502, perform a first mutation operation on the retained candidate learning recommended paths with a first preset probability, and the first mutation operation includes:

[0098] Randomly select a course module node in the candidate learning recommended path;

[0099] Retrieve the knowledge point nodes associated with the course module node in the subgraph atlas but not in the candidate learning recommended path, and screen the knowledge point nodes with the resource type efficiency characteristics higher than the sixth preset threshold;

[0100] Replace the knowledge point node connected to the course module node in the node sequence with the screened knowledge point node;

[0101] S503, recalculate the path fitness score for the candidate learning recommended paths after the first mutation operation;

[0102] S504, repeat the execution of S501 to S503 until the improvement amplitude of the highest path fitness score in the candidate learning path does not exceed the seventh preset threshold in consecutive U rounds of iterations, and use the candidate learning recommended path with the highest path fitness score in the U-th round as the optimal learning recommended path, where U is a positive integer.

[0103] Through the screening based on the path fitness score and the dynamic optimization of the first mutation operation, while maintaining the rationality of the learning recommended path, it can guide learners to preferentially contact the knowledge units with high resource efficiency and excellent learning experience, avoid the path recommendation falling into local optimality, improve the personalized fitness and learning completion quality of the learning recommended path, and improve the learning efficiency and the achievement effect of job capabilities.

[0104] It should be noted that the setting of the interval and threshold sizes is for the convenience of comparison. Among them, the size of the threshold depends on the amount of sample data and the base quantity set by those skilled in the art for each group of sample data, as long as it does not affect the proportional relationship between the parameters and the quantified values. And the above formulas are all calculations of taking the numerical value without dimension, and the formulas are all obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation, and the preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0105] The embodiments of the present invention have been described above. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of this embodiment.

Claims

1. An intelligent learning path recommendation system based on big data analysis of user behavior, characterized in that, Including: A behavior data collection module, which is used to collect the behavior data of learners on the learning platform. Among them, the behavior data includes click behavior, learning duration, learning completion rate, and test scores; A behavior inertia modeling module, which is used to extract the behavior inertia parameters of learners based on the behavior data. The behavior inertia parameters include learning rhythm stability coefficient, cognitive load fluctuation coefficient, and resource type effectiveness characteristics; A task ability map construction module, which is used to construct a task ability map according to the association relationships between the job tasks and ability indicators predefined by the learning platform, between the ability indicators and course modules, and between the course modules and knowledge points. The task ability map includes job task nodes, ability indicator nodes, course module nodes, and knowledge point nodes, as well as the associated edges between the nodes; Calculate the mastery score of each ability indicator node according to the test scores; Extract a sub-map in the task ability map according to the mastery score and behavior inertia parameters; A recommended path generation module, which is used to generate an optimal learning recommendation path based on the behavior inertia parameters and the sub-map according to the path recommendation engine. Among them, the path recommendation engine is used to generate Q candidate learning recommendation paths, and select the optimal learning recommendation path as the recommendation result based on the path fitness score. The candidate learning recommendation path is a sequence of course module nodes and knowledge point nodes arranged in order, and Q is a positive integer.

2. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that, Construct a time rhythm dispersion coefficient through the ratio of the standard deviation of the learning peak time points to the mean of the learning peak time points. Calculate the Pearson coefficient between the learning duration and the test scores of the learner within the first preset time period, and perform normalization processing through the sigmoid function to obtain the first intermediate feature. Combine the time rhythm dispersion coefficient and the first intermediate feature to obtain the learning rhythm stability coefficient; Based on the ratio of the difference between the number of test wrong questions of the learner and the mean of the number of test wrong questions to the standard deviation of the number of test wrong questions, construct an error rate deviation coefficient. Based on the learner's page stay time, the number of page rollbacks in a single learning session, and the average number of page rollbacks in a single learning session of other learners on the platform, construct a behavior operation anomaly coefficient. Combine the error rate deviation coefficient and the operation anomaly coefficient to obtain the cognitive load fluctuation coefficient; Obtain the second intermediate feature through the ratio of the learning duration of the learner on one type of resource to the total learning duration on all resource types, and obtain the third intermediate feature through the ratio of the average test score of the learner to the full score of the test. Combine the second intermediate feature and the third intermediate feature to obtain the resource type effectiveness characteristics.

3. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that, The association relationships in the task ability map include: One-to-many association between the job task node and the ability indicator node; One-to-many association between the ability indicator node and the course module node; One-to-many association between the course module node and the knowledge point node; Directed edges are established between the knowledge point nodes according to the front-to-back dependence relationship of the knowledge system; And generate directed edges between different types of nodes to construct a nested map structure of job task - ability indicator - course module - knowledge point.

4. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that, The mastery score of the ability index node is calculated based on the test scores associated with the ability index node. Among them, the test scores of the ability index node are normalized and weighted and summed to obtain the mastery score.

5. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 4, characterized in that, According to the mastery score and the behavioral inertia parameter, a sub-graph map is extracted from the task ability map. The specific steps include: S201, screening the ability index nodes with a mastery score greater than the first preset threshold and a learning rhythm stability coefficient greater than the second preset threshold; S202, extracting the job task nodes and course module nodes connected to them, and the knowledge point nodes connected to the course module nodes according to the screened ability index nodes; S203, obtaining a sub-graph map according to the screened ability index nodes, the extracted job task nodes, course module nodes and knowledge point nodes, and the edges between the nodes.

6. The intelligent learning path recommendation system based on user behavior big data analysis according to claim 5, wherein, The recommended path generation module imposes the following internal constraint conditions during the execution of the candidate learning recommendation path generation process by the path recommendation engine: The order of the course module nodes and knowledge point nodes is the sequential learning order between the course module nodes and knowledge point nodes in the sub-graph map; There is an edge between adjacent knowledge point nodes in the candidate learning recommendation path of the learner; The estimated total learning time of the candidate learning recommendation path of the learner does not exceed the fourth preset threshold.

7. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that, The recommended path generation module further includes: used to receive the intervention information input by the external port as the external constraint condition imposed during the execution of the candidate learning recommendation path generation process by the path recommendation engine. The intervention information includes: The forced priority learning order of the specified course module node; Specify that the candidate learning recommendation path includes specific knowledge point nodes; Specify that the estimated total learning time of the candidate learning recommendation path is not less than the fifth preset threshold.

8. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that The recommended path generation module performs structured encoding on the candidate learning recommendation path. The specific steps include: S301, hierarchically encoding the course module nodes and knowledge point nodes in the candidate learning recommendation path according to the dependency relationship in the sub-graph map to generate a node sequence with topological serial numbers; S302, mapping the resource type efficiency characteristics in the behavioral inertia parameter to node weight labels, and forming an encoding vector with the node sequence generated in S301.

9. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that, The calculation process of the path fitness score includes: S401, detecting the nodes in the node sequence that violate the dependency relationship in the sub-graph map as conflict nodes, assigning a penalty value to the conflict nodes, and generating an order rationality score in combination with the hierarchical depth of the conflict nodes; S402, setting the difficulty levels of the course module nodes and knowledge point nodes, counting the node pairs with a difficulty level difference of more than 2 levels between adjacent nodes in the candidate learning recommendation path, and calculating the difficulty gradient smoothness score through exponential decay; S403, performing weighted summation on the order rationality score and the difficulty gradient smoothness score to obtain the path fitness score.

10. The intelligent learning path recommendation system based on big data analysis of user behavior according to claim 1, characterized in that, The steps for the recommended path generation module to select the optimal learning recommendation path based on the path fitness score include: S501, arranging the Q candidate learning recommendation paths in descending order according to the path fitness score, and retaining a preset number of candidate learning recommendation paths; S502, perform a first mutation operation on the reserved candidate learning recommendation path with a first preset probability. The first mutation operation includes: Randomly select a course module node in the candidate learning recommendation path; Retrieve knowledge point nodes associated with the course module node from the sub-graph atlas but not in the candidate learning recommendation path, and filter out knowledge point nodes whose resource type efficiency characteristics are higher than the sixth preset threshold; Replace the knowledge point node connected to the course module node in the node sequence with the filtered knowledge point node; S503, recalculate the path fitness score for the candidate learning recommendation path after the first mutation operation; S504, repeatedly execute S501 to S503 until the improvement amplitude of the highest path fitness score in the candidate learning path does not exceed the seventh preset threshold in consecutive U rounds of iteration, and use the candidate learning recommendation path with the highest path fitness score in the U-th round as the optimal learning recommendation path, where U is a positive integer.

Images