A knowledge graph-based learning path generation method, device and medium

Abstract

The application discloses a kind of learning path generation method, equipment and medium based on knowledge graph, it is related to intelligent education data processing technical field, including: acquisition course standard, knowledge point, learning resource, evaluation task and historical learning behavior data, establish version learning knowledge graph, train timing diagram knowledge state model;According to learning path generation request and learner learning evidence sequence, obtain knowledge mastery level and state uncertainty degree, divide pre-evidence satisfied knowledge point, unmastered knowledge point and to be verified knowledge point, form target knowledge blocking subgraph;According to target knowledge blocking subgraph and knowledge state classification result, generate learning execution unit, form candidate learning path record;According to candidate learning path record, determine final learning path, generate learning path execution list.Knowledge state uncertainty and pre-evidence satisfaction state are distinguished, and learning path generation under the constraint of verification branch and predicted learning result is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent education data processing technology, and in particular to a method, device and medium for generating learning paths based on knowledge graphs. Background Technology

[0002] Online education platforms, course resource libraries, and assessment systems continuously accumulate data on curriculum standards, knowledge point structures, learning resources, assessment results, and learning behavior. Learning path generation technology is typically used to arrange subsequent learning content based on learning objectives and learners' historical performance. Conventional methods generally begin by representing the preconditions, inclusions, and resource coverage relationships between knowledge points using a knowledge graph. Then, by combining assessment scores, resource completion status, or knowledge status prediction results, learnable knowledge points and matching resources are selected, and recommended paths are formed according to course sequence, difficulty gradient, or expected learning time.

[0003] The aforementioned conventional methods still have two limitations in personalized path generation: First, learner status is mostly represented by recent scores or average mastery probability, making it difficult to distinguish between knowledge points where prior evidence has been met, knowledge points that have not been mastered, and knowledge points that require verification due to insufficient evidence, resulting in an unclear scope of blocking the target knowledge points; Second, path combinations tend to focus on resource matching and sequential recommendation, lacking joint constraints on verified and unverified branches as well as the expected level of mastery, resulting in insufficient evaluation of the completeness and feasibility of candidate paths. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a knowledge graph-based learning path generation method to solve the problems of learner knowledge state uncertainty, insufficient differentiation of state of prior evidence, and insufficient constraints of candidate learning paths on verification branches and expected learning results in the existing technology.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a learning path generation method based on a knowledge graph, comprising: collecting curriculum standards, knowledge points, learning resources, assessment tasks, and historical learning behavior data; establishing a versioned learning knowledge graph containing precondition groups and precondition evidence satisfaction conditions; and training a temporal graph knowledge state model; based on the learning path generation request and the learner's learning evidence sequence, using the temporal graph knowledge state model to obtain the knowledge mastery level and state uncertainty, classifying precondition evidence satisfied knowledge points, unmastered knowledge points, and unverified knowledge points, and traversing the versioned learning knowledge graph in reverse to form a target knowledge blocking subgraph; based on the target knowledge blocking subgraph and the knowledge state classification results, combining knowledge points, learning resources, and assessment tasks to generate learning execution units, performing constrained path search according to precondition coverage and expected learning results to form candidate learning path records; and determining the final learning path based on the path completeness, precondition coverage, expected target knowledge mastery level, and expected completion time in the candidate learning path records, and generating a learning path execution list containing execution order, learning resources, assessment tasks, and verification branches.

[0007] As a preferred embodiment of the knowledge graph-based learning path generation method described in this invention, the method involves collecting curriculum standard data, knowledge point data, learning resource data, assessment task data, and historical learning behavior data; unifying course identifiers, course versions, knowledge point identifiers, and behavior occurrence times; and establishing knowledge point entities, learning resource entities, and assessment task entities. The knowledge point entities are aligned according to knowledge point definitions, knowledge objectives, and assessment scope to establish resource coverage relationships between learning resource entities and knowledge point entities, and to establish evidence generation relationships between assessment task entities and knowledge point entities. The method also extracts data from the curriculum standard data. Forced prerequisite relationships are established by determining effective auxiliary prerequisite relationships based on the textbook's citation order and historical learning behavior data. Prerequisite knowledge points that jointly constrain the same subsequent knowledge point are written into the corresponding prerequisite condition group. Prerequisite evidence satisfaction conditions are set for forced prerequisite relationships and effective auxiliary prerequisite relationships based on the target mastery requirements, allowed learning evidence types, and learning evidence validity periods. A graph structure is used to combine knowledge point entities, learning resource entities, assessment task entities, prerequisite condition groups, and corresponding relationships, excluding effective auxiliary prerequisite relationships that constitute directed cycles. The course version, data update time, and graph effective time are recorded to form a versioned learning knowledge graph.

[0008] As a preferred embodiment of the knowledge graph-based learning path generation method of the present invention, historical learning behavior data is arranged according to learner identifiers and behavior occurrence times. Knowledge point identifiers, learning resource completion status, assessment results, answer time, prompt usage, and time intervals between adjacent behaviors are combined into a learning event sequence. The learning event sequence prior to the independent assessment is used as model input, and whether the independent assessment meets the corresponding knowledge point mastery requirements is used as a supervision label to generate a knowledge state model training sample. Based on the knowledge state model training sample, a temporal graph knowledge state model is established, including a learning event encoding layer, a temporal state extraction layer, a graph relation aggregation layer, and a knowledge state output layer. The mandatory and effective auxiliary prerequisite relations in the versioned learning knowledge graph are input into the graph relation aggregation layer, and adjacent knowledge point states are aggregated according to the corresponding relation weights to train the temporal graph knowledge state model. The final parameters of the temporal graph knowledge state model are determined based on the model validation results, and the trained temporal graph knowledge state model is associated with the graph version identifier of the corresponding versioned learning knowledge graph.

[0009] As a preferred embodiment of the knowledge graph-based learning path generation method of the present invention, the method includes: receiving a learning path generation request containing learner identifier, target knowledge point, planned learning duration, available learning resource type, and request time; generating a learning path generation request record; obtaining historical answer records, assessment results, learning resource completion records, prompt usage records, and answer time records prior to the request time based on the learner identifier, and arranging them according to the time of occurrence of the behavior to form a learner learning evidence sequence; inputting the learner learning evidence sequence into a time-series graph knowledge state model associated with the current version of the learning knowledge graph, aggregating the states of adjacent knowledge points according to mandatory prerequisite relationships and effective auxiliary prerequisite relationships, obtaining the knowledge mastery level and state uncertainty, and forming a learner knowledge state record; and, based on the knowledge mastery level, state uncertainty, effective learning evidence, and prerequisite evidence satisfaction conditions, writing the knowledge points within the prerequisite dependency range of the target knowledge point into the prerequisite evidence satisfied knowledge point set, the unmastered knowledge point set, and the unverified knowledge point set, respectively.

[0010] As a preferred embodiment of the knowledge graph-based learning path generation method of the present invention, the method involves taking the target knowledge point as the starting point for reverse traversal, obtaining the precondition groups and corresponding precondition points connecting the target knowledge point in the versioned learning knowledge graph; retaining all unsatisfied precondition points within the precondition groups that satisfy all types, and retaining the executable precondition point branches that are not yet satisfied for the precondition groups that select the types that satisfy; continuing to obtain the previous level precondition groups along the unmastered knowledge points and the knowledge points to be verified, and stopping the reverse traversal of the corresponding branch when reaching the knowledge point where the precondition evidence is satisfied and the starting knowledge point where there is no previous level precondition group; deleting knowledge points and precondition relationships that are not connected to the target knowledge point and the unsatisfied precondition groups, and retaining the target knowledge point, the unmastered knowledge point, the knowledge point to be verified, the precondition evidence satisfied at the branch termination position, and the corresponding precondition groups to form a target knowledge blocking subgraph.

[0011] As a preferred embodiment of the knowledge graph-based learning path generation method of the present invention, the method involves: acquiring unmastered knowledge points, knowledge points to be verified, target knowledge points, and a set of preconditions from the target knowledge blocking subgraph; writing knowledge points for which precondition evidence has been satisfied into an initial precondition coverage set; based on the knowledge mastery confidence lower bound of unmastered knowledge points, selecting learning resources from the versioned learning knowledge graph that are consistent with the course version, whose resource type matches the learning path generation request, are available, and whose difficulty is appropriate, and associating them with assessment tasks that meet the precondition conditions, thereby generating a learning verification execution unit; and based on the knowledge points to be verified... The uncertainty of the state of a point determines the verification order, associates it with the assessment tasks that meet the preconditions, generates a preliminary verification execution unit with verification pass and verification fail branches, and connects the verification fail branches to the corresponding learning resources and reassessment tasks; based on the credible lower bound of knowledge mastery of the target knowledge point, it filters learning resources and assessment tasks that cover the target knowledge point, and generates a target learning execution unit that includes the target mastery requirements; the learning verification execution unit, the preliminary verification execution unit, and the target learning execution unit are written into the target knowledge blocking subgraph according to the corresponding knowledge point and precondition group to form a learning execution graph.

[0012] As a preferred embodiment of the knowledge graph-based learning path generation method of the present invention, the initial precondition coverage set is used as the initial state for path search. Learning verification execution units, prior verification execution units, target learning execution units, and their corresponding precondition groups are obtained from the learning execution graph. For all precondition groups that satisfy the type, after all precondition points within the group are covered by the initial precondition coverage set and the learning execution units already added to the path, the corresponding subsequent learning execution units are expanded. For selected precondition groups that satisfy the type, learning execution units are expanded along each executable precondition point branch within the group, and the path branch status is updated based on verified and unverified branches. After each expansion of learning execution units, the precondition coverage, estimated completion time, estimated knowledge mastery level, and used learning parameters of the current path are updated. Resources and executed assessment tasks are identified. Expansion paths are selected based on the following criteria: availability of learning resources, unique learning execution units, unique evidence roles, difficulty matching of adjacent learning resources, and connection of failed verification branches to supplementary learning execution units. The selected expansion paths are then ranked according to the number of mandatory preconditions covered, the expected level of target knowledge mastery, the coverage of auxiliary preconditions, and the expected completion time, retaining intermediate paths that meet the expansion criteria. When both the target learning execution unit and the target assessment task are included, and all failed verification branches are connected to supplementary learning execution units, the corresponding expansion path is identified as a candidate learning path. The order of learning execution units, verification branches, precondition coverage, expected level of target knowledge mastery, and expected completion time in the candidate learning path are recorded to form a candidate learning path record.

[0013] As a preferred embodiment of the knowledge graph-based learning path generation method of the present invention, the method involves: acquiring candidate learning path records and learning path generation request records; performing consistency verification based on learner identifiers, target knowledge points, and graph version identifiers; filtering complete candidate learning paths based on path completeness status, target learning execution unit markers, target assessment task markers, and verification branch connection status; excluding complete candidate learning paths that do not cover mandatory precondition groups based on mandatory precondition coverage status; and filtering target reachable paths based on expected target knowledge mastery level and target mastery requirements; sorting target reachable paths according to expected target knowledge mastery level, auxiliary precondition coverage, expected completion time, number of learning execution units, and number of condition branches to determine the final learning path; and recording the execution order of learning execution units, learning resources, assessment tasks, the next learning execution unit after successful verification, and supplementary learning execution units after unsuccessful verification, based on the precondition relationships and execution connection relationships in the final learning path, thereby generating a learning path execution list.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the knowledge graph-based learning path generation method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the knowledge graph-based learning path generation method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By using a time-series graph knowledge state model to process learners' learning evidence sequences, the level of knowledge mastery and the degree of uncertainty in the state are obtained. Furthermore, the invention categorizes prior knowledge points into those that have been satisfied, those that have not been mastered, and those that need to be verified, thereby enabling the differentiation and judgment of learners' knowledge states and the sufficiency of evidence, and improving the accuracy of identifying states where prior conditions are met. By traversing the versioned learning knowledge graph backwards from the target knowledge point and combining the knowledge state classification results to form a target knowledge blocking subgraph, the invention achieves the centralized extraction of unsatisfied prior relationships and knowledge points that hinder the learning of the target knowledge point, thus improving the targeting of the range of target learning path generation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a knowledge graph-based learning path generation method.

[0019] Figure 2 A flowchart for the formation of knowledge state division and target knowledge blocking subgraph.

[0020] Figure 3 A flowchart for generating learning execution units and forming candidate learning paths.

[0021] Figure 4 A flowchart for determining the final learning path and generating the learning path execution list. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a learning path generation method based on knowledge graphs, including the following steps: S1. Collect curriculum standards, knowledge points, learning resources, assessment tasks and historical learning behavior data, establish a versioned learning knowledge graph that includes precondition groups and precondition evidence satisfaction conditions, and train a time sequence graph knowledge state model. Furthermore, data on curriculum standards, knowledge points, learning resources, assessment tasks, and historical learning behaviors are collected to unify course identifiers, course versions, knowledge point identifiers, and the time of occurrence of behaviors, and to establish knowledge point entities, learning resource entities, and assessment task entities.

[0026] Specifically, the system reads curriculum standard data from the course management platform, knowledge point data and textbook sequence data from the textbook resource library, learning resource data and learning resource completion records from the online learning platform, and assessment task data, assessment results, answer time records, and prompt usage records from the assessment management platform. It also standardizes the format of course identifiers, grade levels, course versions, and time fields in each data source, converts assessment scores into standardized scores between zero and one, converts resource completion rates into values ​​between zero and one, and deletes historical learning behavior data lacking learner anonymity identifiers, knowledge point identifiers, and the time of occurrence of the behavior.

[0027] The curriculum standard data includes curriculum identifier, grade level, curriculum version, curriculum objectives, teaching requirements, knowledge content, knowledge sequence requirements, and objective mastery requirements; knowledge point data includes knowledge point name, knowledge point definition, chapter, knowledge objectives, difficulty level, and applicable curriculum version; learning resource data includes learning resource identifier, resource type, resource content, covered knowledge points, resource difficulty, open time, access conditions, and resource completion records; assessment task data includes assessment task identifier, tested knowledge points, assessment questions, assessment type, passing conditions, answering time limit, and assessment results; historical learning behavior data includes learner anonymity identifier, knowledge point identifier, learning resource identifier, assessment task identifier, behavior occurrence time, resource completion rate, answering results, standardized score, answering time, and prompt usage.

[0028] It should be noted that the curriculum standard data uses the current, valid version of the curriculum standard; the historical learning behavior data uses anonymized data, and learners' names, ID numbers, and contact information are not read. The time range of the historical learning behavior data is usually the most recent one to three full academic years, in order to balance the number of training samples and the timeliness of the teaching content.

[0029] Furthermore, based on the knowledge point definition, knowledge objectives, and assessment scope, knowledge point entities are aligned to establish resource coverage relationships between learning resource entities and knowledge point entities, and evidence generation relationships between assessment task entities and knowledge point entities are established.

[0030] Specifically, the names of knowledge points are normalized using synonyms, and text semantic feature extraction methods are used to extract text semantic features for knowledge point names, definitions, objectives, and assessment scopes. Knowledge points with the same course identifier, grade level, and course version, and whose names are identical after synonym normalization, are directly aligned to the same knowledge point entity. Knowledge points with different names but whose comprehensive semantic similarity reaches the knowledge point alignment threshold are aligned to the same knowledge point entity after course knowledge review. The comprehensive semantic similarity is calculated based on the similarity of four types of text semantic features: knowledge point name, definition, objective, and assessment scope. A higher association weight is assigned based on the importance of the knowledge point definition and objective in describing the knowledge content, and then a comprehensive judgment is made in combination with the similarity of the knowledge point name and assessment scope. When any key feature in the four types of text features is significantly inconsistent, the comprehensive semantic similarity is reduced to avoid incorrect merging of knowledge points with similar names but different knowledge objectives or assessment scopes. Knowledge points with the same name but different knowledge objectives or assessment scopes are established as separate knowledge point entities.

[0031] Among them, the knowledge point entity includes a unique knowledge point identifier, course identifier, grade level, course version, knowledge point name, knowledge point definition, knowledge objective, knowledge point difficulty, and objective mastery requirements; the resource coverage relationship includes learning resource identifier, covered knowledge point identifier, knowledge objective coverage ratio, applicable course version, and resource difficulty; the evidence generation relationship includes assessment task identifier, examined knowledge point identifier, allowed learning evidence types, assessment pass conditions, and learning evidence validity period.

[0032] It should be noted that the knowledge point alignment threshold is determined based on knowledge point pairs that have been manually labeled as the same and different knowledge points. The threshold is selected to maximize the similarity between the harmonic mean of the knowledge point alignment precision and recall. The knowledge point alignment threshold is typically between 0.85 and 0.95. Knowledge point difficulty is represented by five levels, corresponding to beginner, basic, general, intermediate, and comprehensive application difficulty, respectively. When no difficulty label is set, the level is determined based on the standardized score distribution of learners in their first independent assessment of the corresponding knowledge point.

[0033] Furthermore, mandatory prerequisite relationships are extracted from the curriculum standard data, and effective auxiliary prerequisite relationships are determined based on the textbook citation order and historical learning behavior data. Prerequisite knowledge points that jointly constrain the same subsequent knowledge point are written into the corresponding prerequisite condition group.

[0034] Specifically, the curriculum standard data is used to extract the knowledge sequence requirements of "master first", "as a foundation", "learn after completion" and those with the same meaning. The knowledge sequence relationship clearly stipulated in the curriculum standard is written into the mandatory prerequisite relationship. Based on the order of textbook chapters and the learning time sequence of knowledge points in historical learning behavior, candidate auxiliary prerequisite relationships are generated. The pass rate of learners who meet the candidate prerequisite knowledge points and those who do not meet the candidate prerequisite knowledge points in the first independent assessment of subsequent knowledge points is statistically analyzed.

[0035] It should be noted that the historical support value for the auxiliary prerequisite relationships is determined by the difference between the pass rate of learners who have met the candidate prerequisite knowledge points in the first independent assessment of the subsequent knowledge points and the pass rate of learners who have not met the candidate prerequisite knowledge points in the first independent assessment of the subsequent knowledge points, expressed as: ; in, This indicates the historical support value of the auxiliary preceding relationship. This indicates the number of learners who have met the prerequisite knowledge points and achieved the target mastery requirement in the first independent assessment of the subsequent knowledge points. This represents the total number of learners who have met the prerequisite knowledge points and participated in the first independent assessment of subsequent knowledge points. This indicates the number of learners who did not meet the prerequisite knowledge points but achieved the target mastery requirement in the first independent assessment of the subsequent knowledge points. This represents the total number of learners who did not meet the prerequisite knowledge points and participated in the first independent assessment of subsequent knowledge points.

[0036] It should be noted that when zero or When the value is 0, historical support values ​​for auxiliary predecessor relationships are not calculated, and the corresponding candidate auxiliary predecessor relationships are not enabled; when... Less than thirty or When the number is less than 30, the corresponding candidate auxiliary prerequisite relations are retained as relations to be verified and are not written into the currently valid versioned learning knowledge graph; when and When all the required support values ​​for the auxiliary preceding relationships are no less than 30, and the historical support value of the auxiliary preceding relationship reaches the relationship support threshold, the candidate auxiliary preceding relationship is written into the valid auxiliary preceding relationship. The relationship support threshold is determined based on the sample of preceding relationships confirmed by manual review, and is usually between 0.6 and 0.8.

[0037] The minimum number of effective learners, 30, is determined by testing the stability of the sample size based on the historical prior relationship samples that have been manually reviewed and confirmed. The historical support value of the auxiliary prior relationship is repeatedly calculated according to different sample sizes, and the changes in the calculation results with the increase of the sample size are compared. The number of learners in each group when the historical support value of the auxiliary prior relationship begins to stabilize is determined as the minimum number of effective learners. In this embodiment, it is determined to be 30.

[0038] Furthermore, the prerequisite knowledge points that jointly constrain the same subsequent knowledge point are grouped into the corresponding prerequisite condition group, and the type of prerequisite condition group is determined according to the connection semantics in the curriculum standard.

[0039] Specifically, when the curriculum standard requires that multiple prerequisite knowledge points meet the corresponding requirements, the corresponding prerequisite knowledge points are written into the prerequisite condition group of all satisfying types; when the curriculum standard stipulates that learning can continue as long as any one of the multiple prerequisite knowledge points meets the requirements, the corresponding prerequisite knowledge point is written into the prerequisite condition group of the selected satisfying type; when a single prerequisite knowledge point independently constrains subsequent knowledge points, a prerequisite condition group of all satisfying types containing only the prerequisite knowledge points is established.

[0040] The precondition group includes a precondition group identifier, a precondition group type, a precondition knowledge point identifier, a subsequent knowledge point identifier, a precondition relationship type, and a relationship weight. All precondition knowledge points within the group that meet all type requirements satisfy the corresponding precondition evidence condition; at least one precondition knowledge point within the group that meets all type requirements satisfies the corresponding precondition evidence condition.

[0041] It should be noted that the relationship weight of the mandatory preceding relationship is set to 1, and the relationship weight of the effective auxiliary preceding relationship adopts the historical support value of the corresponding auxiliary preceding relationship, which is used for subsequent time series graph knowledge state model execution graph relationship aggregation, and for candidate learning paths to calculate the coverage of auxiliary preceding conditions.

[0042] Furthermore, based on the target mastery requirements, the permitted types of learning evidence, and the validity period of learning evidence, conditions for the fulfillment of prerequisite evidence are set for mandatory prerequisite relationships and effective auxiliary prerequisite relationships.

[0043] Specifically, the minimum mastery level is based on the achievement requirements in the curriculum standards; when the curriculum standards do not specify a value, the passing standardized score of the corresponding assessment task is used as the minimum mastery level; the types of learning evidence allowed are determined according to the assessment method of the assessment task; the validity period of learning evidence is determined based on the retesting pass rate of learners who have completed a valid assessment of a certain knowledge point and then participated in an independent assessment of the same knowledge point again.

[0044] The conditions for meeting the pre-existing evidence include the minimum level of mastery, the types of learning evidence allowed, the validity period of the learning evidence, and the requirement to mark direct learning evidence. The types of learning evidence allowed include independent assessment evidence, homework completion evidence, project task evidence, experimental operation evidence, and post-test evidence of learning resources.

[0045] It should be noted that the minimum mastery level typically ranges from 0.6 to 0.9, with basic knowledge points typically ranging from 0.6 to 0.75. Core knowledge points are determined based on the key mastery requirements in the curriculum standards, the scope of influence of prerequisite relationships in the versioned learning knowledge graph, and key nodes in the target knowledge path. Core knowledge points that directly influence multiple subsequent knowledge points typically range from 0.75 to 0.90. The validity period of learning evidence is typically seven to one hundred and eighty days. When the number of valid retest samples is less than thirty, the retesting cycle specified in the curriculum standards or teaching plan shall be adopted.

[0046] Furthermore, learning resource entities and assessment task entities are established, and resource coverage relationships and evidence generation relationships are established with knowledge point entities respectively.

[0047] Specifically, the learning resource entity records the learning resource identifier, covered knowledge points, resource type, resource difficulty, applicable course version, opening time, access conditions, estimated completion time, and expected learning gain; the assessment task entity records the assessment task identifier, tested knowledge points, allowed learning evidence types, assessment pass conditions, learning evidence validity period, and estimated completion time. The estimated completion time for learning resources is determined using the median of historical learners' effective completion times, and the estimated completion time for assessment tasks is determined using the median of historical learners' effective response times.

[0048] It should be noted that the expected learning gain of learning resources is calculated using the following formula: ; in, This represents the expected learning gain from the learning resources. Indicates the first The percentage of knowledge mastery of each valid paired sample before using the learning resources. Indicates the first The proportion of knowledge mastery of each effective paired sample after using the learning resources. Indicates the number of valid paired samples. This means that the calculation result will be limited to the range of 0 to 1.

[0049] It should be noted that valid paired samples should meet the following conditions: the pre-test and post-test examine the same knowledge point; the post-test occurs within seven days of the completion of the learning resource; the difference between the pre-test knowledge mastery ratio and the full-score mastery ratio is not less than the minimum mastery resolution; and no other learning resources covering the same knowledge point are used between the pre-test and post-test. The minimum mastery resolution is determined based on the smallest distinguishable score interval in the standardized scores of the corresponding assessment task. If the difference between the pre-test knowledge mastery ratio and the full-score mastery ratio is less than the minimum mastery resolution, the corresponding paired samples will not be used to calculate the expected learning gain of the learning resource. When the number of valid paired samples for a single learning resource is less than thirty, only learning resource samples that are applicable to the same course version, resource type, and difficulty level, cover the same knowledge points, have the same knowledge objective coverage, the same content granularity, and are applicable to the same learner's grade level and current ability level will be merged. The lower quartile of the merged sample calculation result will be used as the expected learning gain. Learning resource samples that do not meet the aforementioned merging conditions will not participate in the calculation of the expected learning gain of the current learning resource.

[0050] Furthermore, a graph structure is used to combine knowledge point entities, learning resource entities, assessment task entities, prerequisite condition groups and their corresponding relationships, excluding effective auxiliary prerequisite relationships that constitute directed cycles, and recording course version, data update time and graph effective time to form a versioned learning knowledge graph.

[0051] Specifically, directed cycle detection is performed on the mandatory prerequisite relations and effective auxiliary prerequisite relations in the initial learning knowledge graph. When a directed cycle consisting of effective auxiliary prerequisite relations is detected, the corresponding effective auxiliary prerequisite relations are removed from the current effective graph in order of their historical support values ​​from low to high, until the directed cycle is resolved. When a directed cycle containing mandatory prerequisite relations is detected, the corresponding mandatory prerequisite relations are submitted for course knowledge review, and the corresponding graph version is not released until the correct knowledge sequence is determined.

[0052] The versioned learning knowledge graph includes the graph version identifier, curriculum standard version, textbook version, knowledge point data update time, learning resource data update time, assessment task data update time, graph effective time, and the previous graph version identifier.

[0053] It should be noted that only one versioned learning knowledge graph in a valid state is used during the generation of the same learning path to avoid inconsistencies in the coverage of knowledge point entities, prerequisite groups, and learning resources caused by changes in course versions.

[0054] Furthermore, historical learning behavior data is arranged according to learner identification and the time of occurrence of the behavior, and knowledge point identification, completion status of learning resources, assessment results, answer time, use of prompts, and time intervals between adjacent behaviors are combined into a learning event sequence.

[0055] Specifically, the sequence of learning events prior to the independent assessment is used as the model input, and whether the independent assessment meets the target mastery requirements of the corresponding knowledge point is used as the supervision label to generate training samples for the knowledge state model; the supervision label is one when the target mastery requirements are met, and zero when the target mastery requirements are not met.

[0056] The knowledge state model training samples are grouped according to learner identifiers. All knowledge state model training samples corresponding to the same learner identifier are written into only one set of the model training set, model validation set, and model test set to avoid the same learner's learning behavior sequence participating in model training and model validation simultaneously. The sample size ratio of the model training set, model validation set, and model test set is usually set to seven:2:1. When the number of historical learners is small, grouping is first completed according to learner identifiers, and then the grouping results are adjusted while maintaining a basic consistency in the course version, knowledge point coverage, and supervision label ratio in different sets.

[0057] In addition, the maximum length of the learning event sequence for a single model training sample is determined based on the distribution of historical sequence lengths. The length that can cover no less than 90% of the historical learning event sequences is selected, usually between 20 and 100 learning events. When the number of learning events exceeds the maximum length, the learning event closest to the sample deadline is retained. When the number of learning events is less than the maximum length, the mask position is used to fill in the gaps.

[0058] Furthermore, a temporal graph knowledge state model is established based on the training samples of the knowledge state model, including a learning event encoding layer, a temporal state extraction layer, a graph relation aggregation layer, and a knowledge state output layer.

[0059] Specifically, the learning event encoding layer converts knowledge point identifiers, behavior types, standardized scores, resource completion rates, answering time, prompt usage, and time intervals into learning event features; the temporal state extraction layer adopts a gated loop unit structure to recursively process the learning event features according to the order of learning events to extract the learner's temporal learning state; the graph relationship aggregation layer aggregates the states of adjacent knowledge points according to the preceding relationships in the versioned learning knowledge graph; and the knowledge state output layer outputs the knowledge mastery probability of each knowledge point.

[0060] It should be noted that the temporal state extraction layer reads the features of each learning event sequentially according to their occurrence order, and generates the current hidden state based on the historical hidden state corresponding to the previous learning event, the features of the current learning event, the update gate, the reset gate, and the candidate hidden states. The update gate controls the proportion of historical hidden states retained in the current hidden state, and the reset gate controls the proportion of historical hidden states participating in the candidate hidden state calculation. The current hidden state corresponding to the last valid learning event serves as the learner's temporal learning state. Learning events that are padded using mask positions do not participate in hidden state updates.

[0061] Among them, the graph relation aggregation layer uses the relation weights of the mandatory preceding relation and the relation weights of the effective auxiliary preceding relation together with the graph attention coefficient for the aggregation of adjacent knowledge point states, so that the mandatory preceding relation and the effective auxiliary preceding relation with higher historical support have a greater impact on knowledge state reasoning.

[0062] It should be noted that the graph attention coefficient is not a fixed relation weight preset by humans, but a normalized aggregation coefficient obtained by the graph relation aggregation layer during the training of the temporal graph knowledge state model. It is used to represent the relative influence of different preceding knowledge point states corresponding to the same subsequent knowledge point in this aggregation.

[0063] The graph relationship aggregation layer embeds and calculates attention scores based on the state features of adjacent knowledge points and the type of preceding relationships.

[0064] Specifically, the features of subsequent knowledge point states, preceding knowledge point states, and preceding relation types are embedded and combined. A linear mapping is then performed using trainable parameters to obtain the attention score for each preceding relation. The attention scores for all preceding relations pointing to the same subsequent knowledge point are normalized to ensure a consistent sum of graph attention coefficients for all preceding relations corresponding to the same subsequent knowledge point, resulting in graph attention coefficients. Relation weights do not participate in attention score calculation but are used in adjacent knowledge point state aggregation after graph attention coefficient normalization. This is used to adjust the influence of different preceding relation state information on the inference of subsequent knowledge point states. The value range of a single graph attention coefficient is 0~1, and the sum of graph attention coefficients for all input preceding relations corresponding to the same subsequent knowledge point is 1. The graph attention coefficients are dynamically generated based on the model training parameters and the knowledge state features corresponding to the learner's learning evidence sequence, and participate in adjacent knowledge point state aggregation together with the relation weights of mandatory preceding relations and effective auxiliary preceding relations.

[0065] Among them, the relation weight is only used for weighting the aggregation result corresponding to the graph attention coefficient, and will not participate in the calculation of other aggregation weights again.

[0066] Furthermore, the mandatory prerequisite relations and effective auxiliary prerequisite relations in the versioned learning knowledge graph are input into the graph relation aggregation layer, and the states of adjacent knowledge points are aggregated according to the corresponding relation weights to train the time sequence graph knowledge state model.

[0067] Specifically, five sub-models with identical structures are established. Each sub-model is trained using training samples drawn with replacement from the model training set and uses different initial model parameters. The error between the knowledge state output and the supervision label is calculated using the binary cross-entropy loss function, and the model parameters are updated using the adaptive moment estimation optimization method.

[0068] It should be noted that the temporal graph knowledge state model does not directly call a general off-the-shelf model with a fixed structure. Instead, it is a task-adapted model built based on existing mature temporal feature extraction networks, graph neural network aggregation mechanisms, and supervised learning training frameworks. The combination of the learning event encoding layer, temporal state extraction layer, graph relation aggregation layer, and knowledge state output layer is determined based on the mandatory preconditions, effective auxiliary preconditions, and learner learning evidence sequences in the versioned learning knowledge graph. The learning event feature dimension is typically set to 32 to 128, the temporal state hiding dimension is typically set to 64 to 256, the number of graph relation aggregation layers is typically set to one to three, the random deactivation ratio is typically set to 0.1 to 0.5, the model training batch size is typically set to 32 to 256, and the initial learning rate is typically set to 10%. -4 Up to 10 -3After each training round, the model validation set loss is calculated. Training is stopped when the reduction in model validation set loss is less than 0.001 for ten consecutive training rounds. The maximum number of training rounds is set to two hundred rounds.

[0069] The range of values ​​for the model training parameters is determined based on the parameter tuning results of the model validation set, the scale of historical learning behavior data, and the resource constraints of the training equipment. After training with different combinations of candidate parameters, the final parameters of the time series graph knowledge state model are determined based on the model validation set loss, the prediction accuracy of knowledge mastery level, and the calibration effect of state uncertainty.

[0070] The calibration effect of state uncertainty is determined based on the consistency between the state uncertainty output by the model and the actual independent assessment results. Specifically, the knowledge mastery probability output by the model is divided into multiple confidence intervals according to a preset interval. The deviation between the predicted mastery probability and the actual assessment pass rate within each confidence interval is statistically analyzed, and the results are weighted and summarized according to the sample size of each interval to obtain the state uncertainty calibration result. The state uncertainty calibration effect is used to evaluate whether the uncertainty output by the model can reflect the actual prediction error range and is related to the subsequent calibration process of the uncertainty adjustment coefficient in the knowledge mastery confidence lower bound. After the model parameters are determined, the uncertainty adjustment coefficient is determined based on the state uncertainty calibration result using the model validation set, so that the proportion of the knowledge mastery confidence lower bound covering the actual independent assessment results reaches the preset requirement.

[0071] The range of model training parameters is determined based on the model validation set tuning results, the scale of historical learning behavior data, and the resource constraints of the training device. When there are many learning events, many knowledge points, and sufficient GPU memory on the training device, a larger dimension of learning event features, a larger dimension of temporal state hiding, a larger model training batch, and a larger number of graph relationship aggregation layers are adopted. When the model validation set loss fluctuates greatly or the training set loss decreases while the model validation set loss increases, the random deactivation ratio is increased, and the initial learning rate is reduced within the initial learning rate range. After training with different candidate parameter combinations, the final parameters of the temporal graph knowledge state model are determined based on the model validation set loss, the prediction accuracy of knowledge mastery level, and the calibration effect of state uncertainty.

[0072] Furthermore, based on the model validation results, the final parameters of the time sequence graph knowledge state model are determined, and the trained time sequence graph knowledge state model is associated with the graph version identifier of the corresponding versioned learning knowledge graph.

[0073] Specifically, the structural parameters, final model parameters, input feature order, maximum length of the learning event sequence, and corresponding graph version identifier of the five trained sub-models are combined to form a time-series graph knowledge state model.

[0074] S2. Based on the requests generated by the learning path and the evidence sequence of the learner's learning, the knowledge mastery level and state uncertainty are obtained by using the time sequence graph knowledge state model. The knowledge points that have been satisfied by the prior evidence, the knowledge points that have not been mastered, and the knowledge points that need to be verified are divided into knowledge points that have been satisfied by the prior evidence, knowledge points that have not been mastered, and knowledge points that need to be verified. The versioned learning knowledge graph is traversed in reverse to form the target knowledge blocking subgraph. Furthermore, it receives a learning path generation request containing learner identifier, target knowledge point, planned learning duration, available learning resource type, and request time, and generates a learning path generation request record.

[0075] Specifically, based on the unique knowledge point identifier of the target knowledge point, the corresponding knowledge point entity is read from the versioned learning knowledge graph that is in a valid state, and it is verified whether the course version to which the target knowledge point belongs is consistent with the current course version in the learning path generation request; if they are consistent, the graph version identifier of the versioned learning knowledge graph is written into the learning path generation request record for use by this knowledge state reasoning and the generation of the target knowledge blocking subgraph.

[0076] The learning path generation request record includes learner identifier, target knowledge point, planned learning duration, available learning resource type, current course version, request time, and graph version identifier.

[0077] It should be noted that the versioned learning knowledge graph is not switched during the same learning path generation process; if the target knowledge point does not exist in the current course version, the current processing is terminated and an invalid target knowledge point is generated to avoid using the same knowledge point in other course versions to replace the target knowledge point.

[0078] Furthermore, based on the learner's identifier, historical answer records, assessment results, learning resource completion records, prompt usage records, and answer time records prior to the request time are obtained and arranged according to the time of the behavior to form a sequence of learner learning evidence.

[0079] Specifically, historical answer records and assessment results are linked to assessment task entities and corresponding knowledge point entities, and learning resource completion records are linked to learning resource entities and corresponding knowledge point entities; only learning events whose course version is consistent with the current version of the learning knowledge graph and can be linked to knowledge point entities, learning resource entities, or assessment task entities are retained.

[0080] Each learning event includes a knowledge point identifier, a learning resource identifier or assessment task identifier, behavior type, standardized score, resource completion rate, answer time, prompt usage, behavior occurrence time, and time interval between the previous learning event.

[0081] It should be noted that multiple responses to the same assessment task are recorded, including each response and its occurrence time, to reflect changes in the learner's knowledge status over time. Duplicate records with the same time, the same task, and the same result due to page refresh or repeated reporting are only retained once. When the number of learning events exceeds the maximum length of the learning event sequence specified by the time sequence diagram knowledge state model, the learning event closest to the requested time is retained. If the number of learning events is insufficient, mask positions are used to fill in the gaps.

[0082] Furthermore, the learner's learning evidence sequence is input into a time-series graph knowledge state model associated with the current version of the learning knowledge graph. Adjacent knowledge point states are aggregated according to mandatory preconditions and effective auxiliary preconditions to obtain the knowledge mastery level and the degree of state uncertainty, thus forming a learner knowledge state record.

[0083] Specifically, the temporal graph knowledge state model converts each learning event into learning event features through the learning event encoding layer, extracts the learner's temporal learning state according to the time of behavior occurrence through the temporal state extraction layer, and aggregates the states of adjacent knowledge points along the mandatory and effective auxiliary prerequisite relationships in the versioned learning knowledge graph through the graph relationship aggregation layer.

[0084] It should be noted that the level of knowledge mastery is calculated according to the following formula: ; in, This indicates the current level of knowledge mastery of the knowledge point. Indicates the first Each sub-model outputs the probability of mastering the current knowledge point. The knowledge mastery level ranges from 0 to 1, with a higher value indicating a higher probability that the learner has mastered the current knowledge point.

[0085] It should be noted that the knowledge mastery level is obtained by summing the mastery probabilities output by multiple sub-models for the current knowledge point. Specifically, the average of the mastery probabilities output by each sub-model is used as the knowledge mastery level for the current knowledge point.

[0086] The degree of uncertainty is calculated according to the following formula: ; in, This indicates the degree of uncertainty in the current state of a knowledge point. Indicates the first Each sub-model outputs the probability of mastery for the current knowledge point. This indicates the current level of knowledge mastery for a particular knowledge point. The greater the uncertainty in the state, the greater the difference in the judgments of different sub-models regarding the current level of knowledge mastery.

[0087] It should be noted that the state uncertainty is used to represent the degree of dispersion among the output results of multiple sub-models, and essentially reflects the consistency of the prediction results of different sub-models on the current knowledge mastery state. After normalization, the state uncertainty is used as a dimensionless uncertainty index, and its value range corresponds to the knowledge mastery level, and is used to measure the reliability of the knowledge mastery probability.

[0088] The reliable lower bound of knowledge mastery is calculated according to the following formula: ; in, This represents the lower bound of the current knowledge mastery. Indicates the level of knowledge mastery. Indicates the degree of uncertainty in the state. It represents the uncertainty adjustment coefficient, which is used to calibrate the scale of the influence of the degree of uncertainty in the state, so that the level of knowledge and the degree of uncertainty are on the same adjustment scale.

[0089] It should be noted that the uncertainty adjustment coefficient was calibrated using the model validation set, with candidate values ​​set at intervals of 0.1 within the range of 0.5 to 2.5. The smallest candidate value was selected that ensured the confidence lower bound of knowledge mastery was not higher than the sample proportion of actual mastery results in subsequent independent assessments, reaching 90% to 95%. The confidence lower bound of knowledge mastery was used to deduct the uncertainty caused by differences in model output, avoiding the direct identification of knowledge points with insufficient evidence as having met the preconditions based solely on the average mastery probability.

[0090] Furthermore, based on the level of knowledge mastery, the degree of uncertainty of the state, the effective learning evidence, and the conditions for satisfying the prior evidence, the knowledge points within the scope of the prior dependence of the target knowledge point are respectively written into the set of knowledge points for which the prior evidence has been satisfied, the set of knowledge points for which the prior evidence has not been mastered, and the set of knowledge points for which the prior evidence is to be verified.

[0091] Specifically, the permitted learning evidence type, assessment pass conditions, and learning evidence validity period are retrieved from the corresponding assessment task entity and evidence generation relationship. If the learning evidence type belongs to the permitted learning evidence type, the learning evidence result meets the corresponding assessment pass conditions, and the interval between the request time and the learning evidence occurrence time does not exceed the learning evidence validity period, the learning evidence is determined to be valid learning evidence.

[0092] Direct learning evidence includes independent assessment evidence, project task evidence, and experimental operation evidence formed for the current knowledge point; post-test evidence of learning resources and homework completion evidence are judged according to the allowed learning evidence types in the conditions for satisfying the prior evidence.

[0093] Furthermore, for knowledge points that require marking as valid direct learning evidence, only independent assessment evidence, project task evidence, and experimental operation evidence will be used as the basis for satisfying the prerequisite evidence; for knowledge points that require marking as invalid direct learning evidence, post-test evidence from learning resources and homework completion evidence that conform to the permitted learning evidence types will be used as the basis for satisfying the prerequisite evidence.

[0094] Specifically, when a knowledge point has valid learning evidence that meets the preconditions and the lower bound of knowledge mastery reaches the corresponding minimum mastery level, the knowledge point is added to the set of knowledge points for which the preconditions have been met. When a knowledge point has learning evidence that is within its validity period and the level of knowledge mastery is lower than the corresponding minimum mastery level, the knowledge point is added to the set of knowledge points that have not been mastered. When a knowledge point does not have valid learning evidence that meets the preconditions, or when the level of knowledge mastery reaches the corresponding minimum mastery level but the lower bound of knowledge mastery does not reach the corresponding minimum mastery level, the knowledge point is added to the set of knowledge points to be verified.

[0095] It should be noted that knowledge points for which there is no valid learning evidence should not be included in the set of knowledge points for which prior evidence has been satisfied, solely based on the graph relation aggregation results. When multiple valid learning evidences exist for the same knowledge point, the valid learning evidence closest to the request time is used as the basis for state division, while earlier valid learning evidences continue to be retained in the learner's learning evidence sequence to participate in the state reasoning of the time-series graph knowledge state model. For new learners without any historical learning behavior data, knowledge points for which there is no valid learning evidence within the scope of the target knowledge point's prior dependencies are included in the set of knowledge points to be verified.

[0096] Furthermore, using the target knowledge point as the starting point for reverse traversal, we obtain the precondition groups and corresponding preconditions that connect the target knowledge point in the versioned learning knowledge graph.

[0097] Specifically, for all satisfying type precondition groups, all unsatisfied precondition knowledge points within the group are retained; for selected satisfying type precondition groups, it is first determined whether there are precondition knowledge points that have been satisfied by the evidence within the group. When there is at least one precondition knowledge point that has been satisfied by the evidence within the group, the selected satisfying type precondition group is determined to be satisfied, and other unsatisfied precondition knowledge points within the group are not further expanded; when there is no precondition knowledge point that has been satisfied by the evidence within the group, the branch of the precondition knowledge point that has not yet been satisfied but meets the path expansion condition is retained; the previous level precondition group is obtained along the unmastered knowledge points and the knowledge points to be verified, and the reverse traversal of the corresponding branch is stopped when the precondition knowledge point that has been satisfied by the evidence and the starting knowledge point that does not have a previous level precondition group are reached.

[0098] Among them, multiple prerequisite knowledge points in the type prerequisite condition group are considered as replaceable prerequisite branches. This step does not predetermine a unique branch to avoid deleting prerequisite branches that may have higher expected learning effects based solely on the current estimated completion time.

[0099] It should be noted that the versioned learning knowledge graph has already completed directed cycle detection in step one, so the same knowledge points will not be accessed repeatedly in the same preceding relation path during the reverse traversal.

[0100] Furthermore, knowledge points and prerequisite relationships that are not connected to the target knowledge point and the unmet prerequisite condition group are deleted, while the target knowledge point, the unmastered knowledge point, the knowledge point to be verified, the prerequisite evidence that the branch termination position has been satisfied, and the corresponding prerequisite condition group are retained to form a target knowledge blocking subgraph.

[0101] Specifically, invalid branches formed by knowledge points that have already been satisfied by prior evidence are deleted, while knowledge points, precondition groups, and precondition relationships that still block the learning of the target knowledge point are retained. The target knowledge blocking subgraph is then associated with the learning path generation request record, learner knowledge status record, and graph version identifier.

[0102] The target knowledge blocking subgraph includes the target knowledge point, the unmastered knowledge point, the knowledge point to be verified, the knowledge point whose premise evidence for the branch termination position has been satisfied, the premise condition group, the premise relation type, the relation weight, the traversal level, and the reason for the branch termination.

[0103] It should be noted that when there are no unmastered knowledge points or knowledge points to be verified, the target knowledge blocking subgraph only retains the target knowledge points and the pre-existing evidence states that have been met. Subsequent steps still generate target assessment tasks to confirm whether the target knowledge points have met the target mastery requirements.

[0104] S3. Based on the target knowledge blocking subgraph and knowledge state classification results, combine knowledge points, learning resources and assessment tasks to generate learning execution units, and perform constrained path search according to the coverage of preconditions and expected learning results to form candidate learning path records. Furthermore, the unmastered knowledge points, unverified knowledge points, target knowledge points, and precondition groups in the target knowledge blocking subgraph are obtained, and the knowledge points whose precondition evidence has been satisfied are written into the initial precondition coverage set.

[0105] Specifically, based on the graph version identifier in the learning path generation request record, knowledge point entities, precondition groups, precondition evidence fulfillment conditions, learning resource entities, assessment task entities, resource coverage relationships, and evidence generation relationships in the same version of the learning knowledge graph are read to avoid mixing knowledge points, learning resources, and assessment tasks from different course versions in candidate learning paths.

[0106] The knowledge status classification results include the set of knowledge points for which prior evidence has been satisfied, the set of knowledge points not yet mastered, and the set of knowledge points to be verified; the learner's knowledge status record includes the knowledge mastery level, the degree of uncertainty of the status, and the credible lower bound of knowledge mastery for each knowledge point; the learning path generation request record includes the target knowledge point, the planned learning duration, the type of available learning resources, and the request time.

[0107] Furthermore, based on the credible lower bound of knowledge mastery for the unmastered knowledge points, learning resources with consistent course versions, resource types that match the learning path generation request, are available, and have appropriate difficulty are selected from the versioned learning knowledge graph. These resources are then associated with assessment tasks that meet the pre-existing evidence conditions to generate learning verification execution units.

[0108] Specifically, learning resources that are consistent with the current course version in the learning path generation request record, belong to the available learning resource type, are requested within the learning resource's open time, and whose current access conditions are met by the learner are retained; the learner's current ability level is determined based on the current knowledge mastery level in the learner's knowledge status record, and the resource difficulty level in the learning resource entity is compared with the learner's current ability level, retaining learning resources whose difficulty level differs from the learner's current ability level by no more than one level; from the assessment task entities connected to unmastered knowledge points, assessment tasks whose evidence type belongs to the allowed learning evidence type in the corresponding pre-evidence satisfaction conditions, and whose assessment pass conditions can determine whether the knowledge point has reached the minimum mastery level are selected.

[0109] The learning and verification execution unit includes the learning execution unit identifier, execution unit type, corresponding knowledge point, learning resources, post-test task, expected completion time, knowledge mastery level before execution, expected knowledge mastery level after execution, verification pass conditions, precondition coverage results after verification pass, and supplementary execution relationships after verification fail.

[0110] It should be noted that the learner's current ability level is determined based on the level of knowledge mastery recorded in the learner's knowledge status record. A level of knowledge mastery below 60% is determined as Level 1, 60% to 75% as Level 2, 75% to 90% as Level 3, and above 90% as Level 4. The difficulty level of the learning resources is marked with difficulty levels 1 to 5 in the knowledge point entity. If no difficulty mark is set, the level is determined based on the learner's historical improvement in knowledge mastery after completing the resources.

[0111] It should also be noted that if the learning resource has already been fully studied by the current learner and the completion time is still within the validity period of the learning evidence for the corresponding knowledge point, the same learning resource will not be selected again; the assessment task should cover at least 60% of all knowledge objectives for the corresponding knowledge point, and the assessment task for the core knowledge point should cover at least 80% of all knowledge objectives.

[0112] Furthermore, the expected knowledge mastery level after implementing the learning resources is calculated based on the expected learning gain of the learning resources, and is expressed as follows: ; in, This indicates the expected level of knowledge mastery before using the current learning resources. This indicates the expected level of knowledge mastery after using the current learning resources. This represents the expected learning gain of the current learning resource, and the values ​​mentioned above all range from 0 to 1.

[0113] It should be noted that the expected knowledge mastery level is used for comparison between candidate learning paths and does not directly replace the actual learning evidence generated by the post-test task; only after the post-test task meets the verification conditions will the corresponding knowledge point be written into the precondition coverage set during the path execution process.

[0114] Furthermore, the verification order is determined based on the degree of uncertainty of the state of the knowledge points to be verified, and assessment tasks that meet the conditions for satisfying the prior evidence are associated. Preliminary verification execution units with verification pass branches and verification fail branches are generated, and the verification fail branches are connected to the corresponding learning resources and reassessment tasks.

[0115] Specifically, the knowledge points to be verified are arranged from highest to lowest level of uncertainty. When the level of uncertainty is the same, they are arranged from lowest to highest level of the number of prerequisite relationship layers between the knowledge points to be verified and the target knowledge points. From the assessment task entities corresponding to the knowledge points to be verified, assessment tasks that meet the prerequisite evidence satisfaction conditions, are not lower than the minimum mastery level, and are in a usable state are selected to generate preliminary verification execution units.

[0116] The preliminary verification execution unit includes the knowledge point to be verified, the preliminary assessment task, the estimated completion time of the assessment task, the verification pass conditions, the verification pass branch, and the verification fail branch. The verification pass branch is used to write the knowledge point to be verified into the precondition coverage set, and the verification fail branch is used to connect the learning resources, post-test tasks, and re-verification tasks of the coverage knowledge point.

[0117] It should be noted that when there is no qualified preliminary assessment task for the knowledge point to be verified, but there are available learning resources and post-test tasks, the knowledge point to be verified will be treated as an unmastered knowledge point and a learning verification execution unit will be generated directly. When there is neither a valid assessment task nor available learning resources, an incomplete learning execution unit will not be generated, and the corresponding knowledge point will be written into the path to generate an obstruction record.

[0118] Furthermore, based on the credible lower bound of knowledge mastery of the target knowledge points, learning resources and assessment tasks covering the target knowledge points are selected to generate target learning execution units that include the target mastery requirements.

[0119] Specifically, learning resources covering the target knowledge points and assessment tasks examining the target knowledge points are read from the versioned learning knowledge graph. These resources are then filtered according to course version, learning resource type, opening time, access conditions, resource difficulty, and assessment task availability. Learning resources with the same or slightly different difficulty level as the recommended resources are selected, along with assessment tasks whose knowledge target coverage reaches the specified proportion of the target knowledge points. The target knowledge points, learning resources, and assessment tasks are then combined into a target learning execution unit.

[0120] The target learning execution unit includes target knowledge points, target learning resources, target assessment tasks, expected completion time, expected target knowledge mastery level, and target verification pass conditions; the target verification pass conditions adopt the target mastery requirements in the target knowledge point entity.

[0121] It should be noted that when the target knowledge point already has direct learning evidence within its validity period and the lower bound of knowledge mastery meets the target mastery requirements, only the target assessment task can be set for verification. If the verification is successful, a candidate learning path containing only the target verification process will be formed. If the verification fails, the target learning resources and the reassessment task will be connected.

[0122] Furthermore, the learning verification execution unit, the preliminary verification execution unit, and the target learning execution unit are written into the target knowledge blocking subgraph according to the corresponding knowledge points and precondition groups to form a learning execution graph.

[0123] Specifically, the learning execution unit is used as the execution node in the learning execution graph, and the satisfaction relationship, verification pass relationship, and verification fail relationship of the prerequisite condition group are used as the directed connection relationship between the execution nodes; knowledge points whose prerequisite evidence has been satisfied do not generate learning execution units, but only serve as the covered termination node of the corresponding prerequisite branch; knowledge points that have not been mastered are connected to the learning verification execution unit; knowledge points to be verified are connected to the preceding verification execution unit and its verification fail branch; target knowledge points are connected to the target learning execution unit.

[0124] The learning execution graph includes learning execution units, directed connections between execution nodes, an initial set of precondition coverage, a set of precondition groups to be processed, and records of blocked path generation.

[0125] Furthermore, using the initial precondition coverage set as the initial state for path search, the learning verification execution unit, the prior verification execution unit, the target learning execution unit, and their corresponding precondition groups are obtained in the learning execution graph.

[0126] Specifically, the executable starting node is a learning execution unit that does not meet the previous level's prerequisite group. For all prerequisite groups that meet the type, after all prerequisite knowledge points in the group are covered by the initial prerequisite coverage set and the learning execution units already added to the path, the corresponding subsequent learning execution units are expanded. For prerequisite groups that select the type that meet the type, when there are no prerequisite knowledge points that have been met in the group, the learning verification execution units and preliminary verification execution units corresponding to the prerequisite knowledge points in the group are selected to have available learning resources or available assessment tasks, the course version of the learning resources and assessment tasks is consistent with the current version of the learning knowledge graph, the current learner meets the learning resource access conditions, and the corresponding learning resources, assessment tasks, and evidence functions do not overlap with the learning execution units already added to the current path. The selected prerequisite knowledge points are determined as executable prerequisite knowledge point branches, and the learning execution units are expanded along each executable prerequisite knowledge point branch. The path branch status is updated according to the verified branches and the unverified branches.

[0127] Each intermediate path records the learning execution units that have been added, the knowledge points that have been covered, the preconditions that have been met, the preconditions that have not yet been processed, the cumulative estimated completion time, the estimated knowledge mastery level of each knowledge point, and the learning resources that have been used.

[0128] Furthermore, after each extended learning execution unit, the current path's precondition coverage, estimated completion time, estimated knowledge mastery level, used learning resources, and executed assessment tasks are updated.

[0129] Specifically, the expansion paths are filtered according to the following conditions: availability of learning resources, no duplication of learning execution units, no duplication of evidence roles, difficulty matching of adjacent learning resources, and connection of supplementary learning execution units to branches that have failed verification. All mandatory preconditions of the current learning execution unit have been covered, there is no duplicate addition of the same learning execution unit in the same intermediate path, and there are no multiple assessment tasks targeting the same knowledge point, used to satisfy the same set of preconditions, and with overlapping validity periods of evidence in the same path branch that are repeatedly included in the precondition coverage results. The difficulty difference between adjacent learning resources does not exceed one level, and the same knowledge point only enters the learning state once in the same path branch.

[0130] It should be noted that "no duplication of evidence" means that multiple assessment tasks within the same path branch cannot simultaneously serve as valid learning evidence for the same knowledge point and the same set of preconditions. Subsequent assessment tasks or reassessments performed after a previous assessment task fails to pass verification do not constitute duplicate evidence; the previous assessment task is only used to determine the branch that failed verification and is no longer included in the precondition coverage results. When multiple assessment tasks generate valid learning evidence, only the assessment task closest to the corresponding path execution position and meeting the precondition conditions is included in the precondition coverage results.

[0131] It should also be noted that the planned learning time is not a hard condition for stopping the expansion of intermediate paths, so as to avoid not being able to obtain candidate learning paths that exceed the planned learning time but can fully meet the target mastery requirements; when the cumulative estimated completion time exceeds the planned learning time, a planned learning time exceeding the planned learning time will be written into the intermediate path for subsequent steps to make a decision.

[0132] Furthermore, the number of mandatory preconditions covered is determined based on the number of mandatory preconditions already satisfied in the intermediate path; the degree of coverage of auxiliary preconditions is determined based on the historical support values ​​of effective auxiliary preconditions already covered in the intermediate path.

[0133] Specifically, the degree of coverage of auxiliary preconditions is expressed as: ; in, Indicates the degree of coverage of auxiliary prerequisites. This indicates the number of valid auxiliary preceding relations involved in the current candidate path. Indicates the first The overlay flag for a valid auxiliary preceding relationship is set to 1 if it is overlaid and 0 if it is not overlaid. Indicates the first Historical support values ​​for effective auxiliary preceding relationships.

[0134] It should be noted that when the current candidate path does not involve a valid auxiliary prerequisite relationship, the coverage of the auxiliary prerequisite conditions is set to 1 to avoid the inability to generate calculation results due to the absence of an auxiliary prerequisite relationship.

[0135] Furthermore, according to the execution order of the learning execution units, the expected knowledge mastery level after the execution of each learning resource is calculated sequentially, and the expected coverage result of the preceding knowledge points is passed to the subsequent knowledge points according to the prior relationship. When multiple learning resources are applied to the same knowledge point consecutively, the expected knowledge mastery level after the execution of the previous learning resource is used as the basis for the calculation of the next learning resource, and the expected knowledge mastery level is updated recursively according to the execution order of the learning resources. The expected learning gain of the same learning resource is calculated only once in the same path, and the same assessment evidence that has formed an effective prior condition coverage result is not used again for the calculation of the expected knowledge mastery level improvement.

[0136] Specifically, for branches that pass verification in the preliminary verification execution unit, the expected knowledge mastery level for the corresponding knowledge point is determined to be no less than the value required to pass the verification of the assessment task; for branches that fail verification, the original knowledge mastery level is retained, and the expected knowledge mastery level is recalculated after supplementing learning resources; when the intermediate path contains multiple reachable termination branches, the expected completion time and expected target knowledge mastery level of each termination branch are calculated separately, and the largest expected completion time among the termination branches is taken as the expected completion time of the intermediate path, and the smallest expected target knowledge mastery level among the termination branches is taken as the expected target knowledge mastery level of the intermediate path.

[0137] It should be noted that using the maximum expected completion time and the minimum expected target knowledge mastery level can evaluate paths with conditional branches according to unfavorable validation results, avoiding overestimating the performance of candidate learning paths based solely on validated branches.

[0138] Furthermore, the selected expansion paths are arranged according to the number of mandatory preconditions covered, the expected level of knowledge mastery of the target, the degree of coverage of auxiliary preconditions, and the expected completion time, and intermediate paths that meet the path expansion conditions are retained.

[0139] Specifically, after each layer of path expansion is completed, the intermediate paths are arranged in order of decreasing number of mandatory prerequisites covered, decreasing expected target knowledge mastery level, decreasing auxiliary prerequisites covered, increasing expected completion time, and increasing number of learning execution units. The preset number of intermediate paths at the top are retained for further expansion.

[0140] It should be noted that the preset number is determined based on the number of knowledge points in the target knowledge blocking subgraph and the available computing resources of the learning path generation device, and is usually five to twenty. Too few preset numbers may result in premature deletion of executable pre-branch, while too many preset numbers will increase the computational load of path search.

[0141] Furthermore, when both the target learning execution unit and the target evaluation task are included, and all branches that fail validation are connected to the supplementary learning execution unit, the corresponding extension path is determined as the candidate learning path.

[0142] Specifically, the path search stops when all intermediate paths have formed complete candidate learning paths, or when there are no more learning execution units that can be added; when there are learning execution unit sequences with the same content, resource identifier, evaluation task and execution order in the same candidate learning path, only the candidate learning path with the shorter expected completion time is retained; when the expected completion time is the same, only the candidate learning path generated earlier is retained.

[0143] Furthermore, the sequence of learning execution units, verification branches, coverage of prerequisites, expected level of mastery of target knowledge, and expected completion time in the candidate learning paths are recorded to form a candidate learning path record.

[0144] The candidate learning path record includes the candidate path number, learning execution unit and execution order, verified branches, verified branches, path completeness marker, mandatory precondition coverage marker, auxiliary precondition coverage level, expected target knowledge mastery level, expected completion time, number of learning execution units, number of conditional branches, target learning execution unit marker, target assessment task marker, and planned learning time exceeded marker.

[0145] It should be noted that when no complete candidate learning path exists, a path generation obstruction record is generated based on the knowledge points that failed to establish a learning execution unit or failed to complete the precondition closure. The path generation obstruction record is only generated when no complete candidate learning path exists, and is used to output the path generation obstruction result in subsequent steps. It is not used as a candidate learning path to participate in the adjudication.

[0146] S4. Based on the path integrity status, precondition coverage, expected target knowledge mastery level, and expected completion time in the candidate learning path records, determine the final learning path and generate a learning path execution list that includes execution order, learning resources, assessment tasks, and verification branches.

[0147] Furthermore, candidate learning path records and learning path generation request records are obtained, and consistency verification is performed based on learner identifiers, target knowledge points, and graph version identifiers.

[0148] Specifically, it verifies whether the learner identifier, target knowledge point, and graph version identifier associated with the candidate learning path record are consistent with the learning path generation request record; it verifies whether the learning resource type in the candidate learning path is an available learning resource type, and verifies the availability status of the learning resources and assessment tasks again based on the request time.

[0149] The availability status verification includes whether the learning resources are offline, whether the open period has expired, whether the access permissions have changed, whether the assessment tasks have been discontinued, and whether the course version is still consistent with the current graph version.

[0150] It should be noted that the re-verification of the availability status of learning resources and assessment tasks is used to handle resource decommissioning, expiration of open periods, and changes in access permissions that occur between the completion of the candidate learning path search and the determination of the final learning path. It does not modify the learning resource entities and assessment task entities in the versioned learning knowledge graph.

[0151] Furthermore, complete candidate learning paths are selected based on the path integrity status, target learning execution unit markers, target evaluation task markers, and verification branch connection status.

[0152] Specifically, if the target learning execution unit in the candidate learning path is marked as valid, the target evaluation task is marked as valid, each preliminary verification execution unit has a verification passed branch and a verification failed branch, each verification passed branch can reach the target learning execution unit directly or through subsequent learning execution units, each verification failed branch is connected to the corresponding supplementary learning execution unit and can reach the target learning execution unit through the supplementary learning execution unit, and there are no interrupted branches in the candidate learning path that cannot reach the target learning execution unit, then the corresponding candidate learning path is determined as a complete candidate learning path.

[0153] The supplementary learning execution unit includes learning resources, post-test tasks, and re-verification tasks set up for knowledge points that failed the verification.

[0154] It should be noted that if a candidate learning path lacks a target learning execution unit, lacks a target evaluation task, has a failed verification branch that is not connected to a supplementary learning execution unit, or has an interrupted branch that cannot reach the target learning execution unit, the corresponding candidate learning path will be marked as invalid and will no longer participate in the final learning path decision.

[0155] Furthermore, complete candidate learning paths for groups of mandatory preconditions that are not covered are excluded based on the mandatory precondition coverage status, and target reachable paths are selected based on the expected target knowledge mastery level and target mastery requirements.

[0156] Specifically, all mandatory prerequisite relations in the target knowledge blocking subgraph are compared item by item with the mandatory prerequisite relations already covered by the candidate learning path; when the knowledge point corresponding to the mandatory prerequisite relation already belongs to the set of knowledge points whose prerequisite evidence is satisfied, or when a learning execution unit that can generate the condition that the prerequisite evidence is satisfied has been set, the corresponding mandatory prerequisite relation is determined to be covered; the prerequisite condition group of all satisfying types is determined to be covered only when all mandatory prerequisite knowledge points in the group are covered, and the prerequisite condition group of selecting satisfying types is determined to be covered when at least one mandatory prerequisite knowledge point in the group is covered.

[0157] The coverage of auxiliary preconditions is not used as a rejection criterion for the validity of candidate learning paths, but is only used to rank multiple candidate learning paths that all meet the mandatory preconditions.

[0158] It should be noted that the coverage of auxiliary prerequisites does not replace the mandatory prerequisite relationships explicitly stipulated in the curriculum standards; the expected level of mastery of target knowledge is only used for the adjudication of candidate learning paths and does not serve as actual proof that learners have mastered the target knowledge points. In the final implementation process, actual assessment results still need to be obtained through target assessment tasks.

[0159] Furthermore, the target attainable paths are sorted according to the expected level of knowledge mastery, the coverage of supporting prerequisites, the expected completion time, the number of learning execution units, and the number of conditional branches, to determine the final learning path.

[0160] Specifically, candidate learning paths are first ranked from highest to lowest according to the expected level of knowledge mastery; if the expected level of knowledge mastery is the same, they are ranked from highest to lowest according to the coverage of auxiliary prerequisites; if the coverage of auxiliary prerequisites is still the same, they are ranked from shortest to longest expected completion time; if the expected completion time is still the same, they are ranked from fewest to most learning execution units; if the number of learning execution units is still the same, they are ranked from fewest to most conditional branches; if all the above results are the same, they are ranked from smallest to largest according to the candidate path number, and the candidate learning path ranked first is determined as the final learning path.

[0161] Among them, the candidate path numbers are consecutively numbered according to the generation time of the candidate learning path records; when the generation time of the candidate learning path records is the same, the order of numbering is determined according to the expansion order of the learning execution units during the path search process.

[0162] It should be noted that using a hierarchical sorting method instead of a weighted summation of multiple evaluation indicators can avoid the ambiguous impact of different indicator dimensions and artificial weight settings on the final path, and can make the process of determining the final learning path explain and verify item by item.

[0163] Furthermore, when the estimated completion time of the target reachable path exceeds the planned learning time in the learning path generation request record, the timed-out final learning path is determined, and a flag indicating insufficient planned learning time is generated.

[0164] Specifically, the expected overtime is calculated using the following formula: ; in, This indicates that the duration is expected to exceed the allotted time. This indicates the estimated completion time of the final learning path after the timeout. This indicates the planned learning duration in the learning path generation request.

[0165] It should be noted that when the expected time exceeds 0, the insufficient planned learning time and the expected time exceeding the target will be output together with the learning path execution list to prompt learners to adjust the planned learning time. The mandatory prerequisite knowledge points and learning execution units necessary to achieve the mastery requirements will not be deleted.

[0166] Furthermore, when there is no complete candidate learning path expected to achieve the target mastery requirement, a result indicating that the target mastery requirement is unattainable is generated; when there is no complete candidate learning path, a path-generated obstruction record is read, and a path-generated obstruction result is generated.

[0167] Specifically, when there are complete candidate learning paths but the expected level of mastery of the target knowledge does not meet the target mastery requirements, they are arranged from high to low according to the expected level of mastery of the target knowledge. When the expected level of mastery of the target knowledge is the same, they are arranged from most to least number of mandatory preconditions covered, from high to low degree of coverage of auxiliary preconditions, and from short to long expected completion time. The candidate learning path ranked first is determined as the reference learning path, and the reference learning path, the expected level of mastery of the target knowledge, and the difference between the target mastery requirements and the actual level are written into the result of unachievable target mastery requirements.

[0168] The path generation obstruction results include the obstructed knowledge point, the corresponding prerequisite group, the reason for the lack of learning resources, the reason for the lack of assessment tasks, the reason for the course version mismatch, and the reason for the resource access conditions not being met. When there are multiple obstruction reasons for the same obstructed knowledge point, each obstruction reason is written into the path generation obstruction results.

[0169] It should be noted that the result of "the learning objectives cannot be achieved" means that the existing available learning resources, assessment tasks, and planning conditions are insufficient to form a complete candidate learning path to achieve the expected learning objectives, and does not mean that the learner is actually unable to master the target knowledge points; when there is no complete candidate learning path, an incomplete learning path execution list will not be generated.

[0170] Furthermore, based on the preconditions and execution connections in the final learning path, the execution order, learning resources, assessment tasks, the next learning execution unit after successful verification, and the supplementary learning execution unit after unsuccessful verification are recorded to generate a learning path execution list.

[0171] Specifically, the first learning execution unit is set as the learning execution unit that does not meet the prerequisite group of the previous level. When there are multiple independent and executable learning execution units, they are arranged from the highest to the lowest according to the number of prerequisite relationship layers between the corresponding knowledge point and the target knowledge point, so that the basic knowledge points that are farther away from the target knowledge point are executed first. When the number of prerequisite relationship layers is the same, they are arranged from the shortest to the longest according to the expected completion time of the learning execution units. When the expected completion time is still the same, they are arranged according to the order of the learning execution unit identifiers.

[0172] The learning path execution list includes a list identifier, learner identifier, target knowledge point, map version identifier, final path number, planned learning duration, estimated completion duration, insufficient planned learning duration marker, and learning execution units arranged according to execution sequence number. Each learning execution unit includes an execution unit identifier, execution sequence number, execution unit type, corresponding knowledge point, learning resources, assessment task, estimated completion duration, verification pass conditions, the next learning execution unit after verification pass, and supplementary learning execution units after verification fail.

[0173] Furthermore, the branches that pass verification and those that fail verification are written into the learning path execution list.

[0174] Specifically, when the actual evaluation result of the preliminary verification execution unit or the learning verification execution unit meets the verification pass condition, the user jumps to the corresponding subsequent learning execution unit according to the next learning execution unit field after verification pass, and marks the corresponding knowledge point as the knowledge point whose preconditions have been met in the current learning path execution process; when the actual evaluation result does not meet the verification pass condition, the user enters the corresponding supplementary learning execution unit according to the supplementary learning execution unit field after verification failure, and enters the re-verification task after the supplementary learning execution unit is completed.

[0175] When the re-verification task meets the verification pass conditions, it proceeds to the corresponding subsequent learning execution unit; when the re-verification task still fails to meet the verification pass conditions, it reads other unexecuted branches in the currently selected type prerequisite condition group and checks whether there are alternative learning resources, alternative assessment tasks, or supplementary learning execution units that meet the path expansion conditions; when there are alternative execution branches, it marks the current verification failure branch as an unexecutable branch and returns to the candidate learning path search process to continue expanding alternative execution branches; when there are no alternative execution branches, it stops the continued execution of the current branch and generates a current knowledge point verification failure mark.

[0176] It should be noted that when generating the learning path execution list in this step, only the verification branches and supplementary learning execution units that have already been formed in the candidate learning path record are registered. No new learning resources, assessment tasks, or knowledge point branches are added during the final path determination process.

[0177] Furthermore, the learning path execution list is sent to the learning terminal.

[0178] Specifically, the learning terminal displays the currently executable learning execution units according to the execution sequence number, and hides the subsequent learning execution units that have not yet met the prerequisites. After the current assessment task is completed, the terminal reads the next learning execution unit after verification or the supplementary learning execution unit after verification failure based on the actual assessment results, and displays the corresponding knowledge point name, learning resource name, resource type, estimated completion time, assessment task name, and verification conditions.

[0179] When there is a marker indicating insufficient planned learning time, the learning terminal will also display the estimated time exceeding the limit. The execution order in the learning path execution list is used to control the opening order of learning execution units, and the verification branch is used to control the subsequent execution direction after the assessment is completed. Together, they constitute a learning path that can be executed directly.

[0180] It should be noted that the final output may be one of the following: a learning path execution list, an unattainable target mastery requirement, or a path generation obstruction result. If there is a complete candidate learning path that meets the target mastery requirement, the output will be a learning path execution list. If there is no complete candidate learning path that meets the target mastery requirement but a complete candidate learning path exists, the output will be an unattainable target mastery requirement result. If there is no complete candidate learning path, the output will be a path generation obstruction result.

[0181] This embodiment also provides a computer device applicable to the knowledge graph-based learning path generation method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the knowledge graph-based learning path generation method proposed in the above embodiment.

[0182] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0183] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the knowledge graph-based learning path generation method as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as 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.

[0184] In summary, this invention achieves the following by: using a time-series graph knowledge state model to process learner learning evidence sequences, obtaining knowledge mastery levels and state uncertainty, and classifying prior evidence into satisfied knowledge points, unmastered knowledge points, and unverified knowledge points, thereby distinguishing and judging learners' knowledge states and the sufficiency of evidence and improving the accuracy of identifying precondition satisfaction states; and by traversing the versioned learning knowledge graph backwards from the target knowledge point and combining the knowledge state classification results to form a target knowledge blocking subgraph, thereby achieving the centralized extraction of unsatisfied preconditions and unverified knowledge points that hinder the learning of the target knowledge point, and improving the targeting of the target learning path generation range.

[0185] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A learning path generation method based on knowledge graphs, characterized in that, include: Collect curriculum standards, knowledge points, learning resources, assessment tasks, and historical learning behavior data to establish a versioned learning knowledge graph that includes precondition sets and precondition evidence satisfaction conditions, and train a time-series graph knowledge state model. Based on the requests generated by the learning path and the evidence sequence of learners' learning, a time sequence graph knowledge state model is used to obtain the knowledge mastery level and the degree of state uncertainty. The prior evidence is divided into knowledge points that have been satisfied, knowledge points that have not been mastered, and knowledge points that need to be verified. The versioned learning knowledge graph is traversed in reverse to form a target knowledge blocking subgraph. Based on the target knowledge blocking subgraph and knowledge state classification results, knowledge points, learning resources and assessment tasks are combined to generate learning execution units. Constrained path search is performed according to the coverage of preconditions and the expected learning results to form candidate learning path records. Based on the path completeness status, precondition coverage, expected target knowledge mastery level, and expected completion time in the candidate learning path records, the final learning path is determined, and a learning path execution list containing the execution order, learning resources, assessment tasks, and verification branches is generated.

2. The knowledge graph-based learning path generation method as described in claim 1, characterized in that, The process of establishing a versioned learning knowledge graph that includes precondition sets and precondition evidence satisfaction conditions by collecting curriculum standards, knowledge points, learning resources, assessment tasks, and historical learning behavior data includes: Collect curriculum standard data, knowledge point data, learning resource data, assessment task data, and historical learning behavior data; unify course identifiers, course versions, knowledge point identifiers, and behavior occurrence times; and establish knowledge point entities, learning resource entities, and assessment task entities. Align knowledge point entities according to the knowledge point definition, knowledge objectives and assessment scope, establish resource coverage relationship between learning resource entities and knowledge point entities, and establish evidence generation relationship between assessment task entities and knowledge point entities. Extract mandatory prerequisite relationships from curriculum standard data, determine effective auxiliary prerequisite relationships based on textbook citation order and historical learning behavior data, and write prerequisite knowledge points that jointly constrain the same subsequent knowledge point into the corresponding prerequisite condition group. Based on the target mastery requirements, the permitted types of learning evidence, and the validity period of learning evidence, conditions for the fulfillment of prerequisite evidence are set for mandatory prerequisite relationships and effective auxiliary prerequisite relationships. The knowledge point entities, learning resource entities, assessment task entities, prerequisite groups and their corresponding relationships are combined into a graph structure, and effective auxiliary prerequisite relationships that form directed cycles are excluded. The course version, data update time and graph effective time are recorded to form a versioned learning knowledge graph.

3. The knowledge graph-based learning path generation method as described in claim 1, characterized in that, The training time-series graph knowledge state model includes: Historical learning behavior data is arranged according to learner identification and the time of occurrence of the behavior. Knowledge point identification, completion status of learning resources, assessment results, answer time, use of prompts, and time interval between adjacent behaviors are combined into a learning event sequence. The sequence of learning events prior to the independent assessment is used as the model input, and whether the independent assessment meets the target mastery requirements of the corresponding knowledge point is used as the supervision label to generate training samples for the knowledge state model. A temporal graph knowledge state model is established based on the training samples of the knowledge state model, including a learning event encoding layer, a temporal state extraction layer, a graph relation aggregation layer, and a knowledge state output layer. The mandatory prerequisite relations and effective auxiliary prerequisite relations in the versioned learning knowledge graph are input into the graph relation aggregation layer, and the states of adjacent knowledge points are aggregated according to the corresponding relation weights to train the time sequence graph knowledge state model. The final parameters of the time sequence graph knowledge state model are determined based on the model validation results, and the trained time sequence graph knowledge state model is associated with the graph version identifier of the corresponding versioned learning knowledge graph.

4. The knowledge graph-based learning path generation method as described in claim 3, characterized in that, The process, which involves generating requests based on the learning path and obtaining evidence sequences from learners' learning, uses a time-series graph knowledge state model to acquire knowledge mastery levels and state uncertainty, and categorizes prior evidence-satisfied knowledge points, unmastered knowledge points, and knowledge points to be verified as follows: Receive a learning path generation request containing learner identifier, target knowledge point, planned learning duration, available learning resource type, and request time, and generate a learning path generation request record; Based on the learner's identifier, retrieve historical answer records, assessment results, learning resource completion records, prompt usage records, and answer time records prior to the requested time, and arrange them according to the time of the behavior to form a sequence of learner learning evidence; The learner's learning evidence sequence is input into the time-series graph knowledge state model associated with the current version of the learning knowledge graph. The states of adjacent knowledge points are aggregated according to the mandatory preconditions and the effective auxiliary preconditions to obtain the knowledge mastery level and the degree of state uncertainty, thus forming a learner's knowledge state record. Based on the level of knowledge mastery, the degree of uncertainty of the state, the effective learning evidence, and the conditions for meeting the prior evidence, the knowledge points within the scope of the prior dependence of the target knowledge point are respectively written into the set of knowledge points whose prior evidence has been met, the set of knowledge points whose prior evidence has not been mastered, and the set of knowledge points to be verified.

5. The knowledge graph-based learning path generation method as described in claim 4, characterized in that, The reverse traversal of the versioned learning knowledge graph to form the target knowledge blocking subgraph includes: Starting from the target knowledge point, we obtain the preconditions and corresponding preconditions that connect the target knowledge point in the versioned learning knowledge graph. For all condition groups that satisfy the type, retain all unsatisfied prerequisite knowledge points within the group; for condition groups that select the type that satisfy the type, retain the executable prerequisite knowledge point branches that have not yet been satisfied. Continue to obtain the previous level of prerequisite groups along the knowledge points that are not yet mastered and the knowledge points that need to be verified, and stop the reverse traversal of the corresponding branch when the knowledge point where the prerequisite evidence has been satisfied and the starting knowledge point where there is no previous level of prerequisite group is reached. Delete knowledge points and prerequisite relationships that are not connected to the target knowledge point and the group of prerequisites that are not met, and retain the target knowledge point, the knowledge point that is not mastered, the knowledge point to be verified, the prerequisite evidence that the branch termination position is met, and the corresponding group of prerequisites to form a target knowledge blocking subgraph.

6. The knowledge graph-based learning path generation method as described in claim 1 or 5, characterized in that, The step of generating a learning execution unit by combining knowledge points, learning resources, and assessment tasks based on the target knowledge blocking subgraph and knowledge state classification results includes: Obtain the unmastered knowledge points, unverified knowledge points, target knowledge points, and precondition groups from the target knowledge blocking subgraph, and write the knowledge points whose precondition evidence has been satisfied into the initial precondition coverage set. Based on the credible lower bound of knowledge mastery for unmastered knowledge points, learning resources with consistent course versions, resource types that match the learning path generation request, are available and have appropriate difficulty are selected from the versioned learning knowledge graph. These resources are then associated with assessment tasks that meet the pre-existing evidence conditions to generate learning verification execution units. The verification order is determined based on the uncertainty of the state of the knowledge point to be verified. The assessment tasks that meet the conditions of the prior evidence are associated, and a preliminary verification execution unit with a verification pass branch and a verification fail branch is generated. The verification fail branch is then connected to the corresponding learning resources and the reassessment task. Based on the credible lower bound of knowledge mastery of the target knowledge points, learning resources and assessment tasks covering the target knowledge points are selected to generate target learning execution units that include the target mastery requirements. The learning verification execution unit, the preliminary verification execution unit, and the target learning execution unit are written into the target knowledge blocking subgraph according to the corresponding knowledge points and precondition groups to form the learning execution graph.

7. The knowledge graph-based learning path generation method as described in claim 6, characterized in that, The process of performing constrained path search based on the coverage of preconditions and the expected learning results to form candidate learning path records includes: Using the initial precondition coverage set as the initial state for path search, we obtain the learning verification execution unit, the prior verification execution unit, the target learning execution unit and their corresponding precondition groups in the learning execution graph. For all the precondition groups that meet the type, after all the preconditions in the group are covered by the initial precondition coverage set and the learning execution units that have been added to the path, the corresponding subsequent learning execution units are expanded. For the selected prerequisite condition group, the learning execution unit is expanded along each executable prerequisite knowledge point branch within the group, and the path branch status is updated according to the verified and unverified branches. After each extended learning execution unit, update the current path's precondition coverage, estimated completion time, estimated knowledge mastery level, used learning resources, and executed assessment tasks. The expansion paths are filtered based on the following conditions: availability of learning resources, unique learning execution units, unique evidence role, difficulty matching of adjacent learning resources, and connection of branches that failed verification to supplement learning execution units. The selected expansion paths are arranged according to the number of mandatory preconditions covered, the expected level of knowledge mastery of the target, the degree of coverage of auxiliary preconditions, and the expected completion time, and intermediate paths that meet the path expansion conditions are retained. When both the target learning execution unit and the target assessment task are included, and all branches that fail validation are connected to the supplementary learning execution unit, the corresponding extension path is determined as the candidate learning path. Record the order of learning execution units, verification branches, coverage of prerequisites, expected level of mastery of target knowledge, and expected completion time in the candidate learning path to form a candidate learning path record.

8. The knowledge graph-based learning path generation method as described in claim 1 or 7, characterized in that, The process of determining the final learning path based on the path completeness status, precondition coverage, expected target knowledge mastery level, and expected completion time in the candidate learning path records, and generating a learning path execution list that includes execution order, learning resources, assessment tasks, and verification branches, includes: Obtain candidate learning path records and learning path generation request records, and perform consistency verification based on learner identifier, target knowledge point, and graph version identifier; Based on the path integrity status, target learning execution unit marking, target evaluation task marking, and verification branch connection status, complete candidate learning paths are selected; Complete candidate learning paths for groups of mandatory preconditions that are not covered are excluded based on the coverage status of mandatory preconditions, and target attainable paths are selected based on the expected level of target knowledge mastery and target mastery requirements. The target reachable paths are sorted according to the expected level of knowledge mastery, the coverage of supporting prerequisites, the expected completion time, the number of learning execution units, and the number of condition branches, and the final learning path is determined. Based on the preconditions and execution connections in the final learning path, the execution order, learning resources, assessment tasks, the next learning execution unit after successful verification, and the supplementary learning execution unit after unsuccessful verification are recorded to generate a learning path execution list.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the knowledge graph-based learning path generation method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the knowledge graph-based learning path generation method according to any one of claims 1 to 8.

Images