Self-evolution online learning personalized recommendation method

Abstract

The invention relates to the technical field of network data processing, in particular to a self-evolution online learning personalized recommendation method, which comprises the following steps of: acquiring behavior data of a learner during online learning in real time; constructing a simplified knowledge graph of the knowledge system of each subject by adopting a hierarchical decoupling method; a behavior continuity merging algorithm is adopted to obtain learning behaviors of the learner at each knowledge point based on behavior data mapping; fitting by adopting a time sequence capability vector fitting algorithm based on the behavior data to obtain the learning capability of the learner at each knowledge point; periodically determining a first recommended learning mode of each knowledge point based on the learning behavior and learning ability of the group learner at each knowledge point; the first recommendation learning mode of each knowledge point in the current period is corrected based on the learning behavior and learning ability of an individual learner to serve as a personalized recommendation result, so that'group efficiency + individual adaptation 'dual-drive recommendation is realized, meanwhile, dynamic standard correction and real-time response can be realized, and the consistency and accuracy of recommendation logic are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of network data processing, and in particular to a self-evolving online learning personalized recommendation method. Background Art

[0002] Personalized learning is an effective means of improving the effectiveness of all types of learning and training. With the widespread adoption of the internet, acquiring knowledge and developing skills online has become mainstream. However, the development of technologies supporting personalized learning still lags behind that of online teaching. With the advent of the cloud era, big data has attracted increasing attention. Big data has brought three major changes to online learning: first, it helps identify the factors that truly influence online learning; second, it provides insights into learners' real-world situations; and third, it facilitates the development of personalized learning. By leveraging big data technology, by analyzing learners' online learning behaviors, it is possible to understand individual characteristics to a certain extent, helping teachers and administrators to more objectively understand the learning situation. Based on this, they can recommend online learning methods and content to learners, guiding them to engage in personalized learning, thereby improving the efficiency and effectiveness of online learning.

[0003] At present, knowledge graph technology has been effectively applied in many fields as an important means of visually describing knowledge resources and their carriers. Applying knowledge graph technology to the representation of subject knowledge systems is an effective method, such as Chinese patent CN114861069A, a network learning resource analysis and personalized recommendation method based on knowledge graph, a learning resource concept linking, analysis and evaluation technology based on knowledge graph, and a learner knowledge system evaluation model and learning path intelligent planning based on knowledge graph and learning goals, to achieve knowledge push and personalized learning resource and learning strategy recommendation centered on learners' personalized interests and needs. However, this patent still inevitably has the following technical problems: 1. High complexity in knowledge graph construction: Due to the inherent perception in existing technologies that "full coverage = precise recommendation", existing technologies usually use directed acyclic graphs with full coverage to construct knowledge graphs. This patent relies on transfer learning and multi-domain knowledge dependencies. Its knowledge point centrality algorithm needs to traverse the 1-m order neighbors of each node, resulting in high computational complexity and difficulty in meeting real-time requirements. In addition, the complexity and full domain support characteristics of the existing knowledge graph itself also make it consume too much computing resources, making it unsuitable for the representation of knowledge in small and medium-sized disciplines.

[0004] 2. Static Evaluation Criteria: This patent recommends knowledge points based on their centrality. This centrality calculation relies on a fixed weight matrix and neighboring node contributions, inevitably causing recommendations to lag behind changes in learners' abilities. Alternative technical approaches typically considered by those skilled in the art include deep learning-based recommendations (such as the Transformer model). However, this undoubtedly requires higher computing power and lacks interpretability, making it unsuitable for personalized recommendations in online learning scenarios.

[0005] 3. Broken behavioral data mapping: The patent does not specify the rules for collecting behavioral data. Its explicit data is essentially subjective labels that cannot quantify behavioral details, while implicit data only records isolated events, breaking the behavioral logic chain. Therefore, the patent directly adopts explicit / implicit data classification, which will undoubtedly lead to a broken learning behavior mapping. The resulting fragmented learning behavior makes it easy for the analysis of online learning resources to deviate from the actual level.

[0006] 4. Separation of path recommendation and knowledge dependency: While this patent clarifies the dependencies between knowledge points when constructing the knowledge graph, these dependencies are only used to construct the topological structure of the graph and are not used as mandatory constraints in subsequent path recommendations. At the same time, dynamic heterogeneous information networks mine "semantic associations" through random walks, which can easily include weakly associated or even unnecessary knowledge points into recommended paths, further weakening the structural nature of the knowledge system and leading to confusion between semantic associations and logical dependencies. Summary of the Invention

[0007] In response to the above technical problems, the present invention proposes a self-evolving online learning personalized recommendation method, which aims to avoid the technical problem of "separation of path recommendation and knowledge dependence", improve the logical coherence of recommendation, and realize the dual-driven recommendation of "group effectiveness + individual adaptation". At the same time, it can realize dynamic standard correction and real-time response to improve recommendation accuracy.

[0008] In a first aspect, the present application provides a self-evolving online learning personalized recommendation method, comprising the following steps: Collect learners’ behavioral data during online learning in real time; Adopting the hierarchical decoupling method to construct a simplified knowledge graph of each subject's knowledge system; Adopting the behavior continuity merging algorithm and the preset semantic mapping library, the learner's learning behavior of each knowledge point on the simplified knowledge graph is obtained based on the behavior data mapping; Using a time-series ability vector fitting algorithm and a preset semantic mapping library, the learner's learning ability for each knowledge point on the simplified knowledge graph is obtained based on behavioral data fitting, so as to achieve dynamic updating of learning ability while preserving the historical learning ability trajectory; Based on the quantitative results of group learners' learning behavior and learning ability at each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point is periodically determined; Based on the quantitative results of individual learners' learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point in the current cycle is revised to obtain the learner's second recommended learning method as the personalized recommendation result for the learner's online learning.

[0009] The technical concept of this application is: first, based on the real-time collection of learners' online learning behavior data, a hierarchical decoupling method is used to construct a simplified knowledge graph of each subject's knowledge system, reducing the high complexity and resource consumption caused by the reliance on transfer learning and multi-domain knowledge relations for "full coverage", and solving the problem that complex graphs in traditional methods are difficult to apply in real time; secondly, through the behavior continuity merging algorithm combined with the preset semantic mapping library, discrete behavior data are merged and mapped to the knowledge points of the simplified knowledge graph according to the logical chain, which makes up for the defect of broken behavior data mapping in the existing patents and ensures the quantitative integrity of behavior details. At the same time, the temporal ability vector fitting algorithm is used to fit the learner's temporal learning ability based on the behavior data, while retaining the historical trajectory. The ability value is dynamically updated based on the data, avoiding the lag of traditional static centrality calculation; further, the first recommended learning method for each knowledge point is periodically determined based on the quantitative results of the group's learning behavior and learning ability at each knowledge point on the simplified knowledge graph, and then the second recommended learning method is obtained in combination with the quantitative results of the individual's learning behavior and learning ability at each knowledge point on the simplified knowledge graph. It not only improves the recommendation efficiency by utilizing group rules, but also realizes personalization through individual correction, realizing the dual-driven recommendation of "group effectiveness + individual adaptation", and the dependency relationship between knowledge points is clearly defined when constructing the knowledge graph and used as a mandatory constraint for learning method recommendation, avoiding interference from weakly associated knowledge points and strengthening the structural nature of the knowledge system.

[0010] In some embodiments, a hierarchical decoupling method is used to construct a simplified knowledge graph of a subject knowledge system, including: Acquire the learners’ subject knowledge system for online learning; Extract the ID, name and parent knowledge point of each knowledge point in the subject knowledge system, and construct the tree structure knowledge point system corresponding to the subject; Establish a directed connection relationship for the leaf nodes in the tree-structured knowledge point system, and verify the directed connection relationship to avoid the formation of closed loops between leaf nodes, thus obtaining a directed acyclic graph of leaf nodes; Based on the tree-structured knowledge point system and the directed acyclic graph of leaf nodes, a simplified knowledge graph with dual relationship separation is obtained.

[0011] In some embodiments, the behavior data includes the initiator of the action, the object of the action, the action of the actor, the time when the action occurred, and the result of the action. The behavior continuity merging algorithm and the preset semantic mapping library are used to obtain the learner's learning behavior for each knowledge point on the simplified knowledge graph based on the behavior data mapping, including: Map the action initiator in the behavior data to the learner ID; Map the action objects in the behavior data into knowledge points of a simplified knowledge graph according to the preset semantic mapping library; Mapping the actor's actions in the behavioral data into learning methods according to the preset semantic mapping library; Using the behavior continuity merging algorithm, the continuous learning duration and starting learning time of each knowledge point are obtained based on the action occurrence time mapping; The set of learner ID, knowledge points, learning methods, continuous learning duration and starting learning time constitutes the learner's learning behavior for each knowledge point on the simplified knowledge graph; Using a time-series ability vector fitting algorithm and a preset semantic mapping library, we obtain the learner's learning ability for each knowledge point on the simplified knowledge graph based on behavioral data fitting, thereby achieving dynamic updates of learning ability while preserving historical learning ability trajectories, including: Mapping the action occurrence time in behavioral data to the evaluation time; Adopting the time series ability vector fitting algorithm, the current ability value of each knowledge point is obtained based on the action results, so as to achieve dynamic update of learning ability while preserving the historical learning ability trajectory; The set of the learner ID, evaluation time and current ability value of the knowledge point constitutes the learner's learning ability for each knowledge point on the simplified knowledge graph.

[0012] In some embodiments, a behavior continuity merging algorithm is used to obtain the continuous learning duration and the starting learning time of each knowledge point based on the action occurrence time mapping, including: Traverse the behavior data corresponding to the learner ID in the order of the action occurrence time, and compare two adjacent behavior data one by one. If the knowledge points and learning methods after mapping the two adjacent behavior data are the same and the action occurrence time difference between the two behavior data is less than or equal to the preset threshold, then the two behavior data are merged into the same learning behavior and continue to be compared with the next behavior data. Otherwise, the two behavior data are each mapped to a learning behavior; Repeat the previous step until the behavior data corresponding to the learner ID is traversed. For the merged learning behaviors, calculate the time difference between the action occurrence time of the last behavior data and the action occurrence time of the first behavior data as the duration, and map the action occurrence time of the first behavior data as the start time. For the remaining learning behaviors, query the default duration of the corresponding actor's action as the duration, and map the action occurrence time of the corresponding behavior data to the start time.

[0013] In some embodiments, a time-series capability vector fitting algorithm is used to obtain the current capability value of each knowledge point based on the action results, so as to achieve dynamic update of learning capability while preserving the historical learning capability trajectory, including: Convert the action results in the behavioral data into the measured ability values of the corresponding knowledge points; Constructing a time series database of capability vectors, wherein the time series database of capability vectors includes a sequence of historical capability values corresponding to each knowledge point on the simplified knowledge graph; Perform linear fitting on the historical ability value sequence corresponding to each knowledge point to generate an ability evolution curve. Substitute the action occurrence time in the behavioral data into the ability evolution curve to obtain the predicted ability value of the corresponding knowledge point. Based on the measured ability value and predicted ability value of the knowledge point, the current ability value of the corresponding knowledge point is obtained by fusion; The current capability value of the knowledge point is stored in the time series database, and the time series database of the capability vector is dynamically updated.

[0014] In some embodiments, the current capability value of the knowledge point is stored in a time series database, and the time series database of the capability vector is dynamically updated, including: The current capability value of the knowledge point is stored in the historical capability value sequence corresponding to the knowledge point in chronological order; Based on the hierarchical path of each knowledge point in the simplified knowledge graph, determine the index position of each knowledge point in the capability vector; Update the index position of the knowledge point in the capability vector to the current capability value, and maintain the historical capability value for the remaining index positions at the same level of the capability vector; Based on the updated values of all index positions at the same level, the values of the index positions at the previous level are aggregated and generated; The set of values corresponding to all index positions in the capability vector and the historical capability value sequences constitutes the time series database of the dynamically updated capability vector.

[0015] In some embodiments, based on the quantitative results of the learning behavior and learning ability of group learners at each knowledge point on the simplified knowledge graph, a first recommended learning method for each knowledge point is periodically determined, including: Periodically, based on the quantitative results of the learning ability of group learners at each knowledge point on the simplified knowledge graph, the top n% of learners in learning ability are selected for each knowledge point. Based on the quantitative results of the learning behavior of the selected learners, the statistical training values of each learning method of the selected learners in the corresponding knowledge points are calculated; Based on the statistical training values of each learning method of the selected learner in the corresponding knowledge point, a first recommended learning method for the corresponding knowledge point in the current cycle is determined.

[0016] In some embodiments, the learning behavior includes a learner ID, a knowledge point, a learning method, a continuous learning duration, and a start learning time. Based on the quantified results of the learning behavior of the selected learner at the corresponding knowledge point, the statistical training values of each learning method of the selected learner at the corresponding knowledge point are calculated, including: Traverse the learning behaviors of the selected learners and convert the duration of each learning behavior into the training value of the corresponding learning method; During the traversal process, the training values of each learning method in each knowledge point are accumulated and counted to obtain the statistical training values of each learning method of the selected learner in the corresponding knowledge point, forming a sparse matrix of "knowledge point-learning method-statistical training value".

[0017] In some embodiments, based on the quantitative results of the learning behavior and learning ability of the individual learner for each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point in the current cycle is modified to obtain the second recommended learning method for the learner, including: Based on the quantitative results of the individual learner's learning behavior and learning ability at each knowledge point on the simplified knowledge graph, the learning efficiency vector of the individual learner in the current subject is calculated; Convert the first recommended learning method for each knowledge point into a group weight matrix as the dominant learning method matrix for the subject knowledge point; Multiply the learning style efficiency vector by the advantage learning style matrix to obtain the modified advantage learning style matrix; The modified dominant learning style matrix is numerically ranked, and the learner's second recommended learning style is determined based on the ranking result.

[0018] In some embodiments, based on the quantitative results of the learning behavior and learning ability of the individual learner at each knowledge point on the simplified knowledge graph, the learning efficiency vector of the individual learner in the current subject is calculated, including: Based on the quantitative results of the learning ability of individual learners for each knowledge point on the simplified knowledge graph, the top m% of knowledge points in the current subject are selected; Based on the quantitative results of individual learners' learning behaviors, the statistical training values of each learning method of the individual learners in the selected knowledge points are calculated; For each learning method, calculate the mean of the statistical training values corresponding to all selected knowledge points; The learning methods are sorted based on the mean value, and a fixed individual weight is assigned to each learning method according to the sorting result, and the sum of the individual weights is ensured to be 1, which is used as the learning method efficiency vector of the individual learner in the current subject.

[0019] The beneficial technical effects of the present invention include at least: 1. A self-evolving online learning personalized recommendation method is adopted. The layered decoupling method simplifies the knowledge graph. While reducing computing costs, the preset semantic mapping library and behavior continuity merging algorithm ensure the complete mapping of behavioral data, resolving the contradiction between complex graphs and inefficient computing in traditional methods. The time series ability vector fitting algorithm dynamically updates learning capabilities and combines the dual recommendation mechanism of group and individual. It not only uses group rules to improve recommendation efficiency, but also achieves personalization through individual correction, realizing the dual-driven recommendation of "group effectiveness + individual adaptation", overcoming the technical problem of static evaluation standards leading to recommendation lag in existing technologies. The clear knowledge point dependency relationships in the simplified knowledge graph are used as mandatory constraints in path recommendation to avoid interference from weakly related knowledge points, ensure that the recommendation results conform to the logic of the knowledge system, and solve the problem of the separation of path recommendation and knowledge dependency. Ultimately, the coordinated optimization of "low computational complexity, high recommendation accuracy, strong dynamics, and excellent structure" is achieved, significantly improving the efficiency and effectiveness of personalized recommendation, and achieving the comprehensive technical effect of "1+1>2". 2. A hierarchical decoupling method is used to divide the subject knowledge system into a tree-like trunk (non-leaf nodes) and a directed acyclic graph of leaf nodes (minimum granularity knowledge points). By structurally constraining the connection relationship between leaf nodes, the actual subject needs can be met. This breaks the inherent perception of traditional knowledge graphs that "full coverage = precise recommendation". Simplified full coverage reduces complexity to support real-time computing. This allows the subsequent use of knowledge points as mandatory constraints in path recommendations and the implementation of a dual-driven recommendation strategy of "group effectiveness + individual adaptation" to be completed within limited resources. The two work together to achieve dynamic standard correction and real-time response for online learning recommendation strategies. 3. To address the technical problem of broken behavioral data mapping in the existing technology, this application uses a behavioral continuity merging algorithm and a preset semantic mapping library to associate and simplify the knowledge points in the knowledge graph, map the behavioral data to the learner's learning behavior at each knowledge point on the simplified knowledge graph, and transform the "event flow" into a "learning story", thereby restoring the complete learning plot, effectively solving the semantic ambiguity problem of multi-source behavioral data, avoiding behavioral fragmentation and distortion, ensuring data credibility, and providing high-quality data input for subsequent recommendation steps; 4. The construction of a time-series database provides continuous support for capability evolution in the temporal dimension, resolving the flaw of traditional static evaluations that cannot track capability changes. The evolution curve generated by linear fitting transforms discrete behavioral results into predictable capability development trends, avoiding reliance on high-computing models. The fusion of measured and predicted capability values retains real-time behavioral feedback while leveraging historical patterns to improve the accuracy of capability value updates and enhance interpretability. Ultimately, this achieves the coordinated optimization of "dynamic tracking—trend prediction—real-time feedback," enabling learning capability assessments to be updated in real time with learning behavior and reflect long-term progress. This provides a data foundation for subsequent personalized recommendations that is more aligned with the learner's actual capability level, significantly improving the timeliness and accuracy of recommendation results. 5. By periodically selecting the top n% of learners based on the results of periodic evaluations of a fixed group, we ensure the advancement of group representatives. This solves the problem of traditional recommendations relying on the entire data being susceptible to interference from inefficient behavior. It also drives the self-evolution of recommendation rules, avoiding the obsolescence of static models. The design of statistical training values (covering both cumulative and recent values) reflects long-term learning patterns while capturing recent progress trends, improving the timeliness of recommendation methods. A sparse matrix of "knowledge points, learning methods, and statistical training values" is dynamically fitted to revise recommendation standards. The construction and incremental updates of the sparse matrix balance computational efficiency and data integrity, overcoming the limitations of high-computing-power models. Ultimately, this achieves the coordinated optimization of "group representatives, efficient statistics, and dynamic recommendations." This enables the recommendation learning method to be based on the group's effective learning experience while reducing computational costs through sparse matrix technology, significantly improving the accuracy and deployability of recommendation results. 6. By screening the top m% of knowledge points based on individual abilities, we ensure that we focus on the learner's strengths, solving the problem of traditional recommendations ignoring individual strengths; by constructing a group weight matrix, we retain common learning rules and avoid the limitations of purely individual recommendations; by using a matrix multiplication fusion mechanism, we quantitatively superimpose "group commonalities" and "individual characteristics" through mathematical operations, which not only uses group experience to improve recommendation reliability, but also strengthens personalized adaptation through individual weights; the combination of sparse matrices and mean sorting selection methods further reduces the computational load and ensures recommendation efficiency, ultimately achieving the collaborative optimization of "group experience empowerment - individual advantage amplification - dynamic fusion correction", so that the recommended learning method is in line with the general efficient model of the discipline and accurately matches the learner's own strengths, significantly improving the personalization and accuracy of the recommendation, while enabling dynamic standard correction and real-time response, achieving an individual-group fusion recommendation effect of "1+1>2" under the coordination of various technical steps, and resolving the contradiction of "group optimal ≠ individual optimal".

[0020] Other features and advantages of the present invention will be disclosed in detail in the following specific embodiments and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a flow chart of the self-evolving online learning personalized recommendation method according to an embodiment of the present invention.

[0022] Figure 2 Schematic diagram of a directed acyclic graph of leaf nodes according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following is an explanation and description of the technical solutions of the embodiments of the present invention in conjunction with the drawings of the embodiments of the present invention. However, the following embodiments are only preferred embodiments of the present invention and are not exhaustive. Based on the embodiments in the implementation manner, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0024] In the following description, terms such as "inside", "outside", "up", "down", "left", "right", etc. that indicate directions or positional relationships are only used to facilitate the description of the embodiments and simplify the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0025] Please see the attached Figure 1 , Figure 1 A flowchart of a self-evolving online learning personalized recommendation method provided by an embodiment of this specification is shown.

[0026] like Figure 1 As shown, the self-evolving online learning personalized recommendation method may include at least the following steps: S1, collects learners’ behavioral data during online learning in real time.

[0027] Among them, the behavioral data of learners during online learning is uniformly collected using the XAPI protocol, and the behavioral data of each application system is centrally collected through a data listener. The collection results are recorded as a statement sequence. The behavioral data is a six-tuple (actor, when, where, verb, object, result), where actor represents the initiator of the action, when represents the time when the action occurs, where represents the application system module where the action occurs, verb represents the actor's action, object represents the action object, and result represents the action result.

[0028] S2, uses the hierarchical decoupling method to construct a simplified knowledge graph of each subject's knowledge system.

[0029] Specifically, in this embodiment, a hierarchical decoupling method is used to construct a simplified knowledge graph of the subject knowledge system, including: S21, obtaining the learner’s subject knowledge system for online learning; S22, extracting the ID, name and parent knowledge point of each knowledge point in the subject knowledge system, and constructing a tree-structured knowledge point system corresponding to the subject.

[0030] This embodiment organizes the subject knowledge point system into a tree structure according to a hierarchical relationship. Each node in the tree structure knowledge point system is represented by a triple (knid, knname, parent_knid), where knid represents the unique identifier of the knowledge point, knname represents the name of the knowledge point, parent_knid represents the parent node ID, and the root node has no parent node.

[0031] S23, establishing a directed connection relationship for the leaf nodes in the tree-structured knowledge point system, and verifying the directed connection relationship to avoid forming a closed loop between the leaf nodes, thereby obtaining a directed acyclic graph of the leaf nodes.

[0032] It is understandable that this embodiment is based on the tree structure knowledge point system, combines the resources related to each knowledge point, and uses the knowledge graph tool and simplification strategy to sort out only the leaf nodes (the smallest granularity knowledge points) to form a directed acyclic graph of leaf nodes. Figure 2 As shown, the expression of the directed acyclic graph of leaf nodes allows the existence of disconnected leaf node directed acyclic graphs, but requires that the directed connection relationship between the minimum granularity knowledge points cannot form a closed loop. For example, if a closed loop connection relationship (A-->B-->C-->A) occurs, an error is reported, indicating that the directed acyclic graph of leaf nodes is not compliant.

[0033] S24, based on the tree-structured knowledge point system and the directed acyclic graph of leaf nodes, a simplified knowledge graph with dual relationship separation is obtained.

[0034] It is understandable that the simplified knowledge graph of dual-relation separation includes: 1) Containment relationship: parent node → child node (tree propagation); 2) Connection relationship: leaf node → leaf node (directed acyclic graph constraint, and cannot form a closed loop) Due to the inherent cognition of "full coverage = accurate recommendation" in the existing technology, the existing technology usually uses a directed acyclic graph with full coverage to construct a knowledge graph. For example, Chinese patent CN114861069A relies on transfer learning and multi-domain knowledge dependencies. Its knowledge point centrality algorithm needs to traverse the 1-m order neighbors of each node. For n knowledge points, the complexity of neighbor node extraction is O(n 2), resulting in high computational complexity and difficulty in meeting real-time requirements, and unconstrained closed-loop dependencies may lead to an infinite loop in centrality calculations. The complexity of the existing knowledge graph itself and its global support characteristics also make it consume too much computing resources and cannot be applied to the representation of small and medium-sized subject knowledge. To this end, this embodiment uses a hierarchical decoupling method to divide the subject knowledge system into a tree-shaped trunk (non-leaf nodes) and a directed acyclic graph of leaf nodes (minimum granularity knowledge points). By structurally constraining the relationship between leaf nodes, it can meet the actual subject needs, breaking the inherent cognition of traditional knowledge graphs that "global coverage = precise recommendation", simplifying the global coverage, reducing complexity, and supporting real-time computing, so that the subsequent use of knowledge points as mandatory constraints in path recommendations and realizing the dual-driven recommendation of "group effectiveness + individual adaptation" can be completed under limited resources. The two work together to achieve dynamic standard correction and real-time response of online learning recommendation strategies. It is understandable that the global directed acyclic graph of the prior art is like conducting a global path search in a maze (large computational effort and easy to get lost), while the simplified knowledge graph with dual-relationship separation proposed in this embodiment is like first taking an elevator to the target floor (quick tree-like hierarchical positioning) and then fine-tuning the path in the local area (small-scale calculation of the directed acyclic graph of leaf nodes). This improvement transforms the knowledge graph from a theoretical model into a real-time learning support system that can be implemented in an engineering context, completely solving the problem of "advanced algorithms but difficult to apply" caused by the complexity of existing online learning resource recommendation methods.

[0035] S3 adopts the behavior continuity merging algorithm and the preset semantic mapping library to obtain the learner's learning behavior of each knowledge point on the simplified knowledge graph based on behavior data mapping.

[0036] Specifically, in this embodiment, S3 includes: S31, directly mapping the action initiator in the behavior data to the learner ID; S32, mapping the action objects in the behavior data into knowledge points of the simplified knowledge graph according to the preset semantic mapping library.

[0037] For example, the preset semantic mapping library maps the action object of "lever experiment" to the knowledge point of "torque balance".

[0038] On the other hand, the preset semantic mapping library can also be implemented in a semi-automatic way, that is, through the existing preset mapping relationships, new mapping relationships are derived through synonyms. All mapping relationships need to be manually confirmed, and a supporting system can also support teachers and managers to manually establish the mapping relationship between the two.

[0039] S33, mapping the actor's actions in the behavior data into learning methods according to a preset semantic mapping library.

[0040] For example, the preset semantic mapping library maps the human action of "playing a video" to the learning method of video learning. It is understandable that in this embodiment, a knowledge point may correspond to multiple learning methods, including video learning, exercise training, etc.

[0041] S34, using the behavior continuity merging algorithm, based on the action occurrence time mapping, obtains the continuous learning duration and starting learning time of each knowledge point; S35, the set of the learner ID, knowledge points, learning methods, continuous learning time and starting learning time constitutes the learner's learning behavior for each knowledge point on the simplified knowledge graph.

[0042] It can be understood that learning behavior is a five-tuple (S, W, Ks, M, D), where S represents the learner's unique identifier, W represents the start time of the behavior, that is, the start time of learning, Ks represents the set of associated knowledge points, and the set of associated knowledge points is strictly equivalent to the knowledge point dependency relationship in the simplified knowledge graph (it can contain multiple associated knowledge points, such as an experiment involving {buoyancy principle, density calculation}), M represents the learning method (such as video learning, exercise training, correcting wrong questions, etc.), and D represents the continuous learning time.

[0043] Furthermore, in this embodiment, S34, a behavior continuity merging algorithm is used to obtain the continuous learning duration and the starting learning time of each knowledge point based on the action occurrence time mapping, including: S341, traverse the behavior data corresponding to the learner ID in the order of the action occurrence time, and compare the two adjacent behavior data one by one. If the knowledge points and learning methods of two adjacent behavioral data are the same after mapping, and the time difference between the actions of the two behavioral data is less than or equal to the preset threshold, the two behavioral data will be merged into the same learning behavior and continue to be compared with the latter behavioral data. If the knowledge points and learning methods mapped to two adjacent behavior data are different or the time difference between the actions of the two behavior data is greater than the preset threshold, the two behavior data are each mapped to a learning behavior; S342, repeat the previous step S341 until the behavior data corresponding to the learner ID is traversed. For the merged learning behaviors, calculate the time difference between the action occurrence time of the last behavior data and the action occurrence time of the first behavior data as the duration, and map the action occurrence time of the first behavior data as the start time. For the remaining learning behaviors, query the default duration of the corresponding actor's action as the duration, and map the action occurrence time of the corresponding behavior data as the start time.

[0044] For example, for the learner's behavior data: [09:00 Play Video A] → [09:15 Pause] → [09:30 Continue Playing A] → [09:45 End] Existing technology processing: Recording is 4 independent events (2 plays + 2 pauses), losing continuity; The solution of this embodiment is to merge into a single record through the behavior continuity merging algorithm: the learning behavior = (learner ID, 09:00, {knowledge point A}, video learning, 45 minutes).

[0045] For example, the existing technology classifies the actions of "dragging a 3D model" and "clicking an experiment button" as "interactive operations", but this embodiment uses a preset semantic mapping library to correspond the two to actually different learning methods ("observational learning" and "practical learning").

[0046] For example, for a learner's learning behavior path: video learning → exercise training → correcting wrong questions, Existing technology processing: Record as three independent events, breaking the logical chain of behavior; The solution of this embodiment is to associate the knowledge point set Ks to the same knowledge graph node (such as knowledge point A) to retain the behavior context.

[0047] It can be seen that the former existing technology is like only classifying ingredients (vegetables / meat), while the latter embodiment is like providing recipes (how to cut and cook). The use of existing technology will undoubtedly result in "ingredients cannot be turned into dishes."

[0048] In response to the technical problem of broken behavioral data mapping in the existing technology, this embodiment uses a behavioral continuity merging algorithm and a preset semantic mapping library to associate the knowledge points in the simplified knowledge graph, map the behavioral data into the learner's learning behavior at each knowledge point on the simplified knowledge graph, and convert the "event flow" into a "learning story", thereby restoring the complete learning plot, effectively solving the semantic ambiguity problem of multi-source behavioral data, avoiding behavioral fragmentation and distortion, ensuring data credibility, and providing high-quality data input for subsequent recommendation steps.

[0049] S4 uses a time-series ability vector fitting algorithm and a preset semantic mapping library to obtain the learner's learning ability for each knowledge point on the simplified knowledge graph based on behavioral data fitting, thereby achieving dynamic updating of learning ability while preserving the historical learning ability trajectory; Specifically, in this embodiment, a time-series ability vector fitting algorithm and a preset semantic mapping library are used to obtain the learner's learning ability for each knowledge point on the simplified knowledge graph based on behavioral data fitting, so as to achieve dynamic updating of learning ability while preserving the historical learning ability trajectory, including: S41, directly mapping the action occurrence time in the behavior data to the evaluation time; S42, using a time-series capability vector fitting algorithm, the current capability value of each knowledge point is obtained based on the action results, so as to achieve dynamic update of learning capability while preserving the historical learning capability trajectory; S43, the set of the learner ID, evaluation time and current ability value of the knowledge point constitutes the learner's learning ability for each knowledge point on the simplified knowledge graph.

[0050] It can be understood that learning ability is a triple (S, Ks, X), where S represents the learner's unique identifier, Ks represents the associated knowledge point set, and X represents the current ability value of the knowledge point, which is used to represent the learner's mastery of the knowledge point.

[0051] Furthermore, we can also express X=(X0,X1,X2,...,Xn), where X1-Xn represents the learner's mastery of the i-th ability of the current knowledge point, n is the knowledge ability predefined by the subject (such as memory, calculation, reasoning, etc.), and X0 represents the learner's overall mastery of the current knowledge point. Through this expression, the learner's mastery of the subject is a (n+1)*m matrix, where m is the number of knowledge points in the subject.

[0052] Furthermore, in this embodiment, S42 uses a time-series capability vector fitting algorithm to obtain the current capability value of each knowledge point based on the action results, so as to achieve dynamic update of learning capability while retaining the historical learning capability trajectory, including: S421, converting the action results in the behavior data into the measured ability values of the corresponding knowledge points.

[0053] For example, the method of converting the action results in the behavioral data into the measured ability value of the corresponding knowledge point can be: 60% of the video learning has been played → measured ability value = 0.6, exercise training accuracy score = 80 → measured ability value = 0.8, etc.

[0054] S422, constructing a time series database of capability vectors, wherein the time series database of capability vectors includes a sequence of historical capability values corresponding to each knowledge point on the simplified knowledge graph, and an ordered array representing the learner's historical capability values at each knowledge point.

[0055] Among them, if the historical ability value of a certain knowledge point does not exist, the historical ability value sequence corresponding to the knowledge point can be initialized to a zero vector or initialized to the learner's average ability value for the current subject system.

[0056] S423, performing linear fitting on the historical capability value sequence corresponding to each knowledge point to generate a capability evolution curve, retaining the evolution process, substituting the action occurrence time in the behavior data into the capability evolution curve to obtain the predicted capability value of the corresponding knowledge point; It is understandable that this embodiment records the learner's ability value for each knowledge point in a time series and dynamically fits the curve (such as monthly sampling). Specifically, the implementation method is: First, extract the learner's historical ability value sequence at a certain knowledge point from the time series database of ability vectors (such as [X1, X2, ..., X i ]); Then, the historical capability values are sampled at fixed intervals (such as the 1st of each month) and the sampling points are selected to form a time series T=[t1,t2,...,t j ]; Then, linear regression fitting is performed on the capability value data between sampling points to generate a continuous capability evolution curve f(t)=a×t+b (coefficients a and b are solved by the least squares method); Finally, the action occurrence time in the behavioral data is substituted into the fitted curve f(t) to obtain the prediction ability value.

[0057] S424, based on the measured ability value and the predicted ability value of the knowledge point, the current ability value of the corresponding knowledge point is obtained by fusion.

[0058] The fusion of the measured capability value and the predicted capability value may be achieved by weighted summation or other fusion methods, which is not limited in this embodiment.

[0059] S425 , storing the current capability value of the knowledge point in the time series database, preserving the progress track, and dynamically updating the time series database of the capability vector.

[0060] On the one hand, this embodiment can only update the current ability value and historical ability value sequence of the corresponding knowledge point, capture the learner's progress trend through the time series data of the historical ability value sequence, and support subsequent dynamic recommendations.

[0061] Because existing online learning personalized recommendation technologies have the problem of static evaluation criteria, their evaluation data usually directly overwrites historical values, resulting in the inability to identify the "speed of progress" in ability assessment, and only mechanical matching. There is also a real-time contradiction. That is, in high-frequency evaluation scenarios, directly overwriting historical ability values will cause data jitter, while in low-frequency evaluation scenarios, it will cause delayed ability vector updates. To this end, the time-series ability vector fitting algorithm proposed in this embodiment provides a high-quality data foundation for solving the above problems through multi-step collaboration: first, the action occurrence time in the behavioral data is directly mapped to the evaluation time to ensure accurate recording of the time dimension; then, a time-series database of ability vectors is constructed to store the historical ability value sequence of each knowledge point, fully capturing the learner's progress trajectory on each knowledge point; then, a linear fit is performed on the historical ability value sequence to generate an ability evolution curve, and the behavior occurrence time is substituted into the curve to predict the current ability value, which not only preserves historical data but also supports trend analysis; then, the measured ability value converted from the action result is merged with the predicted ability value, balancing real-time behavior feedback and long-term trend prediction; finally, the fused current ability value is stored and dynamically updated in the time-series database, continuously improving the learner's ability development record.

[0062] It is understandable that the construction of a time series database provides continuity support in the time dimension for capability evolution, solving the defect that traditional static evaluation cannot track capability changes; the evolution curve generated by linear fitting converts discrete behavioral results into predictable capability development trends, avoiding dependence on high-computing power models; the fusion of measured and predicted capability values retains real-time behavioral feedback while using historical laws to improve the accuracy of capability value updates and enhance interpretability, ultimately achieving the collaborative optimization of "dynamic tracking-trend prediction-real-time feedback", so that learning ability assessment can be updated in real time with learning behavior and reflect long-term progress, providing a data basis for subsequent personalized recommendations that is more in line with the learner's actual ability level, thereby significantly improving the timeliness and accuracy of recommendation results.

[0063] On the other hand, to further ensure the interpretability of the capability evolution curve, in this embodiment, S425 stores the current capability value of the knowledge point in the time series database and dynamically updates the time series database of the capability vector, which may also include: S4251, storing the current capability value of the knowledge point in chronological order into the historical capability value sequence corresponding to the knowledge point; S4252, based on the hierarchical path of each knowledge point in the simplified knowledge graph, determine the index position of each knowledge point in the capability vector.

[0064] It can be understood that the ability vector represents an ordered array of learners’ ability values at each knowledge point, and its structure is [Y1, Y2, ..., Y p], the index position (such as Y1) corresponds to the unique code of the node in the knowledge graph (such as the index of Ks=ALG-001 is 1).

[0065] Specifically, the mapping rules between hierarchical paths and indexes can be: 1) Path decomposition: Split the hierarchical path of a knowledge point node into a hierarchical sequence using delimiters (such as “→”). For example, the hierarchical path of “Mathematics → Algebra → Equation” can be decomposed into [“Mathematics”, “Algebra”, “Equation”].

[0066] 2) Index assignment: Assign the lowest index (such as Y1) to the root node (such as "mathematics"), and recursively expand the child nodes. The index increases with each level down, and the hierarchical path of each knowledge point uniquely corresponds to an index to avoid ambiguity.

[0067] Furthermore, if a new level is inserted (e.g., Mathematics → Algebra → Function), only a new index needs to be assigned to the node "Function".

[0068] S4253: Update the index position of the knowledge point in the capability vector to the current capability value, and maintain the historical capability values for the remaining index positions at the same level of the capability vector. S4254: Based on the updated values of all index positions at the same level, the values of the index positions at the previous level are aggregated and generated.

[0069] It is understood that the current capability value of a higher-level node (such as "Algebra") is generated by aggregating the current capability values of lower-level nodes (such as "Equation" and "Function") to achieve interpretability. If the "Algebra" capability declines, the specific weaknesses of its child nodes can be traced back. Aggregation methods can include averaging, weighted summation (weights determined by the importance of knowledge points), and other methods, which are not limited in this embodiment.

[0070] S4255: The values corresponding to all index positions in the capability vector and the set of historical capability value sequences constitute a time series database of the capability vector after dynamic update.

[0071] It can be understood that in this embodiment, hierarchical correlation not only supports vertical analysis of assignments, but the hierarchical comparison of time series (historical ability value series) also makes the evolution curve layered and visualized. Specifically, each knowledge point index corresponds to an independent ability evolution curve, which enables horizontal comparison and vertical tracking. That is, knowledge points at the same level (such as "algebra" and "geometry") can be compared horizontally to determine the learners' strengths. At the same time, the upper-level node curve reflects the comprehensive ability, and the lower-level node curve reveals the detailed changes.

[0072] It can be understood that this embodiment systematically achieves the following technical effects by strictly mapping the hierarchical path of the knowledge graph with the capability vector index and designing a hierarchical adjustment mechanism: 1) Structured representation: The logical relationship between knowledge points is directly mapped to the positional relationship of the capability vector.

[0073] 2) Causal explainability: Changes in abilities can be attributed layer by layer to specific knowledge points, supporting precise teaching.

[0074] 3) Dynamic adaptability: The hierarchical adjustment mechanism ensures that the model remains effective as the knowledge system evolves.

[0075] This design solves the "black box" problem in traditional competency assessment, upgrading learning analytics from data statistics to deep insights driven by knowledge structure.

[0076] Through the above steps, this embodiment converts discrete behavioral data into structured vectors, providing a quantitative basis for subsequent personalized recommendations. The existing technology is like observing and learning with a broken lens (scattered and distorted), while this embodiment is like setting up a high-speed panoramic camera (continuous, accurate, and multi-dimensional). This qualitative change from "information recording" to "learning restoration" upgrades data governance from a cost center to a value engine, providing high-fidelity fuel for personalized learning recommendations.

[0077] S5, based on the quantitative results of the learning behavior and learning ability of group learners for each knowledge point on the simplified knowledge graph, periodically determine the first recommended learning method for each knowledge point.

[0078] The scope of the group of learners may be fixed in advance, for example, periodically determined based on the quantitative results of the learning behavior and learning ability of learners in a previous learning class.

[0079] Specifically, in this embodiment, based on the quantitative results of the learning behavior and learning ability of group learners at each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point is periodically determined, including: S51, periodically based on the quantitative results of the learning ability of group learners for each knowledge point on the simplified knowledge graph, select the top n% learners in learning ability for each knowledge point.

[0080] In this embodiment, learner selection based on quantified learning ability can be based on the current ability value corresponding to a knowledge point. Alternatively, if sufficient learner behavior data is available (e.g., learners have multiple ability evaluations on a knowledge point), the ability improvement index can be used. Specifically, learning methods for learners with rapid progress should be recommended. For example, the value of n can be 10.

[0081] S52: Based on the quantified results of the learning behaviors of the selected learners, statistical training values of the learning methods of the selected learners in the corresponding knowledge points are calculated.

[0082] Furthermore, in this embodiment, the learning behavior includes the learner ID, knowledge point, learning method, continuous learning time, and starting learning time. Based on the quantitative results of the selected learner's learning behavior at the corresponding knowledge point, the statistical training value of each learning method of the selected learner at the corresponding knowledge point is calculated, including: S521, traversing the learning behaviors of the selected learners, and converting the continuous learning duration of each learning behavior into a training value of the corresponding learning method.

[0083] For example, if the continuous learning time is 1.2 hours, the training value of the corresponding learning method is 1.2.

[0084] S522, during the traversal process, the training values of each learning method in each knowledge point are accumulated and counted to obtain the statistical training values of each learning method of the selected learner in the corresponding knowledge point, forming a sparse matrix of "knowledge point-learning method-statistical training value".

[0085] Among them, the statistical training values of each learning method may include cumulative statistical training values, that is, statistical training values obtained by accumulating all learning behaviors to represent the overall training situation, and may also include recent statistical training values, that is, statistical training values within a predetermined period of time from the current time to represent the recent training situation.

[0086] The logic behind the cumulative statistics is to add up the training values for all behaviors within the same knowledge point and learning method. For example, let's say learner A has two learning methods for the "Equation" knowledge point: watching videos and doing exercises. Watching videos three times (1.2 hours + 0.8 hours + 1.5 hours) results in a statistical training value of 3.5, and doing exercises twice (0.5 hours + 0.7 hours) results in a statistical training value of 1.2.

[0087] Sparse matrices are often used to process data with large amounts of zero values. For example, each learner's training values for different knowledge points and learning methods may only be partially non-zero, so using a sparse matrix is more efficient. For example, if a learner has only learned a few knowledge points and the majority of the other knowledge points are not recorded, the matrix will be mostly zero. Sparse matrices can effectively save space and processing resources.

[0088] Specifically, in this embodiment, a sparse matrix can be constructed in such a way that rows represent knowledge points (e.g., "equations," "geometry," and "algebra"), columns represent learning methods (e.g., "video learning," "exercise training," and "interactive experiment"). The value of each cell represents the learner's statistical training value for that knowledge point-learning method combination. Sparsity means that only non-zero values are stored (e.g., if a learner has not been exposed to the "interactive experiment" learning method for the "geometry" knowledge point, the corresponding cell will be 0).

[0089] Furthermore, steps S521-S522 may be executed once a day. To improve performance during execution, an auxiliary data structure may be designed so that only the newly added learning behavior data set is traversed.

[0090] It can be understood that in this embodiment, the learner's "knowledge points-learning methods-statistical training values" adopt sparse matrix storage and periodic incremental updates, which can effectively reduce the computing load and enable the recommendation algorithm (such as group fitting) to be deployed on the online learning platform.

[0091] S53: Determine a first recommended learning method for the corresponding knowledge point in the current cycle based on the statistical training values of each learning method of the selected learner in the corresponding knowledge point.

[0092] Among them, in this embodiment, the implementation method of determining the first recommended learning method for the corresponding knowledge point in the current cycle can be to select several learning methods with the highest statistical training value for each knowledge point as the first recommended learning method for the knowledge point in the current cycle.

[0093] Furthermore, in this embodiment, determining the first recommended learning method for a corresponding knowledge point in the current cycle can also be implemented by, for each knowledge point, extracting the top N learning methods ranked by statistical training value, where N can be preset to 2, then eliminating learning methods with statistical training values less than 1 / 10 of the maximum statistical training value corresponding to the knowledge point, and using the remaining learning methods as the first recommended learning method for the knowledge point in the current cycle. It is understood that the statistical training values corresponding to the first recommended learning method for a knowledge point in the current cycle are still sparse matrices, as exemplified in Table 1 below.

[0094] Table 1 Sparse matrix of the group “knowledge point—first recommended learning method—statistical training value”

[0095] Existing personalized recommendation technologies for online learning have technical problems such as a single recommendation basis, static evaluation criteria, and insufficient utilization of group learning models. Traditional methods mostly rely on knowledge point centrality calculation rules or other fixed rules, which inevitably lead to recommendation results lagging behind changes in learners' abilities. The alternative technical path that technical personnel in this field can think of is usually to adopt recommendations based on deep learning, but this undoubtedly requires higher computing power and lacks explainability. It is not suitable for the scenario of personalized recommendation for online learning, and it is difficult to efficiently reflect the effective learning rules of the group. To this end, this embodiment solves the above-mentioned problem through a multi-step collaborative approach. Specifically: first, by periodically selecting the top n% of learners in the group in terms of learning ability, these high-ability learners are used as group representatives to ensure the advancement of the recommendation basis; then, by traversing the learning behaviors of the selected learners (such as learning methods, duration, etc.), the duration of continuous learning is converted into a training value for the corresponding learning method, and the statistical training values are accumulated according to the knowledge point-learning method dimension to form a sparse matrix of "knowledge point-learning method-statistical training value" (only non-zero values are stored to reduce the computational load), which not only captures the effective learning pattern of the group on each knowledge point, but also optimizes storage and processing efficiency through the sparse matrix; finally, the first recommended learning method for each knowledge point is determined based on the statistical training values, and the high-frequency and efficient learning methods of the group are refined as the recommendation basis. It can be understood that this embodiment regularly selects the top n% of learners based on the periodic evaluation results of a fixed group to ensure the advancement of the group representatives, thereby solving the problem that traditional recommendations rely on all data and are easily interfered with by inefficient behaviors. At the same time, it can drive the self-evolution of recommendation rules and avoid the obsolescence of static models. The design of statistical training values (covering cumulative values and recent values) not only reflects long-term learning patterns but also captures recent progress trends, improving the timeliness of the recommendation method. The sparse matrix of "knowledge points-learning methods-statistical training values" of the group is dynamically fitted to correct the recommendation standards. The construction and incremental update of the sparse matrix balance computational efficiency and data integrity, overcoming the limitations of high-computing-power models, and ultimately achieving the collaborative optimization of "group representatives-efficient statistics-dynamic recommendations". This makes the recommended learning method based on the effective learning experience of the group and reduces the computational cost through sparse matrix technology, significantly improving the accuracy and deployability of the recommendation results, and creating a "1+1>2" group learning model mining and recommendation solution under the coordination of various technical steps.

[0096] S6, based on the quantitative results of the individual learner's learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point in the current cycle is revised to obtain the learner's second recommended learning method as the personalized recommendation result for the learner's online learning.

[0097] It can be understood that this embodiment corrects the group recommendation result (i.e., the first recommended learning method) by combining the individual learners' advantageous knowledge points, thereby combining the commonality of group effectiveness with the differences in individual preferences to achieve accurate recommendation of online learning methods.

[0098] Specifically, in this embodiment, S6 includes: S61, based on the quantitative results of the individual learner's learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the learning efficiency vector of the individual learner in the current subject is calculated.

[0099] Furthermore, in this embodiment, based on the quantitative results of the learning behavior and learning ability of the individual learner at each knowledge point on the simplified knowledge graph, the learning efficiency vector of the individual learner in the current subject is calculated, including: S611, based on the quantitative results of the learning ability of individual learners for each knowledge point on the simplified knowledge graph, select the top m% of knowledge points in the current subject in terms of learning ability.

[0100] In this embodiment, knowledge point selection based on the quantified results of individual learners' learning abilities can be based on the current ability value corresponding to the knowledge point, indicating that the learner has adopted a learning method that suits them and has achieved good results in that knowledge point. Alternatively, if sufficient learner behavior data is available (the learner has multiple ability evaluations for a knowledge point), the selection can be based on the improvement index of the ability value corresponding to the knowledge point, indicating that the learner has adopted a learning method that suits them and has mastered the learning content more quickly at that knowledge point. For example, the value of m can be 30.

[0101] S612, based on the quantified results of the individual learner's learning behavior, calculating the statistical training value of each learning method of the individual learner in the selected knowledge point; The implementation method of "calculating the statistical training values of each learning method of the individual learner in the selected knowledge point based on the quantified results of the learning behavior of the individual learner" is similar to the implementation method of the aforementioned "S52, calculating the statistical training values of each learning method of the selected learner in the corresponding knowledge point based on the quantified results of the learning behavior of the selected learner", and will not be repeated here in this embodiment. This step can obtain the statistical training values of each learning method of the individual learner in the corresponding knowledge point, forming a sparse matrix of "knowledge point-learning method-statistical training value" of the individual learner in the current subject. An example is described in Table 2 below.

[0102] Table 2 Sparse matrix of individual “selected knowledge points - learning methods - statistical training values”

[0103] S613: For each learning method, calculate the mean of the statistical training values corresponding to all selected knowledge points.

[0104] Taking the sparse matrix corresponding to Table 2 as an example, the mean of learning method 1 corresponding to the selected knowledge points is 1.5, the mean of learning method 2 is 4.7, the mean of learning method 3 is 3.5, and the mean of learning method 4 is 0.9.

[0105] S614, sorting the learning methods based on the mean value, assigning a fixed individual weight to each learning method according to the sorting result, and ensuring that the sum of the individual weights is 1, as the learning method efficiency vector of the individual learner in the current subject.

[0106] For example, taking the four learning methods corresponding to Table 2 as an example, the individual weight of the learning method 2 ranked first is 0.4, which serves as the efficiency vector of the individual learner's learning method 1 in the current subject. The learning method 3 ranked second is 0.3, the learning method 4 ranked third is 0.2, and the learning method 1 ranked fourth is 0.1, ensuring that the sum of all individual weights is 1.

[0107] It can be understood that, in this embodiment, based on selecting the top m% of knowledge points in terms of learning ability, the mean ranking weight generation method is used to generate the learning efficiency vector of the individual learner in the current subject.

[0108] S62, converting the first recommended learning method for each knowledge point into a group weight matrix as the advantageous learning method matrix for the subject knowledge point.

[0109] Specifically, in this embodiment, S62 is implemented by first converting the statistical training value of each learning method in the first recommended learning method for each knowledge point into a group weight vector: the statistical training value of a certain learning method / the sum of all statistical training values of the first recommended learning methods. The sum of the group weights of all learning methods in the first recommended learning method corresponding to a knowledge point is ensured to be 1. The collection of group weight vectors corresponding to each knowledge point constitutes the dominant learning method matrix for the subject knowledge point.

[0110] For example, using the sparse matrix corresponding to Table 1, the group weight vector corresponding to "Knowledge Point 1 - Learning Method 1 - Statistical Training Value" is 1.2 / (1.2+2.3+0.9)≈0.3, the group weight vector corresponding to "Knowledge Point 1 - Learning Method 2 - Statistical Training Value" is 2.3 / (1.2+2.3+0.9)≈0.5, and the group weight vector corresponding to "Knowledge Point 1 - Learning Method 4 - Statistical Training Value" is 0.9 / (1.2+2.3+0.9)≈0.2. This ensures that the sum of the group weights of all learning methods 1 / 2 / 4 in the first recommended learning method corresponding to Knowledge Point 1 is 1. The conversion methods for group weight vectors corresponding to other knowledge points can refer to this example.

[0111] S63, multiplying the learning style efficiency vector by the advantage learning style matrix to obtain a modified advantage learning style matrix.

[0112] For example, the individual weight 0.1 corresponding to learning method 1 in the learning method efficiency vector is multiplied by "knowledge point 1-learning method 1-group weight vector 0.3" in the advantageous learning method matrix of the subject knowledge point, and the obtained modified advantageous learning method matrix is "knowledge point 1-learning method 1-modified vector is 0.03".

[0113] S64, numerically sorting the modified dominant learning style matrix, and determining the learner's second recommended learning style based on the sorting result.

[0114] Among them, the implementation method of "S64, numerically sorting the revised advantageous learning method matrix, and determining the learner's second recommended learning method based on the sorting result" is similar to the aforementioned "S53, based on the statistical training values of each learning method of the selected learner in the corresponding knowledge point, determining the first recommended learning method of the corresponding knowledge point in the current cycle", and this embodiment will not be repeated here.

[0115] Existing personalized recommendation technologies for online learning suffer from a core problem: a disconnect between group experience and individual differences. Traditional methods often rely on single-group statistics or direct recommendations based on individual behavior, making it difficult to balance common patterns with individual preferences. This results in recommendations that lack group effectiveness and are inappropriate for learners' strengths. To address this issue, the proposed "group effectiveness + individual adaptation" dual-driven recommendation solution addresses this issue through a multi-step collaborative approach. First, the top m% of individual knowledge points are selected, focusing on the learner's areas of expertise. The statistical training values of each learning method for these knowledge points are calculated, and an individual weight vector is generated using the mean-ranked weight generation method to form a "learning method efficiency vector" that reflects individual learning efficiency. Next, the statistical training values of the first recommended learning method for each knowledge point are converted into a group weight vector to construct a subject "advantage learning method matrix," preserving the group's common experience of high-frequency and efficient learning patterns. Subsequently, the individual efficiency vector is multiplied by the group weight matrix to integrate group commonality with individual characteristics, correcting the universality bias of group recommendations. Finally, by sorting the corrected matrix, a second recommended learning method that best fits the learner's strengths is determined. It can be understood that this embodiment ensures that the focus is on the learner's advantage areas by screening the top m% knowledge points of individual ability, solving the problem of traditional recommendations ignoring individual strengths; the common learning rules are retained by constructing a group weight matrix, avoiding the limitations of pure individual recommendations; the matrix multiplication fusion mechanism is used to quantify and superimpose "group commonalities" and "individual characteristics" through mathematical operations, which not only uses group experience to improve recommendation reliability, but also strengthens personalized adaptation through individual weights; the combination of sparse matrix and mean sorting selection method further reduces the computing load and ensures recommendation efficiency, and finally realizes the collaborative optimization of "group experience empowerment-individual advantage amplification-dynamic fusion correction", so that the recommended learning method is in line with the general efficient model of the discipline and accurately matches the learner's own strengths, significantly improving the personalization and accuracy of the recommendation, and at the same time can realize dynamic standard correction and real-time response, achieving the "1+1>2" individual-group fusion recommendation effect under the coordination of various technical steps, and solving the contradiction of "group optimal ≠ individual optimal".

[0116] The above description is merely a description of the preferred embodiments disclosed in this application and the technical principles employed. Those skilled in the art should understand that the scope of protection provided by this disclosure is not limited to technical solutions formed by a specific combination of the aforementioned technical features, but also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents without departing from the scope of the disclosure. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in this disclosure.

[0117] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details have been included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

Claims

1. A self-evolving online learning personalized recommendation method, characterized by: The following steps are involved: Collect learners’ behavioral data during online learning in real time; Adopting the hierarchical decoupling method to construct a simplified knowledge graph of each subject's knowledge system; Adopting the behavior continuity merging algorithm and the preset semantic mapping library, the learner's learning behavior of each knowledge point on the simplified knowledge graph is obtained based on the behavior data mapping; Using a time-series ability vector fitting algorithm and a preset semantic mapping library, the learner's learning ability for each knowledge point on the simplified knowledge graph is obtained based on behavioral data fitting, so as to achieve dynamic updating of learning ability while preserving the historical learning ability trajectory; Based on the quantitative results of group learners' learning behavior and learning ability at each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point is periodically determined; Based on the quantitative results of individual learners' learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point in the current cycle is revised to obtain the learner's second recommended learning method as the personalized recommendation result for the learner's online learning.

2. A self-evolving online learning personalized recommendation method according to claim 1, characterized in that: A simplified knowledge graph of the subject knowledge system is constructed using a hierarchical decoupling method, including: Acquire the learners’ subject knowledge system for online learning; Extract the ID, name and parent knowledge point of each knowledge point in the subject knowledge system, and construct the tree structure knowledge point system corresponding to the subject; Establish a directed connection relationship for the leaf nodes in the tree-structured knowledge point system, and verify the directed connection relationship to avoid the formation of closed loops between leaf nodes, thus obtaining a directed acyclic graph of leaf nodes; Based on the tree-structured knowledge point system and the directed acyclic graph of leaf nodes, a simplified knowledge graph with dual relationship separation is obtained.

3. The self-evolving online learning personalized recommendation method according to claim 1, characterized in that: The behavior data includes the initiator of the action, the object of the action, the action of the actor, the time when the action occurs, and the result of the action. The behavior continuity merging algorithm and the preset semantic mapping library are used to obtain the learner's learning behavior for each knowledge point on the simplified knowledge graph based on the behavior data mapping, including: Map the action initiator in the behavior data to the learner ID; Map the action objects in the behavior data into knowledge points of a simplified knowledge graph according to the preset semantic mapping library; Mapping the actor's actions in the behavioral data into learning methods according to the preset semantic mapping library; Using the behavior continuity merging algorithm, the continuous learning duration and starting learning time of each knowledge point are obtained based on the action occurrence time mapping; The set of learner ID, knowledge points, learning methods, continuous learning duration and starting learning time constitutes the learner's learning behavior for each knowledge point on the simplified knowledge graph; Using a time-series ability vector fitting algorithm and a preset semantic mapping library, we obtain the learner's learning ability for each knowledge point on the simplified knowledge graph based on behavioral data fitting, thereby achieving dynamic updates of learning ability while preserving historical learning ability trajectories, including: Mapping the action occurrence time in behavioral data to the evaluation time; Adopting the time series ability vector fitting algorithm, the current ability value of each knowledge point is obtained based on the action results, so as to achieve dynamic update of learning ability while preserving the historical learning ability trajectory; The set of the learner ID, evaluation time and current ability value of the knowledge point constitutes the learner's learning ability for each knowledge point on the simplified knowledge graph.

4. A self-evolving online learning personalized recommendation method as claimed in claim 3, characterized in that: Using the behavior continuity merging algorithm, the continuous learning duration and starting learning time of each knowledge point are obtained based on the action occurrence time mapping, including: Traverse the behavior data corresponding to the learner ID in the order of the action occurrence time, and compare two adjacent behavior data one by one. If the knowledge points and learning methods after mapping the two adjacent behavior data are the same and the action occurrence time difference between the two behavior data is less than or equal to the preset threshold, then the two behavior data are merged into the same learning behavior and continue to be compared with the next behavior data. Otherwise, the two behavior data are each mapped to a learning behavior; Repeat the previous step until the behavior data corresponding to the learner ID is traversed. For the merged learning behaviors, calculate the time difference between the action occurrence time of the last behavior data and the action occurrence time of the first behavior data as the duration, and map the action occurrence time of the first behavior data as the start time. For the remaining learning behaviors, query the default duration of the corresponding actor's action as the duration, and map the action occurrence time of the corresponding behavior data to the start time.

5. The self-evolving online learning personalized recommendation method according to claim 3, characterized in that: Using a time-series capability vector fitting algorithm, the current capability value of each knowledge point is obtained based on the action results. This allows for dynamic updates of learning capabilities while preserving historical learning capability trajectories, including: Convert the action results in the behavioral data into the measured ability values of the corresponding knowledge points; Constructing a time series database of capability vectors, wherein the time series database of capability vectors includes a sequence of historical capability values corresponding to each knowledge point on the simplified knowledge graph; Perform linear fitting on the historical ability value sequence corresponding to each knowledge point to generate an ability evolution curve. Substitute the action occurrence time in the behavioral data into the ability evolution curve to obtain the predicted ability value of the corresponding knowledge point. Based on the measured ability value and predicted ability value of the knowledge point, the current ability value of the corresponding knowledge point is obtained by fusion; The current capability value of the knowledge point is stored in the time series database, and the time series database of the capability vector is dynamically updated.

6. A self-evolving online learning personalized recommendation method according to claim 5, characterized in that: The current capability value of the knowledge point is stored in the time series database, and the time series database of the capability vector is dynamically updated, including: The current capability value of the knowledge point is stored in the historical capability value sequence corresponding to the knowledge point in chronological order; Based on the hierarchical path of each knowledge point in the simplified knowledge graph, determine the index position of each knowledge point in the capability vector; Update the index position of the knowledge point in the capability vector to the current capability value, and maintain the historical capability value for the remaining index positions at the same level of the capability vector; Based on the updated values of all index positions at the same level, the values of the index positions at the previous level are aggregated and generated; The set of values corresponding to all index positions in the capability vector and the historical capability value sequences constitutes the time series database of the dynamically updated capability vector.

7. The self-evolving online learning personalized recommendation method according to claim 1, characterized in that: Based on the quantitative results of group learners' learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point is periodically determined, including: Periodically, based on the quantitative results of the learning ability of group learners at each knowledge point on the simplified knowledge graph, the top n% of learners in learning ability are selected for each knowledge point. Based on the quantitative results of the learning behavior of the selected learners, the statistical training values of each learning method of the selected learners in the corresponding knowledge points are calculated; Based on the statistical training values of each learning method of the selected learner in the corresponding knowledge point, a first recommended learning method for the corresponding knowledge point in the current cycle is determined.

8. A self-evolving online learning personalized recommendation method according to claim 7, characterized in that: The learning behavior includes the learner ID, knowledge point, learning method, continuous learning time, and starting learning time. Based on the quantitative results of the selected learner's learning behavior at the corresponding knowledge point, the statistical training value of each learning method of the selected learner in the corresponding knowledge point is calculated, including: Traverse the learning behaviors of the selected learners and convert the duration of each learning behavior into the training value of the corresponding learning method; During the traversal process, the training values of each learning method in each knowledge point are accumulated and counted to obtain the statistical training values of each learning method of the selected learner in the corresponding knowledge point, forming a sparse matrix of "knowledge point-learning method-statistical training value".

9. The self-evolving online learning personalized recommendation method according to claim 1, characterized in that: Based on the quantitative results of the individual learner's learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the first recommended learning method for each knowledge point in the current cycle is modified to obtain the learner's second recommended learning method, including: Based on the quantitative results of the individual learner's learning behavior and learning ability at each knowledge point on the simplified knowledge graph, the learning efficiency vector of the individual learner in the current subject is calculated; Convert the first recommended learning method for each knowledge point into a group weight matrix as the dominant learning method matrix for the subject knowledge point; Multiply the learning style efficiency vector by the advantage learning style matrix to obtain the modified advantage learning style matrix; The modified dominant learning style matrix is numerically ranked, and the learner's second recommended learning style is determined based on the ranking result.

10. The self-evolving online learning personalized recommendation method according to claim 9, characterized in that: Based on the quantitative results of the individual learner's learning behavior and learning ability for each knowledge point on the simplified knowledge graph, the learning efficiency vector of the individual learner in the current subject is calculated, including: Based on the quantitative results of the learning ability of individual learners for each knowledge point on the simplified knowledge graph, the top m% of knowledge points in the current subject are selected; Based on the quantitative results of individual learners' learning behaviors, the statistical training values of each learning method of the individual learners in the selected knowledge points are calculated; For each learning method, calculate the mean of the statistical training values corresponding to all selected knowledge points; The learning methods are sorted based on the mean value, and a fixed individual weight is assigned to each learning method according to the sorting result, and the sum of the individual weights is ensured to be 1, which is used as the learning method efficiency vector of the individual learner in the current subject.

Images