Adaptive optimization and resource personalized recommendation method and system for teaching resource codes

Abstract

The invention relates to the field of online education and intelligent teaching, and discloses a teaching resource code adaptive optimization and resource personalized recommendation method and system. According to the method, dynamic feature codes including dynamic difficulty, quality and timeliness are introduced, and a static feature region and a dynamic feature region are separately managed; constructing a quaternary heterogeneous graph neural network, and generating a high-confidence pseudo tag from three perspectives of group behavior consistency, time sequence behavior rationality and content relevance through a graph model, a time sequence model and a content model; adaptively calculating a resource feature update value to form closed-loop optimization; and realizing personalized recommendation through a deep learning model based on the user portrait. According to the method, the problems of static resource coding, difficulty in mining implicit behaviors, high cost of cold start and manual intervention and the like in a traditional method are solved, recommendation normal form upgrading from a passive response demand to an active prediction demand is realized, and the accuracy and efficiency of teaching resource recommendation are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to teaching resource encoding and personalized recommendation technology based on heterogeneous graph neural networks and semi-supervised learning. Background Art

[0002] With the rapid development of online education and intelligent teaching platforms, digital management and personalized recommendation of teaching resources have become key technical means to improve teaching effectiveness. This demand is particularly prominent in actual teaching application scenarios.

[0003] On online higher education platforms, a computer science course may include hundreds of different types of teaching resources, including video lectures, programming exercises, theoretical documents, and tests. Students vary in their learning abilities, knowledge backgrounds, and progress. The same programming exercise may be too difficult for a beginner but too easy for an experienced student. Traditional resource management systems typically use static labels to categorize resources, such as labeling a programming exercise as "medium difficulty." However, these fixed labels fail to reflect the dynamic nature of resources in real-world use.

[0004] In K12 intelligent teaching systems, managing math problem banks faces similar challenges. A geometry problem originally marked as "difficult" may, after extensive student feedback, become less difficult due to the addition of a high-quality solution video. However, the system still recommends problems based on the original "difficult" label, resulting in a mismatch between recommendations and students' actual needs. Furthermore, students' learning status is dynamic. After mastering the basics of trigonometric functions, their ability to handle related advanced problems improves significantly, but existing systems struggle to capture these changes in learning status.

[0005] In the application scenario of enterprise training platforms, the timeliness of technical training resources is even more pronounced. For example, a tutorial on a JavaScript framework may be highly valuable in its early stages, but its usefulness gradually declines with the rapid evolution of technology. Existing recommendation systems often fail to dynamically detect these timeliness changes and may continue to recommend outdated technical materials to trainees, impacting training effectiveness.

[0006] The technical problems exposed in these application scenarios are mainly concentrated in the following aspects:

[0007] First, there is the issue of static resource feature coding. Existing technologies usually perform one-time labeling when resources are put into the database, including attributes such as difficulty level, quality score, and applicable objects. These labels remain unchanged during subsequent use. However, the actual characteristics of teaching resources will change with factors such as user feedback, content updates, and the passage of time. For example, the perceived difficulty of a beginner programming tutorial may be significantly reduced after it is equipped with an interactive code demonstration tool, but static coding cannot reflect this change, causing the recommendation system to continue matching according to the original difficulty, affecting the accuracy of recommendations.

[0008] Secondly, the depth of the correlation analysis between user behavior and resource characteristics is insufficient. Traditional methods typically employ simple collaborative filtering or content-based recommendation algorithms, which struggle to deeply explore the complex correlations inherent in user behavior data. In practice, user learning behavior often exhibits multi-layered and multi-dimensional characteristics, such as learning duration, frequency of repeat visits, completion quality, and subsequent behavioral paths. These behavioral data and resource characteristics exhibit complex, nonlinear relationships, making it difficult for existing shallow correlation analysis methods to fully utilize this information.

[0009] Third, there's a lack of an effective dynamic feature update mechanism. Even if the system can collect user feedback, existing technologies lack a mechanism to effectively transform this feedback into resource feature updates. Especially when labeled data is scarce, how to combine limited labeled information with extensive user behavior data for feature optimization remains a technical challenge that has yet to be effectively addressed.

[0010] Fourth, the role of time in feature modeling is often overlooked. The value and applicability of educational resources are clearly time-sensitive, and users' learning status is constantly evolving. However, existing technologies lack the ability to accurately model this time factor. For example, after completing a particular knowledge point, a student's receptiveness to related advanced content will significantly improve in the short term. However, this time window effect is difficult to capture in traditional recommendation systems.

[0011] Finally, the accuracy of personalized recommendations needs to be improved. Existing recommendation methods often rely on simple similarity matching based on historical user behavior, making it difficult to accurately capture a user's current learning status and true needs. Especially in scenarios where users have diverse learning paths and personalized learning goals, the traditional "one-size-fits-all" recommendation model can no longer meet the needs of precise teaching.

[0012] In response to the above technical problems, there is an urgent need to develop a teaching resource coding optimization and personalized recommendation technology solution that can realize dynamic updating of resource characteristics, in-depth mining of user behavior associations, and accurate modeling of the influence of time factors, so as to improve the recommendation accuracy and user experience of online education and intelligent teaching systems. Summary of the Invention

[0013] The purpose of this application is to provide a method and system for adaptive optimization of teaching resource coding and personalized resource recommendation to solve the problems raised in the above background technology.

[0014] This application discloses a method for adaptive optimization of teaching resource coding and personalized resource recommendation, comprising the following steps:

[0015] Building a digital resource library, wherein the teaching resources in the digital resource library include basic coding information and dynamic feature coding information, wherein the dynamic feature coding information includes dynamic difficulty, dynamic quality, and dynamic timeliness;

[0016] Collecting user behavior data, deeply analyzing the association between the user behavior data and the dynamic feature coding information of the teaching resources through a semi-supervised learning model, and establishing an intrinsic association model between the user behavior data and the dynamic feature coding information of the teaching resources;

[0017] Based on the established intrinsic correlation model, the updated values ​​of the dynamic characteristics of the teaching resources are calculated in real time by an adaptive dynamic characteristic calculation engine, and the calculated dynamic characteristic update values ​​are fed back to the resource encoding module of the digital resource library to update the dynamic characteristic encoding information of the resources;

[0018] Building a user profile based on the user behavior data and extracting user features;

[0019] Based on the updated resource dynamic feature coding information and the user features, a personalized teaching resource recommendation list is generated through a personalized recommendation engine.

[0020] In a preferred example, the dynamic feature coding information includes: a static feature area, which includes the initially marked difficulty coefficient, quality rating, timeliness and other feature reference values; a dynamic feature area, which includes features such as difficulty correction value, quality evolution value, timeliness attenuation value, etc., and is updated according to the real-time numerical value feedback from the adaptive dynamic characteristic calculation engine.

[0021] In a preferred embodiment, the semi-supervised learning model adopts a hybrid architecture of a quaternary heterogeneous graph neural network and multi-view collaborative learning, including: a graph structure initialization phase, in which a heterogeneous information network is constructed based on teaching resource encoding and initial annotation data; a pseudo-label iterative generation phase, in which the annotation data is dynamically expanded through a fourth-order confidence evaluation mechanism based on multi-view verification; a feature optimization and update phase, in which the pseudo-label is integrated to continuously optimize the dynamic feature representation of the resource, and output a feature association model and a high-confidence pseudo-label dataset;

[0022] Furthermore, the multi-perspective collaborative learning mechanism verifies pseudo labels from three perspectives: group behavior consistency, temporal behavior rationality, and content relevance, using a graph model, a temporal model, and a content model. A fourth-order confidence evaluation is used to dynamically adjust the verification strength of different perspectives.

[0023] Furthermore, the pseudo-label iterative generation stage adopts a four-order confidence evaluation mechanism to classify pseudo-labels into four levels: high confidence, medium confidence, low confidence, and invalid. Each level adopts a different verification strategy and update frequency.

[0024] In a preferred embodiment, the resource dynamic characteristic value update formula adopted by the adaptive dynamic characteristic calculation engine is:

[0025] feat_dyn_new=α×feat_dyn_current+(1-α)×(feat_static+β×Σ[weight(t_i)×Δfeat_behavior_i]) formula (5)

[0026] Where: feat_dyn_new is the latest resource feature value, feat_dyn_current is the current feature value, α is the smoothing coefficient, β is the adjustment coefficient of the total increment, feat_static is the static initial feature value, weight(t_i) is the time decay weight function of the i-th behavior, and Δfeat_behavior_i is the feature change value caused by the i-th behavior.

[0027] In a preferred example, the user portrait includes user characteristics such as user learning style, interest preferences, ability level, learning progress, etc. The user characteristics are dynamically updated through a portrait update model based on user learning behavior data, test results and other information.

[0028] In a preferred example, the personalized recommendation engine adopts a deep learning model, takes the user feature vector and the basic coding vector and dynamic feature coding vector of the teaching resources as input, calculates the resource relevance score based on the principles of matching user ability with resource difficulty, matching preference with resource quality, matching learning needs with resource timeliness, and generates a personalized recommendation list.

[0029] In a preferred embodiment, the quaternary heterogeneous graph neural network comprises three layers:

[0030] The user-behavior layer (L1) aggregates user node features to behavior nodes through the adjacency matrix to generate the initial behavior embedding vector:

[0031]

[0032] Among them A u is the user-behavior adjacency matrix, F uis the user feature tensor, σ is the activation function;

[0033] The resource-behavior layer (L2) integrates resource node features, behavior node features, and time position coding, and calculates the resource impact weight matrix through bilinear mapping:

[0034] W r→b =Softmax(MLP(F r ))⊙PosEmb(t)) Formula (2)

[0035] Among them F r is the resource feature, PosEmb(t) is the temporal position encoding function, ⊙ represents element-wise multiplication, MLP is a multi-layer perceptron, and Softmax is a normalization function;

[0036] The behavior-label layer (L3) calculates the contribution of the behavior to the label through the attention mechanism, and passes it to the label node as the cross-layer influence weight.

[0037] In a preferred embodiment, the behavior-label layer calculates the contribution of the behavior node to the label node through the attention mechanism as follows:

[0038]

[0039] Among them, F b is the behavior node feature, F l is the label feature, Wq and Wk are learnable parameter matrices, LeakyReLU is the leaky rectified linear unit activation function, and exp is the exponential function.

[0040] In a preferred embodiment, the calculation formula of the influence weight w_{ij} of behavior on resource feature attributes in the quaternary heterogeneous graph neural network is:

[0041]

[0042] in: is the attention weight of behavior i on the label, is the time gating value of behavior i, controlling the time sensitivity, is the influence weight of resource feature j on behavior i, γ is the attenuation coefficient, Π(Δt_i>90) is the time threshold indicator function, and the function value is 1 when the number of days Δt_i from behavior i to today is greater than 90 days, otherwise it is 0.

[0043] In a preferred example, the basic coding information in the digital resource library is stored in a relational database, the dynamic feature coding information is stored separately in a time series database, and a coding version tree is established to support bidirectional version jump queries.

[0044] In a preferred example, the temporal position code PosEmb(t) of the resource-behavior layer and the time attenuation coefficient λ in the influence weight calculation formula work together to jointly implement a time sensitivity mechanism in both the feature extraction and feature update stages.

[0045] This application also discloses a system for adaptive optimization of teaching resource coding and personalized resource recommendation, including:

[0046] A digital resource library module is used to construct a digital resource library, wherein the teaching resources in the digital resource library include basic coding information and dynamic feature coding information, wherein the dynamic feature coding information includes dynamic difficulty, dynamic quality and dynamic timeliness;

[0047] User behavior data collection module, used to collect user behavior data;

[0048] A semi-supervised learning module is used to deeply analyze the association between the user behavior data and the dynamic feature information of the teaching resources through a semi-supervised learning model, and establish an intrinsic association model between the user behavior data and the dynamic feature information of the teaching resources;

[0049] An adaptive dynamic characteristic calculation engine, configured to calculate in real time based on the intrinsic correlation model the updated values ​​of the dynamic characteristics of the teaching resources, and feed the calculated updated values ​​of the dynamic characteristics back to the resource encoding module of the digital resource library to update the dynamic characteristic encoding information of the resources;

[0050] A user portrait center, used to build user portraits based on the user behavior data and extract user features;

[0051] The personalized recommendation engine is used to generate a personalized teaching resource recommendation list based on the updated resource dynamic feature coding information and the user characteristics.

[0052] The adaptive optimization of teaching resource encoding and personalized resource recommendation method described in this application builds a digital resource library, collects user behavior data, uses a semi-supervised learning model to analyze the relationship between user behavior and resource characteristics, and updates resource characteristics in real time based on the association model. Ultimately, personalized resource recommendations are achieved based on user characteristics. This method overcomes the problems of static resource encoding and poor recommendation results in the existing technology, significantly improving the real-time and accuracy of resource representation and the precision of personalized recommendations, and has the following technical effects:

[0053] This technology introduces dynamic feature encoding information that includes dynamic difficulty, dynamic quality, and dynamic timeliness. It also employs a structural design that separates the initial feature baseline value from the correction values ​​updated based on user behavior (such as difficulty correction value, quality evolution value, and time decay value). This allows resource features to be dynamically adjusted based on user behavior feedback. This technical solution addresses the static nature of traditional encoding through dynamic encoding structure design, overcomes the representation distortion caused by traditional encoding rigidity, and significantly improves the real-time and accuracy of feature representation.

[0054] A hybrid architecture combining a quaternary heterogeneous graph neural network and multi-perspective collaborative learning is employed. In the three phases of graph structure initialization, pseudo-label iterative generation, and feature optimization and updating, a heterogeneous graph consisting of three types of nodes—users, resources, and behaviors—is constructed. Graph neural networks aggregate multiple types of relationships (such as the "user-error behavior-knowledge point-resource" path). Combined with semi-supervised learning (label propagation + self-training), the semantic information of unlabeled data is converted into vector features. The structural information of unlabeled data is mined, and a deep graph association structure between users, behaviors, resources, and labels is established. Feature association models and pseudo-label data for updating resource features are generated, forming a closed-loop data optimization. Traditional recommendation systems often use collaborative filtering or single-modal models (such as only a user-resource matrix), which are unable to effectively integrate multi-dimensional information such as user behavior, resources (including resource feature information), and user behavior hierarchies. This leads to problems such as "difficulty in mining implicit behavior" and "serious cold start." This technical solution overcomes the bottleneck of implicit behavior mining through a semi-supervised learning model, improves feature association modeling capabilities, and significantly enhances the accuracy and practicality of recommendations in educational scenarios.

[0055] In addition, this technical solution realizes the adaptive evolution of the system through a closed-loop optimization mechanism, forming a complete closed-loop link of "encoding update conditions → semi-supervised model output dynamic feature parameters → encoding update calculation → model retraining". Semi-supervised learning reduces dependence on manual labeling, while the closed-loop mechanism reduces the frequency of manual review. Effectively utilize massive amounts of unlabeled behavioral data to automatically establish a complex correlation model between user behavior and intrinsic characteristics of resources (difficulty, quality, timeliness, etc.) and generate pseudo-label data. Directly quantifying behavioral patterns as the driving force of changes in resource characteristics solves the problem that traditional methods cannot utilize unstructured behavioral data, and realizes the adaptive evolution of the system, which has the advantages of dynamic adaptability, improved recommendation accuracy and reduced labor costs.

[0056] In terms of resource dynamic feature updates, the adaptive dynamic feature calculation engine uses the following formula to update feature values ​​based on the combined effects of behavioral influence and time factors:

[0057] feat_dyn_new=α×feat_dyn_current+(1-α)×(feat_static+β×Σ[weight(t_i)×Δfeat_behavior_i]) formula (5)

[0058] This formula avoids drastic fluctuations in features while maintaining feature stability, reflecting the dynamic trend of resource features evolving with user behavior.

[0059] In terms of user portrait construction, user characteristics including learning style, interest preferences, ability level, learning progress, etc. are extracted, and as user behavior data is continuously updated, a dynamic representation of user status is formed to provide data support for subsequent recommendations.

[0060] The personalized recommendation engine, based on a high-level model, constructs an interactive matching relationship between user feature vectors and resource encoding vectors, achieving a precise match between "real user needs and resource characteristics." Using a deep learning model, the engine outputs personalized recommendations based on principles such as matching user capabilities with resource difficulty, matching user preferences with resource quality, and matching learning needs with resource timeliness. Unlike traditional, one-size-fits-all recommendation methods, this approach deeply explores users' individual learning characteristics, accurately capturing their actual learning needs and status, significantly improving the relevance and accuracy of resource recommendations.

[0061] In the heterogeneous graph neural network structure, a three-layer hierarchical architecture is adopted:

[0062] The user-behavior layer uses the following formula to construct the initial embedding vector of user features and behavior nodes:

[0063]

[0064] The resource-behavior layer fuses resource features with time location encoding using the following formula to form time-sensitive resource behavior weights:

[0065] W r→b =Softmax(MLP(F r ))⊙PosEmb(t))Formula (2):

[0066] The behavior-label layer uses the following attention mechanism to calculate the contribution of the behavior node to the label node:

[0067]

[0068] This three-layer structure realizes the full-chain modeling of users, behaviors, resources and labels, and improves the expressive ability of behavior-driven resource feature updates.

[0069] The final impact weight of resource features is calculated by the following formula, which combines attention weight, time gating value, resource impact weight and time decay mechanism:

[0070]

[0071] This mechanism effectively models the dynamic impact of user behavior on resource characteristics and has time-sensitive regulation capabilities.

[0072] The digital resource library adopts a structure that combines a relational database and a time-series database, separately stores basic coding and dynamic coding information, and introduces a version tree structure to achieve historical tracing and two-way jump of coding status, thereby improving the manageability of resource status.

[0073] In addition, the temporal position encoding PosEmb(t) and the time decay strategy (1-γ·Π(Δt_i>90)) in the influence weight calculation are used in collaboration in the feature extraction and update stages to achieve the front-to-back linkage of time-sensitive modeling.

[0074] The generation of pseudo-labels adopts a four-order confidence evaluation mechanism, which divides pseudo-labels into four levels: high, medium, low, and invalid. Different verification strategies and update frequencies are adopted for different confidence levels. Combined with the three-perspective verification of graph model, time series model, and content model, the quality of pseudo-labels and learning efficiency are improved.

[0075] In summary, this application proposes a closed-loop recommendation system with dynamic resource feature modeling as the core, combined with graph neural network behavior analysis, multi-perspective pseudo-label verification, time-sensitive control and personalized recommendation and other technical means. It improves the dynamics of the resource coding system, the comprehensiveness of user understanding, the accuracy of behavior analysis and the personalization level of recommendation results, and is significantly better than the existing recommendation method based on static labels.

[0076] The specification of this application records a large number of technical features, which are distributed in various technical solutions. If all possible combinations of technical features of this application (i.e., technical solutions) are to be listed, the specification will be too lengthy. In order to avoid this problem, the various technical features disclosed in the above-mentioned invention content of this application, the various technical features disclosed in the various embodiments and examples below, and the various technical features disclosed in the accompanying drawings can be freely combined with each other to form various new technical solutions (these technical solutions are all deemed to have been recorded in this specification), unless such a combination of technical features is technically infeasible. For example, in one example, feature A+B+C is disclosed, and in another example, feature A+B+D+E is disclosed. Features C and D are equivalent technical means that play the same role. Technically, only one of them can be used, and it is impossible to use them at the same time. Feature E can be technically combined with feature C. Then, the solution of A+B+C+D should not be considered as having been recorded because it is technically infeasible, while the solution of A+B+C+E should be considered as having been recorded. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the system architecture of the adaptive optimization of teaching resource coding and personalized resource recommendation solution according to this application.

[0078] Figure 2 It is a flow chart of the adaptive optimization of teaching resource coding and personalized resource recommendation method according to the first embodiment of the present application.

[0079] Figure 3 It is a structural diagram of the adaptive optimization of teaching resource coding and personalized resource recommendation system according to the second embodiment of the present application. DETAILED DESCRIPTION

[0080] In the following description, many technical details are provided to help readers better understand this application. However, those skilled in the art will understand that even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can be implemented.

[0081] Description of some concepts:

[0082] A digital resource library refers to a system or platform that centrally stores, manages, and provides unified access to teaching resources (such as courseware, test questions, videos, exercise explanations, etc.) in digital form.

[0083] User behavior data refers to the various operation records of learners in the process of using teaching resources, including but not limited to clicks, browsing, pausing, collecting, marking, asking questions about knowledge points, concept associations, and attempting exercises.

[0084] The static feature area refers to a set of a priori attributes of resources, such as difficulty level, knowledge point label, content type, etc. These features remain unchanged in the short term and are used to describe the basic characteristics of resources.

[0085] The dynamic feature area refers to resource attributes that are updated in real time based on user behavior feedback, such as difficulty correction value, quality evolution value, time attenuation value, etc., which are used to characterize the dynamic changes of resources with usage.

[0086] Semi-supervised learning models refer to machine learning methods that use a small amount of manually labeled data and a large amount of unlabeled data to mine implicit behavioral patterns through techniques such as pseudo-label generation, self-training, and label propagation.

[0087] Pseudo labels refer to the labels automatically predicted and assigned by the model to unlabeled samples during the semi-supervised learning process, which are used to expand the training set and iteratively optimize the model.

[0088] Heterogeneous GNN refers to a deep learning model built within the graph neural network framework for a variety of node types (such as users, resources, behaviors, and labels) and edge types (such as users-resources, behaviors-knowledge points). It is used to capture complex relationships between multi-dimensional entities and perform feature propagation and aggregation.

[0089] Multi-perspective collaborative learning refers to learning from different data perspectives (such as time series perspective, content perspective, and user perspective) under the same model or multi-model parallel framework, and then collaboratively integrating the results of each perspective to enhance the model's comprehensive understanding of multi-source information.

[0090] The feature association model refers to a mapping model generated based on heterogeneous graph neural networks and semi-supervised learning, which is used to characterize the deep correlation between user behavior and dynamic resource features.

[0091] The closed-loop optimization mechanism refers to the cyclic process of "data collection → model learning → feature update → personalized recommendation". It enhances the model capabilities through continuous iteration and enables the system to have the ability of adaptive evolution.

[0092] Personalized recommendation refers to screening and prioritizing the most suitable and valuable teaching resources in the resource library based on the user's ability level, interest preferences and learning needs, so as to improve learning effects and usage satisfaction.

[0093] The cold start problem refers to the phenomenon that when there is a lack of sufficient historical interaction data, the system has difficulty in accurately inferring the preferences of new users or new resources, resulting in poor recommendation results.

[0094] The following is a summary of some of the innovative features of this application:

[0095] The teaching resource coding adaptive optimization and personalized recommendation method proposed in this application is guided by technical problems existing in the fields of online education and intelligent teaching, such as the static nature of resource feature coding, insufficient depth of user behavior association analysis, lack of dynamic feature update mechanism, inaccurate time factor modeling, and low accuracy of personalized recommendation. It carries out systematic and chain-like technical construction and innovative design from the aspects of digital resource library construction, user behavior data collection, semi-supervised learning association modeling, adaptive dynamic feature calculation, user portrait extraction, and personalized recommendation generation.

[0096] Specifically, in the construction of the digital resource library, this application introduces dynamic feature coding information including dynamic difficulty, dynamic quality and dynamic timeliness, and adopts an innovative coding structure that separates the management of static feature areas and dynamic feature areas, laying a solid foundation for the subsequent dynamic evolution of resource features. In the design of the semi-supervised learning association model, with the four-element heterogeneous graph neural network as the core, through the composite and progressive optimization of the graph structure initialization, pseudo-label iterative generation, feature optimization and update stages, the association features contained in the massive user behavior data are deeply mined. At the same time, a multi-perspective collaborative learning mechanism is introduced to construct a comprehensive and three-dimensional feature association analysis framework from the three dimensions of group behavior consistency, temporal behavior rationality, and content relevance.

[0097] Furthermore, in implementing an adaptive dynamic feature calculation engine, this application innovatively designs a resource feature update algorithm that integrates behavioral influence weights, time position encoding, time decay strategies, and a smooth update mechanism. This ensures the smoothness of feature updates while also taking into account time sensitivity, enabling the real-time synchronous evolution of resource features and user behavior. At the same time, this application also proposes a solution for dynamically constructing user profiles based on user behavior data. By extracting multi-dimensional user features such as learning style, interest preferences, ability level, and learning progress, it forms the key support for accurately portraying the user's real-time status.

[0098] Finally, in the design of the personalized recommendation engine, this application uses a deep learning model to interactively match the extracted user feature vectors with the updated resource encoding vectors, and perform multi-criteria sorting from the perspectives of user capabilities, preferences, learning needs, etc., ultimately achieving a precise match between "user real needs-resource characteristics" and significantly improving the fit of personalized recommendations.

[0099] In summary, this application has formed an organic technical synergy of "data-driven closed-loop optimization, deep behavioral association of graph networks, introduction of time and position coding, generation of multi-perspective pseudo-labels, dynamic user feature profiling, and hybrid sorting personalized recommendations" through system design and innovative technical solutions in key links such as resource library construction, behavioral association modeling, dynamic feature calculation, user portrait extraction, and personalized recommendation generation. Ultimately, it achieved a leapfrog upgrade from "static coding" to "dynamic evolution of features", from "shallow behavioral association" to "deep graph network mining", and from "offline recommendation generation" to "online real-time update", overcoming the industry's difficult problems in personalized resource recommendations in the field of education, and demonstrating significant technical advantages and application value in improving recommendation accuracy, optimizing manual review costs, and enabling smart teaching decision-making.

[0100] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0101] The first embodiment of the present application relates to a method for adaptive optimization of teaching resource coding and personalized resource recommendation, the process of which is as follows: Figure 2 As shown, the method includes the following steps:

[0102] The method for adaptive optimization of teaching resource coding and personalized resource recommendation of this embodiment includes the following steps:

[0103] Step 100: Construct a digital resource library, wherein the teaching resources in the digital resource library include basic coding information and dynamic feature coding information, wherein the dynamic feature coding information includes dynamic difficulty, dynamic quality and dynamic timeliness.

[0104] Optionally, the dynamic feature coding information includes: a static feature area, which contains the initially marked difficulty coefficient, quality rating, timeliness and other feature reference values; a dynamic feature area, which contains features such as difficulty correction value, quality evolution value, timeliness attenuation value, etc., and is updated according to the real-time numerical value feedback from the adaptive dynamic characteristic calculation engine.

[0105] Optionally, the basic coding information in the digital resource library is stored in a relational database, and the dynamic feature coding information is stored separately in a time series database, and a coding version tree is established to support bidirectional version jump queries.

[0106] Specifically, this embodiment uses a two-layer coding architecture that combines basic coding information with dynamic feature coding information when constructing a digital resource library, effectively solving the problem of resource feature solidification in traditional teaching resource management systems. Basic coding information includes relatively stable static attributes such as resource type, format, subject, and knowledge points, and is mainly used for basic retrieval and classification of resources. For example, the basic coding of a calculus teaching video may include the resource type as "video", the subject classification as "mathematics-calculus", and the knowledge point as "derivative concept".

[0107] The core innovation of dynamic feature encoding information lies in the introduction of two interrelated yet relatively independent areas: the static feature area and the dynamic feature area. The static feature area stores the baseline feature values ​​initially annotated by education experts, such as difficulty coefficient, quality rating, and timeliness level, providing a stable reference for subsequent dynamic adjustments. The dynamic feature area stores correction values, such as difficulty correction value, quality evolution value, and timeliness decay value, which are calculated and updated in real time based on actual user feedback, reflecting the dynamic changes in resource characteristics.

[0108] In terms of storage architecture, the system uses a relational database to store basic coding information to ensure data consistency and query efficiency. Dynamic feature coding information is stored separately in a time-series database, leveraging the advantages of time-series databases in processing timestamped data and enabling efficient recording and querying of feature changes over time. Furthermore, the system introduces a coding version tree to implement versioned management of resource feature coding and support bidirectional version jump queries. Users can trace historical changes in resource features and roll back to the coding state at a specific point in time for comparative analysis.

[0109] The technical effects of this two-layer coding architecture are reflected in the following three aspects:

[0110] The separation of static and dynamic features enables the system to reflect the dynamic performance of resources in actual teaching in real time while maintaining the stability of the basic attributes of the resources;

[0111] The use of a time series database significantly improves the efficiency of dynamic feature updates and supports high-frequency feature modification operations;

[0112] The version tree structure provides data support for teaching quality analysis and resource optimization, enabling education managers to timely discover quality trends and applicability changes of teaching resources based on feature change trajectories.

[0113] In summary, this embodiment uses innovative designs such as a two-layer coding architecture, a time series database, and a version tree to build a digital resource library that can fully reflect the dynamic characteristics of teaching resources, support flexible management, and efficient retrieval, providing a solid data foundation for subsequent adaptive optimization and personalized recommendations.

[0114] Step 200: Collect user behavior data, deeply analyze the relationship between the user behavior data and the dynamic feature coding information of the teaching resources through a semi-supervised learning model, and establish an intrinsic correlation model between the user behavior data and the dynamic feature coding information of the teaching resources.

[0115] Optionally, the semi-supervised learning model adopts a hybrid architecture of a quaternary heterogeneous graph neural network and multi-view collaborative learning, including: a graph structure initialization phase, in which a heterogeneous information network is constructed based on the teaching resource encoding and initial annotated data; a pseudo-label iterative generation phase, in which the annotated data is dynamically expanded through a fourth-order confidence evaluation mechanism based on multi-view verification; a feature optimization and update phase, in which the pseudo-label is integrated to continuously optimize the dynamic feature representation of the resource, and output a feature association model and a high-confidence pseudo-label dataset;

[0116] Furthermore, the multi-perspective collaborative learning mechanism verifies pseudo labels from three perspectives: group behavior consistency, temporal behavior rationality, and content relevance, using a graph model, a temporal model, and a content model. It uses a fourth-order confidence evaluation to dynamically adjust the verification strength of different perspectives.

[0117] Furthermore, the pseudo-label iterative generation stage adopts a four-order confidence evaluation mechanism to classify pseudo-labels into four levels: high confidence, medium confidence, low confidence, and invalid. Each level adopts a different verification strategy and update frequency.

[0118] Optionally, the quaternary heterogeneous graph neural network comprises three layers:

[0119] The user-behavior layer (L1) aggregates user node features to behavior nodes through the adjacency matrix to generate the initial behavior embedding vector:

[0120] V b (1) =σ(A u ·F u ) Formula (1)

[0121] Among them A u is the user-behavior adjacency matrix, F u is the user feature tensor, σ is the activation function;

[0122] The resource-behavior layer (L2) integrates resource node features, behavior node features, and time position coding, and calculates the resource impact weight matrix through bilinear mapping:

[0123] W r→b =Softmax(MLP(F r ))⊙PosEmb(t)) Formula (2)

[0124] Among them F ris the resource feature, PosEmb(t) is the temporal position encoding function, ⊙ represents element-wise multiplication, MLP is a multi-layer perceptron, and Softmax is a normalization function;

[0125] The behavior-label layer (L3) calculates the contribution of the behavior to the label through the attention mechanism, and passes it to the label node as the cross-layer influence weight.

[0126] Optionally, the temporal position code PosEmb(t) of the resource-behavior layer and the time decay coefficient λ in the influence weight calculation formula work together to jointly implement a time sensitivity mechanism in both the feature extraction and feature update stages.

[0127] Optionally, the behavior-label layer calculates the contribution of the behavior node to the label node through the attention mechanism using the following specific formula:

[0128]

[0129] Among them, F b is the behavior node feature, F l is the label feature, Wq and Wk are learnable parameter matrices, LeakyReLU is the leaky rectified linear unit activation function, and exp is the exponential function.

[0130] Optionally, the calculation formula for the influence weight w_{ij} of behavior on resource feature attributes in the quaternary heterogeneous graph neural network is:

[0131]

[0132] in: is the attention weight of behavior i on the label, is the time gating value of behavior i, controlling the time sensitivity, is the influence weight of resource feature j on behavior i, γ is the attenuation coefficient, Π(Δt_i>90) is the time threshold indicator function, and the function value is 1 when the number of days Δt_i from behavior i to today is greater than 90 days, otherwise it is 0.

[0133] Specifically, this embodiment adopts a semi-supervised learning model in step 200, which integrates a four-element heterogeneous graph neural network and a multi-perspective collaborative learning mechanism, breaking through the limitation of traditional teaching recommendation systems that only rely on the user-resource binary relationship.

[0134] The heterogeneous graph constructed by the model includes four heterogeneous entities: user nodes (U), resource nodes (R), behavior type nodes (B), and label nodes (L). The introduction of behavior type nodes is a key innovation. It explicitly models various user behaviors (such as browsing, collecting, and doing exercises) as independent nodes in the graph, allowing complex multi-element interaction relationships to be accurately expressed. For example, when student A repeatedly pauses and adds note annotations while learning a video on a certain knowledge point, the system establishes a four-element association path of "student A-repeated learning behavior-knowledge point video-difficulty label" in the heterogeneous graph.

[0135] In the three-level message passing architecture, the system realizes the progressive flow of information from user features to resource features. The user-behavior layer (L1) maps the user's basic features to the behavior nodes using formula (1), generating an initial feature representation of the behavior. The resource-behavior layer (L2) uses the time position encoding function in formula (2) to realize the time-sensitive aggregation of resource features to behavior nodes, ensuring that recent behaviors have higher influence weights. The behavior-label layer (L3) calculates the contribution of each behavior to the resource label through the attention mechanism in formula (3).

[0136] The four-order confidence evaluation mechanism builds a refined pseudo-label quality control system. High-confidence labels need to pass the three-perspective strong verification of the graph model, the time series model, and the content model, and can be directly used for model training. Medium-confidence labels introduce double-feature space correlation calculation for supplementary verification. Low-confidence labels use the "machine preliminary judgment + manual calibration" method, and experts focus on reviewing the difficulty, quality, and timeliness of the three key dimensions. Invalid labels are used as negative samples for contrastive learning to improve the model's discrimination ability.

[0137] The advantage of the multi-perspective collaborative learning mechanism lies in its multi-level verification strategy. The graph model perspective checks the consistency of group behavior, the time series model perspective models the compliance of user behavior sequences, and the content model perspective constructs the correlation index of behavior and resource content from the dimensions of text, vision, and semantics.

[0138] This technical solution can significantly reduce the cost of manual annotation while improving the quality of annotation. With the help of semi-supervised learning, the system can automatically annotate high-confidence labels, reduce the amount of manual review of medium-confidence labels, and improve the overall annotation efficiency. In addition, the introduction of the four-element heterogeneous graph enables the system to mine complex interaction patterns that traditional binary models cannot capture, quickly builds feature representations for new users or new resources through behavior nodes in the cold start scenario, and effectively alleviates the problem of data sparsity.

[0139] Step 300: Based on the internal correlation model established in step 200, the adaptive dynamic characteristic calculation engine calculates the updated value of each dynamic characteristic of the teaching resource in real time, and feeds back the calculated dynamic characteristic update value to the resource encoding module of the digital resource library to update the resource dynamic characteristic encoding information.

[0140] Optionally, the resource dynamic characteristic value update formula used by the adaptive dynamic characteristic calculation engine is:

[0141] feat_dyn_new = a x feat_dyn_current + (1-a) x (feat_static + b x å[weight(t_i) x Dfeat_behavior_i]) Formula (5)

[0142] Wherein: feat_dyn_new is the latest resource characteristic value, feat_dyn_current is the current characteristic value, a is the smoothing coefficient, b is the total increment adjustment coefficient, feat_static is the static initial characteristic value, weight(t_i) is the time decay weight function of the i-th behavior, and Dfeat_behavior_i is the characteristic change value caused by the i-th behavior.

[0143] Specifically, the adaptive dynamic characteristic calculation engine in step 300 realizes the smooth and gradual update of the teaching resource characteristics based on formula (5), effectively balancing the contradiction between characteristic stability and response sensitivity.

[0144] The smoothing coefficient a in the formula controls the weight ratio of the current characteristic value and the newly calculated characteristic change value. Reasonable setting of a value can avoid the drastic fluctuation of resource characteristics caused by individual abnormal behavior. For example, when a programming exercise problem originally labeled as medium difficulty is repeatedly attempted and frequently failed by a large number of beginners in a certain course, the system will not immediately adjust its difficulty to the difficult level, but will gradually adjust through the smoothing mechanism to ensure the rationality and traceability of the characteristic change.

[0145] The setting of the total increment adjustment coefficient b reflects the precise control of the system on the influence degree of different types of behavior. The b value is usually dynamically adjusted according to the behavior type to ensure the dominant role of high-quality behavior data in resource characteristic update. The time decay weight function weight(t_i) adopts an exponential decay model, with higher weight for recent behaviors and gradual decay over time. This time decay mechanism ensures the timeliness of resource characteristic update, making recent user behavior have greater influence on characteristic adjustment.

[0146] The calculation of the feature change value Δfeat_behavior_i is based on analyzing the deviation between user behavior and expected behavior. For example, when the average learning time of a user on a video tutorial significantly exceeds the expected duration, the system calculates a positive difficulty correction value. A calibration function maps the deviation rate to the feature change range, resulting in the corresponding difficulty correction value. Similarly, the change values ​​of the quality and timeliness features are calculated based on comprehensive user feedback.

[0147] The core of the adaptive mechanism lies in its ability to dynamically adjust calculation parameters according to different teaching scenarios. In K12 basic education scenarios, the system tends to use a larger α value to maintain the stability of resource characteristics; in higher education or vocational training scenarios, the α value is relatively small to more quickly respond to learners' ability improvement and knowledge update needs. For resources with strong practical programming, the β value is usually set higher, because users' actual operational feedback can more directly reflect the difficulty of the resource; for theoretical explanation resources, the β value is relatively low to avoid misjudgments due to large individual differences.

[0148] The computing engine also implements a multi-feature collaborative update mechanism. When a resource's difficulty characteristic changes significantly, the system automatically checks whether its quality and timeliness characteristics require corresponding adjustments, avoiding inconsistent resource descriptions caused by isolated changes to a single characteristic.

[0149] This technical solution significantly improves the dynamic accuracy of resource feature representation. By adopting an adaptive dynamic update mechanism, the matching between resource features and users' actual learning experience is significantly improved. More importantly, the system can automatically identify changing trends in resource quality and promptly identify outdated resources that need to be updated or retired. This provides data support for the lifecycle management of teaching resources, effectively improving the accuracy of resource recommendations and the rationality of user learning paths.

[0150] Step 400: Build a user profile based on the user behavior data and extract user features.

[0151] Optionally, the user portrait includes user characteristics such as user learning style, interest preferences, ability level, learning progress, etc. The user characteristics are dynamically updated through a portrait update model based on user learning behavior data, test results and other information.

[0152] Specifically, the user portrait construction system in step 400 achieves accurate modeling and real-time tracking of the user's learning status through multi-dimensional feature extraction and dynamic update mechanism.

[0153] The core of the user portrait is to convert abstract learning behavior into a quantifiable feature vector, providing accurate user representation for subsequent personalized recommendation. Learning style features use a multivariate learning theory model to classify users into four main categories: visual, auditory, hands-on, and reading. The system automatically identifies users by analyzing their preferences for different media types of resources. For example, if a user's average time spent on video resources is significantly higher than that on text resources, and the completion rate of video resources is significantly higher than that of text resources, the system will identify the user as a visual learner and give a corresponding visual learning tendency score.

[0154] Interest preference features are quantitatively modeled based on knowledge point access frequency and learning depth. The system maintains a knowledge point interest matrix, recording the user's attention and mastery level in each knowledge domain. Interest is calculated by considering factors such as user access frequency, average learning time, and interaction behavior frequency, and the final interest score is obtained by weighted average, reflecting the user's learning interest in a specific knowledge domain.

[0155] Ability level features are measured by a multi-dimensional evaluation system, including basic knowledge mastery, problem-solving ability, learning speed, and knowledge transfer ability. Basic knowledge mastery uses an exponential moving average method to achieve dynamic updating, taking into account the user's current mastery level and latest test scores. Problem-solving ability focuses on the user's performance in solving comprehensive problems across knowledge points. Learning speed is quantified by learning efficiency, reflecting the user's learning absorption ability. Knowledge transfer ability is evaluated by analyzing the correlation between the user's learning effects in related knowledge domains.

[0156] Learning progress features use a personalized learning path tracking mechanism to record the user's completion and mastery in the preset curriculum system. The system maintains a learning progress chart for each user, including linear course completion percentage and knowledge network coverage breadth and mastery depth, providing accurate basis for personalized recommendation.

[0157] The portrait update model uses an incremental learning mechanism to ensure that user features can reflect changes in learning status in a timely manner. The update frequency is dynamically adjusted according to user activity, with high-active users updating frequently and low-active users updating relatively slowly. The update algorithm uses a time decay weight mechanism, with recent behavior having a greater impact on portrait features.

[0158] This user profiling system significantly improves the accuracy of personalized recommendations. Multi-dimensional dynamic profiling significantly increases the accuracy of user feature recognition, enabling the system to accurately predict the difficulty of learning new knowledge points, providing a reliable basis for adaptive learning path planning. The dynamic update mechanism enables the system to promptly capture changes in the user's learning status, particularly shifts in learning interests and improvements in ability levels, providing a foundation for continuous personalized service. Furthermore, in cold start scenarios, the system quickly constructs a preliminary profile based on the user's initial test results and learning behavior patterns, significantly shortening the time it takes to build a new user profile and improving the learning experience for new users.

[0159] Step 500: Based on the updated resource dynamic feature coding information and the user features, a personalized teaching resource recommendation list is generated through a personalized recommendation engine.

[0160] Optionally, the personalized recommendation engine adopts a deep learning model, takes the user feature vector and the basic coding vector and dynamic feature coding vector of the teaching resources as input, calculates the resource relevance score based on the principles of matching resource difficulty with user ability, matching resource quality with preference, matching resource timeliness with learning needs, and generates a personalized recommendation list.

[0161] Specifically, the personalized recommendation engine in step 500 adopts a deep neural collaborative filtering model, and by constructing a high-dimensional interactive space of user feature vectors and resource encoding vectors, it realizes a technical paradigm upgrade from traditional "similarity-based recommendation" to "adaptability-based recommendation".

[0162] The engine's core innovation lies in the weighted fusion of a resource's basic encoding vector and its dynamic feature encoding vector to form a comprehensive resource representation vector. The fusion weights are dynamically adjusted based on the resource type. For theoretical knowledge resources, the stability of the underlying attributes is prioritized, while for practical operation resources, the guiding significance of dynamic feedback is emphasized. This design ensures that the recommendation algorithm leverages the inherent properties of resources while fully considering the evolution of features driven by user feedback.

[0163] The technical implementation for matching user capabilities with resource difficulty is based on the construction of an ability-difficulty adaptation function. The system maintains a spectrum of user capabilities and resource difficulty, and calculates the degree of adaptation using a modified cosine similarity formula. It also introduces the concept of the "zone of proximal development" to recommend resources with a difficulty slightly higher than the user's current capabilities, thereby promoting capability improvement.

[0164] The mechanism for matching user preferences with resource quality uses a multi-layer perceptron network to achieve a deep mapping of preference and quality features. The user preference vector encompasses multiple dimensions, including content preferences, media preferences, and teaching style preferences, while the resource quality vector includes quality indicators such as content accuracy, clarity of expression, and interactivity. The network training data is derived from historical user feedback signals.

[0165] The algorithm for matching learning needs with resource timeliness takes into account the urgency of the user's current learning goals and the time sensitivity of the resource content. The learning needs vector contains short-term, medium-term, and long-term goals, and the resource timeliness score is calculated based on factors such as content release time and technology update frequency. The matching algorithm uses a time decay weight, with faster decay for technical resources and slower decay for theoretical resources.

[0166] The generation of personalized recommendation lists uses a multi-objective optimization strategy, considering adaptability, diversity, and novelty in three dimensions. The recommendation list is optimized using a reordering algorithm to ensure that the recommended results cover multiple knowledge subfields and avoid excessive concentration of content.

[0167] The recommendation engine also implements real-time learning and model updating mechanisms. The system continuously collects user feedback on the recommended results and adjusts model parameters in real time through online learning algorithms. The model update frequency is adaptively adjusted based on user activity to ensure that the recommendation system can quickly adapt to changes in user needs and resource feature evolution.

[0168] This technical solution significantly improves recommendation accuracy and user satisfaction. The system can accurately identify the user's knowledge transfer potential, and the acceptance rate of cross-domain recommendations is significantly improved. In the new user cold start scenario, by combining user portraits and resource dynamic characteristics, the system can provide effective recommendations even with only a small amount of learning behavior records. The multi-objective optimization strategy effectively balances the accuracy and diversity of recommendations, and the user's evaluation of the novelty of the recommended content and the rationality of the learning path are significantly improved, with a clear improvement in average learning efficiency.

[0169] In order to better understand the technical solution of the present application, a specific example is given below, which mainly lists the details for the purpose of understanding, and does not limit the protection scope of the present application.

[0170] In this example, an adaptive optimization of teaching resource coding and personalized resource recommendation is provided, mainly including user behavior data collection center, digital resource library, semi-supervised learning model, adaptive dynamic characteristic calculation engine, user portrait center, and personalized recommendation engine. The digital teaching resources constructed by this method contain basic coding information and dynamic difficulty, dynamic quality, and dynamic timeliness, etc. key feature coding information; the semi-supervised learning module deeply analyzes the user teaching and learning behavior, and automatically establishes the internal correlation model between behavior and resource characteristics; the adaptive dynamic characteristic calculation module uses the model to calculate the feature update value in real time; the resource coding module automatically updates the dynamic feature coding accordingly; the recommendation engine is based on the updated resource coding and user portrait to realize highly personalized resource matching. This example significantly improves the real-time, accuracy, and precision of personalized recommendation of resource representation. The specific system block diagram is as follows Figure 1As shown, specifically as follows:

[0171] (I) Construction of digital resource library

[0172] The focus of constructing a digital resource library is the design of a resource coding system. The coding information includes basic information coding and key feature coding. The feature coding is dynamically updated according to the real-time use behavior and feedback data of users. Through the realization of real-time updating and historical tracing of resource features, the shortcomings of the traditional coding system are solved. The resource storage architecture is that the basic coding is stored in a relational database, and the feature coding is stored separately in a time series database. A coding version tree is established to support bidirectional version jump query.

[0173] Specifically, the basic coding components include resource name hash value, file format code, subject classification code, knowledge point identifier, etc., which serve as the unique identifier of the resource and support basic retrieval and classification.

[0174] Specifically, the feature coding includes static feature area containing initial annotated difficulty coefficient, quality rating, timeliness, etc. The static feature area is generated by artificial annotation of education experts and serves as a reference benchmark for dynamic updating. The dynamic feature area is composed of difficulty correction value, quality evolution value, timeliness decay value, etc. The dynamic feature area is updated according to the real-time values fed back by the adaptive dynamic characteristic calculation engine.

[0175] (II) Construction of semi-supervised learning model based on teaching scene

[0176] This model adopts a hybrid architecture of "four-element heterogeneous graph neural network + multi-view collaborative learning". The semi-supervised learning is realized through a three-level processing flow. In the graph structure initialization stage, a heterogeneous information network is constructed based on teaching resource coding and initial annotation data. In the pseudo-label iteration generation stage, the annotation data is dynamically expanded through a four-order confidence evaluation mechanism. In the feature optimization update stage, the pseudo-label is continuously optimized to represent the dynamic features of the resources. Finally, the feature correlation model and high-confidence pseudo-label dataset are output. The overall model forms a closed-loop optimization of "data input → graph construction → feature learning → decision verification → feature update → new round of optimization". The data output is used for the adaptive dynamic feature calculation engine.

[0177] Optionally, the model has the following detailed design method:

[0178] (1) Heterogeneous information network construction

[0179] Abstract teaching entities and their relationships are converted into computable graph data structures, which serve as the basic carrier for semi-supervised learning. A heterogeneous graph network containing "user nodes (U)", "resource nodes (R)", "behavior type nodes (B)", and "label nodes (L)" is constructed. Specifically as follows:

[0180] Node type definition:

[0181]

[0182]

[0183] Edge type and weight mechanism:

[0184]

[0185] (2) Graph Neural Network Design

[0186] Design of a Graph Neural Network Behavior Feature Weight Calculation Mechanism: This graph neural network uses a three-order message passing architecture to model the quantitative impact of behavior on resource feature attributes. It also constructs a weight calculation system that includes hierarchical aggregation, attention enhancement, and dynamic optimization. The specific technical implementation is as follows:

[0187] 1. Hierarchical aggregation architecture design, using a three-level feature transfer link to achieve layer-by-layer accumulation of behavior weights

[0188] ① User-behavior layer (L1): Constructs a feature map from user nodes to behavior nodes, aggregates basic user features (age / gender / historical preferences, etc.) through the adjacency matrix, and generates the initial behavior weight vector:

[0189]

[0190] Among them A u is the “user-behavior” association matrix, F u For the user feature tensor, it lays the foundation for the user-side impact of the behavior.

[0191] ② Resource-Behavior Layer (L2): The core implements the weight injection of resource features into behavior nodes and integrates multi-dimensional resource attributes through bilinear transformation:

[0192] W r→b =Softmax(MLP(F r ))⊙PosEmb(t)), formula (2)

[0193] Among them F r Contains resource features such as text / image / category. PosEmb(t) is the time position code, which realizes the time-sensitive aggregation of resource features at the behavior node and forms the resource impact weight matrix.

[0194] ③ Behavior-label layer (L3): Establishes a semantic transmission channel from behavior to resource labels, and calculates the contribution of behavior to labels through the attention mechanism:

[0195]

[0196] Among them Fb is the behavior node feature, F l is the label feature, and finally generates the weight vector α of the impact of the behavior on the resource label b→l .

[0197] 2. Dynamic Attention Weight Adjustment and Feature Weight Update Mechanism

[0198] A time-decaying attention model is introduced, incorporating an exponential decay factor into the user-behavior attention calculation. Behaviors within the past 90 days are assigned a decaying weight in reverse chronological order, with recent behaviors within the past 30 days receiving a 20% increase in weight. Dynamic weight adjustment is achieved through a time-gating function. A dual-channel weighted aggregation is used when updating behavior nodes to ensure the appropriate contribution of both types of features to behavior representation. The behavior node weight iteration rule uses a gated loop mechanism to filter out noisy features. A node weight decay strategy establishes a 90-day weight decay mechanism.

[0199] 3. Behavior’s impact on resource attribute weight output

[0200] Through the layer-by-layer mapping of the three-level architecture, a multi-dimensional influence weight matrix of behavior on resource characteristics is finally formed. And output the matrix. In which each element w ij Indicates the impact of the i-th behavior on the j-th resource attribute and outputs the matrix. The specific calculation path is:

[0201]

[0202] in, is the attention weight of behavior i to the label (L3 layer output); Time gate value of behavior i (increase by 20% in the past 30 days); is the influence weight of resource feature j on behavior i (L2 layer matrix element); γ is the decay coefficient, which is 0.3 (consistent with the 90-day decay strategy).

[0203] (3) Pseudo-label generation mechanism

[0204] The pseudo-label generation mechanism has built a collaborative system with "fourth-order confidence assessment as the framework and multi-perspective verification as the support". By deeply embedding the three verification perspectives of graph model, time series model, and content model into the processing flow of each confidence level, it forms a closed loop of "automated screening-feature space verification-manual assisted review-negative sample feedback", achieving precise matching of technical means and verification granularity in different confidence intervals, taking into account both annotation efficiency and label quality.

[0205] Based on the prediction probability, pseudo labels are divided into four levels. Each level uses multi-view verification technology as the core judgment basis, forming a step-by-step verification mechanism:

[0206]

[0207]

[0208] Specifically, the multi-view verification technology is implemented as follows:

[0209]

[0210] (3) Design of adaptive dynamic characteristics calculation engine

[0211] Combining the existing feature coding status of the resource with the newly input behavioral data and its implicit feature impact, the updated values ​​of the dynamic features of the resource (difficulty, quality, timeliness, etc.) are quantitatively calculated.

[0212] Specifically, the general method for calculating the latest dynamic characteristic values ​​of various characteristics of resources (such as difficulty characteristics, quality characteristics, etc.) is as follows:

[0213] ① Perform weighted aggregation on all behaviors Δfeat_behavior_i within a time window (or reaching the trigger threshold) (e.g., weight decay by time: recent behaviors have a greater weight, and weight(t_i) is the decay factor):

[0214] Δfeat_total=Σ[weight(t_i)×Δfeat_behavior_i]

[0215] ② To avoid drastic fluctuations, a smoothing strategy is used to update resource characteristic values:

[0216] feat_dyn_new=α×feat_dyn_current+(1-α)×(feat_static+β×Σ[weight(t_i)×Δfeat_behavior_i]) formula (5)

[0217] Among them: feat_dyn_new is the latest resource feature value, feat_dyn_current is the current feature value; α is the smoothing coefficient (such as 0.8), β is the adjustment coefficient of the total increment (taking into account the overall weight scale); feat_static is the static initial feature value (as a benchmark, optional).

[0218] Finally, the calculated latest dynamic feature value of the resource feat_dyn_new is output

[0219] (Diff_dyn_new, Qual_dyn_new,......) are fed back to the resource encoding module of the digital resource library for updating.

[0220] (4) Building a User Portrait Center

[0221] The user portrait center extracts user features (such as learning style, interest preference (based on knowledge point access frequency), ability level (based on answering situation), learning progress, etc.) by analyzing user behavior. The data of the center comes from user behavior logs, basic information (school stage, subject, etc.), optional learning level test results, etc. The user portrait is also dynamically updated as the user behavior data accumulates, and the data output can be used by a semi-supervised learning model (for refining behavior analysis) and a personalized recommendation engine.

[0222] (Five) Personalized recommendation engine

[0223] The personalized recommendation engine takes the user portrait feature vector and the basic encoding vector + dynamic feature encoding vector of the resource as input features by constructing a high-order model (not limited to, such as a deep learning model: deep collaborative filtering, NCFs), calculates the resource ranking score according to the user's ability matching difficulty, preference matching quality, and current learning demand matching timeliness, and generates a highly personalized resource recommendation list.

[0224] The second embodiment of the present application relates to a teaching resource coding adaptive optimization and resource personalized recommendation system, the structure of which is shown in Figure 3 The teaching resource coding adaptive optimization and resource personalized recommendation system comprises:

[0225] A digital resource library module is used to build a digital resource library, wherein the teaching resources in the digital resource library contain basic encoding information and dynamic feature encoding information, and the dynamic feature encoding information includes dynamic difficulty, dynamic quality, and dynamic timeliness.

[0226] A user behavior data acquisition module is used to acquire user behavior data.

[0227] A semi-supervised learning module is used to deeply analyze the correlation between the user behavior data and the dynamic feature information of the teaching resources by a semi-supervised learning model, and to establish an internal correlation model of the user behavior data and the dynamic feature information of the teaching resources.

[0228] An adaptive dynamic characteristic calculation engine is used to calculate the updated values of the dynamic features of the teaching resources in real time based on the internal correlation model, and to feed back the calculated dynamic feature updated values to the resource coding module of the digital resource library to update the resource dynamic feature encoding information.

[0229] A user portrait center is used to build a user portrait based on the user behavior data and to extract user features.

[0230] A personalized recommendation engine is used to generate a personalized teaching resource recommendation list based on the updated resource dynamic feature encoding information and the user features.

[0231] The first embodiment is a method embodiment corresponding to the present embodiment, and the technical details in the first embodiment can be applied to the present embodiment, and the technical details in the present embodiment can also be applied to the first embodiment.

[0232] The above-mentioned embodiments have the following technical effects:

[0233] The adaptive optimization of teaching resource coding and the personalized resource recommendation method described in the above-mentioned embodiments, by constructing a digital resource library, collecting user behavior data, analyzing the association between user behavior and resource characteristics using a semi-supervised learning model, and updating resource characteristics in real time based on the association model, ultimately achieving personalized resource recommendation combined with user characteristics. This method overcomes the problem of static resource coding and poor recommendation effect in the prior art, significantly improving the real-time accuracy and accuracy of personalized recommendation of resource representation. The method has the following technical effects:

[0234] Dynamic feature coding information including dynamic difficulty, dynamic quality and dynamic timeliness is introduced, and a structure design of static feature area and dynamic feature area is adopted to separate the management of initial feature benchmark value and correction value (such as difficulty correction value, quality evolution value and timeliness decay value) updated according to user behavior, so that resource characteristics can be dynamically adjusted according to user behavior feedback. This technical solution solves the problem of staticity of traditional coding through dynamic coding structure design, overcomes the distortion problem caused by the solidification of traditional coding, and significantly improves the real-time accuracy of feature representation.

[0235] A hybrid architecture of four-element heterogeneous graph neural network and multi-view collaborative learning is adopted, and in the three stages of graph structure initialization, pseudo-label iterative generation and feature optimization update, a heterogeneous graph containing user, resource and behavior nodes is constructed. Through the graph neural network, the multi-type association relationship (such as the "user-mistaken behavior-knowledge point-resource" path) is aggregated, and the semi-supervised learning (label propagation + self-training) is combined to convert the semantic information of unlabeled data into vector features, mine the structure information of unlabeled data, establish deep graph association structure between users, behaviors, resources and labels, generate feature association models and pseudo-label data for updating resource characteristics, and form a data closed-loop optimization. Traditional recommendation systems mostly use collaborative filtering or single-modal models (such as only user-resource matrix), which cannot effectively fuse multi-dimensional information such as user behavior, resource (including resource characteristic information), and user behavior level, resulting in the problems of "difficulty in mining implicit behavior" and "serious cold start". The semi-supervised learning model breaks through the bottleneck of implicit behavior mining, improves the feature association modeling capability, and significantly improves the accuracy and practicality of recommendation in the education scene.

[0236] In addition, the technical solution realizes the self-adaptive evolution of the system through a closed-loop optimization mechanism, forming a complete closed-loop link of "coding update condition → semi-supervised model output dynamic feature parameter → coding update calculation → model retraining". Semi-supervised learning reduces the dependence on manual labeling, and the closed-loop mechanism reduces the frequency of manual review. The massive unlabeled behavior data is effectively utilized to automatically establish a complex correlation model between user behavior and intrinsic features of resources (difficulty, quality, timeliness, etc.) and generate pseudo-label data. The behavior pattern is directly quantified as the driving force for the change of resource features, solving the problem that traditional methods cannot utilize unstructured behavior data, realizing the self-adaptive evolution of the system, and having the advantages of dynamic adaptability, recommendation accuracy improvement, and labor cost reduction.

[0237] In terms of dynamic feature updating of resources, the adaptive dynamic characteristic calculation engine updates the feature value based on the comprehensive action of behavior influence and time factor using the following formula:

[0238] feat_dyn_new = α × feat_dyn_current + (1-α) × (feat_static + β × Σ[weight(t_i) × Δfeat_behavior_i]) Formula (5)

[0239] This formula maintains the stability of the feature while avoiding the dramatic fluctuation of the feature, reflecting the dynamic trend of the evolution of resource features with user behavior.

[0240] In terms of user portrait construction, user features including learning style, interest preference, ability level, and learning progress are extracted and continuously updated with user behavior data, forming a dynamic representation of user state and providing data support for subsequent recommendation.

[0241] The personalized recommendation engine constructs the interactive matching relationship between the user feature vector and the resource encoding vector based on the high-order model, realizing the precise matching of "user real demand-resource feature". The engine outputs personalized recommendation results according to the principles of matching resource difficulty with user ability, matching resource quality with user preference, and matching resource timeliness with learning demand through a deep learning model. Unlike traditional "one-size-fits-all" recommendation methods, this method deeply excavates user's personalized learning features, accurately captures user's actual learning needs and learning state, and significantly improves the matching degree and accuracy of resource recommendation.

[0242] In the heterogeneous graph neural network structure, a three-level hierarchical architecture is adopted:

[0243] The user-behavior layer uses the following formula to construct the initial embedding vector of user features and behavior nodes:

[0244] V b (1) = σ(Au ·F u ) Formula (1):

[0245] The resource-behavior layer fuses resource features and time position encoding through the following formula to form a time-sensitive resource behavior weight:

[0246] W r→b = Softmax(MLP(F r ))⊙PosEmb(t)) Formula (2):

[0247] The behavior-label layer uses the following attention mechanism to calculate the contribution of the behavior node to the label node:

[0248]

[0249] This three-layer structure realizes the full-chain modeling of user-behavior-resource-label, and improves the expression ability of behavior-driven resource feature updating.

[0250] The final influence weight of the resource feature is calculated by the following formula, which combines attention weight, time gating value, resource influence weight, and time decay mechanism:

[0251]

[0252] This mechanism effectively models the dynamic influence strength of user behavior on resource features and has time-sensitive regulation ability.

[0253] The digital resource library adopts a structure combining relational databases and time series databases, separates the storage of basic encoding and dynamic encoding information, and introduces a version tree structure to realize historical tracing and bidirectional jumping of encoding states, improving the manageability of resource states.

[0254] In addition, the time position encoding PosEmb(t) and the time decay strategy (1-γ·Π(Δt_i>90)) in the influence weight calculation are used collaboratively in the feature extraction and updating stages, realizing the linkage of time-sensitive modeling before and after.

[0255] The generation of pseudo labels uses a four-order confidence evaluation mechanism, which divides pseudo labels into four levels: high, medium, low, and invalid. Different verification strategies and update frequencies are adopted for different confidence levels, combined with three-perspective verification of graph model, time series model, and content model, to improve the quality of pseudo labels and learning efficiency.

[0256] In summary, the above embodiments propose a closed-loop recommendation system that takes dynamic resource feature modeling as the core, combined with graph neural network behavior analysis, multi-perspective pseudo-label verification, time-sensitive control, and personalized recommendation, etc. The system improves the dynamic nature of the resource encoding system, the comprehensiveness of user understanding, the accuracy of behavior analysis, and the level of personalization of the recommendation results, and is significantly superior to existing recommendation methods based on static labels.

[0257] It should be noted that the implementation functions of each module shown in the above embodiments of the adaptive optimization of teaching resource coding and the resource personalized recommendation system can be understood with reference to the related description of the adaptive optimization of teaching resource coding and the resource personalized recommendation method. The functions of each module shown in the above embodiments of the adaptive optimization of teaching resource coding and the resource personalized recommendation system can be implemented by programs (executable instructions) running on a processor, or by specific logic circuits. The above adaptive optimization of teaching resource coding and the resource personalized recommendation system of the embodiments of the present application, if implemented in the form of software functional modules and sold or used as independent products, can also be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read Only Memory), a magnetic disk or an optical disk, and various program code storage media. Thus, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0258] Correspondingly, the embodiments of the present application also provide a computer storage medium, which stores computer executable instructions. When the computer executable instructions are executed by a processor, the methods of the embodiments of the present application are implemented.

[0259] In addition, the application also provides a teaching resource coding adaptive optimization and resource personalized recommendation system, which comprises a memory for storing computer executable instructions, and a processor; the processor is used to realize the steps in the above method embodiments when executing the computer executable instructions in the memory. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The memory can be a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk or a solid state disk, etc. The steps of the method disclosed in the embodiments of the application can be directly embodied by a hardware processor for execution, or executed by a combination of hardware and software modules in the processor.

[0260] It should be noted that the relationship terms such as first and second in the application file of the present patent are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including one" does not exclude the presence of another identical element in the process, method, article or equipment including the element. In the application file of the present patent, if it is mentioned that a certain behavior is performed according to a certain element, it means that the behavior is performed at least according to the element, which includes two cases: the behavior is performed only according to the element, and the behavior is performed according to the element and other elements. The expressions of multiple, multiple times, multiple kinds, etc. include 2, 2 times, 2 kinds and more than 2, more than 2 times, more than 2 kinds.

[0261] All the documents mentioned in the present application are considered to be included in the disclosure of the present application as a whole, so as to be used as a modification if necessary. In addition, it should be understood that those skilled in the art can make various modifications or changes to the present application after reading the above disclosure of the present application, and these equivalent forms also fall within the scope of the present application.

Claims

1. A method for adaptive optimization of teaching resource coding and personalized resource recommendation, characterized in that: The following steps are involved: Building a digital resource library, wherein the teaching resources in the digital resource library include basic coding information and dynamic feature coding information, wherein the dynamic feature coding information includes dynamic difficulty, dynamic quality, and dynamic timeliness; Collecting user behavior data, deeply analyzing the association between the user behavior data and the dynamic feature coding information of the teaching resources through a semi-supervised learning model, and establishing an intrinsic association model between the user behavior data and the dynamic feature coding information of the teaching resources; Based on the established intrinsic correlation model, the updated values ​​of the dynamic characteristics of the teaching resources are calculated in real time by an adaptive dynamic characteristic calculation engine, and the calculated dynamic characteristic update values ​​are fed back to the resource encoding module of the digital resource library to update the dynamic characteristic encoding information of the resources; Build a user profile based on the user behavior data and extract user features; Based on the updated resource dynamic feature coding information and the user features, a personalized teaching resource recommendation list is generated through a personalized recommendation engine.

2. The method according to claim 1, characterized in that The dynamic feature coding information includes: a static feature area, which contains the initially marked difficulty coefficient, quality rating, timeliness and other feature benchmark values; a dynamic feature area, which contains features such as difficulty correction value, quality evolution value, timeliness attenuation value, etc., and is updated according to the real-time numerical value feedback from the adaptive dynamic characteristic calculation engine.

3. The method according to claim 1, characterized in that The semi-supervised learning model adopts a hybrid architecture of quaternary heterogeneous graph neural networks and multi-view collaborative learning, including: a graph structure initialization phase, which constructs a heterogeneous information network based on teaching resource encoding and initial annotated data; a pseudo-label iterative generation phase, which dynamically expands the annotated data through a fourth-order confidence evaluation mechanism based on multi-view verification; a feature optimization and update phase, which continuously optimizes the dynamic feature representation of resources by fusing pseudo-labels, and outputs a feature association model and a high-confidence pseudo-label dataset; Furthermore, the multi-perspective collaborative learning mechanism verifies pseudo labels from three perspectives: group behavior consistency, temporal behavior rationality, and content relevance, using a graph model, a temporal model, and a content model. A fourth-order confidence evaluation is used to dynamically adjust the verification strength of different perspectives. Furthermore, the pseudo-label iterative generation stage adopts a four-order confidence evaluation mechanism to classify pseudo-labels into four levels: high confidence, medium confidence, low confidence, and invalid. Each level adopts a different verification strategy and update frequency.

4. The method according to claim 1, wherein The resource dynamic characteristic value update formula adopted by the adaptive dynamic characteristic calculation engine is: feat_dyn_new=α×feat_dyn_current+(1- α)×(feat_static+β×Σ[weight(t_i)×Δfeat_behavior_i]) formula (5) Where: feat_dyn_new is the latest resource feature value, feat_dyn_current is the current feature value, α is the smoothing coefficient, β is the adjustment coefficient of the total increment, feat_static is the static initial feature value, weight(t_i) is the time decay weight function of the i-th behavior, and Δfeat_behavior_i is the feature change value caused by the i-th behavior.

5. The method according to claim 1, wherein The user profile includes user characteristics such as user learning style, interest preferences, ability level, learning progress, etc. The user characteristics are dynamically updated through a profile update model based on user learning behavior data, test results and other information.

6. The method according to claim 1, characterized in that The personalized recommendation engine adopts a deep learning model, takes the user feature vector and the basic coding vector and dynamic feature coding vector of the teaching resources as input, calculates the resource relevance score based on the principles of matching user ability with resource difficulty, matching preference with resource quality, and matching learning needs with resource timeliness, and generates a personalized recommendation list.

7. The method according to claim 3, characterized in that The four-element heterogeneous graph neural network consists of three layers: The user-behavior layer (L1) aggregates user node features to behavior nodes through the adjacency matrix to generate the initial behavior embedding vector: Among them A u is the user-behavior adjacency matrix, F u is the user feature tensor, σ is the activation function; The resource-behavior layer (L2) integrates resource node features, behavior node features, and time position coding, and calculates the resource impact weight matrix through bilinear mapping: W r→b =Softmax(MLP(F r ))⊙PosEmb(t)) Formula (2) Among them F r is the resource feature, PosEmb(t) is the temporal position encoding function, ⊙ represents element-wise multiplication, MLP is a multi-layer perceptron, and Softmax is a normalization function; The behavior-label layer (L3) calculates the contribution of the behavior to the label through the attention mechanism, and passes it to the label node as the cross-layer influence weight.

8. The method according to claim 7, characterized in that The behavior-label layer calculates the contribution of the behavior node to the label node through the attention mechanism as follows: Among them, F b is the behavior node feature, F l is the label feature, Wq and Wk are learnable parameter matrices, LeakyReLU is the leaky rectified linear unit activation function, and exp is the exponential function.

9. The method according to claim 7 or 8, characterized in that The calculation formula for the influence weight w_{ij} of behavior on resource feature attributes in the quaternary heterogeneous graph neural network is: in: is the attention weight of behavior i on the label, is the time gating value of behavior i, controlling the time sensitivity, is the influence weight of resource feature j on behavior i, γ is the attenuation coefficient, Π(Δt_i>90) is the time threshold indicator function, and the function value is 1 when the number of days Δt_i from behavior i to today is greater than 90 days, otherwise it is 0.

10. The method according to claim 1, characterized in that The basic coding information in the digital resource library is stored in a relational database, and the dynamic feature coding information is stored separately in a time series database. A coding version tree is established to support bidirectional version jump queries.

11. The method according to claim 7 or 10, characterized in that The temporal position code PosEmb(t) of the resource-behavior layer and the time decay coefficient λ in the influence weight calculation formula work together to jointly implement a time sensitivity mechanism in both the feature extraction and feature update stages.

12. A self-adaptive optimization and personalized resource recommendation system for teaching resource coding, characterized in that: include: A digital resource library module is used to construct a digital resource library, wherein the teaching resources in the digital resource library include basic coding information and dynamic feature coding information, wherein the dynamic feature coding information includes dynamic difficulty, dynamic quality and dynamic timeliness; User behavior data collection module, used to collect user behavior data; A semi-supervised learning module is used to deeply analyze the association between the user behavior data and the dynamic feature information of the teaching resources through a semi-supervised learning model, and establish an intrinsic association model between the user behavior data and the dynamic feature information of the teaching resources; An adaptive dynamic characteristic calculation engine, configured to calculate in real time based on the intrinsic correlation model the updated values ​​of the dynamic characteristics of the teaching resources, and feed the calculated updated values ​​of the dynamic characteristics back to the resource encoding module of the digital resource library to update the dynamic characteristic encoding information of the resources; A user portrait center, used to build user portraits based on the user behavior data and extract user features; The personalized recommendation engine is used to generate a personalized teaching resource recommendation list based on the updated resource dynamic feature coding information and the user characteristics.

Images