Education content personalized recommendation system based on deep learning

Abstract

The invention discloses an education content personalized recommendation system based on deep learning, and relates to the field of education technology and deep learning, and the system comprises a multi-modal data processing module which collects and processes four types of data, and forms a standardized vector; the target disassembling and mapping module is used for disassembling a macroscopic target into microscopic sub-targets and generating a list and a mapping table; the ability and interest modeling module evaluates ability and constructs an interest model; the resource recombination recommendation module screens and pushes combined resources; the feedback iterative optimization module collects feedback, updates data, optimizes parameters and reversely transmits the parameters; according to the method, a personalized learning path is constructed through a multi-modal data processing and target dynamic disassembly technology, and the requirements of learners are accurately evaluated in combination with multi-dimensional capability quantification and interest modeling; and meanwhile, a resource intelligent recombination recommendation mechanism is adopted, a comprehensive and high-timeliness resource list is generated based on an adaptation degree algorithm, and a feedback closed loop is formed to continuously optimize system parameters, so that the learning experience and effect are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of educational technology and deep learning, specifically to a personalized recommendation system for educational content based on deep learning. Background Technology

[0002] With the deepening development of educational informatization, the demand for personalized learning is becoming increasingly prominent. Traditional education models are unable to meet the differentiated needs of learners in terms of knowledge base, learning pace, and interest preferences. Breakthroughs in artificial intelligence technology have provided new solutions for the education field. Deep learning-based educational content recommendation systems can dynamically adapt learning resources by analyzing learners' multi-dimensional data, becoming a key technology for improving learning efficiency. Currently, the technology for collecting and analyzing multimodal learning data has gradually matured, but how to transform this data into structured learning objectives and achieve intelligent resource recommendation remains an urgent problem to be solved in the field of educational technology.

[0003] Traditional educational recommendation systems often rely on single data sources or static rules, which have several limitations: First, the goals are broken down in a coarse manner, lacking a structured breakdown of macro-level learning objectives, making it difficult for learners to grasp the key points at each stage. Second, resource recommendations are disconnected from learners' abilities and interests, failing to dynamically assess learners' mastery of specific knowledge points, resulting in insufficient adaptability of recommended content. Third, the system lacks a closed-loop optimization mechanism, unable to adjust recommendation strategies based on learners' real-time feedback, leading to a decline in long-term effectiveness. In addition, traditional methods have limited ability to integrate multimodal data, making it difficult to fully capture learners' interactive behaviors and implicit needs, further limiting the accuracy of personalized recommendations. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a personalized educational content recommendation system based on deep learning. This system integrates text, video, audio, and interactive data through a multimodal data processing module to construct a standardized learning feature library. Combining goal decomposition and mapping techniques, macro-level learning goals are transformed into dynamic priority sub-goals, which are then linked to knowledge points and time planning. An ability and interest modeling module accurately quantifies learners' levels and dynamically adjusts interest weights. A resource reorganization and recommendation module filters suitable resources based on a multi-dimensional fit algorithm to generate a personalized recommendation list. Finally, a feedback iterative optimization module continuously updates system parameters.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a personalized recommendation system for educational content based on deep learning, the system comprising:

[0006] Multimodal data processing module: Collects four types of multimodal learning data: text, video, audio, and interactive data; receives and standardizes the macro-learning objectives set by learners; extracts the original feature vectors of each modality; and merges them to form a standardized multimodal feature vector.

[0007] The goal decomposition and mapping module decomposes the macro learning goal into structured micro sub-goals, labels each sub-goal with a progress evaluation threshold, associated knowledge points and time period, and determines the recommendation order of sub-goals through a dynamic sub-goal priority algorithm, generating a list of micro sub-goals and a sub-goal-resource mapping table.

[0008] Ability and Interest Modeling Module: Based on multimodal feature vectors and a list of sub-goals, the module evaluates learners’ ability levels in each sub-goal through a multidimensional ability quantification algorithm. Combining learners’ interaction data with resources, the module constructs an interest model through a dynamic interest weighting algorithm and outputs ability level labels and interest models.

[0009] Resource reorganization and recommendation module: Based on the sub-goal-resource mapping table, ability level tags and interest model, candidate resources are screened from the education resource database and combined. The suitability of the resource combination is evaluated through a multi-dimensional resource suitability algorithm, a recommended resource list is generated and pushed to the learner's end.

[0010] Feedback Iterative Optimization Module: Collects explicit and implicit feedback data from learners on recommended resources, updates the sub-goal progress and resource mapping table based on the feedback data, optimizes the parameters of each module of the system through a feedback-driven parameter iterative algorithm, and transmits the results back to the relevant modules.

[0011] Furthermore, in the multimodal data processing module, text data includes homework answers, notes, and question records, which are collected through input via the system's interactive interface; video data includes course viewing footage, operation demonstration recordings, and pause / fast-forward behavior records, which are collected through the terminal's camera or video playback logs; audio data includes spoken practice recordings and voice Q&A segments, which are collected through microphone recordings; and interactive data includes answer records, click paths, learning duration, and content completion rate, which are collected in real time through the system's background logs.

[0012] Furthermore, in the target decomposition and mapping module, when the macro-level learning target is decomposed into structured micro-level sub-targets, the following are specifically included: Progress evaluation threshold: For each micro-level sub-target, multiple evaluation levels are marked, and each level is determined based on the consecutive answer accuracy rate and the difficulty coefficient of the corresponding sub-target's question type; Related knowledge points: Through matching the knowledge point map in the education field, two types of related knowledge points are marked for each sub-target. Core knowledge points are the knowledge points directly corresponding to the sub-target, and prerequisite knowledge points are the preparatory knowledge points that need to be mastered to learn the sub-target, and the dependence strength between knowledge points is recorded; Time cycle: Includes the total cycle and the stage cycle. The total cycle is the planned total time to complete the sub-target, and the stage cycle is dynamically divided into multiple learning stages according to the priority of the sub-target. Each stage cycle is associated with a specific time node.

[0013] Furthermore, in the target decomposition and mapping module, the expression for the dynamic sub-target priority algorithm is: Among them, P j Let T be the priority score for the j-th micro-sub-objective, ranging from 0 to 1, where α, β, and γ are weighting coefficients, and α + β + γ = 1. total For the overall macroeconomic target period, T used,j Let m be the time consumed for the j-th sub-objective, m be the number of subsequent sub-objectives that depend on the j-th sub-objective, and D be the time consumed for the j-th sub-objective. j,k Let P be the dependency coefficient of the j-th sub-objective on the k-th subsequent sub-objective. k For the current priority score of the k-th subsequent sub-objective, A init,j Let A be the initial capability score for the user's j-th sub-goal. thr,j Let be the capability threshold for the j-th sub-target.

[0014] Furthermore, the target decomposition and mapping module generates a list of micro-sub-targets and a sub-target-resource mapping table, specifically including:

[0015] The list of micro-level sub-objectives is a structured data table containing: a unique identifier for each sub-objective, the name of the sub-objective, hierarchical attributes, progress assessment thresholds, a set of associated knowledge point IDs, the total number of days in the cycle, and the time nodes of each stage cycle.

[0016] The sub-target-resource mapping table is a related data table, which contains: sub-target unique identifier, resource unique identifier, resource type, resource quantity percentage, resource association strength, and resource applicable stage label.

[0017] Furthermore, in the capability and interest modeling module, the expression for the multidimensional capability quantification algorithm is: Among them, A j Let be the final ability score for the j-th sub-objective, ranging from 0 to 1, δ be the stability weight, q be the number of tests or answers for the j-th sub-objective, and D be the final ability score for the j-th sub-objective.k,j R represents the difficulty coefficient of the question type for the k-th question and the j-th sub-objective. k,j Let p be the answer to the k-th question, p be the number of learning cycles for the j-th sub-goal in the last 30 days, and S be the answer to the k-th question. t,j For the ability score in the t-th period, W t,j Let t be the time weight for the t-th period.

[0018] Furthermore, in the capability and interest modeling module, the expression for the dynamic interest weight algorithm is: W u,x (t)=W u,x (t-1)×(1-λ)+λ×(I u,x ×D u,x ), where W u,x (t) represents the interest weight of user u on feature x at time t, with a value ranging from 0 to 1. u,x (t-1) represents the historical interest weight at time t-1, λ is the update coefficient, and I u,x Let D be the interaction strength of user u with feature x. u,x The interaction depth of user u with feature x is calculated based on the ratio of learning time to content length.

[0019] Furthermore, in the resource reorganization recommendation module, the expression for the multi-dimensional resource suitability algorithm is: S comb =ω1×M t +ω2×A fit +ω3×I pref +ω4×T res , of which S comb The overall fit score for the resource combination ranges from 0 to 1, where ω1, ω2, ω3, and ω4 are weighting coefficients, and ω1 + ω2 + ω3 + ω4 = 1. M t A represents the degree of target matching between the resource combination and the current sub-target. fit To ensure the suitability of resource combinations with user capabilities, I pref T represents the degree of matching between resource combinations and user interests and preferences. res The timeliness score is given for the resource combination.

[0020] Furthermore, in the resource reorganization and recommendation module, the educational resource database includes: structured resource entries, each of which contains: a unique resource identifier, resource type, content characteristics, difficulty level, applicable stage tag, update time, associated knowledge point ID, weight, and quality indicators.

[0021] Furthermore, in the feedback iterative optimization module, the expression for the feedback-driven parameter iterative algorithm is: Where, θ t+1 For the model parameters after iteration, θ tHere are the model parameters before iteration, η is the learning rate, and L is the loss function. The loss function with respect to parameter θ t The partial derivative of F strength For the feedback intensity coefficient, C θ This is the confidence coefficient for the current parameter.

[0022] Compared with existing technologies, this deep learning-based personalized recommendation system for educational content has the following advantages:

[0023] I. This invention, by constructing a multimodal data processing module, achieves comprehensive collection and standardized processing of learners' text, video, audio, and interactive data, effectively integrating diverse learning data sources. Simultaneously, by combining goal decomposition and mapping techniques, it refines macro-level learning goals into structured micro-level sub-goals and dynamically determines the priority of these sub-goals, providing learners with clear and personalized learning path planning. This design not only improves the attainability and relevance of learning goals but also accurately assesses learners' ability levels and constructs personalized interest models through multi-dimensional ability quantification and dynamic interest modeling, ensuring the adaptability and attractiveness of recommended resources, thereby significantly enhancing the learning experience and effectiveness.

[0024] Second, this invention uses a resource reorganization recommendation mechanism to intelligently filter and combine educational resources based on a sub-goal-resource mapping table, ability level, and interest model, generating a highly suitable recommendation list. This mechanism employs a multi-dimensional resource suitability algorithm, comprehensively considering goal matching, ability suitability, interest preference matching, and timeliness to ensure the comprehensiveness and timeliness of the recommended resources. In addition, the feedback iterative optimization module continuously collects learner feedback and dynamically adjusts system parameters to form a closed-loop optimization system.

[0025] This innovative design enables continuous optimization and personalization of educational content recommendations, providing learners with more precise and efficient learning support.

[0026] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0028] Figure 1 A flowchart illustrating the workflow of a deep learning-based personalized recommendation system for educational content.

[0029] Figure 2 This is a module framework diagram for a deep learning-based personalized recommendation system for educational content. Detailed Implementation

[0030] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0031] Example 1:

[0032] A scenario for middle school students learning quadratic functions in junior high school.

[0033] Multimodal data processing module:

[0034] The system collects multimodal learning data in four categories: text, video, audio, and interactive data. Text data includes students' answers to quadratic function assignments, class notes, and questions submitted to the system after class. This data is collected through the system's interactive interface and directly reflects students' understanding of quadratic function concepts and areas of confusion, providing textual evidence for subsequent ability assessments. Video data includes footage of online quadratic function courses, teacher demonstrations of problem-solving, and records of students pausing and fast-forwarding during viewing. This data is collected through the terminal's camera and video playback logs, capturing students' attention focus and weak points during learning, helping to determine learning needs. Audio data primarily consists of recordings of students asking questions about difficult points in quadratic functions. Recorded through microphones, this data supplements the level of confusion conveyed by tone of voice, further refining learning questions. Interactive data covers quadratic function-related answer records, paths to clicked course chapters, duration of each learning session, and assignment completion rates. This data is collected in real-time through the system's backend logs, dynamically tracking students' learning behavior and reflecting their learning engagement and knowledge absorption. The system receives the student's macro-level learning objective, "Master the core knowledge and problem-solving methods of quadratic functions in junior high school within one month," and standardizes this objective. This standardization process eliminates ambiguity in the objective's wording and provides a unified benchmark for subsequent breakdown. Then, it extracts the original feature vectors from four modalities: text, video, audio, and interactive. These vectors are then fused to form a standardized multimodal feature vector. This fusion process integrates multi-dimensional learning data, avoiding the limitations of single-modal data and providing comprehensive data support for subsequent modules. Figure 1 As shown.

[0035] Target decomposition and mapping module:

[0036] The macro-level learning objective of "mastering the core knowledge and problem-solving methods of quadratic functions in junior high school within one month" is broken down into four structured micro-level sub-objectives: "understanding the definition and expression of quadratic functions," "drawing the graph and analyzing the properties of quadratic functions," "the relationship between quadratic functions and quadratic equations in one variable," and "solving problems through practical applications of quadratic functions." This breakdown transforms the abstract and large macro-level objective into concrete and actionable smaller goals, reducing student learning pressure and facilitating phased assessment and resource recommendations by the system. Progress assessment thresholds are assigned to each micro-level sub-objective: Basic level: ≥60% accuracy in two consecutive questions; Intermediate level: ≥80% accuracy in three consecutive questions; Advanced level: ≥90% accuracy in five consecutive questions. These thresholds are weighted based on the difficulty level of the questions corresponding to the sub-objective. This threshold assignment clarifies the phased learning standards for each sub-objective, allowing students to clearly understand whether their learning progress has met the standards and providing a basis for the system to determine whether to proceed to the next stage. Related knowledge points are also marked: for example, the core knowledge point for "understanding the definition and expression of quadratic functions" is "the definition and general expression of quadratic functions," The annotation method is "vertex-based," with the prerequisite knowledge point being "definition and expression of linear functions." It also records the dependency strength between knowledge points. This annotation helps clarify the logical connections between knowledge points, preventing students from skipping necessary prerequisite knowledge and causing learning gaps. It also provides a reference for the system to recommend related supplementary resources. The annotation time period is one month in total, divided into three phases: basic, reinforcement, and assessment. Each phase is associated with specific time nodes, and the phase duration can be dynamically adjusted according to the priority of sub-goals. This periodic annotation allows students to create a clear learning time plan, avoiding a chaotic learning rhythm. The recommended order of sub-goals, determined by the dynamic sub-goal priority algorithm, is: "Understanding the definition and expression of quadratic functions → Drawing and analyzing the properties of quadratic function graphs → Relationship between quadratic functions and quadratic equations → Solving problems through practical applications of quadratic functions." The expression for the dynamic sub-goal priority algorithm is: P j = Among them, P j Let T be the priority score for the j-th micro-sub-objective, ranging from 0 to 1, where α, β, and γ are weighting coefficients, and α + β + γ = 1. total For the overall macroeconomic target period, T used,j Let m be the time consumed for the j-th sub-objective, m be the number of subsequent sub-objectives that depend on the j-th sub-objective, and D be the time consumed for the j-th sub-objective. j,k Let P be the dependency coefficient of the j-th sub-objective on the k-th subsequent sub-objective. k For the current priority score of the k-th subsequent sub-objective, A init,j Let A be the initial capability score for the user's j-th sub-goal. thr,jLet be the capability threshold for the j-th sub-goal. The purpose of this algorithm is to combine factors such as time consumption, knowledge dependence, and initial ability to prioritize students according to their learning patterns, ensuring that the learning process proceeds step by step. Finally, it generates a micro-level sub-goal list containing fields such as sub-goal unique identifier, sub-goal name, hierarchical attribute, progress evaluation threshold, set of associated knowledge point IDs, total cycle days, and stage cycle time nodes, as well as a sub-goal-resource mapping table containing fields such as sub-goal unique identifier, resource unique identifier, resource type, resource quantity percentage, resource association strength, and resource applicable stage label.

[0037] Ability and Interest Modeling Module:

[0038] Based on multimodal feature vectors and a list of micro-level sub-objectives, a multidimensional ability quantification algorithm is used to evaluate students' ability levels in each sub-objective. The expression for the multidimensional ability quantification algorithm is as follows: Among them, A j Let be the final ability score for the j-th sub-objective, ranging from 0 to 1, δ be the stability weight, q be the number of tests or answers for the j-th sub-objective, and D be the final ability score for the j-th sub-objective. k,j R represents the difficulty coefficient of the question type for the k-th question and the j-th sub-objective. k,j Let p be the answer to the k-th question, p be the number of learning cycles for the j-th sub-goal in the last 30 days, and S be the answer to the k-th question. t,j For the ability score in the t-th period, W t,j The core function of this algorithm is to comprehensively analyze multi-dimensional learning data to objectively and quantitatively calculate student abilities, avoiding biases caused by judging abilities based on a single data point. Taking the sub-objective of "understanding the definition and expression of quadratic functions" as an example, the algorithm combines the number of tests or answers for this sub-objective, the difficulty coefficient of each question in relation to the sub-objective, the answer results, the number of learning cycles for this sub-objective in the past 30 days, the ability score for each cycle, and the time weight for each cycle to ultimately calculate the student's ability score for this sub-objective and output an ability level label of "Basic," "Intermediate," or "Proficient." This label allows the system and the student to quickly understand their current ability level, providing a direct basis for matching the difficulty of subsequent resource recommendations. Simultaneously, combining the interaction data between the student and quadratic function learning resources, an interest model is constructed through a dynamic interest weighting algorithm. The expression for the dynamic interest weighting algorithm is: W u,x (t)=W u,x (t-1)×(1-λ)+λ×(I u,x ×D u,x ), where W u,x (t) represents the interest weight of user u on feature x at time t, with a value ranging from 0 to 1. u,x (t-1) represents the historical interest weight at time t-1, λ is the update coefficient, and I u,xLet D be the interaction strength of user u with feature x. u,x The algorithm calculates the interaction depth of user u with feature x based on the ratio of learning time to content length. Its function is to capture students' interest tendencies in real-time interactions and dynamically update interest weights by combining historical interests with current interactions, preventing interest judgments from becoming outdated. The algorithm calculates the student's interest weights for different types of resources; for example, it ultimately determines that the student's interest weight for "animation-related educational resources" is 0.8. This interest model provides interest guidance for resource selection, ensuring that recommended resources match student preferences and improving learning initiative and participation.

[0039] Resource reorganization recommendation module:

[0040] Based on the sub-goal-resource mapping table, ability level labels, and interest models, candidate resources are selected from the educational resource database. This selection logic ensures that candidate resources align with the current learning objectives and match the student's abilities and interests, avoiding the recommendation of irrelevant resources or resources with unsuitable difficulty or format. For example, if a student's sub-goal "Understanding the Definition and Expression of Quadratic Functions" is labeled "Basic Level" and the interest model shows a preference for animated resources, the system will select two animated micro-lessons explaining the definition of quadratic functions, three sets of basic-level definition-clarification exercises, and one interactive courseware on definition and expression conversion as candidate resources from the database, and combine them according to the resource quantity ratio in the sub-goal-resource mapping table. Subsequently, the suitability of this resource combination is evaluated using a multi-dimensional resource suitability algorithm. The expression for the multi-dimensional resource suitability algorithm is: S comb =ω1×M t +ω2×A fit +ω3×I pref +ω4×T res , of which S comb The overall fit score for the resource combination ranges from 0 to 1, where ω1, ω2, ω3, and ω4 are weighting coefficients, and ω1 + ω2 + ω3 + ω4 = 1. M t A represents the degree of target matching between the resource combination and the current sub-target. fit To ensure the suitability of resource combinations with user capabilities, I pref T represents the degree of matching between resource combinations and user interests and preferences. resTo score the timeliness of resource combinations, this algorithm comprehensively assesses whether the resource combination is suitable for students from multiple dimensions, avoiding situations where a single dimension is suitable but the overall combination is unsuitable. The algorithm considers the matching degree between the resource combination and the sub-goal of "understanding the definition and expression of quadratic functions," the matching degree with students' abilities, the matching degree with students' interests, and the timeliness score of the resource combination, ultimately calculating the total suitability score of the resource combination. Based on the total suitability score, a recommended resource list is generated and pushed to students' learning terminals. This push service accurately delivers high-quality, selected and evaluated resources to students, reducing the time cost for students to independently select resources and improving learning efficiency. Furthermore, each resource entry in the educational resource database includes fields such as a unique resource identifier, resource type, content characteristics, difficulty level, applicable stage tag, update time, associated knowledge point ID and weight, and quality indicators. These fields provide standardized attribute criteria for resource selection and evaluation, ensuring that the selection and evaluation process is quantifiable and traceable.

[0041] Feedback Iterative Optimization Module

[0042] We collected both explicit and implicit feedback data from students regarding the recommended resources. Explicit feedback included student ratings of animated micro-lessons and their markings of the difficulty of practice questions. This type of feedback directly obtained students' subjective evaluations of the resources, reflecting their satisfaction with them. Implicit feedback included the time students spent watching animated micro-lessons in their entirety, a comparison of the time spent completing practice questions with the average time spent, and whether they repeatedly watched certain segments. This type of feedback indirectly judged the applicability of the resources through student behavior, supplementing information that might have been missed by explicit feedback. Based on this feedback data, the system first updates the progress of sub-goals. For example, based on students' performance in completing resources related to the sub-goal "Understanding the Definition and Expression of Quadratic Functions," the progress of this sub-goal is increased from 30% to 50%. This progress update tracks learning progress in real time, providing a basis for adjusting sub-goal priorities or moving to the next stage. Simultaneously, the sub-goal-resource mapping table is updated. For instance, based on students' high satisfaction with animated micro-lessons, the resource association strength between the animated micro-lesson and the "Understanding the Definition and Expression of Quadratic Functions" sub-goal is increased from 0.7 to 0.9. This mapping table update optimizes the association between sub-goals and resources, providing a more accurate basis for recommending resources for similar sub-goals in the future. Subsequently, the system optimizes the parameters of each module through a feedback-driven parameter iteration algorithm. The expression for the feedback-driven parameter iteration algorithm is: θ t+1 = Where, θ t+1 For the model parameters after iteration, θ t Here are the model parameters before iteration, η is the learning rate, and L is the loss function. The loss function with respect to parameter θ t The partial derivative of F strengthFor the feedback intensity coefficient, C θ This represents the confidence coefficient of the current parameters. The algorithm's role is to transform feedback data into a basis for adjusting system parameters, enabling the system to self-iterate and optimize, and preventing parameters from becoming outdated and causing a decline in recommendation effectiveness after long-term use. For example, the algorithm may adjust the stability weight in the multi-dimensional ability quantification algorithm and the interest preference weight in the multi-dimensional resource suitability algorithm, and then pass the optimized parameters back to the ability and interest modeling module and the resource reorganization recommendation module to ensure that subsequent ability assessments are more accurate and resource recommendations are more in line with students' needs, forming a closed loop of "recommendation-feedback-optimization" and continuously improving the system's personalized recommendation effect.

[0043] In summary, in the context of middle school students learning quadratic functions, the deep learning-based personalized educational content recommendation system achieves personalized learning support through the collaborative operation of five modules. The multimodal data processing module comprehensively collects four types of learning data and integrates feature vectors, laying the foundation for subsequent analysis; the goal decomposition and mapping module breaks down macro-level goals into micro-level sub-goals, combining them with a dynamic sub-goal priority algorithm to clarify the learning path; the ability and interest modeling module uses a multi-dimensional ability quantification algorithm to evaluate abilities and a dynamic interest weighting algorithm to build interest models, accurately matching needs; the resource reorganization and recommendation module filters resources based on the mapping table, ability tags, and interest models, and pushes them after evaluation by a multi-dimensional resource suitability algorithm; the feedback iterative optimization module collects feedback updates on the progress and mapping table, and optimizes the system through feedback-driven parameter iterative algorithms, forming a "recommendation-feedback-optimization" closed loop, helping students efficiently master quadratic function knowledge.

[0044] Example 2:

[0045] Scenarios for improving adults' everyday English conversation skills.

[0046] Multimodal data processing module:

[0047] Four types of multimodal learning data are collected: text data includes adult learners' daily English conversation assignment scripts, study notes, and questions about conversation scenarios, collected through the system's interactive interface; video data includes footage of foreign teachers watching daily conversation teaching videos, recordings of conversation scenarios, and records of pause, fast forward, and rewind actions, collected through the terminal's camera and video playback logs; audio data includes learners' oral practice recordings and voice Q&A segments initiated to the system, collected through microphone recordings; and interactive data includes daily English listening comprehension answer records, paths to clicked conversation scenario categories, duration of each learning session, and completion rate of conversation exercises, collected in real time through the system's backend logs. The system receives learners' macro-level learning goals, such as "Improve daily English conversation skills within 3 months, able to handle common scenarios like shopping and asking for directions," standardizes these goals, extracts the original feature vectors of the four modalities, and fuses them to form a standardized multimodal feature vector, such as... Figure 1 As shown.

[0048] Target decomposition and mapping module:

[0049] The macro goal of "improving daily English conversation skills within 3 months to handle common scenarios such as shopping and asking for directions" is broken down into four structured micro-goals: "Mastering core vocabulary for daily shopping scenarios," "Using basic sentence patterns for asking for directions," "Expressing short conversational expressions in shopping scenarios," and "Providing fluent conversational responses for asking for directions." Progress assessment thresholds are assigned to each micro-goal: ≥60% accuracy in two consecutive tests for the basic level, ≥80% accuracy in three consecutive tests for the intermediate level, and ≥90% accuracy in five consecutive tests for the advanced level. These thresholds are determined by weighting the difficulty level of the questions corresponding to the sub-goal. Related knowledge points are also marked; for example, the core knowledge point for "Mastering core vocabulary for daily shopping scenarios" is "shopping-related nouns and verbs," and the prerequisite knowledge point is "basic English phonetics and word pronunciation." The strength of the dependency between these knowledge points is recorded. A time period of 3 months is also marked, divided into a basic stage, an intensive stage, and an assessment stage, with each stage linked to specific time nodes. The duration of each stage can be dynamically adjusted based on the priority of the sub-goals. The recommendation order of sub-goals, determined by the dynamic sub-goal prioritization algorithm, is: "Mastering core vocabulary for daily shopping scenarios → Using basic sentence patterns for asking for directions in daily scenarios → Expressing oneself in short conversations during shopping scenarios → Responses in fluent conversations when asking for directions." The expression for the dynamic sub-goal prioritization algorithm is: Generate a list of micro-subtargets containing fields such as subtarget unique identifier, subtarget name, and hierarchical attributes, as well as a subtarget-resource mapping table containing fields such as subtarget unique identifier and resource unique identifier.

[0050] Ability and Interest Modeling Module:

[0051] Based on multimodal feature vectors and a list of micro-level sub-objectives, a multidimensional ability quantification algorithm is used to evaluate learners' ability levels on each sub-objective. The expression for the multidimensional ability quantification algorithm is: A j = Taking the sub-goal of "Mastering core vocabulary for daily shopping scenarios" as an example, the final ability score is calculated and an ability level label is output by combining the number of vocabulary tests for this sub-goal, the difficulty coefficient of each test question relative to the sub-goal, the answer results, the number of learning cycles for this sub-goal in the past 30 days, the ability score for each cycle, and the time weight for each cycle. Combining learner interaction data with English learning resources, such as learners frequently clicking on resources related to "shopping scenarios" and spending twice as much time learning shopping vocabulary audio as other resources, an interest model is constructed using a dynamic interest weighting algorithm. The expression for the dynamic interest weighting algorithm is: W u,x (t)=W u,x (t-1)×(1-λ)+λ×(I u,x ×D u,x The study calculated learners' interest weights for different scenarios and types of resources, and determined that learners had a higher interest weight for "audio resources in shopping scenarios".

[0052] Resource reorganization recommendation module:

[0053] Based on the sub-objective-resource mapping table, ability level tags, and interest models, candidate resources are selected from the educational resource database. These include three audio clips of core vocabulary related to shopping scenarios, two sets of basic-level shopping vocabulary matching exercises, and one short video combining shopping vocabulary with simple sentence patterns. These resources are then combined. The suitability of the resource combination is evaluated using a multi-dimensional resource suitability algorithm. The expression for the multi-dimensional resource suitability algorithm is: S comb =ω1×M t +ω2×A fit +ω3×I pref +ω4×T res The overall fit score is calculated by comprehensively considering the resource combination's relevance to the sub-goal of "mastering core vocabulary in daily shopping scenarios," its suitability to learners' abilities, its alignment with learners' interests, and the timeliness of resource updates. A recommended resource list is generated based on the overall fit score and pushed to learners' mobile phones or computers. All resource entries in the educational resource database include fields such as a unique resource identifier, resource type, content characteristics, and difficulty level, ensuring that resource selection is based on clear criteria.

[0054] Feedback Iterative Optimization Module:

[0055] The system collects explicit and implicit feedback from learners regarding recommended resources. Explicit feedback includes learners' ratings of shopping vocabulary audio clips and their feedback on the difficulty of practice questions. Implicit feedback includes the number of times learners repeatedly listen to a particular shopping vocabulary audio clip, their accuracy rate in completing vocabulary practice questions, and the completeness of watching shopping vocabulary short videos. Based on the feedback data, the progress of the sub-goal is updated, increasing the progress of the sub-goal "Mastering Core Vocabulary in Daily Shopping Scenarios" from 40% to 70%. Simultaneously, the sub-goal-resource mapping table is updated to enhance the resource association strength between the shopping vocabulary audio clips and this sub-goal. The parameters of each module in the system are optimized using a feedback-driven parameter iteration algorithm. The expression for the feedback-driven parameter iteration algorithm is: For example, adjusting the update coefficients in the dynamic interest weighting algorithm and the capability adaptation weights in the multi-dimensional resource adaptation algorithm, and then passing the optimized parameters back to the target decomposition and mapping module and the capability and interest modeling module, can achieve continuous system optimization.

[0056] In summary, in the scenario of improving adults' daily English conversation skills, this recommendation system also functions around five core modules. The multimodal data processing module collects four types of data (text, video, etc.) and fuses feature vectors to capture adults' learning status; the goal decomposition and mapping module breaks down macro-level goals, labels thresholds, knowledge points, and cycles, and uses a dynamic sub-goal priority algorithm to determine the learning order; the ability and interest modeling module uses a multi-dimensional ability quantification algorithm to assess ability levels and a dynamic interest weighting algorithm to lock in interest preferences; the resource reorganization and recommendation module filters resources based on multi-dimensional information and evaluates suitability before pushing them; and the feedback iterative optimization module updates data and parameters based on feedback to continuously optimize recommendations. The system aligns with adults' learning needs, helping them gradually improve their daily English conversation skills and efficiently handle common scenarios.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based educational content personalized recommendation system, characterized by, The system comprises: A multi-modal data processing module: collects four types of multi-modal learning data including text, video, audio and interaction, receives and standardizes the macro learning objectives set by the learner, extracts original feature vectors of various modalities, and fuses them to form standardized multi-modal feature vectors; A target disassembly and mapping module: disassembles the macro learning objectives into structured micro objectives, labels progress evaluation thresholds, associated knowledge points and time periods for each sub-objective, determines the recommended order of sub-objectives through a dynamic sub-objective priority algorithm, and generates a micro objective list and a sub-objective-resource mapping table; A capability and interest modeling module: based on multi-modal feature vectors and sub-objective lists, the learner's ability level at each sub-objective is evaluated through a multi-dimensional capability quantification algorithm, and an interest model is constructed through a dynamic interest weight algorithm based on the interaction data of the learner and the resources, and the capability level label and the interest model are output; A resource reorganization recommendation module: according to the sub-objective-resource mapping table, the capability level label and the interest model, candidate resources are selected and combined from the educational resource database, the adaptability of the resource combination is evaluated through a multi-dimensional resource adaptation algorithm, a recommended resource list is generated, and is pushed to the learner end; A feedback iteration optimization module: collects explicit and implicit feedback data of the recommended resources from the learner, updates the sub-objective progress and the resource mapping table based on the feedback data, optimizes the parameters of each module of the system through a feedback-driven parameter iteration algorithm, and reversely transmits to the related modules. 2.The deep learning-based educational content personalized recommendation system of claim 1, wherein, In the multi-modal data processing module, the text data includes homework answering text, note content and question recording, which are collected through the system interaction interface; the video data includes course watching screen, operation demonstration video and pause, fast forward and other records, which are collected through the terminal camera or video playing log; the audio data includes oral practice recording and voice answering fragment, which are collected through microphone recording; the interaction data includes answering record, click path, learning time and content completion rate, which are collected in real time through system background log. 3.The deep learning-based educational content personalized recommendation system of claim 1, wherein, In the target disassembly and mapping module, when the macro learning objectives are disassembled into structured micro objectives, it specifically includes: progress evaluation threshold: for each micro objective, multiple evaluation levels are labeled, each level is determined based on the weighted continuous answering accuracy and the difficulty coefficient of the corresponding sub-objective; associated knowledge points: through the education knowledge point graph matching, two types of associated knowledge points are labeled for each sub-objective, the core knowledge point is the knowledge point directly corresponding to the sub-objective, and the prerequisite knowledge point is the preparatory knowledge point required to learn the sub-objective, and the dependency strength between knowledge points is recorded; time period: including total period and stage period, the total period is the planned total time length to complete the sub-objective, and the stage period is dynamically divided into multiple learning stages according to the sub-objective priority, and each stage period is associated with specific time nodes. 4.The deep learning-based educational content personalized recommendation system of claim 1, wherein, The expression of the dynamic sub-target priority algorithm in the target disassembly and mapping module is: Wherein, P j is the priority score of the jth micro sub-target, the value range is 0-1, α, β, γ are weight coefficients, and α+β+γ=1, T total is the total period of the macro target, T used,j is the consumed time of the jth sub-target, m is the number of subsequent sub-targets dependent on the jth sub-target, D j,k is the dependency coefficient of the jth sub-target to the kth subsequent sub-target, P k is the current priority score of the kth subsequent sub-target, A init,j is the initial ability score of the jth sub-target of the user, A thr,j is the ability threshold of the jth sub-target. 5.The deep learning-based educational content personalized recommendation system of claim 1, wherein, The micro objective list and the sub-objective-resource mapping table generated in the target disassembly and mapping module specifically include: The micro-sub-target list is a structured data table, and the content contained includes a sub-target unique identifier, a sub-target name, a hierarchical attribute, a progress evaluation threshold, an associated knowledge point ID set, a total cycle day number, and a stage cycle time node; The sub-target-resource mapping table is an associated data table, and the content contained includes a sub-target unique identifier, a resource unique identifier, a resource type, a resource quantity proportion, a resource association strength, and a resource applicable stage label. 6.The deep learning-based educational content personalized recommendation system of claim 1, wherein, The expression of the multi-dimensional ability quantification algorithm in the ability and interest modeling module is: Wherein, A j is the final ability score of the jth sub-goal, the value range is 0-1, δ is the stability weight, q is the test or answer number of the jth sub-goal, D k,j is the question type difficulty coefficient of the kth question and the jth sub-goal, R k,j is the answer result of the kth question, p is the learning period number of the jth sub-goal in the last 30 days, S t,j is the ability score of the tth period, W t,j is the time weight of the tth period. 7.The deep learning-based educational content personalized recommendation system of claim 1, wherein, The expression of the dynamic interest weight algorithm in the ability and interest modeling module is: W u,x (t) = W u,x (t-1) x (1 - λ) + λ x (I u,x x D u,x ), wherein W u,x (t) is the interest weight of the user u for the feature x at the time t, the value range is 0-1, W u,x (t-1) is the historical interest weight at the time t-1, λ is an update coefficient, I u,x is the interaction intensity of the user u for the feature x, and D u,x is the interaction depth of the user u for the feature x, which is calculated based on the ratio of the learning duration to the content length. 8.The deep learning-based education content personalized recommendation system of claim 1, wherein, The expression of the multi-dimensional resource adaptation degree algorithm in the resource reorganization recommendation module is: S comb = ω1×M t + ω2×A fit + ω3×I pref + ω4×T res , wherein S comb is a total adaptation degree score of the resource combination, the value range is 0-1, ω1, ω2, ω3, and ω4 are weight coefficients, and ω1+ω2+ω3+ω4=1, M t is a target matching degree of the resource combination and the current sub-target, A fit is an adaptation degree of the resource combination and the user ability, I pref is a preference matching degree of the resource combination and the user interest, and T res is a timeliness score of the resource combination. 9.The deep learning-based educational content personalization recommendation system of claim 1, wherein, In the resource reorganization recommendation module, the educational resource database includes structured resource entries, and each resource entry contains the following content: a resource unique identifier, a resource type, a content feature, a difficulty level, an applicable stage label, an update time, an associated knowledge point ID and weight, and a quality indicator. 10.The deep learning-based education content personalized recommendation system of claim 1, wherein, The expression of the feedback driving parameter iterative algorithm in the feedback iterative optimization module is: Wherein, θ t+1 is the model parameter after iteration, θ t is the model parameter before iteration, η is the learning rate, L is the loss function, is the partial derivative of the loss function to the parameter θ t , F strength is the feedback intensity coefficient, C θ is the confidence coefficient of the current parameter.

Images