Multi-dimensional academic situation analysis ai large model course recommendation and adaptive learning path optimization system

Abstract

The present application relates to the technical field of large language model, specifically, to an AI large model course recommendation and adaptive learning path optimization system for multi-dimensional academic situation analysis, comprising: an academic situation data processing unit; a multi-dimensional academic situation analysis unit; an AI large model course recommendation unit; and an adaptive learning path optimization unit. Through the matching degree calculation module and the parameter optimization module of the AI large model course recommendation unit, the cosine similarity is first calculated based on the course joint features and the academic situation features, the dynamic matching degree is obtained by combining the course-geographical attention weight, the courses with a dynamic matching degree greater than a threshold value are screened, the top N personalized recommendation list is generated in descending order, and then the real-time learning effect data of the user learning the recommendation list is received, the model parameters are iteratively updated by using the Adam optimizer after constructing the optimization loss, and the problem of lacking iterative optimization of model parameters based on real-time learning effect data in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, specifically to an AI large model course recommendation and adaptive learning path optimization system for multi-dimensional learning analysis. Background Technology

[0002] With the intelligent development of online education, multi-source heterogeneous learning data (structured data from academic affairs systems, semi-structured data from online learning platforms, and unstructured data from classroom interaction tools) has become a key support for personalized learning services. However, existing technologies have not fully integrated authorized geographic location information, and there are shortcomings in the collaborative closed-loop construction of model adaptability to educational scenarios, course recommendation, and learning path optimization, making it difficult to maximize the value of data to meet personalized needs.

[0003] In existing technologies, relevant patents have explored the direction of personalized course recommendation. For example, Chinese patent CN202510601679.8 discloses an online course recommendation method and system based on a large language model. It extracts skill demand correlations through policy text parsing, adjusts recommendation priorities through cross-modal alignment, constructs a cross-skill domain matrix using LLM, and relies on virtual interactive data to maintain system stability when data is insufficient. Another example is Chinese patent CN202411076583.6, which discloses a personalized learning recommendation method based on big data analysis and artificial intelligence. It obtains course tags through LLM, constructs multi-dimensional profiles by combining user data, and generates a recommendation list based on a content value model to avoid problems such as learning task accumulation and reduced efficiency.

[0004] Despite the design advantages of the above technical solutions, they also have the following technical shortcomings: First, they lack a technical process for fine-tuning pre-trained large models based on education-specific samples: The large language model optimization in CN202510601679.8 only revolves around policy texts and does not design complete technical steps for preprocessing, parameter adjustment, and loss optimization for education-specific course texts, learning cases, and geographically related samples, making it impossible for the model to adapt to the course recommendation needs in education scenarios; The model construction in CN202411076583.6 relies on user profiles and course tags and does not incorporate fine-tuning design based on education-specific samples, resulting in insufficient adaptability of the model to education scenarios. Secondly, there is a lack of coding technology solutions that integrate course content, student learning characteristics, and geographical features: CN202510601679.8's cross-modal alignment only focuses on the association between policy and course content, without constructing a single-modal extraction-cross-modal association coding technology process for course content, student learning characteristics, and geographical features, nor is there a design for a multi-head attention mechanism to integrate multi-modal features; CN202411076583.6 only recommends based on tag and profile matching, lacking a multi-modal feature association coding technology module, and thus failing to achieve deep fusion between features to improve recommendation matching accuracy. Thirdly, there is a lack of technical mechanisms for iteratively optimizing model parameters based on real-time learning effect data: CN202510601679.8 does not design a technical link for receiving learning effect data and constructing an optimization loss and updating model parameters through gradient descent; CN202411076583.6 lacks a technical design for dynamic parameter iteration after generating the recommendation list, resulting in the model being unable to continuously optimize based on real-time learning feedback, making it difficult to improve recommendation accuracy. In view of this, we propose an AI large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis. Summary of the Invention

[0005] The purpose of this invention is to provide an AI large-scale model course recommendation and adaptive learning path optimization system with multi-dimensional learning analysis to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention aims to provide an AI large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis, comprising:

[0007] The learning data processing unit collects structured, semi-structured, and unstructured data from the teaching affairs system, online learning platform, and classroom interaction tools. It also acquires authorized geographic location data (regional coordinates and learning venue type) simultaneously. The unit cleans the data using an outlier detection algorithm and outputs a standardized learning dataset with associated geographic location tags after feature standardization.

[0008] The multi-dimensional learning assessment unit extracts knowledge mastery quantification, learning behavior characteristics, and learning preference characteristics based on the standard learning dataset, and generates a multi-dimensional learning assessment result that integrates geographical location-related features.

[0009] The AI-powered large-scale model course recommendation unit employs a pre-trained large-scale model finely tuned from educational course texts, learning case studies, and geographic location-related samples. It uses an improved geo-content fusion encoding algorithm to associate and encode the structured features of the course content, multi-dimensional quantitative features of learning outcomes, and geographic location features. Dynamic matching is calculated based on cosine similarity and attention weights to select highly suitable courses and form a personalized recommendation list. Furthermore, based on real-time learning performance data generated during the user's learning of the personalized recommendation list, the model parameters are iteratively optimized using a gradient descent algorithm.

[0010] An adaptive learning path optimization unit, based on real-time learning interaction data from the personalized recommendation list, dynamically generates content sequence priority, difficulty adjustment coefficient, and single knowledge point learning time through a reinforced learning path optimization model (with learning efficiency improvement as the reward function), and feeds back learning effect data to the AI ​​large model course recommendation unit to support parameter iteration of the AI ​​large model course recommendation unit.

[0011] As a further improvement to this technical solution, the learning data processing unit includes a data acquisition module, a data cleaning module, a feature standardization module, and a tag association module that are connected in sequence via communication, wherein:

[0012] The data acquisition module collects structured data, semi-structured data, unstructured data, and authorized geographic location data through the academic affairs system database interface, online learning platform log interface, classroom interaction tool multimedia interface, and location service authorization interface, respectively.

[0013] The data cleaning module receives the raw data and uses an outlier detection algorithm based on statistical distribution to identify and remove data that exceeds a reasonable range.

[0014] The feature standardization module performs scale unification processing on the cleaned data, converting feature values ​​of different dimensions into standardized features that can be directly calculated.

[0015] The tag association module converts geographic location data into structured tags and binds them to standardized learning data.

[0016] As a further improvement to this technical solution, the tag association module converts geographic location data into structured tags and binds them to standardized learning data, including the following steps:

[0017] S14.1. Use the GeoHash algorithm to encode the original latitude and longitude coordinates in the authorized regional coordinate data to generate a 6-10 character string label (the encoding length is dynamically adjusted based on the geographic accuracy requirements). This string label corresponds one-to-one with a specific geographic area.

[0018] S14.2 Based on the preset location type classification model (a decision tree model trained through historical location data), the original location type descriptions in the geographic location data are mapped into structured category labels;

[0019] S14.3 Using the user's unique identifier and data collection timestamp as a joint index, the distributed association algorithm is used to associate the regional coordinate labels generated in S14.1, the location type labels generated in S14.2, and the standardized geospatial data output by the feature standardization module to form a standardized geospatial dataset containing complete geographic labels.

[0020] As a further improvement to this technical solution, the multi-dimensional learning analysis unit includes a knowledge feature extraction module, a behavioral feature extraction module, a preference feature extraction module, a geographic feature fusion module, and an evaluation result generation module, which are sequentially connected in communication.

[0021] The knowledge feature extraction module calculates a quantitative value of knowledge mastery based on the answer records and knowledge point application records in the standard chemistry dataset and through statistical learning algorithms.

[0022] The behavioral feature extraction module extracts learning behavioral features based on the learning duration distribution and operation sequence logs in the standard learning dataset, using a time-series feature extraction algorithm.

[0023] The preference feature extraction module extracts learning preference features based on resource type access records and content dwell time data in the standard learning dataset, through frequency statistics and proportion analysis.

[0024] The geographic feature fusion module receives the structured geographic location labels from the standard chemical intelligence dataset and uses a feature splicing algorithm to associate and fuse the regional coordinate labels, location type labels, and knowledge, behavior, and preference features.

[0025] The evaluation result generation module performs structured integration of the fused multi-dimensional features to generate multi-dimensional learning assessment results that include quantitative values ​​of knowledge mastery, learning behavior characteristics, learning preference characteristics, and geographical location correlation characteristics.

[0026] As a further improvement to this technical solution, the AI ​​large model course recommendation unit includes a model fine-tuning module, a feature encoding module, a matching degree calculation module, and a parameter optimization module that are sequentially connected in communication; wherein:

[0027] The model fine-tuning module receives course texts, learning cases, and geographical location-related samples from the education field, and performs targeted fine-tuning on the pre-trained large model through a process of data preprocessing, parameter adjustment, and loss optimization.

[0028] The feature encoding module adopts an improved geo-content fusion encoding algorithm, and encodes course content, learning situation, and geographical location features according to the process of "single-modal feature extraction - cross-modal association encoding";

[0029] The matching degree calculation module calculates the dynamic matching degree based on feature vector similarity and attention weight, and generates a personalized recommendation list by filtering according to the threshold.

[0030] The parameter optimization module receives real-time learning performance data and iteratively optimizes the model parameters through a process of loss function construction, gradient calculation, and parameter update.

[0031] As a further improvement to this technical solution, the model fine-tuning module's fine-tuning process for the pre-trained large model includes the following steps:

[0032] S31.1 Perform word segmentation on course texts in the education field to generate a word index sequence. , Represents a structured index sequence of course texts, containing A word index, The number of words in the text; feature vector transformation is performed on the quantitative values ​​of knowledge mastery and learning behavior characteristics in the learning situation cases to obtain the learning situation feature vector. ,Include Each feature component The number of dimensions for learning characteristics; numerical encoding of regional coordinate labels and location type labels in geographic location-related samples to obtain geographic feature vectors. ,Include Each feature component The number of dimensions of geographical features;

[0033] S31.2. Adopting the "freeze-fine-tuning" strategy, let the total number of layers in the pre-trained large model be... , This represents the total number of layers in the neural network of a pre-trained large model, with 80% of the bottom layers frozen. Parameters, only for the top 20% Adjust the parameters;

[0034] S31.3, using cross-entropy loss To optimize the objective, This represents the loss metric that measures the difference between the model's predicted values ​​and the actual values; the mini-batch gradient descent method is used, with the batch size set to... , This indicates the number of samples used for each model parameter update. The pre-trained large model is trained and iterated until... Stop fine-tuning when convergence occurs.

[0035] As a further improvement to this technical solution, the implementation of the improved geo-content fusion coding algorithm of the feature coding module includes the following steps:

[0036] S32.1 Extracting visual feature vectors from video frames using CNN , dimension Extracting semantic feature vectors from text content using BERT , dimension Extracting logically related feature vectors from interactive exercises using graph neural networks. , dimension ;Will , , splicing as a joint feature of the course ;

[0037] S32.2. The knowledge mastery quantification, learning behavior characteristics, and learning preference characteristics output by the multi-dimensional learning analysis unit are converted into learning feature vectors through a fully connected layer. , dimension ;

[0038] S32.3 Convert the region coordinate labels and location type labels output by the label association module into geographic feature vectors through the embedding layer. , dimension ;

[0039] S32.4, Employing a multi-head attention mechanism Integrating the characteristics of curriculum, student learning, and geography, the output is presented in S32.2. As a query vector, output in S32.1 With the output of S32.3 The concatenated result is used as the key vector and value vector, and the number of attention heads is set to [value]. ;

[0040] S32.5, will The outputs of each attention head are concatenated and then transformed by a linear layer to obtain a fused feature vector. , dimension .

[0041] As a further improvement to this technical solution, the matching degree calculation module calculates the dynamic matching degree and generates a personalized recommendation list, including the following steps:

[0042] S33.1, Based on the characteristics of curriculum collaboration With learning characteristic vector Calculate the cosine similarity to obtain the course-student similarity. ;

[0043] S33.2 Multi-head attention mechanism based on fusion of trimodal features The attention weights of course features and geographical features are extracted, and the average of these weights is taken to obtain the course-geographic attention weight. ;

[0044] S33.3, will and The dynamic matching degree is obtained through weighted fusion. ;

[0045] S33.4, Filtering Greater than the threshold The courses, according to Take first in descending order This creates a personalized recommendation list.

[0046] As a further improvement to this technical solution, the iterative optimization of model parameters in the parameter optimization module includes the following steps:

[0047] S34.1, Combine real-time learning performance data corresponding to the personalized recommendation list, including knowledge mastery improvement rate. With course completion rate The optimized loss is obtained through weighted calculation. ;

[0048] S34.2, Using the Adam optimizer Update model parameters and set the initial learning rate. ,Every Round iteration according to decay coefficient Adjustment ;

[0049] S34.3, continued iteration to Less than the stopping threshold When this happens, stop optimizing the model parameters.

[0050] As a further improvement to this technical solution, the adaptive learning path optimization unit includes a real-time interactive data acquisition module, a reinforcement learning model processing module, a path parameter generation module, and a feedback data output module, which are sequentially connected in communication.

[0051] The real-time interactive data acquisition module receives the personalized recommendation list output by the AI ​​large model course recommendation unit and synchronously collects real-time interactive data generated when the user learns from the personalized recommendation list, including knowledge point learning progress, practice answer accuracy, single learning interruption duration and knowledge point test score.

[0052] The reinforcement learning model processing module calls the reinforcement learning path optimization model, takes real-time interactive data as input, and uses the learning efficiency improvement value as the reward function for model inference.

[0053] The path parameter generation module dynamically generates three types of path optimization parameters—content sequence priority, difficulty adjustment coefficient, and single knowledge point learning time—based on the output of the reinforcement learning path optimization model.

[0054] The feedback data output module integrates the path optimization parameters and the corresponding learning effect data into a standardized feedback vector, which is then pushed to the parameter optimization module to support iterative calculation of model parameters.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. This invention uses the model fine-tuning module of the AI ​​large model course recommendation unit to fine-tune the pre-trained large model based on course texts, learning cases, and geographical location-related samples in the education field. It follows the process of data preprocessing, parameter adjustment, and loss optimization to fine-tune the pre-trained large model. This solves the problem of existing technologies lacking education-specific samples for fine-tuning pre-trained large models, enabling the pre-trained large model to better adapt to the course recommendation needs in education scenarios and laying the foundation for generating personalized recommendation lists that fit education scenarios in the future.

[0057] 2. This invention relies on the feature encoding module of the AI ​​large-scale model course recommendation unit and adopts an improved geo-content fusion encoding algorithm. First, it extracts visual features of course videos through CNN, semantic features of course texts through BERT, and logical association features of course interactive exercises through graph neural networks, and concatenates them into joint course features. Then, it converts multi-dimensional learning information features into learning information feature vectors through fully connected layers and geolocation labels into geo feature vectors through embedding layers. Finally, it fuses the three-modal features of course, learning information, and geography through a multi-head attention mechanism. This solves the problem of existing technologies lacking course and learning information and geography feature fusion encoding, and realizes a deep association between course content, learner learning status, and learning geography scene features, providing technical support for improving course recommendation matching accuracy.

[0058] 3. This invention utilizes the matching degree calculation module and parameter optimization module of the AI ​​large-scale model course recommendation unit. First, it calculates the cosine similarity based on the joint features of the courses and the learning characteristics, and obtains the dynamic matching degree by combining the course-geographic attention weight. Courses with a dynamic matching degree greater than a threshold are selected and the top N in descending order are used to generate a personalized recommendation list. Then, it receives real-time learning effect data of the user learning the recommended list, constructs an optimized loss, and uses the Adam optimizer to iteratively update the model parameters. This solves the problem of existing technologies lacking iterative optimization of model parameters based on real-time learning effect data, enabling course recommendations to be continuously adjusted according to the user's real-time learning feedback, thereby improving the dynamic optimization capability of recommendation accuracy. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0060] The meanings of the labels in the diagram are as follows:

[0061] 1. Learning data processing unit; 11. Data acquisition module; 12. Data cleaning module; 13. Feature standardization module; 14. Tag association module;

[0062] 2. Multi-dimensional learning analysis unit; 21. Knowledge feature extraction module; 22. Behavioral feature extraction module; 23. Preference feature extraction module; 24. Geographic feature fusion module; 25. Evaluation result generation module;

[0063] 3. Recommended modules for AI large-scale model courses; 31. Model fine-tuning module; 32. Feature encoding module; 33. Matching degree calculation module; 34. Parameter optimization module;

[0064] 4. Adaptive learning path optimization unit; 41. Real-time interactive data acquisition module; 42. Reinforcement learning model processing module; 43. Path parameter generation module; 44. Feedback data output module. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0066] like Figure 1 As shown, this embodiment provides an AI large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis, including:

[0067] This embodiment adopts a hybrid deployment architecture of "cloud server as the main component and edge terminal as the auxiliary component", and the specific configuration is as follows:

[0068] The cloud server uses Intel Xeon E5-2680v4 series CPUs (≥20 cores), NVIDIA A100 GPUs (≥40GB VRAM, ≥2 GPUs deployed), ≥128GB of memory, and ≥2TB of storage (using a RAID5 array, paired with Samsung PM983 enterprise-grade solid-state drives to accelerate data read and write). It is equipped with an Intel X710 10 Gigabit network adapter to ensure high-speed data transmission. The operating system is CentOS 7.9, and it relies on technologies such as Python 3.8, TensorFlow 2.8, and PyTorch 1.12.

[0069] The edge terminal supports Android 10+ or ​​iOS 14+ mobile devices (CPU performance ≥ Snapdragon 865 or Apple A14 level, memory ≥ 6GB, local reserved storage space ≥ 5GB), equipped with a Sony IMX586 sensor for collecting image data of learning scenarios, and a Qualcomm QCC5125 Bluetooth module to realize near-field communication between devices; communication between the cloud and the edge terminal is carried out through the HTTPS protocol, requiring a network bandwidth ≥ 100Mbps to ensure real-time synchronization of learning data and recommendation results.

[0070] Learning data processing unit 1 collects structured, semi-structured, and unstructured data from the academic affairs system, online learning platform, and classroom interaction tools. It also acquires authorized geographic location data (regional coordinates and learning venue type) simultaneously. The data is cleaned using an outlier detection algorithm and outputs a standardized learning dataset with associated geographic location tags after feature standardization.

[0071] In this embodiment, the learning data processing unit 1 includes a data acquisition module 11, a data cleaning module 12, a feature standardization module 13, and a tag association module 14, which are connected in sequence via communication.

[0072] The data acquisition module 11 collects structured data, semi-structured data, unstructured data, and authorized geographic location data through the database interface of the academic affairs system, the log interface of the online learning platform, the multimedia interface of the classroom interaction tool, and the location service authorization interface, respectively.

[0073] Specifically, the data acquisition module 11 collects multi-source data through multiple interfaces in a collaborative manner, and the specific logic is as follows:

[0074] Academic Affairs System Database Interface: The system uses the JDBC protocol to connect to the MySQL database of the academic affairs system, collects basic student information, course registration information, and academic performance data, and the collection frequency can be set to 2:00 AM daily (non-peak teaching hours).

[0075] Online learning platform log interface: calls the platform's RESTful API to collect semi-structured data in JSON format (learning behavior logs, learning progress data), and adopts real-time incremental collection (pulling data as soon as new logs are generated);

[0076] Multimedia interface for classroom interaction tools: Collects unstructured data (such as interactive recordings, whiteboard writing content, and classroom video clips) via the RTSP protocol, and performs lightweight compression of multimedia data (such as MP3 audio format and H.264 video format).

[0077] Location service authorization interface: After user authorization, call the device's GPS and base station positioning services to collect latitude and longitude coordinates (accurate to 6 decimal places) and original location description data (such as Wi-Fi name and surrounding POIs). The collection frequency is 30 minutes / time. Collection is paused when the device is stationary (position change < 10 meters).

[0078] The data cleaning module 12 receives the raw data and uses an outlier detection algorithm based on statistical distribution to identify and remove data that exceeds the reasonable range.

[0079] Specifically, the data cleaning module 12 aims to "remove invalid data and retain valid samples," and the specific implementation process of the outlier detection algorithm based on statistical distribution is as follows:

[0080] Data Dimension Segmentation: The collected raw data is segmented into dimensions such as "Learning Behavior", "Academic Performance", and "Geographic Location". Outlier detection is performed separately for each dimension (e.g., "Learning Behavior" includes two sub-dimensions: learning duration and number of questions answered; "Academic Performance" includes two sub-dimensions: exam score and homework score).

[0081] Statistical distribution characteristic calculation: For the raw data of each sub-dimension, a sliding window (window size set to 7 days) is used to calculate the statistical distribution characteristics, including the mean of the data in that dimension. Standard deviation For example, calculating the mean of the "daily study time" dimension. =120 minutes, standard deviation =60 minutes;

[0082] Outlier identification and removal: using The principle (common outlier criteria used in industry, requiring no unverified precision data) will exceed " Data within a certain range is considered outlier. Specific scenarios include:

[0083] Daily study time exceeds (like =120 minutes =When the daily study time exceeds 3 minutes when the total study time is 60 minutes, it is judged as "abnormally long duration" (possibly due to device miscounting).

[0084] The accuracy rate of answering questions was lower than (like =70% If the accuracy rate is less than 25% (when the accuracy rate is 15%) and there are no corresponding records of incorrect answers, it is judged as "invalid answer data".

[0085] If the geographical location's latitude and longitude exceed the target service area (e.g., the system only serves a certain city, but latitude and longitude data outside that city appears), it is judged as "abnormal location data";

[0086] Data completion (optional): For local data missing due to outlier removal, if the missing percentage is less than 5%, the average of the previous 3 days for that dimension will be used for completion; if the missing percentage is greater than 5%, it will be marked as a "data missing sample" and will not proceed to the subsequent standardization process.

[0087] The feature standardization module 13 performs scale unification processing on the cleaned data, converting feature values ​​of different dimensions into standardized features that can be directly calculated.

[0088] Specifically, feature standardization module 13 uses the "Z-score standardization" method (a scale unification scheme applicable to most machine learning scenarios) to unify features of different scales for the cleaned and effective data. The specific steps are as follows:

[0089] Feature dimension recognition: First, identify the original dimension of each feature. For example, the dimension of "answer accuracy" is "percentage (0-1)", the dimension of "study time" is "minutes (0-1440)", and the dimension of "exam score" is "score (0-1)".

[0090] Z-score standardization calculation: raw values ​​for each feature. According to the formula " "to perform the conversion, where" This is the mean of the feature across the entire sample. This represents the standard deviation of the feature across the entire sample. After transformation, the mean of all features is 0 and the standard deviation is 1. An example is shown below:

[0091] Original "Accuracy Rate" =80%, if =70% =15%, then after standardization ;

[0092] Original "learning duration" =150 minutes, if =120 minutes =60 minutes, then after standardization ;

[0093] Standardization result verification: Perform range verification on the transformed feature values ​​to ensure that more than 99% of the standardized feature values ​​fall within the range of "[-3,3]" (consistent with normal distribution characteristics). If the percentage of samples with a certain feature outside this range is higher than 1%, then the feature value is recalculated. and (After eliminating extreme value interference), the standardized process is executed again.

[0094] The tag association module 14 converts geographic location data into structured tags and binds them to the standardized learning data.

[0095] In this embodiment, the tag association module 14 converts geographic location data into structured tags and binds them to the standardized learning data, including the following steps:

[0096] S14.1. Use the GeoHash algorithm to encode the original latitude and longitude coordinates in the authorized regional coordinate data to generate a 6-10 character string label (the encoding length is dynamically adjusted based on the geographic accuracy requirements). This string label corresponds one-to-one with a specific geographic area.

[0097] Specifically, the code uses the open-source GeoHash library, and the length is adjusted according to the geographic precision requirements.

[0098] City-level recommendation: 6-digit code (corresponding to a range of 1.2km × 0.6km);

[0099] Campus-level recommendation: 8-digit code (corresponding to a 150m×150m range);

[0100] Recommended for classrooms / study rooms: 10-bit encoding (corresponding to a range of 19m×19m); the encoding result is stored as the "geo_hash_code" field.

[0101] S14.2 Based on the preset location type classification model (a decision tree model trained through historical location data), the original location type descriptions in the geographic location data are mapped into structured category labels;

[0102] Specifically, the preset location type classification model is trained based on the C4.5 decision tree algorithm (a classic and easily implemented classification algorithm), and the specific implementation is as follows:

[0103] Training dataset construction: Collect historical location data (at least one sample), each sample containing "input features" and "output labels", where:

[0104] Input characteristics: Device Wi-Fi name, surrounding POI type, collection time;

[0105] Output tags: Structured venue type (divided into four categories: "Campus Learning Venue", "Home Learning Venue", "Public Learning Venue", and "Non-Learning Venue");

[0106] Model training and validation: Divide the dataset into training and validation sets in a 7:3 ratio. Use the training set to train the decision tree model (set the maximum tree depth to 8 to avoid overfitting). Use the validation set to validate the model's accuracy (the accuracy should be no less than 90%). If the accuracy is not met, add samples and retrain.

[0107] Location type mapping: Input the real-time collected original location type description (e.g., "connected to 'XX Cafe Wi-Fi', nearby POI is 'shopping mall', collection time is 'weekend 10:00'") into the trained decision tree model, and the model outputs a structured category label such as "public learning place", which is stored as the "place_type" field.

[0108] S14.3 Using the user's unique identifier and the data collection timestamp as a joint index, the distributed association algorithm is used to associate the regional coordinate labels generated in S14.1, the location type labels generated in S14.2, and the standardized geographic information data output by the feature standardization module 13 to form a standardized geographic information dataset containing complete geographic labels.

[0109] Specifically, the Spark distributed computing framework (an engineering tool suitable for large-scale data association) is used, with "user unique identifier + data collection timestamp" as a composite index, to associate geographic tags with standardized meteorological data, including:

[0110] Data sharding: Sharding is done by "user ID", and data of the same user is allocated to the same computing node;

[0111] Composite index matching: Using "User ID (e.g., '202401') + collection timestamp (accurate to the second)" as the composite key, match learning data with geographical tags;

[0112] Field integration output: Data is integrated according to "user ID, collection timestamp, standardized learning features, geo_hash_code, place_type" and stored in HDFS for subsequent unit calls.

[0113] It is worth noting that before collecting sensitive personal data such as geographical location and learning behavior, the purpose, scope, and storage period of data collection are clearly communicated through an interaction process with the user's terminal (such as a learning app or web application). After the user confirms their consent, the system records and retains the authorization information, which includes a timestamp. During data transmission, the JSON format data is encrypted using the AES-256 encryption algorithm, relying on the communication component of the learning data processing unit 1. The encryption key is dynamically generated through a secure channel. During the storage phase, the learning data processing unit 1 performs SHA-256 hash desensitization and geographical location generalization (aggregating specific coordinates into regional codes) on personally identifiable information (such as precise geographical coordinates) to ensure that the data cannot be directly associated with a specific user. At the same time, the storage management logic sets the data storage period for learning data to 180 days. After the expiration date, the learning data processing unit 1 automatically triggers a data cleanup operation to completely delete the original data and backups.

[0114] Multi-dimensional Learning Analysis Unit 2: Based on the standard learning dataset, Multi-dimensional Learning Analysis Unit 2 extracts quantitative values ​​of knowledge mastery, learning behavior characteristics, and learning preference characteristics to generate multi-dimensional learning assessment results that integrate geographical location-related features.

[0115] In this embodiment, the multi-dimensional learning analysis unit 2 includes a knowledge feature extraction module 21, a behavior feature extraction module 22, a preference feature extraction module 23, a geographic feature fusion module 24, and an evaluation result generation module 25, which are connected in sequence.

[0116] The knowledge feature extraction module 21 calculates the quantitative value of knowledge mastery based on the answer records and knowledge point application records in the standard chemistry dataset through statistical learning algorithms.

[0117] Specifically, the knowledge feature extraction module 21 takes the answer records (including the answer results of classroom exercises, homework, and unit tests, labeled with the knowledge points associated with the questions, the correctness of the answers, and the error type) and knowledge point application records (including the number of times knowledge points are called in homework / tests, the number of times they are called correctly, and cross-knowledge point integrated application cases) in the standard chemistry learning dataset as input, and calculates the quantitative value of knowledge mastery through a weighted average statistical learning algorithm, which specifically includes:

[0118] Knowledge point system mapping: First, establish a hierarchical system of course knowledge points (e.g., "Advanced Mathematics" is divided into first-level knowledge points such as "Function Limits", "Derivatives and Differentials", and "Integrals", with each level of knowledge point further subdivided into second-level sub-knowledge points). Bind the questions in the answer records and knowledge point application records to the corresponding sub-knowledge points to form a "question-knowledge point" mapping table.

[0119] Single knowledge point basic score calculation: For a single sub-knowledge point, the basic score is calculated according to the formula "Basic score = (Number of correct answers + Number of correct applications) / (Total number of answers + Total number of applications)", where "Number of correct answers" is the number of correct answers in the questions related to this knowledge point, and "Number of correct applications" is the number of times this knowledge point is correctly applied in comprehensive questions;

[0120] Score weighting adjustment: The basic score is weighted according to the importance of the knowledge points (e.g., the weight of core knowledge points is set to 1.2, and the weight of secondary knowledge points is set to 0.8) to obtain a quantitative value of mastery of a single knowledge point (the value ranges from 0 to 1, where 0 indicates not mastered and 1 indicates fully mastered).

[0121] Knowledge point mastery results integration: According to the knowledge point hierarchy, the mastery quantification values ​​of the second-level sub-knowledge points are summarized into the first-level knowledge points to form a structured knowledge feature table of "first-level knowledge point - second-level sub-knowledge point - mastery quantification value" for subsequent modules to call.

[0122] The behavior feature extraction module 22 extracts learning behavior features based on the learning duration distribution and operation sequence logs in the standard learning dataset using a time-series feature extraction algorithm.

[0123] Specifically, in the behavior feature extraction module 22, the extraction of learned behavior features through the temporal feature extraction algorithm includes the following steps:

[0124] Data preprocessing: The operation sequence logs are sorted by user unique identifier and data collection timestamp, and misclick data with a single operation dwell time of less than 1 second are removed; the learning time distribution is classified and statistically analyzed according to the time periods of "early (6:00-12:00), noon (12:00-18:00), and evening (18:00-24:00)" and the learning scenarios of "fragmented (single learning time <30 minutes) and continuous (single learning time ≥30 minutes)".

[0125] Duration distribution feature extraction: Calculate the percentage of learning time in each time period (learning time in a certain time period / total daily learning time), the percentage of continuous learning (total continuous learning time / total learning time), and the percentage of time for each course type (learning time for a certain type of course / total learning time), and generate a duration distribution feature vector;

[0126] Operation sequence pattern mining: A sliding window with a window size of 5 consecutive operations is used to segment and sort the operation sequence logs. The frequency of each operation sequence pattern is counted, and high-frequency patterns with a frequency of ≥5 times are filtered out and encoded as operation sequence feature labels such as "video learning - practice consolidation" and "wrong question review - second practice".

[0127] Behavioral feature integration: Associate the duration distribution feature vector with the operation sequence feature label to form structured learning behavior feature data containing "time period preference", "learning continuity", "course type focus" and "operation habits".

[0128] The preference feature extraction module 23 extracts learning preference features based on resource type access records and content dwell time data in the standard learning dataset through frequency statistics and proportion analysis.

[0129] Specifically, extracting learning preference features through frequency statistics and percentage analysis includes the following steps:

[0130] Resource type classification statistics: The resource types in the standard learning dataset are divided into "video (recorded courses, short videos of knowledge points), text (courseware documents, exercise solutions), interactive (online exercises, mock tests), and audio (audio courses, audio explanations of knowledge points)". The number of visits and total dwell time for each type of resource are counted (records with invalid dwell time <5 seconds are removed).

[0131] Preference percentage calculation: Calculate "Resource type access percentage = number of visits to a certain type of resource / total number of visits" and "Resource type dwell percentage = total dwell time of a certain type of resource / total dwell time". Mark the resources as "core preferred resources", "general preferred resources" and "non-preferred resources" according to the percentages ≥40%, 20%-40%, and <20%, respectively.

[0132] Content theme preference analysis: Frequency statistics are performed on the content themes of core preferred resources (such as the "Machine Learning Basics" theme in video resources) to extract the core content themes that users pay attention to;

[0133] Preference feature integration: "Resource type preference" (core / general / non-preference) is associated with "content theme preference" to form structured learning preference feature data, which can be called by the subsequent geographic feature fusion module 24.

[0134] The geographic feature fusion module 24 receives the geographic location structured labels from the standard chemical intelligence dataset and uses a feature splicing algorithm to associate and fuse the regional coordinate labels, location type labels, and knowledge, behavior, and preference features.

[0135] Specifically, the process of associating and fusing regional coordinate labels, location type labels, and knowledge, behavioral, and preference features using a feature concatenation algorithm includes the following steps:

[0136] Geographic tag digitization: Convert regional coordinate tags (GeoHash strings) into fixed-dimensional (e.g., 128-dimensional) numerical vectors through an embedding layer; Encode location type tags into numerical values ​​according to preset rules (e.g., "Campus learning location" = 1, "Home learning location" = 2, "Public learning location" = 3, "Non-learning location" = 0);

[0137] Unified format of multiple features: knowledge features (knowledge mastery quantification value vector), behavioral features (duration distribution feature vector + operation sequence feature label encoding vector), and preference features (resource type preference + content topic preference encoding vector) are uniformly adjusted into numerical vectors of the same length (e.g., 256 dimensions).

[0138] Feature concatenation: Consolidate the features into a single fused feature vector in the order of "knowledge feature vector - behavior feature vector - preference feature vector - geographic coordinate vector - location type value";

[0139] Fusion feature verification: Check whether the fusion feature vector has missing dimensions or abnormal values ​​(such as null values ​​or values ​​outside the reasonable range). If abnormal, backtrack to the corresponding module to re-extract features to ensure the integrity of the fusion features for subsequent evaluation result generation module 25 to call.

[0140] The assessment result generation module 25 performs structured integration of the fused multi-dimensional features to generate multi-dimensional learning assessment results that include quantitative values ​​of knowledge mastery, learning behavior characteristics, learning preference characteristics, and geographical location correlation characteristics.

[0141] Specifically, generating multi-dimensional learning assessment results through structured integration technology includes the following steps:

[0142] Feature dimension decomposition: The fused feature vector is decomposed into "knowledge dimension", "behavioral dimension", "preference dimension" and "geographical dimension", and the original feature indicators corresponding to each dimension are clarified (such as the knowledge dimension including the quantitative value of mastery of each knowledge point).

[0143] The features within each dimension are structured as follows: the knowledge dimension is presented as "knowledge point name - mastery quantification value - mastery level (≥0.8 mastery, 0.5-0.8 need improvement, <0.5 weak)"; the behavior dimension is presented as "time period preference - learning continuity - operation habits"; the preference dimension is presented as "core preferred resource type - core content theme"; and the geographic dimension is presented as "GeoHash encoding - location type - scene adaptation suggestion".

[0144] Evaluation results are integrated and output: Multi-dimensional learning assessment results are generated according to the structure of "user unique identifier - data collection time period - knowledge dimension details - behavior dimension details - preference dimension details - geographical dimension details - comprehensive learning conclusion" and pushed to AI big model course recommendation unit 3.

[0145] AI Large Model Course Recommendation Unit 3 employs a pre-trained large model finely tuned from course texts, learning case studies, and geographic location-related samples in the education field. An improved geo-content fusion encoding algorithm associates and encodes the structured features of the course content, multi-dimensional quantitative features of learning outcomes, and geographic location features. Dynamic matching is calculated based on cosine similarity and attention weights to select highly suitable courses and form a personalized recommendation list. Furthermore, based on real-time learning performance data generated during the user's learning of the personalized recommendation list, the model parameters are iteratively optimized using a gradient descent algorithm.

[0146] In this embodiment, the AI ​​large model course recommendation unit 3 includes a model fine-tuning module 31, a feature encoding module 32, a matching degree calculation module 33, and a parameter optimization module 34, which are sequentially connected in communication; wherein:

[0147] The model fine-tuning module 31 receives course texts, learning cases, and geographical location-related samples from the education field. It performs targeted fine-tuning on the pre-trained large model through a process of data preprocessing, parameter adjustment, and loss optimization. The core of the model fine-tuning module 31 is to adapt the pre-trained large model to the education scenario. In actual implementation, it is carried out through three steps: data preprocessing, parameter adjustment, and loss optimization. Each step is designed in combination with the characteristics of education data and engineering efficiency requirements.

[0148] The feature encoding module 32 adopts an improved geo-content fusion encoding algorithm to encode course content, learning situation, and geographical location features according to the process of "single-modal feature extraction - cross-modal association encoding". The feature encoding module 32 converts the three heterogeneous features of course, learning situation, and geography into a unified dimension fusion vector through the improved geo-content fusion encoding algorithm. The process is divided into two steps: "single-modal feature extraction" and "cross-modal association encoding". Each step is designed around "feature compatibility" and "association capture capability".

[0149] The matching degree calculation module 33 calculates the dynamic matching degree based on feature vector similarity and attention weight, and generates a personalized recommendation list by filtering according to the threshold. The matching degree calculation module 33 judges the course suitability from two dimensions: "learning situation suitability" and "scenario suitability", and finally generates a personalized recommendation list that users can understand. The process takes into account both calculation accuracy and user experience.

[0150] The parameter optimization module 34 receives real-time learning performance data and iteratively optimizes the model parameters through a process of loss function construction, gradient calculation, and parameter update. Based on the user's real-time learning performance, the parameter optimization module 34 continuously adjusts the model parameters, forming a closed loop of "recommendation-learning-optimization," ensuring that recommendation accuracy continuously improves as the user learns.

[0151] In this embodiment, the fine-tuning process of the model fine-tuning module 31 for the pre-trained large model includes the following steps:

[0152] S31.1 Perform word segmentation on course texts in the education field to generate a word index sequence. , Represents a structured index sequence of course texts, containing A word index, The number of words in the text; feature vector transformation is performed on the quantitative values ​​of knowledge mastery and learning behavior characteristics in the learning situation cases to obtain the learning situation feature vector. ,Include Each feature component The number of dimensions for learning characteristics; numerical encoding of regional coordinate labels and location type labels in geographic location-related samples to obtain geographic feature vectors. ,Include Each feature component The number of dimensions of geographical features;

[0153] When processing course texts, a word segmentation tool based on the jieba extended educational terminology dictionary was used: First, 50+ educational professional terms such as "fundamental theorem of calculus" and "linear regression model" were added to avoid the fragmentation of professional vocabulary. Then, the course outline, knowledge point analysis, and other texts were segmented, and each word was mapped to an index value with reference to a general vocabulary list in the education field (index range 1-50), forming a word index sequence. .here A structured index sequence representing the course text. The number of words in the text changes dynamically with the text length. For example, a 15-word lesson text can typically generate [a certain number of words]. = Approximately 8 indices.

[0154] When processing learning cases, "quantitative values ​​of knowledge mastery" (such as mastery of "function limits" 0.75 and mastery of "application of derivatives" 0.62) and "learning behavior characteristics" (such as "percentage of evening learning 0.6% and percentage of continuous learning 0.4%) are obtained from the multi-dimensional learning analysis unit 2, and arranged into an original sequence according to the fixed order of "knowledge characteristics - behavior characteristics". Then, Z-score standardization is used to eliminate dimensional differences, resulting in the learning performance feature vector. .in It is a vector form of learning characteristics. It is a feature dimension, determined by the number of knowledge points and the number of behavioral indicators. For example, when it includes 12 knowledge points and 6 behavioral indicators, =18; To eliminate dimensional differences, Z-score standardization is used in the standardized formula. , It is the first The mean of all samples for each original feature. That is the corresponding standard deviation.

[0155] The geographic location sample processing involves converting the "regional coordinate labels (GeoHash strings)" output by the label association module 14 into vectors through a 32-dimensional embedding layer. The "location type label" (such as "campus learning location" or "home learning location") is converted into a 4-dimensional vector using one-hot encoding (e.g., "campus learning location" corresponds to [1,0,0,0]). The two are then concatenated to form a geographic feature vector. , It is a vector form of geographical features. It is the feature dimension, which is exactly the sum of 32-dimensional GeoHash embedding plus 4-dimensional location encoding.

[0156] S31.2. Adopting the "freeze-fine-tuning" strategy, let the total number of layers in the pre-trained large model be... , This represents the total number of layers in the neural network of a pre-trained large model, with 80% of the bottom layers frozen. Parameters, only for the top 20% Adjust the parameters;

[0157] During the parameter tuning phase, an educationally adapted BERT-base-uncased pre-trained model (with a total number of neural network layers) was selected. =12), employing a "freeze-fine-tune" strategy: freezing the underlying layer. The parameters (rounded to 9 layers) are responsible for basic semantic understanding. Retaining general capabilities can reduce reliance on educational data and computational costs. Only the parameters of the top 3 layers are fine-tuned to learn the correlation between "learning situation-curriculum-geography" and avoid overfitting caused by insufficient data.

[0158] S31.3, using cross-entropy loss To optimize the objective, This represents the loss metric that measures the difference between the model's predicted values ​​and the actual values; the mini-batch gradient descent method is used, with the batch size set to... , This indicates the number of samples used for each model parameter update. The pre-trained large model is trained and iterated until... Stop fine-tuning when convergence occurs.

[0159] In this loss optimization stage, cross-entropy loss is used. To measure the difference between the model's predicted values ​​and the actual values, mini-batch gradient descent is used for training. The calculation combines the sample size of each update. (Balancing GPU memory usage and training efficiency), the formula is: ,in The model outputs "learning situation-course matching probability". These are genuine labels marked by humans. =1 indicates adaptation. =0 indicates an incompatible match.

[0160] Meanwhile, the training termination condition is 5 consecutive rounds. fluctuation range ( (For the current round), at this point, it is determined that the loss has converged, fine-tuning is stopped, and the model parameters are saved for subsequent feature encoding.

[0161] In this embodiment, the implementation of the improved geo-content fusion coding algorithm of the feature coding module 32 includes the following steps:

[0162] S32.1 Extracting visual feature vectors from video frames using CNN , dimension Extracting semantic feature vectors from text content using BERT , dimension Extracting logically related feature vectors from interactive exercises using graph neural networks. , dimension ;Will , , splicing as a joint feature of the course ;

[0163] For the multimodal data of the course, features were extracted using different networks and then concatenated: keyframes were extracted from the course video at a frequency of "1 frame every 10 frames" and input into the convolutional layer of the ResNet50 network to obtain a 2048-dimensional visual feature vector. The course documents and subtitle text are input into the BERT-uncased model, and the output of the [CLS] token is used as a 768-dimensional semantic feature vector. The interactive exercises first construct a graph structure of "question-knowledge point-type of wrong question", and then extract a 512-dimensional logical association feature vector through a graph attention network (GAT). Then, these three vectors are superimposed and concatenated along their dimensions to form the joint feature of the course. (The dimensions are 2048+768+512=3328, covering the visual, semantic and logical connection information of the course).

[0164] S32.2. The knowledge mastery quantification, learning behavior characteristics, and learning preference characteristics output by the multi-dimensional learning analysis unit 2 are converted into learning feature vectors through a fully connected layer. , dimension ;

[0165] To adapt to subsequent fusion requirements, the learning characteristics processing concatenates the "knowledge mastery quantification value + learning behavior characteristics + learning preference characteristics" output from the multi-dimensional learning analysis unit 2 into the original learning sequence. This sequence is then input into a fully connected layer with two hidden layers (using ReLU activation function) and processed using the formula... This is converted into a fixed-dimensional vector. It is a standard chemical feature vector (dimension) =512), For the original learning sequence, , The weights and biases of the first fully connected layer. , The weights and biases of the second layer are used to ensure that the learning characteristics are compatible with other feature dimensions.

[0166] S32.3. Convert the region coordinate labels and location type labels output by the label association module 14 into geographic feature vectors through the embedding layer. , dimension ;

[0167] The geographic feature processing converts the GeoHash string output by the tag association module 14 into a vector through a 32-dimensional embedding layer. The location type label is converted into a vector through an 8-dimensional embedding layer. The two are concatenated to form a 40-dimensional geographic feature vector. ; The standardized geographic feature vector, dimension =40, ensuring integration with the curriculum and student learning characteristics in subsequent stages.

[0168] S32.4, Employing a multi-head attention mechanism Integrating the characteristics of curriculum, student learning, and geography, the output is presented in S32.2. As a query vector, output in S32.1 With the output of S32.3 The concatenated result is used as the key vector and value vector, and the number of attention heads is set to [value]. ;

[0169] Cross-modal association coding employs a multi-head attention mechanism Capturing the correlation between three types of features, attention head number (Balancing correlation capture capability with computational cost). In specific operations, the learning characteristic vector is used. As a query vector, the course's joint features are used. With geographic feature vectors The concatenated result serves as the key vector (Key) and value vector (Value). First, the Query, Key, and Value are passed through eight independent linear layers to obtain the input for each attention head. (h=1 to 8); then calculate the output for each attention head. , The total dimension of the Key. This is used to mitigate gradient vanishing; finally, the outputs of the eight attention heads are concatenated to obtain... (Dimensions remain at 3368).

[0170] S32.5, will The outputs of each attention head are concatenated and then transformed by a linear layer to obtain a fused feature vector. , dimension .

[0171] Finally, a fused feature vector is generated. At that time, The input linear compression layer is converted into a 512-dimensional vector using the following formula: .in It is a cross-modal fusion feature vector (dimension) =512, to meet subsequent matching degree calculation requirements). and These are the weights and biases of the linear compression layer, respectively, to ensure the effectiveness of the compressed features.

[0172] In this embodiment, the matching degree calculation module 33 calculates the dynamic matching degree and generates a personalized recommendation list, including the following steps:

[0173] S33.1, Based on the characteristics of curriculum collaboration With learning characteristic vector Calculate the cosine similarity to obtain the course-student similarity. ;

[0174] Calculate the similarity between course and student learning. At that time, cosine similarity is used to measure the fit between course content and students' learning needs, and the formula is: For example, a calculus course Learning from users The dot product is 2. Its 2-norm is 20. The 2-norm is 10. Substituting this into the formula, we get... This indicates that the course is perfectly suited to the user's learning situation; the value range of this indicator is [-1, 1], and the closer the value is to 1, the higher the suitability.

[0175] S33.2 Multi-head attention mechanism based on fusion of trimodal features The attention weights of course features and geographical features are extracted, and the average of these weights is taken to obtain the course-geographic attention weight. ;

[0176] Course - Geographical Attention Weight Extracting information from the output of multi-head attention mechanisms reflects the adaptability of the course to the learning scenario. Specifically, starting from... Select course features With geographical features The association weight matrix is ​​used to calculate the average weight of each attention head; for example, the average weights of the 8 attention heads are 0.85, 0.88, 0.90, 0.86, 0.87, 0.89, 0.84, and 0.86, respectively. The average of these averages is then calculated to obtain... The value of this indicator ranges from [0,1]. The closer the value is to 1, the more suitable the course is for the current learning scenario.

[0177] S33.3, will and The dynamic matching degree is obtained through weighted fusion. ;

[0178] Dynamic matching degree It is a combined result of both, taking into account that "matching learning needs" is the core requirement for recommendations, and therefore... Set a weight of 0.7 to give Set a weight of 0.3, and the calculation formula is as follows: For example, a certain course , After substituting, we get\ This indicates that the course has a high degree of comprehensive adaptability; the value range of this indicator is [0,1], and the larger the value, the stronger the comprehensive adaptability.

[0179] S33.4, Filtering Greater than the threshold The courses, according to Take first in descending order This creates a personalized recommendation list.

[0180] When generating a personalized recommendation list, first set a matching threshold. =0.6 (Adjusted based on historical recommendation performance, effectively filtering out poorly adapted courses); then filter out The courses, according to Sort by highest to lowest; select the last sorted items. One course (10 courses are provided to meet users' short-term learning needs while avoiding the decision-making burden of too many choices). In addition to the course name, the list also includes labels... Values ​​and reasons for suitability, such as "[Advanced Mathematics (Calculus)]: "Adapted to your weak points in calculus knowledge + campus learning scenarios", helping users understand the recommendation logic and improve the user experience.

[0181] In this embodiment, the iterative optimization of model parameters in the parameter optimization module 34 includes the following steps:

[0182] S34.1, Combine real-time learning performance data corresponding to the personalized recommendation list, including knowledge mastery improvement rate. With course completion rate The optimized loss is obtained through weighted calculation. ;

[0183] Construct optimization loss At that time, the main reference is the knowledge mastery improvement rate. and course completion rate .in This is the difference in a user's mastery of knowledge points before and after learning the recommended course (e.g., a mastery of "Applications of Derivatives" was 0.5 before learning, and improved to 0.8 after learning). =0.3); It is the ratio of the number of recommended course lessons actually completed by the user to the total number of recommended lessons (e.g., 8 out of 10 recommended lessons are completed). =0.8).

[0184] At the same time, considering that "knowledge acquisition and improvement" is the core goal of learning, we give Set a weight of 0.6 to give A weight of 0.4 is set, and a "negative loss" design is used (the better the learning effect, the closer the loss value is to 0). The calculation formula is as follows: For example, in the case above, The value range of this indicator is [-1, 0].

[0185] S34.2, Using the Adam optimizer Update model parameters and set the initial learning rate. ,Every Round iteration according to decay coefficient Adjustment ;

[0186] Model parameter updates use the Adam optimizer To balance parameter update speed and stability. Initial learning rate. Set as (To avoid excessively large initial update increments that could cause model oscillations); each iteration Wheel, according to attenuation coefficient Adjust the learning rate (e.g., the learning rate in round 10) Round 20 (To avoid parameter oscillations during later training).

[0187] At the same time, during each update, enter Real-time learning performance data (with the same batch size as during model fine-tuning to ensure training consistency) is used to calculate... Gradients of key parameters such as "feature encoding layer weights" and "attention layer weights" Then, the parameters are updated using the Adam optimizer's formula:

[0188] ;in, This is the cumulative value of first-order momentum;

[0189] ;in, This is the cumulative value of second-order momentum;

[0190] ; For the Adam optimizer and second momentum Perform deviation correction;

[0191] Used for parameter updates. For the first Wheel parameters;

[0192] In the above formula, For momentum term, This is a numerical stability term to avoid the denominator being zero.

[0193] S34.3, continued iteration to Less than the stopping threshold When this happens, stop optimizing the model parameters.

[0194] The criterion for stopping the iteration is to set a stopping threshold. (A common choice in engineering practice to balance convergence accuracy and computational efficiency), calculation after each iteration. The absolute value, when it satisfies for 3 consecutive rounds When the model parameters converge and further iterations show no significant improvement, optimization is stopped. The optimized parameters are then synchronized to the model fine-tuning module 31 and the feature encoding module 32 to ensure that subsequent recommendation processes are dynamically adjusted based on the latest learning results, continuously improving recommendation accuracy.

[0195] In summary, the AI ​​large-scale model course recommendation unit 3 constructs a system of "educational scenario-based model adaptation - deep fusion of cross-modal features - dynamic matching and closed-loop optimization," achieving a comprehensive breakthrough in model adaptability, feature processing, matching logic, and optimization mechanisms compared to traditional course recommendation methods. Specifically, it adapts the pre-trained large-scale model to educational scenarios through a "freeze-fine-tuning" strategy, extracts multimodal features of the course using CNN, BERT, and graph neural networks, and deeply fuses these features with learning and geographical features through multi-head attention. Then, it dynamically calculates the comprehensive matching degree based on "course-learning cosine similarity" and "course-geographical attention weights." Finally, based on the knowledge mastery improvement rate and course completion rate, it iterates the model parameters in real time through the Adam optimizer, forming a "recommendation-learning-optimization" closed loop. This not only solves the pain points of traditional recommendation models such as poor adaptability, reliance on single-dimensional features, static matching, and offline optimization, but also achieves a comprehensive improvement in recommendation accuracy, scenario adaptability, and dynamism. For example, it can accurately recommend basic calculus theory courses with video resources to users who are "weak in calculus and prefer video learning, and often learn in campus settings." Furthermore, if the user's knowledge improvement is not significant after learning, it can quickly adjust the recommendation difficulty or content type, ultimately helping the user focus on highly adaptable learning content and significantly improving learning efficiency.

[0196] Furthermore, to clarify the data interaction logic between the various units of the system, this embodiment adopts the RESTful API protocol to realize the entire data flow of "Learning Data Processing Unit 1 - Multi-dimensional Learning Analysis Unit 2 - AI Large Model Course Recommendation Unit 3 - Adaptive Learning Path Optimization Unit 4", and uniformly adopts the JSON format. Specifically, when the Learning Data Processing Unit 1 transmits data to the Multi-dimensional Learning Analysis Unit 2, it uses the POST request method and must carry the Authorization token generated by the JWT mechanism; when the Multi-dimensional Learning Analysis Unit 2 transmits data to the AI ​​Large Model Course Recommendation Unit 3, it uses the POST request method and includes the core fields of "Multi-dimensional Learning Feature Matrix" and "Knowledge Graph Related Data"; when the AI ​​Large Model Course Recommendation Unit 3 transmits data to the Adaptive Learning Path Optimization Unit 4, it uses the POST request method and includes the core fields of "Course Recommendation List" and "Recommendation Confidence Matrix"; when the Adaptive Learning Path Optimization Unit 4 provides feedback on the adjusted path data to the Learning Data Processing Unit 1, Multi-dimensional Learning Analysis Unit 2, and AI Large Model Course Recommendation Unit 3, it uses the PUT request method and includes the core fields of "Optimized Learning Path" and "Path Adjustment Description".

[0197] The Adaptive Learning Path Optimization Unit 4, based on real-time learning interaction data from the personalized recommendation list, dynamically generates content sequence priority, difficulty adjustment coefficient, and single knowledge point learning time through a reinforced learning path optimization model (with learning efficiency improvement as the reward function). It then feeds back the learning effect data to the AI ​​Large Model Course Recommendation Unit 3 to support parameter iteration of the AI ​​Large Model Course Recommendation Unit 3.

[0198] In this embodiment, the adaptive learning path optimization unit 4 includes a real-time interactive data acquisition module 41, a reinforcement learning model processing module 42, a path parameter generation module 43, and a feedback data output module 44, which are sequentially connected in communication.

[0199] The real-time interactive data acquisition module 41 receives the personalized recommendation list output by the AI ​​large model course recommendation unit 3, and synchronously collects the real-time interactive data generated when the user learns the personalized recommendation list, including the learning progress of knowledge points, the accuracy rate of answering practice questions, the duration of a single learning interruption, and the test score of knowledge points.

[0200] Specifically, the real-time interactive data acquisition module 41 first receives the personalized recommendation list output by the AI ​​large model course recommendation unit 3, generates a unique association identifier of "recommendation ID-course ID-user ID" for each recommended course, and then collects the knowledge point learning progress (recording the completion rate as an integer from 0 to 1), the practice answer accuracy rate (recording the ratio of the number of correct answers to the total number of questions as a percentage), the duration of a single learning interruption (recording the time of leaving the learning session in minutes), and the knowledge point test score (recording the test result as an integer from 0 to 1) through the course learning platform progress tracking interface, interactive practice system, client behavior monitoring SDK, and knowledge point test system. Finally, the data is organized in the format of "recommendation ID-user ID-data type-value" and stored in the time-series database to ensure that the data can accurately match the recommended content and be traced back by time.

[0201] The reinforcement learning model processing module 42 calls the reinforcement learning path optimization model, takes real-time interactive data as input, and uses the learning efficiency improvement value as the reward function for model inference.

[0202] Specifically, the reinforcement learning model processing module 42 achieves dynamic optimization of the learning path through the process of "state representation - reward design - policy iteration", and the specific implementation is as follows:

[0203] State representation: Mapping real-time interaction data to reinforcement learning state vectors. ,Include:

[0204] Normalized values ​​for knowledge point learning progress (e.g., 60% converted to 0.6); normalized values ​​for practice question answering accuracy (e.g., 80% converted to 0.8);

[0205] Normalized value of single learning interruption duration (e.g., 20 minutes is converted to 0.2, and 60 minutes is used as the maximum value for linear normalization);

[0206] Normalized scores for knowledge point tests (e.g., 75 points converted to 0.75); final It is a 4-dimensional vector, ensuring that the values ​​of each dimension are within the range of 0-1, which facilitates unified processing in the model.

[0207] Reward Design: Using the improvement in learning efficiency as the reward function The calculation method is as follows:

[0208] Learning efficiency is defined as "improvement in knowledge mastery / learning time"; if, after learning a current knowledge point, the mastery level increases from 0.5 to 0.7, and the learning time is 30 minutes, then the learning efficiency is... ;

[0209] award Take the learning efficiency value directly; the larger the value, the better the learning effect and the higher the reward the model receives.

[0210] Policy iteration: A deep Q-network (DQN) is used as the reinforcement learning path optimization model. Its network structure includes:

[0211] Input layer: Receives a 4-dimensional state vector Hidden layers: 2 fully connected layers, each containing 64 neurons, with ReLU activation function;

[0212] Output layer: Outputs the Q-value (action value) of the three types of actions: "content sequence adjustment", "difficulty adjustment" and "duration adjustment".

[0213] The model updates its parameters using mini-batch gradient descent, sampling 32 interaction data points for training each time, with the goal of maximizing the cumulative reward. ,in As a discount factor, it balances current and future rewards; This represents the total number of time steps.

[0214] The path parameter generation module 43 dynamically generates three types of path optimization parameters based on the output of the reinforcement learning path optimization model: content sequence priority, difficulty adjustment coefficient, and single knowledge point learning time.

[0215] Specifically, the path parameter generation module 43 generates three types of path optimization parameters based on the action Q-values ​​output by the reinforcement learning path optimization model: content sequence priority sorts the knowledge points in the recommendation list according to the "content sequence adjustment" action Q-values ​​output by the model, adjusting the original sequence priority (e.g., adjusting the original ["function limit", "derivative", "integral"] to ["derivative", "function limit", "integral"]); difficulty adjustment coefficient maps the "difficulty adjustment" action Q-values ​​to a coefficient in the range of 0.8-1.2 (a high Q-value results in a coefficient close to 1.2, while a low Q-value results in a coefficient close to 0.8, such as setting the coefficient to 0.8 when the accuracy rate is 60%); and single knowledge point learning time combines the "time adjustment" action Q-values ​​with the knowledge point complexity (judged according to the historical average learning time) to generate a suggested time (e.g., 40 minutes is suggested when the complexity is high and the Q-value is high).

[0216] The feedback data output module 44 integrates the path optimization parameters and the corresponding learning effect data into a standardized feedback vector, which is then pushed to the parameter optimization module 34 to support iterative calculation of model parameters.

[0217] Specifically, the feedback data output module 44 first collects three types of parameters from the path parameter generation module 43: content sequence priority, difficulty adjustment coefficient, and single knowledge point learning time, as well as corresponding learning effect data such as the improvement in knowledge point mastery and course completion rate. Then, according to the structure of "user ID-recommendation ID-path parameter type-value-learning effect value", it converts these data into a standardized feedback vector containing a content sequence priority vector, a difficulty adjustment coefficient vector, and a learning effect vector, ensuring that the data structure matches the input requirements of the parameter optimization module 34 of the AI ​​large model course recommendation unit 3. Finally, it pushes the standardized feedback vector to the parameter optimization module 34 through the RESTful API interface, supporting its iterative optimization of model parameters to form a "recommendation-learning-optimization" closed loop.

[0218] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0219] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional learning analysis-based AI large-scale model course recommendation and adaptive learning path optimization system, characterized in that, include: The learning data processing unit (1) collects structured, semi-structured, and unstructured data from the teaching affairs system, online learning platform, and classroom interaction tools, and simultaneously acquires authorized geographic location data. It cleans the data using an outlier detection algorithm and outputs a standardized learning data set with associated geographic location tags after feature standardization. The learning data processing unit (1) includes a tag association module (14), which converts geographic location data into structured tags and binds them to the standardized learning data, including the following steps: S14.

1. Use the GeoHash algorithm to encode the original latitude and longitude coordinates in the authorized regional coordinate data and generate string labels, which correspond one-to-one with a specific geographical area. S14.2 Based on the preset location type classification model, the original location type description in the geographic location data is mapped into structured category labels; S14.

3. Using the user's unique identifier and the data collection timestamp as a joint index, the distributed association algorithm is used to associate the regional coordinate labels generated in S14.1, the location type labels generated in S14.2, and the standard chemical information data output by the feature standardization module (13) to form a standard chemical information dataset containing complete geographic labels. The multi-dimensional learning analysis unit (2) extracts knowledge mastery quantification, learning behavior characteristics, and learning preference characteristics based on the standard learning dataset, and generates a multi-dimensional learning assessment result that integrates geographical location-related characteristics. The AI ​​large model course recommendation unit (3) adopts a pre-trained large model that has been fine-tuned by course texts, learning cases and geographical location-related samples in the field of education. It uses an improved geo-content fusion coding algorithm to associate and encode the structured features of course content, multi-dimensional quantitative features of learning situation and geographical location features. It calculates the dynamic matching degree based on cosine similarity and attention weight, selects highly suitable courses to form a personalized recommendation list, and iteratively optimizes the model parameters through gradient descent algorithm based on the real-time learning effect data generated during the user's learning of the personalized recommendation list. The adaptive learning path optimization unit (4) dynamically generates content sequence priority, difficulty adjustment coefficient and single knowledge point learning time through the enhanced learning path optimization model based on the real-time learning interaction data of the personalized recommendation list, and feeds back the learning effect data to the AI ​​large model course recommendation unit (3) to support the parameter iteration of the AI ​​large model course recommendation unit (3).

2. The AI ​​large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis according to claim 1, characterized in that, The learning data processing unit (1) further includes a data acquisition module (11), a data cleaning module (12), and a feature standardization module (13) connected in sequence, wherein: The data acquisition module (11) collects structured data, semi-structured data, unstructured data and authorized geographic location data through the database interface of the teaching affairs system, the log interface of the online learning platform, the multimedia interface of the classroom interaction tool and the location service authorization interface, respectively. The data cleaning module (12) receives the raw data and uses an outlier detection algorithm based on statistical distribution to identify and remove data that exceeds the reasonable range; The feature standardization module (13) performs scale unification processing on the cleaned data, converting feature values ​​of different dimensions into standardized features that can be directly calculated.

3. The AI ​​large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis according to claim 2, characterized in that, The multi-dimensional learning analysis unit (2) includes a knowledge feature extraction module (21), a behavior feature extraction module (22), a preference feature extraction module (23), a geographic feature fusion module (24), and an evaluation result generation module (25), which are connected in sequence. The knowledge feature extraction module (21) calculates the knowledge mastery quantification value based on the answer records and knowledge point application records in the standard chemistry dataset and through statistical learning algorithms. The behavioral feature extraction module (22) extracts learning behavioral features based on the learning duration distribution and operation sequence logs in the standard learning dataset, using a time-series feature extraction algorithm. The preference feature extraction module (23) extracts learning preference features based on resource type access records and content dwell time data in the standard learning dataset through frequency statistics and proportion analysis. The geographic feature fusion module (24) receives the geographic location structured labels in the standard chemical intelligence dataset and uses a feature splicing algorithm to associate and fuse the regional coordinate labels, location type labels, and knowledge, behavior, and preference features. The evaluation result generation module (25) performs structured integration of the fused multi-dimensional features to generate a multi-dimensional learning assessment result that includes quantitative values ​​of knowledge mastery, learning behavior features, learning preference features, and geographical location-related features.

4. The AI ​​large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis according to claim 3, characterized in that, The AI ​​large model course recommendation unit (3) includes a model fine-tuning module (31), a feature encoding module (32), a matching degree calculation module (33), and a parameter optimization module (34) that are connected in sequence; wherein: The model fine-tuning module (31) receives course texts, learning cases and geographical location-related samples in the field of education, and performs targeted fine-tuning on the pre-trained large model through the process of data preprocessing-parameter adjustment-loss optimization; The feature encoding module (32) adopts an improved geo-content fusion encoding algorithm and encodes course content, learning situation and geographical location features according to the process of "single-modal feature extraction - cross-modal association encoding"; The matching degree calculation module (33) calculates the dynamic matching degree based on feature vector similarity and attention weight, and generates a personalized recommendation list by filtering according to the threshold. The parameter optimization module (34) receives real-time learning effect data and iteratively optimizes the model parameters through a process of loss function construction, gradient calculation, and parameter update.

5. The AI ​​large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis according to claim 4, characterized in that, The model fine-tuning module (31) performs the following steps in fine-tuning the pre-trained large model: S31.1 Perform word segmentation on course texts in the education field to generate a word index sequence. , Represents a structured index sequence of course texts, containing Word index, The number of words in the text; feature vector transformation is performed on the quantitative values ​​of knowledge mastery and learning behavior characteristics in the learning situation cases to obtain the learning situation feature vector. ,Include Each feature component The number of dimensions for learning characteristics; numerical encoding of regional coordinate labels and location type labels in geographic location-related samples to obtain geographic feature vectors. ,Include Each feature component The number of dimensions of geographical features; S31.

2. Adopting the "freeze-fine-tuning" strategy, let the total number of layers in the pre-trained large model be... Freeze 80% of the bottom layer Parameters, only for the top 20% Adjust the parameters; S31.3, using cross-entropy loss To optimize the objective, we use mini-batch gradient descent with a batch size set to... , This indicates the number of samples used for each model parameter update. The pre-trained large model is trained and iterated until... Stop fine-tuning when convergence occurs.

6. The AI ​​large-scale model course recommendation and adaptive learning path optimization system based on multi-dimensional learning analysis according to claim 5, characterized in that, The implementation of the improved geo-content fusion coding algorithm of the feature coding module (32) includes the following steps: S32.1 Extracting visual feature vectors from video frames using CNN , dimension Extracting semantic feature vectors from text content using BERT , dimension Extracting logically related feature vectors from interactive exercises using graph neural networks. , dimension ;Will , , splicing as a joint feature of the course ; S32.

2. The knowledge mastery quantification, learning behavior characteristics, and learning preference characteristics output by the multi-dimensional learning analysis unit (2) are converted into standard learning feature vectors through a fully connected layer. , dimension ; S32.

3. Convert the area coordinate labels and location type labels output by the label association module (14) into standardized geographic feature vectors through the embedding layer. , dimension ; S32.4, Employing a multi-head attention mechanism Integrating the characteristics of curriculum, student learning, and geography, the output is presented in S32.

2. As a query vector, output in S32.1 With the output of S32.3 The concatenated result is used as the key vector and value vector, and the number of attention heads is set to [value]. ; S32.5, will The outputs of each attention head are concatenated and then transformed by a linear layer to obtain a fused feature vector. , dimension .

7. The AI ​​large-scale model course recommendation and adaptive learning path optimization system for multi-dimensional learning analysis according to claim 6, characterized in that, The matching degree calculation module (33) calculates the dynamic matching degree and generates a personalized recommendation list, including the following steps: S33.1, Based on the characteristics of curriculum collaboration With standard chemical eigenvectors Calculate the cosine similarity to obtain the course-student similarity. ; S33.2, Multi-head attention mechanism based on fusion of trimodal features The attention weights of course features and geographical features are extracted, and the average of these weights is taken to obtain the course-geographic attention weight. ; S33.3, will and The dynamic matching degree is obtained through weighted fusion. ; S33.4, Filtering Greater than the threshold The courses, according to Take first in descending order This creates a personalized recommendation list.

8. The AI ​​large-scale model course recommendation and adaptive learning path optimization system for multi-dimensional learning analysis according to claim 7, characterized in that, The iterative optimization of model parameters in the parameter optimization module (34) includes the following steps: S34.1, Combine real-time learning performance data corresponding to the personalized recommendation list, including knowledge mastery improvement rate. With course completion rate The optimized loss is obtained through weighted calculation. ; S34.2, Using the Adam optimizer Update model parameters and set the initial learning rate. ,Every Round iteration according to decay coefficient Adjustment ; S34.3, continued iteration to Less than the stopping threshold When this happens, stop optimizing the model parameters.

9. The AI ​​large-scale model course recommendation and adaptive learning path optimization system for multi-dimensional learning analysis according to claim 8, characterized in that, The adaptive learning path optimization unit (4) includes a real-time interactive data acquisition module (41), a reinforcement learning model processing module (42), a path parameter generation module (43), and a feedback data output module (44), which are connected in sequence. The real-time interactive data acquisition module (41) receives the personalized recommendation list output by the AI ​​large model course recommendation unit (3) and synchronously collects the real-time interactive data generated when the user learns the personalized recommendation list, including the knowledge point learning progress, the accuracy rate of answering practice questions, the duration of a single learning interruption and the knowledge point test score. The reinforcement learning model processing module (42) calls the reinforcement learning path optimization model, takes real-time interactive data as input, and uses the learning efficiency improvement value as the reward function for model inference; The path parameter generation module (43) dynamically generates three types of path optimization parameters—content sequence priority, difficulty adjustment coefficient, and single knowledge point learning time—based on the output of the reinforcement learning path optimization model. The feedback data output module (44) integrates the path optimization parameters and the corresponding learning effect data into a standardized feedback vector and pushes it to the parameter optimization module (34) to support iterative calculation of model parameters.

Images