Teaching data adaptive recommendation system

Abstract

The invention relates to the field of artificial intelligence, and discloses a teaching data adaptive recommendation system, which comprises a data layer, a recommendation engine layer and an application layer, the data layer comprises a multi-module education resource library and a user behavior log module; the recommendation engine layer comprises a feature extraction module, a collaborative filtering module and a mixed recommendation engine; the application layer comprises a personalized recommendation interface and a dynamic feedback module; the system provided by the invention can realize teaching data adaptive recommendation.

Description

Technical Field

[0001] The present invention relates to the field of teaching data adaptive recommendation, and in particular to a teaching data adaptive recommendation system. Background Art

[0002] There are disparities in cognitive patterns and knowledge density between K12 and higher education, but existing systems use a unified recommendation strategy, resulting in the coexistence of "oversimplification of elementary school teaching data" and "excessive university content." For example, when recommending college preparatory courses to high school students, the system is unable to perceive the differences in their subject foundations, resulting in a mismatch between the difficulty of the teaching data and the learner's ability. This lack of awareness of the academic stage reduces the accuracy of the recommendation system in mixed-stage scenarios. Furthermore, traditional recommendation systems have a significant response lag. Educational scenarios have a short cycle of demand changes. Scenarios such as pre-exam sprints and competition preparations require immediate adjustment of recommendation strategies, but traditional models rely on batch training, with update cycles of hours or even days. This lag is particularly fatal in sudden educational scenarios such as epidemics, resulting in the system's inability to adapt in a timely manner to dynamic changes such as "surges in demand for online teaching" or "experimental courses moving online." Summary of the Invention

[0003] The purpose of the present invention is to provide a teaching data adaptive recommendation system, aiming to solve a teaching data adaptive recommendation system.

[0004] The present invention provides a teaching data adaptive recommendation system, comprising: Teaching data adaptive recommendation system, including, Data layer, recommendation engine layer and application layer; The data layer includes: a multi-module educational resource library and a user behavior log module; The recommendation engine layer includes: feature extraction module, collaborative filtering module and hybrid recommendation engine; The application layer includes: personalized recommendation interface and dynamic feedback module; The multi-module educational resource library is used to store teaching data, and the user behavior log module is used to record user behavior records; The feature extraction module is used to extract teaching features from the teaching data and send them to the hybrid recommendation engine and collaborative filtering module, wherein the teaching features include: image features and text features; The collaborative filtering module is used to obtain teaching data and user behavior records, build user profiles based on the teaching data and user behavior records, obtain user preference vectors based on the user profiles, and input the preference vectors into the hybrid recommendation engine; The hybrid recommendation engine is used to generate a recommendation list based on the preference vector to find the recommended content in the teaching features; The personalized recommendation interface obtains the recommendation list and displays it to the user; The dynamic feedback module is used to collect user operation information on the recommendation list in real time and update user behavior records.

[0005] By adopting the embodiment of the present invention, dynamic and agile recommendation of teaching data can be achieved.

[0006] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it is implemented in accordance with the contents of the specification, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0008] Figure 1 4 is a schematic diagram of a teaching data adaptive recommendation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0010] Example According to an embodiment of the present invention, a teaching data adaptive recommendation system is provided. Figure 1 Schematic diagram of a teaching data adaptive recommendation system according to an embodiment of the present invention. Figure 1 As shown, specifically including: Data layer, recommendation engine layer and application layer; The data layer includes: a multi-module educational resource library and a user behavior log module; The recommendation engine layer includes: feature extraction module, collaborative filtering module and hybrid recommendation engine; The application layer includes: personalized recommendation interface and dynamic feedback module; The multi-module educational resource library is used to store teaching data, and the user behavior log module is used to record user behavior records; The teaching data specifically includes: image data, video data and text data.

[0011] Image teaching data: stores formulas, experimental schematics, anatomical diagrams, etc., supports PNG and JPEG formats, and adds metadata (such as the knowledge point label "quadratic function image" and the difficulty coefficient "0.8").

[0012] Text teaching data: Contains lesson plans, test analysis, and academic literature, stored in PDF / DOCX format, and the text content is extracted for processing by models such as BERT.

[0013] Video teaching data: Integrates micro-class clips and experimental demonstrations in MP4 format, and segments key frames (such as "Ohm's Law Experiment, 3rd Minute") to support content retrieval.

[0014] The feature extraction module is used to extract teaching features from the teaching data and send them to the hybrid recommendation engine and collaborative filtering module, wherein the teaching features include image features and text features; the image features are resized and normalized; The functions of the feature extraction module in the embodiment of the present invention are as follows: Image preprocessing: Standardize (adjust the resolution to 224×224 pixels and convert the color space to grayscale) and enhance (randomly crop, rotate, and adjust contrast) image-based educational data.

[0015] Feature extraction: The feature extraction module model is used to extract high-dimensional semantic features of the image, such as the vertex coordinates and opening direction of the "quadratic function image" and the chromosome arrangement and cell membrane morphology changes of the "cell division diagram".

[0016] Feature output: The extracted feature vector is transmitted to the hybrid recommendation engine and saved to the feature library for subsequent analysis.

[0017] The collaborative filtering module is used to obtain teaching data and user behavior records, build user profiles based on the teaching data and user behavior records, obtain user preference vectors based on the user profiles, and input the preference vectors into the hybrid recommendation engine; The collaborative filtering module is specifically used to: collect users' basic attributes and behavior data based on user behavior records, build user portraits based on basic attributes, behavior data and teaching data, and calculate the teaching data similarity matrix based on user behavior data and knowledge point labels and difficulty levels in the teaching data.

[0018] The collaborative filtering module is specifically used to establish content-aware weights and user preference weights based on user portraits, and fuse the content-aware weights and user preference weights using a gated linear unit to obtain a user preference vector. The formula is as follows: =σ( )⊙( ); in, represents multiplication, σ is the Sigmoid activation function, is the learnable gating parameter, is the preference vector, is a content-aware weight matrix generated by the self-attention mechanism, which reflects the attention paid to the objective content in the teaching data. The objective content includes: the blackboard writing and experimental demonstration in the teaching video, is the user preference weight matrix, which reflects the user's personalized operations on the teaching data. The personalized operations include: user history clicks and collections; for and The vector concatenation operation concatenates the content-aware weight and the preference-aware weight in the feature dimension to form a dual-channel input; ⊙ is the Hadamard product, which is an element-wise multiplication that multiplies the gating coefficient by the weighted sum of the splicing weights to achieve selective feature enhancement.

[0019] In an embodiment of the present invention, the user's preference vector can be precise to a certain type of teaching data in a certain subject, or it can abstract the logic and thinking methods in the teaching data and recommend teaching data with the same thinking mode adopted in multiple subjects to the user together to avoid the monotony of the questions.

[0020] The preference vector also involves the customer's habit of doing exercises or watching videos. For example, some customers are accustomed to doing 10 physics problems, then 10 math problems, and then other subjects such as Chinese and English.

[0021] Objective content also includes sentiment and knowledge point feedback in user reviews. Page tracking analyzes user behavior patterns, such as skipping advanced resources three times in a row, to infer potential preferences.

[0022] The collaborative filtering module is further configured to: Calculate the similarity between teaching data and send it to the hybrid recommendation engine, calculate the user-user similarity matrix, and divide the groups according to the user-user similarity matrix; Calculate similarity between users (such as cosine similarity) to find similar user groups. For example, the similarity between user E and user F is 0.92 because both frequently access "electromagnetism" resources.

[0023] The hybrid recommendation engine is also used to recommend similar teaching data suitable for the group based on the similarity between the teaching data.

[0024] In an embodiment of the present invention, the collaborative filtering module is further configured to calculate a similarity matrix between teaching data and teaching data based on common user groups or content tags. For example, the similarity between "Faraday's law" and "Lenz's law" is 0.88 because both are frequently accessed by the same user.

[0025] The collaborative filtering module is specifically used to calculate the user-user similarity matrix using the cosine similarity formula, as follows: Where λ is the time decay coefficient, Create a timestamp for user behavior. Sim(i, j) is the similarity score between user behavior i and user behavior j, ranging from [−1, 1]. A larger value indicates a stronger similarity. u∈U is a user u in the user set U that has interacted with both user behavior i and user behavior j. Score the interaction intensity of user u on teaching data i, such as the normalized value of click count and browsing time, Score the interaction intensity of user u on teaching data j, The average rating of user u on all interactive teaching data is used for centralized processing.

[0026] The hybrid recommendation engine is used to generate a recommendation list based on the preference vector to find the recommended content in the teaching features; The personalized recommendation interface obtains the recommendation list and displays it to the user; In an embodiment of the present invention, the recommended content is returned to the front-end interface through an API, supporting multiple forms of display, such as lists, cards, and carousels, to meet the browsing needs of users in different scenarios.

[0027] The dynamic feedback module is used to collect user operation information on the recommendation list in real time and update user behavior records.

[0028] The multi-module educational resource library is further used to: acquire video data, segment the video data to obtain video data frames, and extract features from the video data frames to obtain image features.

[0029] In the multi-module education and teaching database, the video segmentation and representative frame extraction process based on the shot change detection algorithm (K-Means clustering) is as follows: First, a shot change detection algorithm is used to identify shot boundaries in the video. A deep learning model (such as ResNet50 or CSPdarknet53) is used to classify video frames. The current frame is concatenated with the previous three frames in the channel dimension to form a 12-channel input. The classification model then determines whether a shot cut has occurred. This method incorporates a sliding window mechanism, using a higher threshold for sudden shot changes and a lower threshold for gradual shot changes. Interference factors such as flash are eliminated by using the inter-frame difference ratio to ensure accurate shot boundary detection.

[0030] After shot segmentation, the representative frame extraction phase begins. The K-Means clustering algorithm is used to cluster the video frames within a shot. First, high-dimensional visual features are extracted from the video frames using a pretrained model (such as InceptionV3), or a 72-dimensional feature vector is generated using a color space transformation (such as HSV) combined with non-uniform block weighting. Subsequently, a hierarchical clustering algorithm is used to perform preliminary clustering of the video frames, generating initial clustering results. Based on this initial clustering result, the K-Means algorithm is used to optimize the initial clustering results. Cluster centers are iteratively calculated to divide the video frames into several clusters. Finally, the frame closest to the cluster center in each cluster is selected as the representative frame for that shot.

[0031] This process uses a shot change detection algorithm to achieve accurate video segmentation, and then uses the K-Means clustering algorithm to extract representative frames, effectively reducing redundant data and providing high-quality data input for subsequent video analysis and processing.

[0032] In the embodiment of the present invention, the cooperation of each module is as follows: The user behavior log system is the core data collection component of the AI ​​agile teaching goal adaptive recommendation system. It is responsible for fully recording user learning behavior and building behavior profiles, providing real-time feedback and long-term behavior analysis support for the recommendation engine.

[0033] The following describes its workflow in detail from four dimensions: (1) Data input Explicit operation records: capture user clicks, browsing time, collection / sharing and other direct interaction behaviors.

[0034] Implicit feedback mining: Analyze user behavior patterns (such as "skipping advanced teaching data three times in a row") through page embedding, and infer potential preferences (such as "currently need to consolidate the basics").

[0035] Interactive content analysis: Extract sentiment and knowledge point feedback from user comments and generate semantic tags (such as "too fast" and "needs step-by-step demonstration").

[0036] (2) Context Enhancement Processing Environmental annotation: Attach metadata such as timestamp, device type, and learning scenario to each behavior to build a scenario-based behavior sequence.

[0037] Data cleaning: Filter out invalid operations and normalize data (e.g., map “30-second browsing” to a 0.2 attention score) to ensure data quality.

[0038] (3) Data output The cleaned behavior sequence is fed to the collaborative filtering module. The dynamic feedback module is fed with real-time behavior streams (e.g., "User B just completed exercise C") to trigger immediate recommendation adjustments.

[0039] In the embodiment of the present invention, the model needs to be trained; Training the hybrid model is a multi-stage, dynamic optimization process that requires three core components: the collaborative feature extraction module, the collaborative filtering module, and the hybrid recommendation engine. The following is a detailed training plan: 1. Data Preparation and Preprocessing Multimodal educational data processing: Video, image, and text data are collected. Video data is segmented into keyframes using shot change detection algorithms (such as K-Means clustering) to reduce redundancy. Image data is resized to a uniform size (e.g., 224×224 pixels) and normalized.

[0040] User behavior log construction: Record user interactions such as clicks, browsing time, and favorites / shares to construct a user-teaching data interaction matrix. This log data is cleaned to remove noise and perform feature enhancement, such as adding a time decay factor to reflect the timeliness of user interests.

[0041] User behavior log privacy protection requires the following: 1) collection layer desensitization (hashing user ID / IP and filtering sensitive fields); 2) transport layer encryption (HTTPS+TLS channel); 3) storage layer shard encryption (AES-256+dynamic key rotation); 4) computing layer differential privacy (adding Gaussian noise to clickstreams); 5) federated learning support (local model training without uploading original logs); and 6) compliance auditing (GDPR log retention policy + two-factor authentication for access rights).

[0042] 2. Feature extraction module training Architecture Design: A feature extraction module (ViT) architecture is used to segment the image into fixed-size patches (e.g., 16×16 pixels) and convert them into serialized token embeddings through a linear projection layer. Positional encoding is added to preserve spatial information, and classification labels ([CLS]) are used as the overall representation of the image.

[0043] Pretraining and fine-tuning: Pretraining is performed on large-scale datasets (such as ImageNet-21k or JFT-300M) to learn general visual feature representations. Fine-tuning is performed on target educational datasets to optimize the model for specific tasks (such as educational data classification and object detection). For example, the pretraining dataset includes OpenImages Education Edition (including courseware screenshots and video frames) combined with EdNet (user course click and exercise behavior logs) to align visual features with learning behavior. The fine-tuning dataset includes a K-12 mathematics micro-course dataset (teaching videos and knowledge point labels) overlaid with ASSISTments (student answer records) to enhance subject adaptation. The dynamic feedback module can integrate real-time interaction logs from MOOC platforms to optimize recommendations.

[0044] Loss function and optimizer: The cross entropy loss function is used for classification tasks, and the optimizer is AdamW, combined with weight decay and gradient clipping to stabilize the training process.

[0045] 3. Collaborative filtering module training Data representation: Use the user-teaching data interaction matrix, where the elements represent the user's interaction intensity with the teaching data (such as the number of clicks, browsing time, etc.).

[0046] Similarity Calculation: Cosine similarity or modified cosine similarity (taking into account the time decay factor) is used to calculate the similarity between users or teaching data. For user-user collaborative filtering, the K users most similar to the target user are found; for teaching data-teaching data collaborative filtering, the K teaching data most similar to the target teaching data are found.

[0047] Model training: Use matrix factorization (e.g., SVD, ALS) or deep learning methods (e.g., neural collaborative filtering) to learn the latent feature vectors of user and teaching data. Optimize model parameters by minimizing the mean squared error (MSE) or cross-entropy loss between predicted and actual ratings.

[0048] 4. Hybrid Recommendation Engine Training Feature fusion: The visual feature vector generated by the feature extraction module and the user preference feature vector generated by the collaborative filtering module are combined. This fusion is performed using weighted fusion, an attention mechanism, or a deep learning model (such as a multi-layer perceptron) to generate a comprehensive recommendation list.

[0049] Dynamic weight adjustment: A dynamic feedback mechanism is introduced to adjust the fusion weights based on user feedback on recommendation results (such as click-through rate, browsing time, and ratings). Initial weights are determined. For example, if the accuracy of the collaborative feature extraction module on the validation set is 0.7 and the collaborative filtering accuracy is 0.6, the initial weights can be proportionally distributed, such as 0.7 / (0.7 + 0.6) and 0.6 / (0.7 + 0.6), and then normalized. Reinforcement learning or online learning algorithms (such as PID controllers and Bayesian optimization) are used to update the weight parameters in real time to optimize the recommendation results.

[0050] Multi-objective optimization: Consider multiple objectives (such as accuracy, diversity, novelty, etc.) in the recommendation process and balance the conflicts between different objectives through methods such as multi-task learning or Pareto optimization.

[0051] The beneficial effects of the present invention are as follows: Recommendation accuracy is significantly improved: The feature extraction module's ability to deeply analyze image-based educational data, combined with the collaborative filtering module's precise capture of user behavior, enables the system to generate a more comprehensive representation of user and teaching data.

[0052] User experience is greatly enhanced: The personalized recommendation interface provides a user-friendly interactive interface, supports real-time feedback and customized recommendation lists, allowing users to find the required teaching data more easily.

[0053] Optimized utilization of educational data: The intelligent scheduling of the hybrid recommendation engine has increased the exposure of less popular but high-quality teaching data by 50%, promoting the balanced utilization of teaching data. Through the rich tagging system and dynamic update mechanism of the multi-module educational database, the system can continuously introduce new teaching data, maintaining the freshness and diversity of recommended content.

[0054] Enhanced system adaptability: A closed-loop optimization mechanism formed by a dynamic feedback module enables the system to continuously learn from changes in user behavior and teaching data characteristics, automatically adjusting recommendation strategies. The modular layered architecture design decouples system components, facilitating independent upgrades and maintenance, improving system scalability and maintainability.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. These modifications or replacements of the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of this solution.

Claims

1. A teaching data adaptive recommendation system, characterized by: include, Data layer, recommendation engine layer and application layer; The data layer includes: a multi-module educational resource library and a user behavior log module; The recommendation engine layer includes: feature extraction module, collaborative filtering module and hybrid recommendation engine; The application layer includes: personalized recommendation interface and dynamic feedback module; The multi-module educational resource library is used to store teaching data, and the user behavior log module is used to record user behavior records; The feature extraction module is used to extract teaching features from the teaching data and send them to the hybrid recommendation engine, wherein the teaching features include: image features and text features; The collaborative filtering module is used to obtain teaching data and user behavior records, build user profiles based on the teaching data and user behavior records, obtain user preference vectors based on the user profiles, and input the preference vectors into the hybrid recommendation engine; The hybrid recommendation engine is used to generate a recommendation list based on user preference vectors to find recommended content in teaching features; The personalized recommendation interface obtains the recommendation list and displays it to the user; The dynamic feedback module is used to collect user operation information on the recommendation list in real time and update user behavior records.

2. The system according to claim 1, wherein: The teaching data specifically includes: image data, video data and text data.

3. The system according to claim 1, wherein: The collaborative filtering module is specifically used to establish content-aware weights and user preference weights based on user portraits, and fuse the content-aware weights and user preference weights using a gated linear unit to obtain a user preference vector. The formula is as follows: =σ( )⊙( ); in, represents multiplication, σ is the Sigmoid activation function, is the learnable gating parameter, is the preference vector, is a content-aware weight matrix generated by the self-attention mechanism, which reflects the attention paid to the objective content in the teaching data. The objective content includes: the blackboard writing and experimental demonstration in the teaching video, is the user preference weight matrix, which reflects the user's personalized operations on the teaching data. The personalized operations include: user history clicks and collections; for and The vector splicing operation splices the content-aware weight and the preference-aware weight in the feature dimension to form a dual-channel input; ⊙ is the Hadamard product, that is, element-wise multiplication, which multiplies the gating coefficient and the weighted sum of the splicing weights to achieve selective feature enhancement.

4. The system according to claim 3, characterized in that The collaborative filtering module is further configured to: Calculate the user-user similarity matrix to obtain users with high similarity, and divide the users with high similarity into a group; The hybrid recommendation engine is also used to recommend advanced goals suitable for the group based on the similarity between the teaching data.

5. The method according to claim 4, characterized in that The collaborative filtering module is specifically used to calculate the user-user similarity matrix using the cosine similarity formula, as follows: Where λ is the time decay coefficient, Create a timestamp for user behavior. Sim(i, j) is the similarity score between user behavior i and user behavior j, with a value range of [−1, 1]. A larger value indicates a stronger similarity. u∈U is a user u in the user set U that has interacted with both user behavior i and user behavior j. Score the interaction intensity of user u on teaching data i, such as the normalized value of click count and browsing time, Score the interaction intensity of user u on teaching data j, The average rating of user u on all interactive teaching data is used for centralized processing.

6. The system according to claim 1, wherein: The multi-module educational resource library is further used to unify the size of image features and perform normalization processing.

7. The system according to claim 1, wherein: The multi-module educational resource library is further used to: acquire video data, segment the video data to obtain video data frames, and extract features from the video data frames to obtain image features.

Images