A Personalized Learning Recommendation Method for E-commerce

Abstract

This invention relates to a personalized learning recommendation method for e-commerce, belonging to the field of information recommendation technology. The method integrates user basic attributes and historical behavior data, deeply integrating multi-dimensional user profiles through adaptive weighting; it utilizes the association mining between products and knowledge points, establishing a product knowledge graph based on graph neural networks to achieve product topic clustering and multi-level knowledge point summarization; based on user profiles and product knowledge associations, it leverages deep reinforcement learning to dynamically plan personalized learning paths, intelligently recommending optimal resource sequences according to the user's current interests and knowledge level. The system collects user feedback information in real time and completes real-time adaptive adjustment of the recommendation strategy through local model updates and federated learning, achieving highly accurate and privacy-preserving personalized learning resource recommendations under interest shifts and behavioral changes.

Description

Technical Field

[0001] This invention belongs to the field of information recommendation technology, and more specifically relates to a personalized learning recommendation method for e-commerce. Background Technology

[0002] With the rapid development of online education, online learning platforms have accumulated a vast amount of learning resources and a diverse user base. Traditional recommendation systems, primarily based on collaborative filtering, content matching, or superficial user feature analysis, struggle to consider individual knowledge structures, interests, and dynamic learning needs. This results in highly personalized recommendations that fail to provide effective individualized guidance and systematic improvement of knowledge. Furthermore, existing methods generally lack in-depth modeling of learning path structures, failing to ensure the gradual progression and coherent integrity of learning content, and exacerbating issues such as fragmented learning information and content jumps.

[0003] On the other hand, with increasing awareness of user privacy and data security regulations, large-scale recommendation model training that relies solely on centralized data faces numerous limitations. The centralized data collection and model optimization of traditional recommendation systems struggle to respond in real-time to changes in individual learning states and behaviors, and also pose security risks such as data leaks. Furthermore, users' interests and learning behaviors are constantly evolving, and existing recommendation algorithms often fail to flexibly and adaptively adjust their strategies, leading to deviations between resource recommendations and users' actual needs.

[0004] Therefore, there is an urgent need for a comprehensive learning resource recommendation method that integrates technologies such as knowledge graphs, deep reinforcement learning and federated learning, and online transfer learning, while taking into account personalization, structure, dynamic adaptation, and user data privacy and security, in order to better meet the continuous growth and intelligent development needs of different users. This places higher technical demands on the algorithmic innovation and practical application of intelligent education recommendation systems. Summary of the Invention

[0005] This invention aims to address the technical problems of existing online learning resource recommendation methods, such as insufficient personalization, incomplete learning path structure modeling, poor dynamic adaptability to user interests and behaviors, and weak data privacy protection, in order to achieve intelligent, structured, dynamically adaptive, and data-secure personalized learning resource recommendation.

[0006] To achieve the above objectives, the present invention employs the following technical solution; the method includes: By integrating user personality attributes and behavior into a model, a multi-granular user profile is created by combining basic user attributes and historical behavior data. Knowledge point and product association mining establishes a network of associations between products and knowledge points. Nodes represent products, and edges represent the similarity relationships of knowledge points between products. A product theme clustering algorithm is introduced to automatically discover knowledge point categories and classify products into different knowledge themes. Personalized learning path planning employs an enhanced deep reinforcement learning algorithm to intelligently plan users' personalized learning paths, dynamically generating the optimal learning path based on the user's current knowledge level, interests, and the order in which products are associated. Real-time feedback and adaptive adjustment of recommendation strategies: Feedback information is collected in real time during the user's learning and consumption process, and an adaptive recommendation strategy adjustment mechanism is introduced. When a shift in user interests or a change in learning pace is detected, the recommendation path and resources are adjusted in a timely manner.

[0007] In one approach, the user personality attributes and behavior fusion modeling includes: systematically collecting and processing user basic attributes and historical behavior data. Basic attributes include static characteristics such as age, gender, and occupation, while behavior data includes dynamic interaction data such as user browsing, purchasing, rating, and collection. The user's attribute feature vector and behavioral feature vector are deeply integrated through an adaptive weight fusion algorithm. The adaptive weight is dynamically optimized through a weight learning algorithm, and an attention mechanism can be used to assign dynamic weights to each sub-dimension of the attribute features and behavioral features.

[0008] In one approach, the knowledge point and product association mining includes: extracting knowledge attributes from the product database of an e-commerce platform, generating knowledge feature vectors for the products, treating each product as a node in a graph neural network, establishing edges between product nodes based on knowledge point similarity, and obtaining a product knowledge graph structure. A graph neural network algorithm is used to aggregate and optimize information embedding in the product knowledge graph to obtain the embedding vector of each product node. Based on node embedding, a product topic clustering algorithm is used to cluster all products and identify knowledge point topic clusters. The topic clusters can be dynamically updated to realize the dynamic reflection of multi-level relationships between products and knowledge points.

[0009] In one approach, the personalized learning path planning includes: modeling the learning path planning as a sequential decision-making process based on user profiles and product knowledge association networks; the environmental state includes the user's current knowledge level, interests and preferences, the set of products already learned, and candidate product theme clusters; and the action space is to recommend specific products at the current moment.

[0010] In one approach, the adaptive adjustment of the real-time feedback and recommendation strategy includes: continuously collecting real-time feedback information from users during the learning and product consumption process, including learning progress, satisfaction rating, exit behavior, and resource usage time, and encoding it into a dynamic feature stream to reflect changes in user interests, behavior, and learning pace; A mini recommendation model is trained or some parameters are updated locally on each user terminal. Optimization is carried out based on local feedback data, and the model parameters of all users are aggregated through a federated averaging algorithm to achieve global lossless updates without data leaving the terminal, thus ensuring user privacy and security.

[0011] In one approach, the personalized learning path planning includes: defining a reward function based on state-behavior modeling, combining user feedback and the overall coherence of the learning path, and using a deep reinforcement learning algorithm to optimize it through experience replay and gradient update; The algorithm combines sequence information embedding and multi-objective reward structure, and uses product association networks to achieve automatic identification and navigation of topic clusters and knowledge chains, dynamically generating and optimizing personalized learning resource recommendation sequences.

[0012] In one approach, the real-time feedback and recommendation strategy adaptive adjustment includes: combining an online transfer learning mechanism to introduce a personalized transfer layer for rapid adaptation and parameter fine-tuning when a significant change in user interests or learning behavior is detected, dynamically adjusting the recommendation path and resources to achieve personalized and high-precision learning resource recommendations.

[0013] Beneficial effects of this invention: 1. This invention clearly subdivides the interaction behavior between users and learning resources into categories such as entry, learning, marking, and submission, and uses these behavior types as entity nodes in the knowledge graph. It not only treats users and resources as nodes, but also incorporates actual behavioral data into the graph modeling, which significantly improves the ability to model user behavior in a fine-grained manner.

[0014] 2. This invention employs data augmentation strategies from both behavioral and resource perspectives. It utilizes TransR technology to obtain embeddings separately, and then introduces a contrastive learning loss function to compare the embedded learning resources, behavioral nodes, and knowledge points across views. This enhances multi-angle, cross-view information identification and feature representation. This approach effectively improves the discriminative power of the embedded representations. Compared to existing technologies such as traditional cosine similarity weighting, dimensionality reduction feature representation, and simple embedding comparison, it exhibits higher tolerance for abnormal behavior and better model generalization ability.

[0015] 3. Graph Neural Network and Attention Mechanism Joint Feature Extraction for Personalized State Modeling After Behavior and Resource Fusion: This invention employs a Graph Convolutional Neural Network (GCN) with an attention mechanism to extract multi-view features from the fused knowledge graph. The output is used as the user state, significantly improving the recommendation model's ability to understand complex behavioral sequences and resource attributes. Compared to K-means and XGBoost, LSTM / Transformer and knowledge tracing models, and single GCNs, this invention achieves deep fusion of multi-source, dynamic, and heterogeneous features.

[0016] 4. Enhanced Learning Mechanism: The reward function integrates knowledge-level and path-level feedback to improve dynamic adaptive recommendation capabilities. This invention employs a deep Q-network as the agent, combined with a comprehensive reward function (knowledge-level, path-level, with adjustable weights). This allows the recommendation strategy to focus on both the user's mastery of individual knowledge points and the optimization of the overall path structure, adapting to the user's learning state and actual performance. This composite reward mechanism, compared to single Q-network feedback, reinforcement learning applications, and hierarchical reward mechanisms, further enhances the personalization and dynamic adaptability of recommendations.

[0017] 5. This invention employs data augmentation methods such as probabilistic modification from a behavioral perspective and random masking and local swapping from a resource perspective, which significantly enhances the model's tolerance to abnormal behavior and data noise, and improves the recommendation accuracy for complex behavioral sequences in real-world scenarios. This is a comprehensive data augmentation strategy not seen in existing technologies.

[0018] In summary, this invention overcomes the limitations of existing technologies in behavior modeling, recommendation model generalization and robustness, and dynamic adaptive capabilities by focusing on fine-grained user behavior modeling, knowledge graph fusion enhancement, embedded representation contrastive learning, joint feature extraction, compound reward reinforcement learning, and multi-angle data augmentation. It demonstrates significant innovation and technological progress and has significant application value for personalized learning path recommendation in large-scale and complex scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A flowchart is provided to illustrate the association network that dynamically reflects the multi-level relationship between products and knowledge points in this invention. Detailed Implementation

[0020] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0021] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0022] like Figure 1 As shown, a personalized learning recommendation method for e-commerce includes the following steps: Step 1: Modeling the Fusion of User Personality Attributes and Behaviors First, a multi-granular user profile is built by innovatively integrating basic user attributes (such as age, gender, and occupation) with historical behavioral data (browsing, purchasing, rating, and saving). Unlike traditional simple feature concatenation methods, this step employs an adaptive weighted fusion algorithm. This algorithm dynamically adjusts the fusion weights based on the impact of attributes and behaviors on personalized recommendations in different scenarios, accurately reflecting the user's true interests and preferences. This provides a deeper and more multi-dimensional understanding of the user in the subsequent recommendation process.

[0023] First, a systematic collection and processing of user basic attributes and historical behavioral data is required. Basic attributes include static features such as age, gender, and occupation, while behavioral data covers dynamic interactions such as browsing, purchasing, rating, and saving. To achieve innovative personalized modeling, these two types of information need to be deeply integrated using an adaptive weighted fusion algorithm. Specifically, let the attribute feature vector of user u be... The behavioral feature vector is The fusion process employs the following adaptive weighting mechanism: Where αu and βu are the weights of the attribute and behavioral features, respectively, and satisfy the following conditions: These weights are not fixed but dynamically optimized using an innovative weight learning algorithm. For example, weights can be learned by minimizing the recommendation loss function: Where yu represents the user's actual preference label, and f(Fu) represents the prediction output by the recommendation model. The weights are determined by combining the scenario and historical performance. Gradient descent can be used for optimization, and even an attention mechanism can be introduced to assign dynamic weights to different feature sub-dimensions: Here, wi and vj represent the attention weights for the respective dimensions of attribute and behavioral features, which are continuously adjusted through self-supervised or semi-supervised learning. The resulting multi-granular profile not only reflects the user's static characteristics but also dynamically captures the user's real-time interests and preferences. This fusion not only enhances the expressive power of the profile but also provides a more suitable and in-depth user understanding foundation for downstream recommendation modules, thereby significantly improving the personalization accuracy of the recommendation system.

[0024] Step Two: Discovering the Relationship Between Knowledge Points and Products Based on the knowledge attributes of products on e-commerce platforms, an innovative Graph Neural Network (GNN) algorithm is employed to establish a network linking products with knowledge points. Nodes represent products, and edges represent the similarity relationships between knowledge points among products. Simultaneously, a product topic clustering algorithm is introduced to automatically discover knowledge point categories and group products into different knowledge topics. This breaks the traditional classification system, enabling more detailed and dynamic extraction of knowledge point and product associations, facilitating personalized recommendations of subsequent learning resources.

[0025] like Figure 2 As shown, a relational network is established that can dynamically reflect the multi-level relationships between products and knowledge points. First, knowledge attributes are extracted from the product database of the e-commerce platform, such as product descriptions, tags, reviews, and expert annotations, to generate knowledge feature vectors for the products. Here, i represents the product number. Each product is considered a node in a Graph Neural Network (GNN), and the edges between nodes are defined as the knowledge point similarity, such as using cosine similarity. If the similarity exceeds a set threshold, an edge is created between items i and j. This results in a graph structure. It reflects the connections between knowledge points of products.

[0026] Next, a graph neural network algorithm is used to aggregate and optimize the information embedding of the product knowledge graph. The embedding of each product node... Calculated from aggregated neighbor information: Wherein, AGG is the neighbor aggregation function (mean, attention weighted). and Here are the network parameters for the k-th layer, and σ is the activation function. This represents the neighbors of product i. In multi-level aggregation, the embedding vector of each node progressively integrates the breadth and depth of knowledge points.

[0027] After generating the knowledge graph structure, a product-topic clustering algorithm (K-means) is introduced based on node embedding. This is applied to all product embeddings. Clustering is performed to automatically identify knowledge point topic clusters. Once products are grouped into different thematic clusters, the relationships between products within a single cluster become closer, and these clusters can be dynamically updated, breaking away from the traditional static classification system.

[0028] Finally, the product-knowledge point graph and topic clustering results lay a solid data structure foundation for subsequent personalized learning recommendations. This mechanism not only brings continuous semantic representation of products and knowledge points and topic structure, but also enables recommendations to perform more effective resource matching and path planning based on knowledge chains and topic levels.

[0029] Step 3: Personalized Learning Path Planning Based on user profiles and product knowledge association networks, an enhanced deep reinforcement learning (DRL) algorithm is designed to intelligently plan personalized learning paths for users. For example, for multi-stage products (such as tutorials, reference books, and practice tools) needed to learn new skills, the model can dynamically generate the optimal learning path based on the user's current knowledge level, interests, and the order of product associations, achieving personalized resource recommendations from basic to advanced levels. This step effectively improves the coherence and educational value of the recommendation sequence.

[0030] Based on user profiles (integrating attribute and behavioral features) and product knowledge association networks, this study employs Deep Reinforcement Learning (DRL) algorithms to intelligently generate optimal learning resource recommendation sequences for each user. First, each learning path planning is viewed as a sequential decision-making process—environment state... It includes information such as the user's current knowledge level, interests, previously learned product collection, and candidate product theme clusters. Action Space This indicates recommending a specific product to the user at the current moment. Based on this state-behavior modeling, a reward function is defined. The system quantifies user feedback after product recommendations (such as learning progress, satisfaction, and product interaction behavior) and introduces a multi-dimensional reward mechanism, which includes both immediate user feedback and consideration of the overall coherence of the learning path and the completeness of the knowledge system.

[0031] The specific algorithm uses a deep Q-network (DQN) structure for modeling. Value function Approximated by deep neural networks, weights Optimization is achieved through experience replay and gradient updates. At each step, the agent selects an action based on the current state. Feedback and rewards after environment execution recommendation and transition to a new state. The network weights are updated according to the following loss function: in, For future reward discount factors, The target network parameters are defined as follows. To enhance the coherence and personalized adaptability of the learning path, sequence information embedding (RNN / LSTM) or multi-objective reward structures can be used to ensure that the final learning recommendation path gradually covers the knowledge points needed by the user, progressing from simple to complex. Simultaneously, by combining product association networks, the agent can automatically identify and jump between topic clusters and knowledge chains, effectively guiding users to learn step-by-step and avoiding content jumps and information fragmentation.

[0032] Ultimately, the model continuously optimizes its strategies through online interaction and real-time feedback to ensure that each user receives the most suitable and educationally valuable personalized learning resource path based on their own profile, knowledge structure, and dynamic interests, effectively improving the intelligent planning capabilities and growth guidance effect of the recommendation system.

[0033] Step 4: Real-time feedback and adaptive adjustment of recommendation strategies Finally, feedback information (such as learning progress, satisfaction ratings, and bounce rates) is collected in real time during the user's learning and consumption process, and an adaptive recommendation strategy adjustment mechanism is introduced. This mechanism utilizes innovative federated learning and online transfer learning algorithms to update the recommendation model without loss, dynamically optimizing the recommendation strategy while ensuring user data privacy. When a shift in user interests or a change in learning pace is detected, the system can promptly adjust the recommendation path and resources, fully realizing continuous personalization and high-precision recommendation efficiency.

[0034] Continuously collect real-time feedback information from users during the learning and product consumption process, such as learning progress, satisfaction ratings, bounce rate, and resource usage time. This feedback is encoded into a dynamic feature stream. This is used to reflect changes in user interests, behaviors, and learning pace. To achieve rapid and lossless adjustment of the recommendation strategy while protecting user data privacy, the system employs an innovative mechanism combining federated learning and online transfer learning algorithms.

[0035] First, a mini recommendation model is trained locally on each user terminal, or some parameters are updated. For example, let the local model parameters be... Based on local feedback data, updates are performed by optimizing the loss function: in It is user u's local dataset. It is a recommendation function. This is the loss function. Every so often, each user client uploads its modified model parameters (not the original data) to the server. The central server aggregates all local models using a federated averaging algorithm (FedAvg). This achieves global lossless updates to the model, ensuring that data remains on the source and protecting user privacy and security.

[0036] To further address the rapid changes in user interests and learning behaviors, the system introduces an online transfer learning mechanism. When a shift in user interests or a significant change in learning pace is detected (such as a significant drop in satisfaction, a surge in bounce rate, or a lag in learning progress), the algorithm automatically adjusts its strategy: introducing a personalized transfer layer (such as adaptive multi-head attention or meta-learning) to quickly adapt to new interest distributions or learning objectives. The transfer learning objective function can be expressed as: in Optimized based on the latest feedback data, D represents the distance between the parameters and the global model (L2 regularization). Controlling the intensity of migration. Through multiple rounds of migration and online fine-tuning, the recommendation system can adjust the recommendation path and resources for each user in real time, minimizing interest gaps, content mismatches, and information lags, and ensuring continuous personalized and high-precision recommendation results.

[0037] Ultimately, the combination of federated learning and online transfer learning not only enhances real-time adaptability and individual model optimization, but also ensures user data security throughout the entire process, providing flexible, dynamic, and efficient intelligent upgrade capabilities for large-scale personalized education recommendation systems.

[0038] Example: I. Implementation Background An e-commerce learning platform, targeting adult continuing education, sells various online courses and textbooks. The platform has a user base of over 100,000 and more daily active users (over 5,000). The platform urgently needs to improve its personalized recommendation capabilities to encourage continued user learning and increase user purchase rates and satisfaction.

[0039] II. Data System and Sample Description The structure of the platform's historical user data and product information is as follows: Table 1 User Basic Attributes Table 2 User Historical Behavior Data III. User Profile Fusion Modeling User profiles are formed by integrating user basic attributes and behavioral data into feature vectors and then weighting them (e.g., using adaptive attention mechanisms).

[0040] Table 3 User Feature Vectors Note: The fusion weight vectors correspond to the weights of attributes, browsing, purchase, and rating features, respectively.

[0041] IV. Knowledge Point and Product Association Mining First, knowledge attributes and their vector representations are extracted for each product. Then, a product knowledge graph is constructed based on knowledge similarity. Finally, a graph neural network is used to optimize the embedding to achieve topic clustering.

[0042] Table 4 Product Knowledge Characteristics Table 5. Similarity of Product Knowledge Points Table 6. Product Theme Clustering Results V. Personalized Learning Path Planning By combining user profiles and product knowledge graphs, the recommendation task is modeled as a sequential decision-making process, and deep reinforcement learning is used for policy optimization. Taking user U1001 as an example: Table 7 Personalized Learning Path Sequence for User U1001 The reward function combines satisfaction and thematic coherence, accumulating rewards to guide subsequent product recommendations.

[0043] VI. Real-time Feedback and Adaptive Strategy Adjustment The system continuously acquires user behavior data and dynamically adjusts its recommendation strategy. For example, after learning G105, user U1001 exhibits bounce behavior. The system detects the interest shift and uses federated learning to fine-tune the recommendation model locally.

[0044] Table 8 User Real-Time Feedback Stream Table 9 Local Model Update and Federated Aggregation Process The system uses local model optimization and federated averaging to quickly respond to changes in individual interests while protecting user privacy.

[0045] VII. Transfer Learning and High Dynamic Adaptability When user interests change significantly (such as career change or frequent theme changes), the platform will insert a personalized migration layer in real time to fine-tune the network weights, enabling rapid adaptation of personalized resource recommendation paths.

[0046] Table 10 Examples of Transfer Learning Applications and Recommendation Shifts VIII. Statistics on Actual Results and Effectiveness After launching this personalized learning recommendation system, platform data shows that key metrics have improved as follows: Table 11 Comparison of Recommendation Effectiveness among User Groups 1. Utilize adaptive weighting to fuse attributes and behaviors to create refined, multi-granular user profiles; 2. Graph Neural Networks (GNNs) are used to mine product knowledge connections and topic clustering to form a clear knowledge network map; 3. Learning path planning incorporates deep reinforcement learning, integrating interests, abilities, and product themes to achieve differentiated recommendations; 4. Real-time feedback streams link local model fine-tuning and federated learning, effectively protecting user privacy while greatly improving the system's dynamic adaptability; 5. The transfer learning mechanism ensures that when an individual's trajectory changes abruptly, the recommendation system can respond immediately, enabling continuous recommendations of learning products tailored to each individual.

[0047] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0048] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An e-commerce personalized learning recommendation method, characterized in that: The method includes: By integrating user personality attributes and behavior into a model, a multi-granular user profile is created by combining basic user attributes and historical behavior data. Knowledge point and product association mining establishes a network of associations between products and knowledge points. Nodes represent products, and edges represent the similarity relationships of knowledge points between products. A product theme clustering algorithm is introduced to automatically discover knowledge point categories and classify products into different knowledge themes. Personalized learning path planning employs an enhanced deep reinforcement learning algorithm to intelligently plan users' personalized learning paths, dynamically generating the optimal learning path based on the user's current knowledge level, interests, and the order in which products are associated. Real-time feedback and adaptive adjustment of recommendation strategies: Feedback information is collected in real time during the user's learning and consumption process, and an adaptive recommendation strategy adjustment mechanism is introduced. When a shift in user interests or a change in learning pace is detected, the recommendation path and resources are adjusted in a timely manner. 2.The method of claim 1, wherein: The aforementioned user personality attribute and behavior fusion modeling includes: systematically collecting and processing user basic attributes and historical behavior data. Basic attributes include static characteristics such as age, gender, and occupation, while behavioral data includes dynamic interaction data such as user browsing, purchasing, evaluation, and collection. The user's attribute feature vector and behavioral feature vector are deeply integrated through an adaptive weight fusion algorithm. The adaptive weight is dynamically optimized through a weight learning algorithm, and an attention mechanism can be used to assign dynamic weights to each sub-dimension of the attribute features and behavioral features. 3.The method of claim 1, wherein: The knowledge point and product association mining includes: extracting knowledge attributes from the product database of the e-commerce platform, generating knowledge feature vectors for the products, treating each product as a node in a graph neural network, establishing edges between product nodes based on knowledge point similarity, and obtaining a product knowledge graph structure. A graph neural network algorithm is used to aggregate and optimize information embedding in the product knowledge graph to obtain the embedding vector of each product node. Based on node embedding, a product topic clustering algorithm is used to cluster all products and identify knowledge point topic clusters. The topic clusters can be dynamically updated to realize the dynamic reflection of multi-level relationships between products and knowledge points.

4. The method of claim 1, wherein: The personalized learning path planning includes: based on user profiles and product knowledge association networks, modeling the learning path planning as a sequential decision-making process, with the environmental state including the user's current knowledge level, interests and preferences, the set of products already learned and candidate product theme clusters, and the action space being the recommendation of specific products at the current moment.

5. The method of claim 1, wherein: The aforementioned real-time feedback and recommendation strategy adaptive adjustment includes: continuously collecting real-time feedback information from users during the learning and product consumption process, including learning progress, satisfaction rating, exit behavior, and resource usage time, and encoding it into a dynamic feature stream to reflect changes in user interests, behavior, and learning pace; A mini recommendation model is trained or some parameters are updated locally on each user terminal. Optimization is carried out based on local feedback data, and the model parameters of all users are aggregated through a federated averaging algorithm to achieve global lossless updates without data leaving the terminal, thus ensuring user privacy and security.

6. The method of claim 4, wherein: The personalized learning path planning includes: defining a reward function based on state-behavior modeling, combining user feedback and the overall coherence of the learning path, and using a deep reinforcement learning algorithm to optimize it through experience replay and gradient update; The algorithm combines sequence information embedding and multi-objective reward structure, and uses product association networks to achieve automatic identification and navigation of topic clusters and knowledge chains, dynamically generating and optimizing personalized learning resource recommendation sequences.

7. The method of claim 5, wherein: The aforementioned real-time feedback and adaptive adjustment of the recommendation strategy include: combining an online transfer learning mechanism, when a significant change in user interests or learning behavior is detected, introducing a personalized transfer layer for rapid adaptation and parameter fine-tuning, dynamically adjusting the recommendation path and resources, and achieving personalized and high-precision learning resource recommendations.

Images