Neural network-based student personalized learning path generation method

Abstract

A neural network-based method for generating personalized learning paths for students involves deploying data acquisition modules on teaching platforms, learning terminals, and interactive interfaces to collect multi-dimensional data in real time, including student learning time, interaction frequency, and error records. A knowledge graph is constructed based on the curriculum syllabus and prior textbook structure, and a graph neural network (GNN) is used to dynamically adjust the association weights of its nodes and edges to reflect students' weak knowledge areas. A Bi-LSTM encoder is constructed from the collected data to generate behavioral feature vectors, which are then aligned and fused with the knowledge graph node embedding vectors through an attention mechanism to obtain a personalized multimodal representation of the student. Using this representation and the knowledge graph node weights as the environmental state, and "selecting the next knowledge point to learn" as the action space, a reward function is designed by integrating multiple indicators. A deep Q-network is used to update the objective and train a policy network to generate personalized learning paths. The trained policy is deployed as a microservice, displaying paths sorted by knowledge modules, and iteratively updating relevant parameters based on student feedback to improve recommendation effectiveness.

Description

Technical Field

[0001] This invention relates to the field of educational technology, and in particular to a method for generating personalized learning paths for students based on neural networks. Background Technology

[0002] In today's era of rapid development in digital education, online learning platforms and various learning terminals are widely used, providing students with a wealth of learning resources and flexible learning methods. However, with the massive increase in learning data and the significant differences among students, how to effectively utilize this data to tailor personalized learning paths for each student to improve learning outcomes and efficiency has become a critical issue that urgently needs to be addressed in the education field.

[0003] Traditional learning path recommendation methods are often based on simple rules or general learning models, failing to fully consider students' individual characteristics and real-time learning status. For example, common learning recommendations based on a fixed order of the course syllabus ignore the differences in students' mastery of different knowledge points. For knowledge points that students have already mastered, they may have to repeat the learning, wasting time; while for weak knowledge points, they may not receive sufficient reinforcement.

[0004] Meanwhile, although some education systems are currently able to collect some student learning data, such as study time and interaction frequency, this data is mostly analyzed in isolation, failing to be organically combined and deeply explored to uncover the underlying student learning behavior patterns and knowledge mastery. For example, simply analyzing study time cannot reveal the actual learning outcomes of students during that period; focusing solely on interaction frequency makes it difficult to determine whether students have truly improved their understanding of knowledge through interaction.

[0005] Furthermore, knowledge graphs, as a tool that can effectively represent knowledge and its relationships, are gradually gaining importance in the field of education. However, existing knowledge graph-based learning recommendation systems often lack an effective mechanism for dynamically updating the knowledge graph, failing to adjust the association weights between knowledge points in the knowledge graph in a timely manner based on students' learning performance. This results in recommendation results that do not accurately reflect students' current weak knowledge areas, thus affecting the accuracy and effectiveness of the recommendations.

[0006] While algorithms such as Deep Q-Networks (DQNs) have achieved significant results in many fields of reinforcement learning, defining the environment, action space, and reward function appropriately in educational settings to better adapt to the needs of personalized learning path recommendations remains a challenging problem. For example, existing reward function designs may be too simplistic, failing to fully consider factors such as improved answer accuracy, learning time efficiency, and learning interest, thus potentially preventing the recommended paths from comprehensively promoting student learning.

[0007] In summary, current educational technology still has many shortcomings in achieving accurate and personalized learning path recommendations. There is an urgent need for a personalized learning path recommendation system that can integrate multi-source data, dynamically adjust knowledge graphs, effectively integrate multimodal data, and make optimization decisions based on reinforcement learning, in order to meet students' growing personalized learning needs and improve the quality of education and teaching. Summary of the Invention

[0008] To address the above problems, this invention proposes a method for generating personalized learning paths for students based on neural networks. The specific steps are as follows, characterized by:

[0009] Step 1: Multi-source data collection. Deploy data collection modules in the teaching platform, learning terminals, and interactive interfaces to collect multi-dimensional data such as students' learning time, interaction frequency, and wrong question records in real time.

[0010] Step 2: Build a knowledge graph and GNN dynamic modeling module. Based on the course outline and prior textbook structure, construct an initial knowledge graph, treating knowledge points as nodes and prerequisite relationships as edges. Use a graph neural network (GNN) to perform multi-layer propagation and aggregation on the knowledge graph. Periodically incorporate students' answer accuracy and mastery error information on each knowledge point into the weight update, dynamically adjusting the association weights of nodes and edges to reflect students' current weak knowledge areas.

[0011] Step 3: Implement behavioral feature extraction and multimodal embedding alignment. Construct a Bi-LSTM encoder for learning duration, interaction frequency, and incorrect question record data to generate behavioral feature vectors. Reduce the dimensionality of the node embedding vectors in the knowledge graph through a projection layer. Use an attention mechanism to align and fuse the two types of vectors to output personalized multimodal representations of students.

[0012] Step 4: Design a path decision and optimization strategy based on DQN, construct the environment state as the current student representation and knowledge graph node weights; define the action space as "select the next knowledge point to be learned"; the reward function comprehensively considers the improvement of answer accuracy, learning time efficiency and learning interest indicators, and uses deep Q network to update the goal, train and update the strategy network online;

[0013] Step 5: Deploy the trained DQN strategy as a microservice, display the learning path according to the recommendation results sorted by knowledge module, and present the mastery level, knowledge weakness heatmap and recommendation reasons in real time;

[0014] Step 6: Collect students' clicks on the recommended path, learning completion rate, and post-test scores. Based on the evaluation results, iteratively update the GNN architecture, attention fusion layer, and DQN parameters to continuously improve the recommendation effect.

[0015] As a further improvement to the present invention, the multi-source data acquisition in step 1 can be represented as:

[0016] The system collects learning duration, interaction frequency, and incorrect answer records. Learning duration is the duration for which a student opens the course page or video playback interface each time. Interaction frequency is the number of times a student posts in the comment section, asks questions, likes / replies, and sends private messages with teachers or peers. Incorrect answer records capture the question ID, the relevant knowledge point, the answer timestamp, and the time taken to answer each question incorrectly in each exercise or quiz.

[0017] As a further improvement to the present invention, the construction of the knowledge graph and GNN dynamic modeling module in step 2 can be represented as follows:

[0018] Step 2.1 Define the knowledge graph metamodel

[0019] Node set Each node The "smallest teachable learning unit" corresponding to a knowledge point;

[0020] Edge set : Among them are directed edges Indicates the precedence relation, undirected edge. Indicates horizontal semantic association;

[0021] Initial weight matrix , The element in the i-th row and j-th column is defined as:

[0022]

[0023] in

[0024]

[0025] For nodes The title, definition, and explanatory text corresponding to each knowledge point are concatenated into a paragraph and mapped by the encoder to form a feature vector.

[0026] Node feature matrix :

[0027] Step 2.2 Knowledge Point Node Extraction

[0028] Export the outline and chapter titles, and add directed edges according to the learning sequence specified in the textbook. ;when Add undirected edges ;

[0029] Step 2.3: Multi-layer propagation and dynamic weight update based on GNN

[0030] Step 2.3.1 Calculate the student's performance vector for each knowledge point.

[0031] Knowledge points Calculate the accuracy within the window [t−Δ,t]. :

[0032]

[0033] Where Δ is the size of the time window.

[0034] And calculate the mastery error. :

[0035]

[0036] Step 2.3.2 Define the GNN message propagation layer

[0037] Input: For the l-th layer, let The node feature matrix, Here is the symmetric normalized adjacency matrix for the (t-1)th round:

[0038]

[0039] in, Let be the adjacency matrix of the previous time step (t-1). For degree matrix, It is a node The sum of all outgoing edge weights at time t−1 for The element in the i-th row and j-th column;

[0040] Dissemination and updating:

[0041]

[0042] in, Let be the node representation matrix of the l-th layer of the GNN. For trainable weights, It is the ReLU activation function;

[0043] Step 2.3.3 Design of Weight Adjustment Function

[0044] To inject students' latest learning performance into the graph structure, an edge weight update function is defined:

[0045]

[0046] Where λ=0.8 is the smoothing coefficient, balancing prior structure and new information; g adopts...

[0047]

[0048] Reflects the strengthening of "weak-weak" node pairs;

[0049] Step 2.4 Periodic Update Process

[0050] Step 2.4.1 Set the update cycle: Triggered after each learning session is completed;

[0051] Step 2.4.2 Retrieve the latest answer log: Calculation and ;

[0052] Step 2.4.3 Update according to the above formula ;

[0053] Step 2.4.4 Re-execute the GNN forward propagation. The last layer of the GNN, i.e., the Kth layer, outputs a vector Z reflecting the student's current knowledge gaps. .

[0054] As a further improvement to the present invention, the behavioral feature extraction and multimodal embedding alignment in step 3 are represented as follows:

[0055] Step 3.1 Data Preprocessing and Sequence Construction

[0056] Step 3.1.1 Define the time window

[0057] Let each student's latest T learning sessions be a sequence sample, and T=20;

[0058] Step 3.1.2 Extracting raw indicators

[0059] For each session t∈{1,…,T}, three metrics are recorded sequentially: the learning duration d of the t-th session. t Unit: seconds; Number of interactions in the t-th session (i) t Unit: Item; Number of incorrect answers in the t-th session: e t Unit road;

[0060] Step 3.1.3 Normalization and Filling

[0061] Perform Min-Max normalization on each metric separately:

[0062]

[0063] in, and These represent the minimum and maximum learning durations across all historical sessions. and These are the minimum and maximum number of interactions in all historical sessions, respectively. These are the minimum and maximum numbers of incorrect answers across all historical sessions; For normalized learning time, This represents the normalized number of interactions. This represents the number of incorrect answers after normalization.

[0064] If the number of historical sessions is less than T, fill the front end with a vector of 0; if it exceeds T, truncate the most recent T sessions.

[0065] Step 3.1.4 generates three sets of input sequences.

[0066]

[0067] Where D, I, and E represent three time series inputs;

[0068] Step 3.2 Bi-LSTM Encoder Design

[0069] For each type of sequence, a bidirectional LSTM module is defined with ℎ hidden units, and the forward and reverse hidden states at the last time step are concatenated as the sequence encoding.

[0070] Taking the input sequence D of the session information duration as an example, the forward calculation is as follows:

[0071]

[0072] in, Let be the hidden state vector of the forward LSTM at time t. It is a forward LSTM network;

[0073] Reverse calculation:

[0074]

[0075] in, Let be the hidden state vector of the inverse LSTM at time t. It is a reverse LSTM network;

[0076] Two-way splicing:

[0077] Run forward and reverse LSTM on each sequence to obtain the hidden state at the last moment, and then concatenate them to generate the sequence code.

[0078]

[0079] Similarly, we can obtain , ;

[0080] Step 3.3 Extraction of Three-Way Behavior Vectors

[0081] Applying the same Bi-LSTM structure to the three sets of data (D, I, and E) respectively yields the corresponding behavior encoding vectors, which are then concatenated. :

[0082]

[0083] The dimension 6h is mapped to the unified behavior embedding dimension d through a fully connected layer:

[0084]

[0085] in, For behavior embedding vectors, and These are the parameters for the fully connected layer;

[0086] Step 3.4 Multimodal Alignment and Fusion

[0087] This section utilizes the knowledge graph embedding vector 𝑍 obtained in step 2, first performing dimensionality reduction mapping, and then aligning it with the behavior vector through an attention mechanism;

[0088] Step 3.4.1 Dimensionality Reduction of Embedded Vectors

[0089] Define a trainable parameter matrix W z With bias b z Dimensionality reduction is performed on all nodes simultaneously, reducing the dimension to d dimensions:

[0090]

[0091] in, This is the original embedding of the i-th node output by the GNN in step 1. Embedding of the dimensionality-reduced nodes;

[0092] Step 3.4.2 Attention Alignment

[0093] Constructing queries, keys, and values

[0094] Query q: using behavior vectors ,in, and The linear layer parameters for constructing the Query in the attention mechanism;

[0095] Key K: , N is the total number of points output by the GNN;

[0096] Value V: Node embedding after dimensionality reduction ;

[0097] Calculate attention weights :

[0098]

[0099] in, It is the i-th row K and the i-th key vector;

[0100] Weighted sum

[0101]

[0102] Step 3.4.3 Fusion Output

[0103] Finally, the behavior vector and attention output are concatenated and mapped to the representation space:

[0104]

[0105] By mapping the fusion layer to the fusion representation, we obtain :

[0106]

[0107] in, and These are the linear mapping parameters for the fusion layer.

[0108] As a further improvement to the present invention, the path decision and optimization strategy based on DQN in step 4 is represented as follows:

[0109] Step 4.1 Definition of Reinforcement Learning Environment

[0110] Step 4.1.1 Define the state

[0111] The state is defined as the fused representation output in step 3. This vector comprehensively reflects the student's most recent T learning behaviors and knowledge graph embedding;

[0112] Step 4.1.2 Define the action space

[0113] At time t, the set of available actions is taken from a subset of the "current weak node":

[0114]

[0115] in, For the node in step 2 The latest mastery error, with a threshold δ set to 0.2; action That is, from Select the next learning point from the list;

[0116] Step 4.1.3 Define Instant Rewards

[0117] Each time starting from the previous knowledge point Jump to select knowledge points After learning, an immediate reward is given based on the student's performance before and after learning, as well as their connectivity on the graph:

[0118]

[0119] , These are the nodes before and after learning. The accuracy rate of answering questions; , ;d graph (⋅,⋅) represents the shortest path length on the knowledge graph; weights α=0.5, β=0.3, γ=0.2;

[0120] Step 4.2 Deep Q-Network Structure

[0121] Step 4.2.1 Input Features

[0122] For each candidate action Construct merged vector :

[0123]

[0124] in ;

[0125] Step 4.2.2 Network Layers

[0126] The Q-value is estimated using a two-layer fully connected network:

[0127]

[0128] in, and These are the activation outputs of the first and second hidden layers of the DQN network, respectively. , and , These are the parameters of the first and second hidden layers of the DQN network, respectively. The output scalar Q(⋅) represents the value estimate of the action. and These are the output layer parameters;

[0129] Step 4.2.3 Action Selection Strategy

[0130]

[0131] Step 4.3 DQN Training Process

[0132] Step 4.3.1 Experience Replay

[0133] Initialize playback pool D and store interaction samples. ;

[0134] Step 4.3.2 Target Network

[0135] Copy the main network parameters To the target network Synchronization occurs every C steps, where C is the target network synchronization period.

[0136] Step 4.3.3 Batch Update

[0137] Random sampling small batch B is the number of mini-batch samples during training;

[0138] Calculate the target value of the j-th training sample

[0139]

[0140] in, Let be the immediate reward obtained by the j-th training sample during the transfer process. As a discount factor, For the next state The set of optional actions below, For action, For the target network parameters, The estimation function of the target network

[0141] Minimize mean square error loss

[0142]

[0143] in, For the main network parameters loss function, With the goal, This represents the estimated value Q of the main network for the j-th sample, with the sample state as the input. and the actions taken ; and use the Adam optimizer to update θ, train and update the policy network online to obtain the trained DQN network; then use s t and ei As input, execute the DQN network to generate a personalized learning path P for the student.

[0144] This invention relates to a method for generating personalized learning paths for students based on neural networks. The beneficial effects of this invention are as follows:

[0145] 1. This invention constructs a knowledge graph based on the curriculum outline and prior textbook structure, and uses a graph neural network (GNN) to periodically incorporate students' answer accuracy and mastery error information on each knowledge point into the weight update, which can dynamically and accurately reflect the students' current weak knowledge areas.

[0146] 2. This invention constructs a Bi-LSTM encoder to generate behavioral feature vectors from learning duration, interaction frequency, and incorrect answer records, and then aligns and fuses these vectors with node embedding vectors in a knowledge graph through an attention mechanism to output personalized multimodal representations of students. This multi-dimensional data fusion approach fully utilizes different types of data, enabling a more comprehensive and in-depth portrayal of students' learning characteristics and states.

[0147] 3. This invention designs a path decision and optimization strategy based on a deep Q-network. Using the student's fusion representation and knowledge graph node weights as the environmental state, it comprehensively considers factors such as improved answer accuracy, learning time efficiency, and learning interest indicators to define a reward function. This allows for online training and updating of the strategy network to generate personalized learning paths. This approach fully considers the unique learning situation and characteristics of each student, resulting in learning paths that better meet individual student needs and help improve learning efficiency and outcomes. Attached Figure Description

[0148] Figure 1 This is a flowchart of the present invention;

[0149] Figure 2 This is a diagram of the Bi-LSTM structure of the present invention. Detailed Implementation

[0150] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0151] This invention proposes a method for generating personalized learning paths for students based on neural networks. It collects student learning time and other data from multiple sources, models these paths using knowledge graphs and GNNs, aligns them with multimodal embeddings, and uses a DQN-designed path decision-making strategy to generate personalized learning paths. These paths are then displayed in real-time and iteratively optimized through closed-loop feedback to improve learning relevance and efficiency. The invention flowchart is shown below. Figure 1 As shown, the steps of the present invention will be described in detail below.

[0152] Step 1: Multi-source data acquisition

[0153] Deploy data collection modules in teaching platforms, learning terminals, and interactive interfaces to collect multi-dimensional data such as students' learning time, interaction frequency, and wrong answer records in real time.

[0154] Learning duration is the total time a student spends opening the course page or video playback interface. Interaction frequency records the number of times a student posts in the comments section, asks questions, likes / replies, and sends private messages to teachers or peers. Error record captures the question ID, related knowledge point, answer timestamp, and time taken for each incorrect answer in each exercise or quiz.

[0155] Step 2: Building the Knowledge Graph and GNN Dynamic Modeling Module

[0156] An initial knowledge graph is constructed based on the course syllabus and prior textbook structure, with knowledge points as nodes and prerequisite relationships as edges. A graph neural network (GNN) is used to propagate and aggregate the knowledge graph in multiple layers. The error information of students' answer accuracy and mastery on each knowledge point is periodically incorporated into the weight update, and the association weights of nodes and edges are dynamically adjusted to reflect the students' current weak knowledge areas.

[0157] Step 2.1 Define the knowledge graph metamodel

[0158] Node set Each node The "smallest teachable learning unit" that corresponds to a knowledge point.

[0159] Edge set : Among them are directed edges Indicates the precedence relation, undirected edge. It indicates horizontal semantic association.

[0160] Initial weight matrix , The element in the i-th row and j-th column is defined as:

[0161]

[0162] in

[0163]

[0164] For nodes The title, definition, and explanatory text corresponding to a knowledge point are concatenated into a paragraph and mapped by an encoder to form a feature vector.

[0165] Node feature matrix :

[0166] Step 2.2 Knowledge Point Node Extraction

[0167] Export the outline and chapter titles, and add directed edges according to the learning sequence specified in the textbook. .when Add undirected edges .

[0168] Step 2.3: Multi-layer propagation and dynamic weight update based on GNN

[0169] Step 2.3.1 Calculate the student's performance vector for each knowledge point.

[0170] Knowledge points Calculate the accuracy within the window [t−Δ,t]. :

[0171]

[0172] Where Δ is the size of the time window.

[0173] And calculate the mastery error. :

[0174]

[0175] Step 2.3.2 Define the GNN message propagation layer

[0176] Input: For the l-th layer, let The node feature matrix, Here is the symmetric normalized adjacency matrix for the (t-1)th round:

[0177]

[0178] in, Let be the adjacency matrix of the previous time step (t-1). For degree matrix, It is a node The sum of all outgoing edge weights at time t−1 for The element in the i-th row and j-th column.

[0179] Dissemination and updating:

[0180]

[0181] in, Let be the node representation matrix of the l-th layer of the GNN. For trainable weights, This is the ReLU activation function.

[0182] Step 2.3.3 Design of Weight Adjustment Function

[0183] To inject students' latest learning performance into the graph structure, an edge weight update function is defined:

[0184]

[0185] Where λ=0.8 is the smoothing coefficient, balancing prior structure and new information; g adopts...

[0186]

[0187] It reflects the strengthening of the "weak-weak" node pair.

[0188] Step 2.4 Periodic Update Process

[0189] Step 2.4.1 Set the update cycle: Triggered after each learning session is completed.

[0190] Step 2.4.2 Retrieve the latest answer log: Calculation and .

[0191] Step 2.4.3 Update according to the above formula .

[0192] Step 2.4.4 Re-execute the GNN forward propagation. The last layer of the GNN, i.e., the Kth layer, outputs a vector Z reflecting the student's current knowledge gaps. .

[0193] Step 3: Implement behavioral feature extraction and multimodal embedding alignment

[0194] Bi-LSTM encoders were constructed for the learning duration, interaction frequency, and incorrect question records, respectively. The Bi-LSTM structure diagram is shown below. Figure 2 As shown, behavioral feature vectors are generated; the node embedding vectors in the knowledge graph are reduced in dimensionality through a projection layer; an attention mechanism is used to align and fuse the two types of vectors to output a personalized multimodal representation of the student.

[0195] Step 3.1 Data Preprocessing and Sequence Construction

[0196] Step 3.1.1 Define the time window

[0197] Let each student's latest T learning sessions be a sequence sample, and T=20.

[0198] Step 3.1.2 Extracting raw indicators

[0199] For each session t∈{1,…,T}, three metrics are recorded sequentially: the learning duration d of the t-th session. t Unit: seconds; Number of interactions in the t-th session (i) tUnit: Item; Number of incorrect answers in the t-th session: e t , unit road.

[0200] Step 3.1.3 Normalization and Filling

[0201] Perform Min-Max normalization on each metric separately:

[0202]

[0203] in, and These represent the minimum and maximum learning durations across all historical sessions. and These are the minimum and maximum number of interactions in all historical sessions, respectively. These are the minimum and maximum numbers of incorrect answers across all historical sessions. For normalized learning time, This represents the normalized number of interactions. This represents the number of incorrect answers after normalization.

[0204] If the number of historical sessions is less than T, fill the front end with a 0 vector; if it exceeds T, truncate the most recent T sessions.

[0205] Step 3.1.4 generates three sets of input sequences.

[0206]

[0207] Where D, I, and E represent three time series inputs.

[0208] Step 3.2 Bi-LSTM Encoder Design

[0209] For each type of sequence, a custom bidirectional LSTM module is defined with ℎ hidden units, and the forward and reverse hidden states at the last time step are concatenated as the sequence encoding.

[0210] Taking the input sequence D of the session information duration as an example, the forward calculation is as follows:

[0211]

[0212] in, Let be the hidden state vector of the forward LSTM at time t. It is a forward LSTM network.

[0213] Reverse calculation:

[0214]

[0215] in, Let be the hidden state vector of the inverse LSTM at time t. It is a reverse LSTM network.

[0216] Two-way splicing:

[0217] Run forward and reverse LSTM on each sequence to obtain the hidden state at the last moment, and then concatenate them to generate the sequence code.

[0218]

[0219] Similarly, we can obtain , .

[0220] Step 3.3 Extraction of Three-Way Behavior Vectors

[0221] Applying the same Bi-LSTM structure to the three sets of data (D, I, and E) respectively yields the corresponding behavior encoding vectors, which are then concatenated. :

[0222]

[0223] The dimension 6h is mapped to the unified behavior embedding dimension d through a fully connected layer:

[0224]

[0225] in, For behavior embedding vectors, and These are the parameters for the fully connected layer.

[0226] Step 3.4 Multimodal Alignment and Fusion

[0227] This section utilizes the knowledge graph embedding vector 𝑍 obtained in step 2, first performing dimensionality reduction mapping, and then aligning it with the behavior vector through an attention mechanism.

[0228] Step 3.4.1 Dimensionality Reduction of Embedded Vectors

[0229] Define a trainable parameter matrix W z With bias b z Dimensionality reduction is performed on all nodes simultaneously, reducing the dimension to d dimensions:

[0230]

[0231] in, This is the original embedding of the i-th node output by the GNN in step 1. Embedded nodes after dimensionality reduction.

[0232] Step 3.4.2 Attention Alignment

[0233] Constructing queries, keys, and values

[0234] Query q: using behavior vectors ,in, and The linear layer parameters for constructing the Query in the attention mechanism.

[0235] Key K: , N represents the total number of points output by the GNN.

[0236] Value V: Node embedding after dimensionality reduction .

[0237] Calculate attention weights :

[0238]

[0239] in, It is the i-th row K and the i-th key vector.

[0240] Weighted sum

[0241]

[0242] Step 3.4.3 Fusion Output

[0243] Finally, the behavior vector and attention output are concatenated and mapped to the representation space:

[0244]

[0245] By mapping the fusion layer to the fusion representation, we obtain :

[0246]

[0247] in, and These are the linear mapping parameters for the fusion layer.

[0248] Step 4: Design a path decision and optimization strategy based on DQN

[0249] The environment state is constructed as the current student representation and the weights of knowledge graph nodes; the action space is defined as "selecting the next knowledge point to be learned"; the reward function comprehensively considers the improvement of answer accuracy, learning time efficiency and learning interest indicators, and uses a deep Q network to update the goal, and trains and updates the policy network online;

[0250] Step 4.1 Definition of Reinforcement Learning Environment

[0251] Step 4.1.1 Define the state

[0252] The state is defined as the fused representation output in step 3. This vector comprehensively reflects the student's most recent T learning behaviors and knowledge graph embedding.

[0253] Step 4.1.2 Define the action space

[0254] At time t, the set of available actions is taken from a subset of the "current weak node":

[0255]

[0256] in, For the node in step 2 The latest mastery error, with a threshold δ set to 0.2; action That is, from Select the next learning point from the list.

[0257] Step 4.1.3 Define Instant Rewards

[0258] Each time starting from the previous knowledge point Jump to select knowledge points After learning, an immediate reward is given based on the student's performance before and after learning, as well as their connectivity on the graph:

[0259]

[0260] , These are the nodes before and after learning. The accuracy rate of answering questions; , ;d graph (⋅,⋅) represents the shortest path length on the knowledge graph; weights α=0.5, β=0.3, γ=0.2.

[0261] Step 4.2 Deep Q-Network Structure

[0262] Step 4.2.1 Input Features

[0263] For each candidate action Construct merged vector :

[0264]

[0265] in .

[0266] Step 4.2.2 Network Layers

[0267] The Q-value is estimated using a two-layer fully connected network:

[0268]

[0269] in, and These are the activation outputs of the first and second hidden layers of the DQN network, respectively. , and , These are the parameters of the first and second hidden layers of the DQN network, respectively. The output scalar Q(⋅) represents the value estimate of the action. and These are the output layer parameters.

[0270] Step 4.2.3 Action Selection Strategy

[0271]

[0272] Step 4.3 DQN Training Process

[0273] Step 4.3.1 Experience Replay

[0274] Initialize playback pool D and store interaction samples. .

[0275] Step 4.3.2 Target Network

[0276] Copy the main network parameters To the target network Synchronization occurs every C steps, where C is the target network synchronization period.

[0277] Step 4.3.3 Batch Update

[0278] Random sampling small batch B is the number of mini-batch samples during training;

[0279] Calculate the target value of the j-th training sample

[0280]

[0281] in, Let be the immediate reward obtained by the j-th training sample during the transfer process. As a discount factor, For the next state The set of optional actions below, For action, For the target network parameters, The estimation function of the target network

[0282] Minimize mean square error loss

[0283]

[0284] in, For the main network parameters loss function, With the goal, This represents the estimated value Q of the main network for the j-th sample, with the sample state as the input. and the actions taken Then, the Adam optimizer is used to update θ, the policy network is trained and updated online, and the trained DQN network is obtained. Then, s... t and e i As input, execute the DQN network to generate a personalized learning path P for the student.

[0285] Step 5: Integrating the real-time recommendation engine with the user interface

[0286] The trained DQN strategy is deployed as a microservice, and the learning path is displayed in order of knowledge module according to the recommendation results. The mastery level, knowledge weakness heatmap and recommendation reasons are presented in real time.

[0287] Step 6: Closed-loop feedback and performance evaluation iteration

[0288] Collect students' clicks on recommended paths, learning completion rates, and post-test scores. Based on the evaluation results, iteratively update the GNN architecture, attention fusion layer, and DQN parameters to continuously improve the recommendation effect.

[0289] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for generating personalized learning paths for students based on neural networks, comprising the following specific steps, characterized by: Step 1: Multi-source data collection. Deploy data collection modules in the teaching platform, learning terminals, and interactive interfaces to collect multi-dimensional data such as students' learning time, interaction frequency, and wrong question records in real time. Step 2: Build a knowledge graph and GNN dynamic modeling module. Based on the course outline and prior textbook structure, construct an initial knowledge graph, treating knowledge points as nodes and prerequisite relationships as edges. Use a graph neural network (GNN) to perform multi-layer propagation and aggregation on the knowledge graph. Periodically incorporate students' answer accuracy and mastery error information on each knowledge point into the weight update, dynamically adjusting the association weights of nodes and edges to reflect students' current weak knowledge areas. Among them, knowledge points Calculation in window accuracy within And calculate the mastery error. : ， To inject students' latest learning performance into the graph structure, an edge weight update function is defined: ， g adopts: ， in: For students to understand the knowledge points In the time window The accuracy rate of answering questions within the scope; The size of the time window; For knowledge points Error in mastery; For the first After round of updates, the adjacency matrix of the th... Line number The elements of the column, i.e., knowledge point nodes. and The association weight of the edges between them; For the previous round, i.e. The associated weight of that edge; This is a smoothing coefficient with a value of 0.8, used to balance prior structure and new information; This is a weight adjustment function that reflects the strengthening of weak-weak node pairs; Knowledge point nodes Error in mastery; Step 3: Implement behavioral feature extraction and multimodal embedding alignment. Construct a Bi-LSTM encoder for learning duration, interaction frequency, and incorrect question record data to generate behavioral feature vectors. Reduce the dimensionality of the node embedding vectors in the knowledge graph through a projection layer. Use an attention mechanism to align and fuse the two types of vectors to output a personalized multimodal representation of the student. The attention mechanism constructs a query using the behavioral feature vectors and constructs keys and values ​​using the dimensionality-reduced knowledge graph node embeddings. Step 4: Design a path decision and optimization strategy based on DQN, construct the environment state as the current student representation and knowledge graph node weights; define the action space as selecting the next knowledge point to be learned; use a deep Q-network to update the goal, and train and update the policy network online; Among them, at time The set of optional actions is taken from a subset of the current weak nodes: ， in: For a moment A set of optional actions; Knowledge point nodes; For the node in step 2 The latest level of mastery error; The threshold is set to 0.2; Action That is, from Select the next learning point from the list; Each time starting from the previous knowledge point Jump to select knowledge points After learning, an immediate reward is given based on the student's performance before and after learning, as well as their connectivity on the graph: ， in: For a moment Instant rewards; , These are the nodes before and after learning. The accuracy rate of answering questions; , These are the nodes before and after learning. The degree of mastery error, and , ; The shortest path length on the knowledge graph; This refers to the knowledge point node from the previous lesson; The selected knowledge point node for the current jump; weight , , ; Step 5: Deploy the trained DQN strategy as a microservice, display the learning path according to the recommendation results sorted by knowledge module, and present the mastery level, knowledge weakness heatmap and recommendation reasons in real time; Step 6: Collect students' clicks on the recommended path, learning completion rate, and post-test scores. Based on the evaluation results, iteratively update the GNN architecture, attention fusion layer, and DQN parameters to continuously improve the recommendation effect.

2. The method for generating personalized learning paths for students based on neural networks according to claim 1, characterized in that: The multi-source data acquisition in step 1 can be represented as: Learning duration, interaction frequency, and incorrect question records are collected separately. Learning duration is the duration for which a student opens the course page or video playback interface each time. Interaction frequency records the number of times students post in the comment section, ask questions, like / reply, and send private messages with teachers or peers; incorrect question record captures the question ID, related knowledge point, answer timestamp, and answer time for each exercise or test.

3. The method for generating personalized learning paths for students based on neural networks according to claim 1, characterized in that: In step 3, when performing behavioral feature extraction and multimodal embedding alignment, the knowledge graph embedding vector obtained in step 2 is used. First, perform dimensionality reduction mapping, and then align it with the behavior vector through an attention mechanism; Dimensionality reduction is performed simultaneously on all nodes, down to... dimension: ， Constructing queries, keys, and values: Query Using behavior vectors ;key , ;value Using dimensionality-reduced node embedding ; Calculate attention weights : ， Weighted summation: ， in: The first output of GNN in step 2 Original embedding of each node; Embedding of the dimensionality-reduced nodes; For a trainable parameter matrix, For bias; Embedded dimensions for unified behavior; For behavior embedding vectors; , Parameters for constructing the linear layer of the query in the attention mechanism; This is the matrix formed by the embedding of all nodes after dimensionality reduction; This represents the total number of nodes output by the GNN. For the first Attention weights for each node; For key The Middle One key vector; For key The Middle One key vector; For query vector; The sum of attention-weighted values.

4. The method for generating personalized learning paths for students based on neural networks according to claim 1, characterized in that: The deep Q-network structure and training process described in step 4 are as follows: For each candidate action Construct merged vector : ， The Q-value is estimated using a two-layer fully connected network: ， ， ， Experience Replay: Initialize the replay pool Store interaction samples ; Target network: Copy the parameters of the main network To the target network ,Every One-step synchronization; batch update: random sampling of small batches. Calculate the first The target value of each training sample : ， Minimize mean squared error loss: ， in: For a moment A set of optional actions; These are the knowledge point nodes corresponding to the candidate actions; For candidate actions The constructed merged vector; The fusion representation output from step 3; Candidate actions Corresponding dimensionality reduction node embedding ; , These are the activation outputs of the first and second hidden layers of the DQN network, respectively. , and , These are the parameters of the first and second hidden layers of the DQN network, respectively. Estimate the value of this action; , For output layer parameters; For experience replay pool; For interactive samples, For a moment The actions taken The next state; Main network parameters; For target network parameters; The target network synchronization period; This refers to the number of mini-batch samples during training. For the first The target value for each training sample; For the first The instantaneous reward obtained by each training sample during the transfer process; Discount factor; For the next state The set of optional actions below; For action; Let be the estimation function of the target network; For the main network parameters The loss function; Indicates the main network to the first The estimated value of a sample, with the sample state as the input. and the actions taken .

Images