Student personalized learning path generation method based on neural network

Abstract

According to the student personalized learning path generation method based on the neural network, a data acquisition module is deployed on a teaching platform, a learning terminal and an interactive interface, and multi-dimensional data such as student learning duration, interactive frequency and wrong question records are acquired in real time. A knowledge graph is constructed based on a course outline and a prior textbook structure, and a graph neural network GNN is utilized to dynamically adjust association weights of nodes and edges of the graph neural network GNN to reflect weak knowledge areas of students. A Bi-LSTM encoder is constructed for the collected data to generate a behavior feature vector, and the behavior feature vector and a knowledge graph node embedded vector are aligned and fused through an attention mechanism to obtain student personalized multi-modal representation. And taking the representation and the knowledge graph node weight as an environment state, selecting a next knowledge point to be learned as an action space, designing a reward function by integrating multiple indexes, updating a target by using a deep Q network, and training a strategy network to generate a personalized learning path. The trained strategy is deployed as micro-service, a path is displayed according to knowledge module sorting, and related parameters are iteratively updated according to student feedback to improve the recommendation effect.

Description

Technical Field

[0001] The present invention relates to the field of educational technology, and in particular to a method for generating a student's personalized learning path based on a neural network. Background Art

[0002] In today's rapidly developing digital education era, online learning platforms and various learning terminals are widely used, providing students with a rich variety of learning resources and flexible learning methods. However, with the massive growth of learning data and the significant individual 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 key issue in the education field that needs to be addressed.

[0003] Traditional learning path recommendations are often based on simple rules or universal learning models, failing to fully consider students' individual characteristics and real-time learning status. For example, the common practice of recommending learning paths based on a fixed sequence within the syllabus ignores students' varying levels of mastery of different knowledge points. This can lead to students wasting time repetitively studying already mastered knowledge points, while also failing to adequately reinforce weaker areas.

[0004] Meanwhile, while some education systems are currently able to collect some student learning data, such as study time and interaction frequency, this data is often analyzed in isolation, failing to organically integrate and deeply explore the underlying learning behavior patterns and knowledge acquisition. For example, simply analyzing study time fails to reveal the true learning outcomes of students during this period; focusing solely on interaction frequency makes it difficult to determine whether students have truly improved their understanding of the knowledge through interaction.

[0005] Furthermore, knowledge graphs, as a tool for effectively representing knowledge and its relationships, are gaining increasing attention in the education field. However, existing knowledge graph-based learning recommendation systems often lack effective mechanisms for dynamically updating knowledge graphs. They are unable to timely adjust the weights of associations between knowledge points in the knowledge graph based on student learning performance, resulting in recommendations that fail to accurately reflect students' current weak areas of knowledge, thus affecting the accuracy and effectiveness of recommendations.

[0006] While algorithms like the Deep Q-Network (DQN) have achieved remarkable results in various fields, in educational settings, the challenge remains to rationally define the environment state, action space, and reward function for reinforcement learning to better meet the needs of recommending personalized learning paths for students. For example, existing reward function designs may be overly simplistic, failing to fully consider factors such as accuracy, learning time efficiency, and learning interest. This can result in recommended paths failing to fully promote student learning.

[0007] To sum up, 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 optimized decisions based on reinforcement learning, so as to meet students' growing personalized learning needs and improve the quality of education and teaching. Summary of the Invention

[0008] To solve the above problems, the present invention proposes a method for generating personalized learning paths for students based on a neural network. The specific steps are as follows, which are 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 on students' learning time, interaction frequency, and error records in real time;

[0010] Step 2: Build a knowledge graph and GNN dynamic modeling module. Construct an initial knowledge graph based on the course syllabus and prior textbook structure, 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 information about students' accuracy and mastery errors on each knowledge point into weight updates, dynamically adjusting the association weights of nodes and edges to reflect students' current weak knowledge areas.

[0011] Step 3: Extract behavioral features and align multimodal embeddings. A Bi-LSTM encoder is constructed for learning time, interaction frequency, and error record data to generate behavioral feature vectors. Dimensionality reduction is performed on the node embedding vectors in the knowledge graph through a projection layer. The attention mechanism is used to align and fuse the two vectors to output a personalized multimodal representation of the student.

[0012] Step 4: Design a DQN-based path decision and optimization strategy, constructing the environment state as the current student representation and knowledge graph node weights; 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, uses the deep Q network to update the target, and trains and updates the policy network online;

[0013] Step 5: Deploy the trained DQN strategy as a microservice. Based on the recommendation results, the learning path is sorted by knowledge module and displayed in real time, including mastery, knowledge weakness heatmaps, and recommendation reasons.

[0014] Step 6: Collect students’ clicks on the 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.

[0015] As a further improvement of the present invention, the multi-source data collection in step 1 can be expressed as:

[0016] The learning time, interaction frequency, and wrong question records are collected separately. The learning time is the duration each time students open the course page or video playback interface; the interaction frequency is the number of times students speak in the comment area, ask questions, like / reply, and the number of private messages with teachers or peers; the wrong question record is to capture the wrong question ID, the knowledge point to which it belongs, the answer timestamp, and the time taken to answer in each exercise or test.

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

[0018] Step 2.1 Define the knowledge graph metamodel

[0019] Node Set : Each node The "minimum teachable learning unit" corresponding to a knowledge point;

[0020] Edge Set :There are directed edges Indicates a prerequisite relationship, an undirected edge Indicates horizontal semantic association;

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

[0022]

[0023] in

[0024]

[0025] For nodes The title, definition, and description text corresponding to the 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 outlines and chapter titles, and add directed edges based on the learning order specified in the textbook. ;when Add undirected edges ;

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

[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 time window size

[0034] And calculate the mastery error :

[0035]

[0036] Step 2.3.2 Define the GNN message propagation layer

[0037] Input: For the lth layer, let is the node feature matrix, is the symmetric normalized adjacency matrix of the t-1th round:

[0038]

[0039] in, is the adjacency matrix of the t-1th round at the previous moment, is the degree matrix, is a node The sum of all outgoing edge weights at time t−1, for The element in row i and column j;

[0040] Dissemination and Update:

[0041]

[0042] in, is the node representation matrix of the GNN layer l, are trainable weights, is the ReLU activation function;

[0043] Step 2.3.3 Design of weight adjustment function

[0044] In order to inject the student's latest learning performance into the graph structure, the edge weight update function is defined:

[0045]

[0046] Among them, λ=0.8 is the smoothing coefficient, which balances the prior structure and new information; g is adopted

[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: trigger after each learning session is completed;

[0051] Step 2.4.2 Pull the latest answer log: Calculate and ;

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

[0053] Step 2.4.4 re-executes the GNN forward propagation. The last layer of GNN, i.e., the 𝐿th layer, outputs a vector Z reflecting the student’s current weak knowledge area. .

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

[0055] Step 3.1 Data preprocessing and sequence construction

[0056] Step 3.1.1 Define the time window

[0057] Assume that each student's latest T learning sessions are a sequence sample, T=20;

[0058] Step 3.1.2 Extracting raw indicators

[0059] For each session t∈{1,…,T}, three indicators are recorded in sequence: the learning time d of the tth session t , unit is seconds; number of interactions in the tth session i t , unit bar; number of wrong questions in the tth session e t , unit road;

[0060] Step 3.1.3 Normalization and padding

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

[0062]

[0063] in, and are the minimum and maximum learning durations in all historical sessions, and are the minimum and maximum number of interactions in all historical sessions, are the minimum and maximum number of wrong questions in all historical sessions respectively; is the normalized learning time, is the normalized number of interactions, is the number of wrong questions after normalization;

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

[0065] Step 3.1.4 Form three sets of input sequences

[0066]

[0067] Among them, D, I, and E represent three time series inputs respectively;

[0068] Step 3.2 Bi-LSTM encoder design

[0069] For each type of sequence, define a bidirectional LSTM module with ℎ hidden units, and take the concatenation of the forward and reverse hidden states at the last moment as the sequence encoding;

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

[0071]

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

[0073] Reverse calculation:

[0074]

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

[0076] Two-way splicing:

[0077] Run forward and backward LSTM on each sequence to obtain the hidden state at the last moment, and concatenate them to generate the sequence encoding:

[0078]

[0079] The same logic applies 、 ;

[0080] Step 3.3 Three-way behavior vector extraction

[0081] Apply the same structure Bi-LSTM to the three groups of data D, I, and E respectively to obtain the corresponding behavior encoding vectors, and then concatenate the vectors. :

[0082]

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

[0084]

[0085] in, is the behavior embedding vector, and is the fully connected layer parameter;

[0086] Step 3.4 Multimodal alignment and fusion

[0087] This section uses the knowledge graph embedding vector 𝑍 obtained in step 2 to first perform dimensionality reduction mapping, and then align it with the behavior vector through the attention mechanism;

[0088] Step 3.4.1 Dimensionality reduction of embedding vector 𝑍

[0089] Define the trainable parameter matrix W z With bias b z , reduce the dimension of all nodes simultaneously to d dimensions:

[0090]

[0091] in, is the original embedding of the i-th node output by step GNN, is the node embedding after dimensionality reduction;

[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 Query in the attention mechanism;

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

[0096] Value V: Use node embedding after dimensionality reduction ;

[0097] Calculating attention weights :

[0098]

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

[0100] Weighted sum value

[0101]

[0102] Step 3.4.3 Fusion Output

[0103] Finally, the action vector is concatenated with the attention output and mapped to the representation space:

[0104]

[0105] Mapped to the fusion representation through the fusion layer, we get :

[0106]

[0107] in, and is the linear mapping parameter of the fusion layer.

[0108] As a further improvement of the present invention, the DQN-based path decision and optimization strategy designed in step 4 is expressed as follows:

[0109] Step 4.1 Reinforcement Learning Environment Definition

[0110] Step 4.1.1 Define the state

[0111] The state is defined as the fused representation of the output of 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 optional actions is taken from the subset of “current weak nodes”:

[0114]

[0115] in, The node in step 2 The latest mastery error of action is set to 0.2. That is, from Select the next learning point;

[0116] Step 4.1.3 Define immediate rewards

[0117] Each time from the previous knowledge point Jump to selected knowledge point After learning, immediate rewards are given based on their previous and subsequent performance and connectivity on the graph:

[0118]

[0119] 、 They are respectively the nodes before and after learning Correct answer rate; , ;d graph (⋅,⋅) is 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 the merge vector :

[0123]

[0124] in ;

[0125] Step 4.2.2 Network Level

[0126] Use a two-layer fully connected network to estimate the Q value:

[0127]

[0128] in, and are the first and second hidden layer activation outputs of the DQN network, 、 and 、 are the first and second hidden layer parameters of the DQN network, and the output scalar Q(⋅) is the value estimate of the action. and is the output layer parameter;

[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 the replay pool D to store interaction samples ;

[0134] Step 4.3.2 Target Network

[0135] Copy the main network parameters To the target network , synchronize once every C steps, where C is the number of synchronization periods of the target network;

[0136] Step 4.3.3 Batch Update

[0137] Randomly sample mini-batches , B is the number of mini-batch samples during training;

[0138] Calculate the target value of the jth training sample

[0139]

[0140] in, is the immediate reward obtained by the j-th training sample during the transfer process, is the discount factor, For the next state The following optional action set, For action, are the target network parameters, is the estimation function of the target network

[0141] Minimize the mean squared error loss

[0142]

[0143] in, For the main network parameters The loss function is For the goal, Represents the estimated value Q of the main network for the jth sample, and the input is the state of the sample 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, the DQN network is executed to generate the student's personalized learning path P.

[0144] The method for generating personalized learning paths for students based on neural networks of the present invention has beneficial effects. The technical effects of the present invention are:

[0145] 1. This invention constructs a knowledge graph based on the course syllabus and prior textbook structure, and uses a graph neural network (GNN) to periodically integrate students' answer accuracy and mastery error information on each knowledge point into weight updates, dynamically and accurately reflecting students' current weak knowledge areas.

[0146] 2. This invention constructs a Bi-LSTM encoder based on learning time, interaction frequency, and error record data to generate behavioral feature vectors. This vector is then aligned and fused with the node embedding vectors in the knowledge graph through an attention mechanism to output a personalized multimodal representation of the student. This multi-dimensional data fusion approach fully utilizes different types of data to more comprehensively and deeply characterize a student's learning characteristics and status.

[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 environment state, the reward function is defined by comprehensively considering the improvement in answer accuracy, learning time efficiency, and learning interest indicators. This approach then trains and updates the policy network online to generate personalized learning paths. This approach fully considers each student's unique learning situation and characteristics, resulting in a learning path that better meets individual student needs and helps improve learning efficiency and outcomes. BRIEF DESCRIPTION OF THE DRAWINGS

[0148] Figure 1 is a flow chart of the present invention;

[0149] Figure 2 This is the Bi-LSTM structure diagram of the present invention. DETAILED DESCRIPTION

[0150] The present invention is 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. By collecting data such as student learning time from multiple sources, the method uses knowledge graphs, GNN modeling, and multimodal embedding alignment to generate personalized learning paths using a DQN-designed path decision strategy. This path is presented through real-time recommendations and iteratively optimized through closed-loop feedback, improving learning relevance and efficiency. The invention flowchart is shown below. Figure 1 As shown, the steps of the present invention are described in detail below.

[0152] Step 1: Multi-source data collection

[0153] Deploy data collection modules in the teaching platform, learning terminals, and interactive interfaces to collect multi-dimensional data on students' learning time, interaction frequency, and wrong question records in real time;

[0154] Study time measures the duration of each session when students open a course page or video playback interface. Interaction frequency records the number of comments, questions, likes / replies, and private messages with teachers or peers. Error records capture the ID, corresponding knowledge point, answer timestamp, and time spent on each incorrect question in each exercise or quiz.

[0155] Step 2: Build the knowledge graph and GNN dynamic modeling module

[0156] An initial knowledge graph is constructed based on the course syllabus and prior textbook structure, treating knowledge points as nodes and prerequisite relationships as edges. A graph neural network (GNN) is used to perform multi-layer propagation and aggregation on the knowledge graph. Information on students' accuracy and mastery errors in each knowledge point is periodically incorporated into weight updates, dynamically adjusting the association weights of nodes and edges to reflect students' current weak knowledge areas.

[0157] Step 2.1 Define the knowledge graph metamodel

[0158] Node Set : Each node The "minimum teachable learning unit" corresponding to a knowledge point.

[0159] Edge Set :There are directed edges Indicates a prerequisite relationship, an undirected edge Indicates horizontal semantic association.

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

[0161]

[0162] in

[0163]

[0164] For nodes The title, definition, and description text corresponding to the knowledge point are concatenated into a feature vector of a paragraph after being mapped by the encoder.

[0165] Node feature matrix :

[0166] Step 2.2 Knowledge point node extraction

[0167] Export outlines and chapter titles, and add directed edges based on the learning order specified in the textbook. .when Add undirected edges .

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

[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 time window size

[0173] And calculate the mastery error :

[0174]

[0175] Step 2.3.2 Define the GNN message propagation layer

[0176] Input: For the lth layer, let is the node feature matrix, is the symmetric normalized adjacency matrix of the t-1th round:

[0177]

[0178] in, is the adjacency matrix of the t-1th round at the previous moment, is the degree matrix, is a node The sum of all outgoing edge weights at time t−1, for The element in row i and column j.

[0179] Dissemination and Update:

[0180]

[0181] in, is the node representation matrix of the GNN layer l, are trainable weights, is the ReLU activation function.

[0182] Step 2.3.3 Design of weight adjustment function

[0183] In order to inject the student's latest learning performance into the graph structure, the edge weight update function is defined:

[0184]

[0185] Among them, λ=0.8 is the smoothing coefficient, which balances the prior structure and new information; g is adopted

[0186]

[0187] Reflects the strengthening of “weak-weak” node pairs.

[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 Pull the latest answer log: Calculate and .

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

[0192] Step 2.4.4 re-executes the GNN forward propagation. The last layer of GNN, i.e., the 𝐿th layer, outputs a vector Z reflecting the student’s current weak knowledge area. .

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

[0194] The Bi-LSTM encoder is constructed for learning time, interaction frequency, and wrong question record data respectively. The Bi-LSTM structure diagram is shown as follows: Figure 2 As shown in the figure, a behavioral feature vector is generated; the node embedding vector in the knowledge graph is reduced in dimension through the projection layer; the 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] Assume that each student's latest T learning sessions are a sequence sample, T=20.

[0198] Step 3.1.2 Extracting raw indicators

[0199] For each session t∈{1,…,T}, three indicators are recorded in sequence: the learning time d of the tth session t , unit is seconds; number of interactions in the tth session i t, unit bar; number of wrong questions in the tth session e t , unit road.

[0200] Step 3.1.3 Normalization and padding

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

[0202]

[0203] in, and are the minimum and maximum learning durations in all historical sessions, and are the minimum and maximum number of interactions in all historical sessions, They are the minimum and maximum number of wrong questions in all historical sessions respectively. is the normalized learning time, is the normalized number of interactions, is the normalized number of wrong questions.

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

[0205] Step 3.1.4 Form three sets of input sequences

[0206]

[0207] Among them, D, I, and E represent three time series inputs respectively.

[0208] Step 3.2 Bi-LSTM encoder design

[0209] For each type of sequence, a bidirectional LSTM module is defined with the number of hidden units being ℎ, and the concatenation of the forward and reverse hidden states at the last moment is taken as the sequence encoding.

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

[0211]

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

[0213] Reverse calculation:

[0214]

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

[0216] Two-way splicing:

[0217] Run forward and backward LSTM on each sequence to obtain the hidden state at the last moment, and concatenate them to generate the sequence encoding:

[0218]

[0219] The same logic applies 、 .

[0220] Step 3.3 Three-way behavior vector extraction

[0221] Apply the same structure Bi-LSTM to the three groups of data D, I, and E respectively to obtain the corresponding behavior encoding vectors, and then concatenate the vectors. :

[0222]

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

[0224]

[0225] in, is the behavior embedding vector, and are the parameters of the fully connected layer.

[0226] Step 3.4 Multimodal alignment and fusion

[0227] In this section, we use the knowledge graph embedding vector 𝑍 obtained in step 2 to first perform dimensionality reduction mapping and then align it with the behavior vector through the attention mechanism.

[0228] Step 3.4.1 Dimensionality reduction of embedding vector 𝑍

[0229] Define the trainable parameter matrix W z With bias b z , reduce the dimension of all nodes simultaneously to d dimensions:

[0230]

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

[0232] Step 3.4.2 Attention Alignment

[0233] Constructing queries, keys, and values

[0234] Query q: Using behavior vectors ,in, and Parameters of the linear layer used to construct the Query in the attention mechanism.

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

[0236] Value V: Use node embedding after dimensionality reduction .

[0237] Calculating attention weights :

[0238]

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

[0240] Weighted sum value

[0241]

[0242] Step 3.4.3 Fusion Output

[0243] Finally, the action vector is concatenated with the attention output and mapped to the representation space:

[0244]

[0245] Mapped to the fusion representation through the fusion layer, we get :

[0246]

[0247] in, and is the linear mapping parameter of the fusion layer.

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

[0249] The environment state is constructed as the current student representation and the knowledge graph node weights; 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, uses a deep Q network to update the target, and trains and updates the policy network online;

[0250] Step 4.1 Reinforcement Learning Environment Definition

[0251] Step 4.1.1 Define the state

[0252] The state is defined as the fused representation of the output of 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 optional actions is taken from the subset of “current weak nodes”:

[0255]

[0256] in, The node in step 2 The latest mastery error of action is set to 0.2. That is, from Select the next learning point.

[0257] Step 4.1.3 Define immediate rewards

[0258] Each time from the previous knowledge point Jump to selected knowledge point After learning, immediate rewards are given based on their previous and subsequent performance and connectivity on the graph:

[0259]

[0260] 、 They are the nodes before and after learning Correct answer rate; , ;d graph (⋅,⋅) is 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 the merge vector :

[0264]

[0265] in .

[0266] Step 4.2.2 Network Level

[0267] Use a two-layer fully connected network to estimate the Q value:

[0268]

[0269] in, and are the first and second hidden layer activation outputs of the DQN network, 、 and 、 are the first and second hidden layer parameters of the DQN network, and the output scalar Q(⋅) is the value estimate of the action. and 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 the replay pool D to store interaction samples .

[0275] Step 4.3.2 Target Network

[0276] Copy the main network parameters To the target network , synchronization is performed every C steps, where C is the number of synchronization periods of the target network.

[0277] Step 4.3.3 Batch Update

[0278] Randomly sample mini-batches , B is the number of mini-batch samples during training;

[0279] Calculate the target value of the jth training sample

[0280]

[0281] in, is the immediate reward obtained by the j-th training sample during the transfer process, is the discount factor, For the next state The following optional action set, For action, are the target network parameters, is the estimation function of the target network

[0282] Minimize the mean squared error loss

[0283]

[0284] in, For the main network parameters The loss function is For the goal, Represents the estimated value Q of the main network for the jth sample, and the input is the state of the sample and the actions taken And use Adam optimizer to update θ, train and update the policy network online to obtain the trained DQN network. t and e i As input, the DQN network is executed to generate the student's personalized learning path P.

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

[0286] Deploy the trained DQN strategy as a microservice, sort the learning paths by knowledge module based on the recommendation results, and present the mastery level, knowledge weakness heat map, and recommendation reasons 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, and iteratively update the GNN architecture, attention fusion layer, and DQN parameters based on the evaluation results to continuously improve the recommendation effect.

[0289] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation 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 a neural network, comprising the following specific steps, characterized in that: Step 1: Multi-source data collection: deploy data collection modules in the teaching platform, learning terminals, and interactive interfaces to collect multi-dimensional data on students' learning time, interaction frequency, and error records in real time; Step 2: Build a knowledge graph and GNN dynamic modeling module. Construct an initial knowledge graph based on the course syllabus and prior textbook structure, 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 information about students' accuracy and mastery errors on each knowledge point into weight updates, dynamically adjusting the association weights of nodes and edges to reflect students' current weak knowledge areas. Step 3: Extract behavioral features and align them with multimodal embeddings. Construct Bi-LSTM encoders for learning time, interaction frequency, and error record data to generate behavioral feature vectors. Reduce the dimension of the node embedding vectors in the knowledge graph through a projection layer. The attention mechanism is used to align and fuse the two types of vectors to output a personalized multimodal representation of the student; Step 4: Design a DQN-based path decision and optimization strategy, constructing the environment state as the current student representation and knowledge graph node weight; The action space is defined as "selecting the next knowledge point to learn." The reward function comprehensively considers the improvement of answer accuracy, learning time efficiency, and learning interest indicators. It uses a deep Q network to update the target and train and update the policy network online. Step 5: Deploy the trained DQN strategy as a microservice. Based on the recommendation results, the learning path is sorted by knowledge module and displayed in real time, including mastery, knowledge weakness heatmaps, and recommendation reasons. Step 6: Collect students’ clicks on the 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.

2. The method for generating a personalized learning path for students based on a neural network according to claim 1, characterized in that: Multi-source data collection in step 1 can be expressed as: The learning time, interaction frequency, and wrong question records are collected separately. The learning time is the duration of each time the student opens the course page or video playback interface; The interaction frequency records the number of times students speak in the comment area, ask questions, like / reply, and private messages with teachers or peers; the wrong question record captures the wrong question ID, knowledge point, answer time stamp, and answer time in each exercise or test.

3. The method for generating a personalized learning path for students based on a neural network according to claim 1, characterized in that: The construction of the knowledge graph and GNN dynamic modeling module in step 2 can be expressed as follows: Step 2.1: Define the knowledge graph metamodel; Node Set : Each node The "minimum teachable learning unit" corresponding to a knowledge point; Edge Set :There are directed edges Indicates a prerequisite relationship, an undirected edge Indicates horizontal semantic association; Initial weight matrix , The element in row i and column j is defined as: ; in; ; For nodes The title, definition, and description text corresponding to the knowledge point are concatenated into a paragraph and mapped by the encoder to form a feature vector. Node feature matrix : ; Step 2.2: Extract knowledge point nodes; Export outlines and chapter titles, and add directed edges based on the learning order specified in the textbook. ;when Add undirected edges ; Step 2.3: GNN-based multi-layer propagation and dynamic weight update; Step 2.3.1 Calculate the student's performance vector for each knowledge point; Knowledge points , calculate the accuracy within the window [t−Δ,t] : ; Where, Δ is the time window size; And calculate the mastery error : ; Step 2.3.2 defines the GNN message propagation layer; Input: For the lth layer, let is the node feature matrix, is the symmetric normalized adjacency matrix of the t-1th round: ,in, is the adjacency matrix of the t-1th round at the previous moment, is the degree matrix, is a node The sum of all outgoing edge weights at time t−1, for The element in row i and column j; Dissemination and Update: ,in, is the node representation matrix of the GNN layer l, are trainable weights, is the ReLU activation function; Step 2.3.3: Design of weight adjustment function; In order to inject the student's latest learning performance into the graph structure, the edge weight update function is defined: , where λ=0.8 is the smoothing coefficient to balance the prior structure and new information; g is adopted Reflects the strengthening of "weak-weak" node pairs; Step 2.4 Periodic update process; Step 2.4.1 Set the update cycle: trigger after each learning session is completed; Step 2.4.2 Pull the latest answer log: Calculate and ; Step 2.4.3 Update according to the above formula ; Step 2.4.4 re-executes the GNN forward propagation. The last layer of GNN, i.e., the 𝐿th layer, outputs a vector Z reflecting the student’s current weak knowledge area. .

4. The method for generating a personalized learning path for students based on a neural network according to claim 1, characterized in that: The behavioral feature extraction and multimodal embedding alignment implemented in step 3 are expressed as follows: Step 3.1 Data preprocessing and sequence construction; Step 3.1.1 defines the time window; Assume that each student's latest T learning sessions are a sequence sample, T=20; Step 3.1.2 extracts the original indicators; For each session t∈{1,…,T}, three indicators are recorded in sequence: the learning time d of the tth session t , unit is seconds; number of interactions in the tth session i t , unit bar; number of wrong questions in the tth session e t , unit road; Step 3.1.3 Normalization and padding; Perform Min-Max normalization on each indicator separately: ; in, and are the minimum and maximum learning durations in all historical sessions, and are the minimum and maximum number of interactions in all historical sessions, are the minimum and maximum number of wrong questions in all historical sessions respectively; is the normalized learning time, is the normalized number of interactions, is the number of wrong questions after normalization; If the number of historical sessions is less than T, fill the front end with a 0 vector; if it exceeds T, intercept the most recent T entries; Step 3.1.4 forms three sets of input sequences; ; Among them, D, I, and E represent three time series inputs respectively; Step 3.2 Bi-LSTM encoder design; For each type of sequence, define a bidirectional LSTM module with ℎ hidden units, and take the concatenation of the forward and reverse hidden states at the last moment as the sequence encoding; Taking the session information duration input sequence D as an example, the forward calculation is: ; in, is the hidden state vector of the forward LSTM at time t, It is a forward LSTM network; Reverse calculation: ,in, is the hidden state vector of the reverse LSTM at time t, It is a reverse LSTM network; Two-way splicing: Run forward and backward LSTM on each sequence to obtain the hidden state at the last moment, and concatenate them to generate the sequence encoding: The same logic applies 、 ; Step 3.3: Extract three-way behavior vectors; Apply the same structure Bi-LSTM to the three groups of data D, I, and E respectively to obtain the corresponding behavior encoding vectors, and then concatenate the vectors. : ; The dimension is mapped from 6h to a unified behavioral embedding dimension d through a fully connected layer: in, is the behavior embedding vector, and is the fully connected layer parameter; Step 3.4: Multimodal alignment and fusion; Using the knowledge graph embedding vector 𝑍 obtained in step 2, we first perform dimensionality reduction mapping and then align it with the behavior vector through the attention mechanism; Step 3.4.1: Reduce the dimension of the embedding vector 𝑍; Define the trainable parameter matrix W z With bias b z , reduce the dimension of all nodes simultaneously to d dimensions: in, is the original embedding of the i-th node output by step GNN, is the node embedding after dimensionality reduction; Step 3.4.2: Attention alignment; Construct query, key, and value; Query q: Using behavior vectors ,in, and Parameters of the linear layer used to construct the Query in the attention mechanism; Key K: , , N is the total number of points output by GNN; Value V: Use node embedding after dimensionality reduction ; Calculating attention weights : in, is the i-th row K, i-th key vector; Weighted sum value ; Step 3.4.3 Fusion output; Finally, the behavior vector is concatenated with the attention output and mapped to the representation space: Mapped to the fusion representation through the fusion layer, we get : in, and is the linear mapping parameter of the fusion layer.

5. The method for generating a personalized learning path for students based on a neural network according to claim 1, characterized in that: The DQN-based path decision and optimization strategy designed in step 4 is expressed as follows: Step 4.1 Reinforcement learning environment definition; Step 4.1.1 Define the state; The state is defined as the fused representation of the output of step 3 , this vector comprehensively reflects the student’s most recent T learning behaviors and knowledge graph embedding; Step 4.1.2 defines the action space; At time t, the set of optional actions is taken from the subset of "current weak nodes": in, The node in step 2 The latest mastery error of action is set to 0.

2. That is, from Select the next learning point; Step 4.1.3 defines the immediate reward; Each time from the previous knowledge point Jump to selected knowledge point After learning, immediate rewards are given based on their previous and subsequent performance and connectivity on the graph: , 、 They are the nodes before and after learning The correct answer rate of , ;d graph (⋅,⋅) is the shortest path length on the knowledge graph; weights α=0.5, β=0.3, γ=0.2; Step 4.2 Deep Q network structure; Step 4.2.1 Input features; For each candidate action , construct the merge vector : ,in ; Step 4.2.2 Network level; Use a two-layer fully connected network to estimate the Q value: ; in, and are the first and second hidden layer activation outputs of the DQN network, 、 and 、 are the first and second hidden layer parameters of the DQN network, and the output scalar Q(⋅) is the value estimate of the action. and is the output layer parameter; Step 4.2.3 action selection strategy; ; Step 4.3 DQN training process; Step 4.3.1 Experience replay; Initialize the replay pool D to store interaction samples ; Step 4.3.2 Target network; Copy the main network parameters To the target network , synchronize once every C steps, where C is the number of synchronization periods of the target network; Step 4.3.3 Batch update; Randomly sample mini-batches , B is the number of mini-batch samples during training; Calculate the target value of the jth training sample ; ; in, is the immediate reward obtained by the j-th training sample during the transfer process, is the discount factor, For the next state The set of optional actions below, For action, are the target network parameters, is the estimation function of the target network Minimize the mean squared error loss; ; in, For the main network parameters The loss function is For the goal, Represents the estimated value Q of the main network for the jth sample, and the input is the state of the sample 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 e i As input, the DQN network is executed to generate the student's personalized learning path P.

Images