Dynamic key value memory network knowledge tracking method and system integrating graph guidance and behavior self-adaption

Abstract

The invention discloses a dynamic key value memory network knowledge tracking method and system fusing graph guidance and a behavior adaptive mechanism, and belongs to the technical field of artificial intelligence and education data mining. The system comprises an input coding module, a graph structure modeling module, a behavior modeling module, a fusion attention mechanism module, a dynamic memory updating module, a prediction output module and a visual interface module. Wherein the input coding module is used for performing vectorization representation on question numbers, answering results, time intervals and student portrait information in student and question interaction logs, and generating capability vectors including absorption capability, structure preference and predictive regulation and control; the graph structure modeling module is used for modeling knowledge graph relationships among the topics by using a graph attention network to generate structure perception representation; the behavior modeling module adopts a Transform structure to carry out modeling on a student answering sequence, and behavior feature representation is extracted.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and education technology, and more specifically, to a method and system for knowledge tracing using a dynamic key-value memory network that integrates graph guidance and behavioral adaptation. Background Technology The rapid development of online education platforms and Massive Open Online Courses (MOOCs) has led to an exponential increase in the time series, scale, and dimensions of student interaction logs. Real-time analysis of this interaction data to dynamically estimate students' mastery of each knowledge point is a key prerequisite for personalized teaching and precise intervention; this task is known as Knowledge Tracing (KT).

[0002] Early Bayesian knowledge tracing models (BKT) used hidden Markov processes to describe student state transitions with a small number of parameters at a single knowledge point, exhibiting good interpretability and computational efficiency. However, this model is based on the assumption of "knowledge point independence," failing to capture prerequisites and hierarchical relationships between concepts and struggling to handle complex behavioral features in long sequences. Subsequent deep knowledge tracing models (DKT) used recurrent neural networks for behavioral sequence modeling, significantly improving prediction accuracy. However, by embedding student knowledge states into a high-dimensional latent space, they lack explicit representations of knowledge structure and individual differences, resulting in limited interpretability.

[0003] Dynamic Key-Value Memory Networks (DKVMN) have achieved a new balance between interpretability and accuracy: key slots represent the semantic prototypes of knowledge points with fixed values, while value slots dynamically store students' mastery of the corresponding concepts. However, existing DKVMN and its derivative methods still face problems such as insufficient structure awareness, limited representation of behavioral sequences, lack of personalized differences, and inadequate time decay mechanisms.

[0004] With the development of technologies such as graph neural networks and Transformers, researchers have attempted to extend DKVMN from different dimensions. For example, graph convolution or graph attention mechanisms have been introduced to explicitly characterize the graph structure between questions / knowledge points; Transformer encoders are used to capture the long-range dynamics in student answer sequences; and designs such as ability vectors and time decay strategies are used to characterize individual differences and forgetting patterns. However, these improvements are mostly scattered across individual modules, lacking end-to-end collaborative integration, making it difficult to further improve prediction accuracy and generalization ability while ensuring interpretability.

[0005] Against this technological backdrop, a dynamic key-value memory network knowledge tracing method (DKVMN-GBA) integrating graph guidance and behavioral adaptation is proposed. Within the DKVMN framework, this method introduces a graph attention network to capture structural dependencies between questions, utilizes Transformer to deeply mine the dynamic features of answer sequences, and explicitly models individual student differences and forgetting effects through multi-dimensional ability vectors and an adaptive time decay mechanism. Through the synergistic fusion of structural, behavioral, and personalized information, the model can provide higher accuracy, stronger interpretability, and better generalization performance in complex teaching scenarios, meeting the urgent need for highly reliable key-value memory (KT) technology in online education. Summary of the Invention

[0006] The DKVMN-GBA system proposed in this invention addresses the ever-growing student-question interaction logs in online education platforms. Without sacrificing real-time performance, it estimates students' mastery of each knowledge point in a fine-grained and interpretable manner, providing data support for decision-making scenarios such as question ranking, learning path adjustment, and classroom intervention. Compared to BKT, DKT, and the original DKVMN, this system integrates four types of information within the same neural framework: knowledge graph structure, long-range behavioral sequences, individual ability differences, and dynamic forgetting curves. This avoids the reliance of traditional methods on the assumption of question independence, short-sighted sequence windows, and static memory parameters, enabling the model to maintain stable prediction accuracy in courses with complex structures or large sequence spans.

[0007] Before data enters the model, the backend uses a unified embedding pipeline to fuse and map discrete and continuous features such as question identifiers, answer results, answer time, and intervals between adjacent answers. The same pipeline also reads student profile fields, such as grade level, historical accuracy, average learning interval, and daily learning time, and adaptively generates absorption ability vectors, prediction and regulation vectors, and structural preference vectors during training. Absorption ability reflects the speed at which students internalize new concepts, prediction and regulation describes preferences for question types and difficulty, and structural preference characterizes the degree of attention paid to prerequisite dependencies. These three types of ability vectors permeate the model's read and write gating, allowing individual differences to be explicitly presented at the lowest level.

[0008] The feature modeling phase comprises parallel graph structure modeling and behavior modeling modules. The graph structure modeling module uses the course knowledge graph as its topological foundation, placing question nodes and implicit knowledge point nodes in the same embedding space. It dynamically measures the importance of different prerequisite paths through learnable adjacency weights, thereby outputting a structure-aware representation. This representation not only includes explicit topological relationships but also preserves local semantic context, enabling the model to correctly distinguish concept levels in multi-skill, multi-prerequisite courses. The behavior modeling module employs a two-layer multi-head self-attention network to globally model the question-answer joint sequence. Position encoding records sequence information, residual connections and layer normalization stabilize gradients, and a sparse sliding window strategy reduces the peak memory latency to about 70% of the native Transformer while maintaining the receptive field, thus maintaining linear inference latency even on long sequence datasets such as ASSIST2017.

[0009] The fusion and memory update phases employ a gating mechanism to softly weight the structural and behavioral representation vectors. The gating signal is jointly generated by the structural preference vector and the behavioral tendency vector, automatically adjusting the focus based on the learning scenario: when students practice cross-concept integrated problems, the focus shifts to the graph structure modeling module; when continuously working through similar problems, the focus shifts to the behavioral modeling module. The key-value memory network first performs an erasure operation to simulate forgetting, then incrementally updates the memory according to the written vector. The decay factor, combined with the response interval and absorption capacity vector, adaptively controls the contribution of old memories through an exponential-sigmoid mixture function, thus presenting an individualized memory curve characterized by "high efficiency with slow forgetting and low efficiency with rapid decay."

[0010] The prediction layer concatenates the latest knowledge state vector, structural representation, behavioral representation, and three types of ability vectors, then inputs them into a two-layer feedforward network to output the student's mastery probability of the current question. The training process employs FocalLoss to enhance the learning of low-frequency error samples, label smoothing to suppress overfitting, Dropout and L2 regularization to reduce redundancy, and mixed precision and gradient clipping to ensure numerical stability.

[0011] Evaluation was conducted using AUC, F1, and RMSE metrics, with five-fold cross-validation performed on the ASSIST2009, ASSIST2015, ASSIST2017, and STATICS2011 datasets. Compared to DKVMN reproduced in the same batch, the AUC improvement ranged from 3.8% to 5.2%, with a 4.4% improvement in AUC and a 6.3% reduction in RMSE on the long and sparse ASSIST2017 dataset. Ablation experiments further confirmed that removing the graph structure modeling module weakens the ability to model structural dependencies; removing the Transformer branch reduces the effectiveness of capturing long-range dependencies; and removing the capability vector or time decay both result in a performance drop of one to three percentage points. All four mechanisms are indispensable.

[0012] The system provides a standard REST API interface for the front-end teaching platform to call the prediction results. The server is deployed using Docker containers, supporting horizontal scaling. To adapt to different deployment environments, the system can adjust the number of Transformer layers and GAT heads according to computing power, and also supports weighted deployment on edge devices. The inference process maintains a response time of less than 200ms in a CPU environment.

[0013] The system pushes mastery curves and diagnoses of weak knowledge points to the teaching hall in real time via a REST API interface, and provides an adaptive question recommendation list to the intelligent question bank. Results show that the average number of passes decreased by nearly 10%, classroom quiz time was shortened by about eight minutes, students' subjective dissatisfaction with question repetition decreased, and backend latency stabilized within 50 milliseconds. Because the framework adopts a standardized modular design, deployers can tailor channels or quantize weights according to computing power conditions. They can also introduce semantic vectors of question text, classroom video emotion scores, or voice reading feedback signals to further enrich the input modality and improve the model's robustness to comprehensive interactive scenarios.

[0014] In summary, DKVMN-GBA unifies graph attention, temporal attention, capability vectors, and a learnable temporal decay mechanism within a key-value memory network, forming a four-dimensional collaborative knowledge tracking approach encompassing structure, behavior, capability, and time. In tests on public datasets and small-scale real-world environments, the system demonstrates stable prediction accuracy, reasonable interpretability, and acceptable inference overhead. It serves as a practical technical option for online education platforms and provides a replicable engineering paradigm for subsequent multimodal fusion research in the field of knowledge tracking. Attached Figure Description

[0015] Figure 1 This is the overall architecture diagram of the DKVMN-GBA model; Figure 2 This is a schematic diagram of the problem structure relationship modeling and graph attention aggregation mechanism; Figure 3 This is a flowchart of the Transformer encoder behavior modeling process. Detailed Implementation

[0016] This invention is specifically implemented based on a dynamic key-value memory network knowledge tracking system that integrates a graph-guided mechanism and a behavior-adaptive strategy. It aims to comprehensively improve the accuracy, personalization, and interpretability of student knowledge state modeling. To ensure a clear technical path and engineering feasibility, this paper will elaborate on the system module structure, data flow path, mathematical modeling and optimization mechanisms, general applicability, and boundaries of application, forming a technical solution that can be directly applied in educational scenarios. Figure 1 As shown, the DKVMN-GBA model consists of six parts: an input encoding layer, a graph structure modeling module, a Transformer behavior modeling module, a fusion attention mechanism module, a dynamic key-value memory update mechanism, and a prediction output module. The model takes the question number and the answer as input, extracts structural and behavioral representation vectors respectively, and fuses individual ability vectors with a time decay mechanism to achieve personalized, interpretable knowledge state updates and answer prediction.

[0017] The system's overall architecture is based on a Dynamic Key-Value Memory Network (DKVMN). It incorporates graph neural networks for structural information modeling and integrates a Transformer behavior modeling module to model the temporal nature of student responses. Furthermore, it combines a time decay mechanism with three-dimensional ability vectors to adapt and regulate individual differences. From the outset, the system design fully considers the continuity of data flow and the closed-loop nature of knowledge state modeling, ultimately achieving real-time prediction of students' mastery of questions and assisting teachers in decision-making and recommendation system design.

[0018] The Transformer module contains two layers of Encoders, each configured as follows: - Number of heads with multi-head attention: 4; - Attention dimension: 64; - Feedforward network dimensions: 256; - Activation function: ReLU; - Dropout rate: 0.1; - Residual connectivity and layer normalization are used to stabilize the training process; In addition, a sparse sliding window attention strategy is introduced, which only retains local window + global token calculation in each layer, reducing memory usage and improving the ability to model long sequences.

[0019] The system is divided into five main modules: input encoding and initialization layer, graph structure modeling module, behavior modeling module, fusion attention and memory mechanism layer, and prediction output module. First, the input encoding layer maps the question numbers and answers from student interaction data into a high-dimensional dense embedding, and constructs a three-dimensional ability vector at the student level: absorption capacity, structural preference, and predictive regulation, to reflect the differences in students' learning characteristics. In the graph structure modeling module, the system introduces a Graph Attention Network (GAT) to model the structural relationships between questions, such as prerequisite paths, parallel topics, and semantic adjacency, thereby generating context-aware question representations. For example... Figure 2 As shown, in the question structure modeling process, the adjacency relationships between questions are first constructed based on the knowledge graph, and the structure-aware vector of each question is calculated based on the graph attention mechanism. This mechanism explicitly models the semantic dependencies and importance of prerequisite paths between questions through attention-weighted aggregation of neighboring nodes, realizing the propagation and fusion of structural information in the graph space. The behavior sequence modeling part uses the Transformer behavior modeling module to perform contextual modeling of the question-response sequence, and combines a positional encoding mechanism to mine the temporal correlation and long-term dependencies of answering behaviors, outputting a global behavior representation vector for subsequent fusion. For example... Figure 3 As shown, the behavior sequence modeling process jointly embeds the question number sequence and the corresponding answer result, adds positional encoding, and then inputs it into a multi-layer Transformer encoder. The model models the long-term dependencies of student behavior through a multi-head self-attention mechanism and a feedforward network, and generates context-enhanced behavior representation vectors for guiding behavior in knowledge state memory reading and writing.

[0020] To enhance the interpretability of the system, the model can output the following at each time step: - Activation distribution of the memory slot; - Current erase vector and write vector; - Attenuation factor value; Teachers can view this information through a visual interface to determine which knowledge points students have forgotten and which question types they are slow to grasp, thus assisting in teaching decisions.

[0021] To integrate information from structural and behavioral paths, the system introduces a fusion attention mechanism module, dynamically adjusting the weight ratio of structural perception representation and behavioral sequence representation in memory operations. The fusion coefficient is generated from the structural preference ability vector, and at each time step, it is jointly weighted based on the structural position of the current question and the behavioral context, resulting in a personalized attention distribution. This attention distribution is then used to access memory slots, enabling dynamic reading and writing of the student's knowledge state. Value memory uses an erasure and write mechanism to update the state, where the erasure vector represents knowledge forgetting and the write vector represents the absorption of new knowledge, both controlled by the student's absorption ability. Simultaneously, the system generates a time decay mechanism based on the answer time interval and the student's absorption ability to model the forgetting curve and naturally regulate the decay of knowledge state.

[0022] The prediction output module takes a comprehensive feature vector as input, concatenates the knowledge readout vector, structure awareness vector, and prediction control capability vector, and then inputs the concatenation into a feedforward neural network for binary classification prediction. During system training, Focal Loss is used as the primary loss function to improve the fitting ability to difficult samples, and label smoothing techniques and memory regularization terms are introduced to avoid overfitting. The training optimization process combines AdamW and SGD strategies to achieve rapid convergence and fine-grained adjustment.

[0023] The data flow interfaces between the modules are as follows: - Input encoding layer: Input dimension [batch_size, seq_len, feature_dim] → Output embedding representation [batch_size, seq_len, d_model] - Graph Modeling Path (GAT): Input embedding → Output structure vector [batch_size, seq_len, g_dim] - Behavioral modeling module (Transformer): Input embedding + positional encoding → Output context vector [batch_size, seq_len, t_dim] - Fusion Attention Mechanism Module: Outputs the fusion vector [batch_size, seq_len, f_dim] - Memory module: Input erase / write vector, output current memory state - Prediction module: Input concatenated vector → Output mastery probability [batch_size, 1] At the data flow level, the entire system follows this path: student interaction data is first input into the embedding layer to generate embedded representations and ability vectors. This data is then fed in parallel into the structural modeling path and behavior modeling module, which output structural perception vectors and behavior representation vectors, respectively. After structural preference adjustment, fused attention is generated for selective reading and writing of memory slots, forming a representation of the student's knowledge mastery state at the current time step. Finally, the prediction and control ability vectors are concatenated to predict the next question, and the prediction results are compared with the actual labels for backpropagation. The entire system ensures data flow consistency and a closed-loop prediction logic, fully perceiving the dynamics of student knowledge mastery at each time step.

[0024] The symbols are defined as follows: - : The question number at the current time step; - :topic Neighbor set in a knowledge graph; - : Questions or knowledge points Embedded representation; - Linear transformation matrix; - : Activation function; this paper uses LeakyReLU; - Learnable attention vectors; - Value memory matrix; - Time decay mechanism; - Adjustment factor in Focal Loss; - : Hyperparameters for controlling attention to difficult samples.

[0025] At the mathematical level, the system employs the following core formulas to construct a logical closed loop. In the graph structure modeling part, let... For the current topic, Given its set of neighboring nodes, the graph attention mechanism is defined as follows: in It is a linear transformation matrix. It is a non-linear activation function. This is a learnable attention vector.

[0026] In the behavioral modeling part, the multi-head attention mechanism in Transformer is represented as follows: After the input sequence is embedded, it is linearly mapped into three sets of vectors: query, key, and value. Then, the attention score is calculated and the context is aggregated.

[0027] The fusion attention distribution is calculated as follows: For structural preference vectors, For the graph path attention distribution, Attention distribution along behavioral paths. Memory erasure and writing: Time decay factor : in For learnable parameters, This represents the absorption capacity vector. Predicted probability. : Focal Loss: in Controlling the degree of suppression for easily classified samples, Adjust the balance between positive and negative samples.

[0028] In terms of system applicability, DKVMN-GBA can be widely used in different types of courses: - It can capture detailed prerequisite dependencies for courses with dense graph structures (such as mathematics and physics); - It also exhibits strong structural robustness to courses with sparse graph structures (such as history and Chinese); - Supports short sequence tasks (fragmented practice) and long sequence tasks (unit assessment); - It can adapt to differences in student behavior and support differentiated instruction and ability-based grouping.

[0029] The system possesses good versatility and scalability: - The graph modeling module supports replacing GAT with GCN, GATv2, Relational GCN, etc.; - The behavior modeling module supports replacing Transformer with Informer, Performer, Longformer, etc.; - Supports multimodal input, such as semantic vectors of the question text, student attention scores in videos, and speech reading performance; - In the future, it can be expanded into a multi-task learning framework to realize composite tasks such as answer prediction, next question recommendation, and learning resource recommendation.

[0030] From a system boundary perspective, DKVMN-GBA is suitable for various question types and educational scenarios. First, it is highly adaptable to question structures, accepting graph inputs that are either dense (e.g., in math courses) or sparse (e.g., in history courses). Second, the system is compatible with varying lengths of student response sequences, supporting both short-sequence tasks (e.g., fragmented practice) and long-sequence tasks (e.g., periodic assessments). Third, personalized ability modeling and time decay mechanisms support students with different learning habits and exhibit good individual generalization ability.

[0031] In terms of versatility, the system can be extended to scenarios involving multimodal learning data, such as introducing a text understanding module to process the language features of question content, or integrating audio and video processing modules to capture multimodal response behaviors. On the graph modeling path, the GAT module can be extended to structures such as GCN, GATv2, and Relational GCN to adapt to more complex knowledge graph types. On the behavior modeling module, the Transformer can be replaced with more efficient structures (such as Performer, Informer, and Longformer) to handle modeling tasks with extremely long sequences.

[0032] Furthermore, the prediction module can be further expanded with a multi-task prediction architecture, which can not only predict the correctness of answers, but also predict the next learning time, cognitive state classification, or recommend suitable teaching resources, thus achieving more comprehensive intelligent teaching assistance.

[0033] Overall, DKVMN-GBA possesses numerous advantages such as high modularity, strong scalability, wide applicability, and strong interpretability. It can serve as a key component in large-scale educational models, providing deep cognitive modeling capabilities, personalized path decision support, and real-time learning status feedback services in future educational scenarios. It is an important technological path to promote the intelligent development of education.

Claims

1. A dynamic key-value memory network knowledge tracing system integrating graph guidance and behavior adaptation mechanisms, characterized in that, The system includes: an input encoding module, a graph structure modeling module, a behavior modeling module, a fusion attention mechanism module, a dynamic memory update module, a prediction output module, and a visualization interface module. The input encoding module receives interaction data between students and questions. The graph structure modeling module, based on the course knowledge graph, constructs an adjacency matrix between questions and introduces a graph attention network to achieve structural modeling, outputting a structural perception representation of each question. The behavior modeling module uses a Transformer structure to capture long-term dependencies in the interaction sequence and generates behavior representation vectors. The fusion attention mechanism module adjusts the information fusion of graph structure paths and behavior paths through structure preference vectors to dynamically generate fusion representations. The dynamic memory module combines attention weights and ability adjustment factors to perform erasure and write operations. The prediction module predicts the correctness of answers by splicing fusion representations, ability vectors, and memory readout vectors. The visualization module displays the dynamic evolution of students' knowledge states.

2. The dynamic key-value memory network knowledge tracking system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, The input encoding module employs a multi-dimensional embedding mapping strategy to convert the question numbers in the student's interactive response sequence. Mapped to a fixed-dimensional vector Answer results Encoding as behavior tag embedding Answering time interval The time vector is then obtained through numerical normalization and nonlinear transformation. Student profile information includes grade level, knowledge point mastery rate, answer frequency, and study time. A student ability vector is generated using feature aggregation. This vector consists of three parts: an absorption capacity vector, used to adjust the memory write strength; Structural preference vector, used for regulating the fusion attention mechanism; prediction regulation vector, used for individual adjustments to the prediction module; input vector. As the foundation for subsequent modules.

3. The dynamic key-value memory network knowledge tracking system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, The graph structure modeling module uses a graph attention network to model the structure of the question nodes. The construction method is as follows: First, using the question nodes in the course knowledge graph as vertices and the prerequisite or parallel relationships between questions as edges, a directed graph G=(V,E) is constructed; then, for each node... Construct its neighbor set The structural neighborhood of nodes is extracted using an adjacency matrix. Attention weights between node pairs are calculated using an attention mechanism. The structure-aware representation is constructed as follows: for any two adjacent nodes i and j, their embedding vectors are concatenated and then normalized using a learnable linear transformation and the LeakyReLU function to obtain the attention weights. Then, the structural representation of the target node is obtained by weighting and summing the neighbor node information; the final graph representation. Robustness can be further enhanced through multi-head attention; This module can capture the upstream and downstream dependencies and semantic adjacency between questions, providing a structural semantic basis for personalized knowledge updates.

4. The dynamic key-value memory network knowledge tracking system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, The behavior modeling module employs a sequence modeling mechanism based on the Transformer architecture, with its input being a pre-encoded sequence of interaction vectors. The output is a vector of behavior representations. The Transformer module consists of several layers of multi-head self-attention mechanisms and feedforward neural networks stacked together. Each layer captures the interaction dependencies between any positions in the sequence through query, key, and value mapping. The calculation formula is as follows: To preserve the temporal information of the sequence, the system adds positional encoding to the input sequence, which can be a fixed sine function or a trainable vector. This module can extract the behavioral evolution patterns and cognitive rhythms of students in the long-term learning process, providing deep semantic support for dynamic memory updates and mastery prediction.

5. The dynamic key-value memory network knowledge tracing system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, The fusion attention mechanism module constructs a learnable fusion gating mechanism based on the degree of matching between the structural preference vector and the structural and behavioral representations. This mechanism controls the weighting ratio of the two types of information and generates the final fusion representation. The fusion process employs the following mechanism: the structural preference vector is input into a linear transformation and a sigmoid activation function to obtain the fusion weights. Then, the fusion operation is completed through linear combination, that is: This mechanism can adaptively adjust according to individual students' preferences for structural or behavioral paths, effectively improving the generalization ability of the attention mechanism under different learning styles; by introducing a structural preference ability vector, the system realizes personalized control of the fused information flow.

6. The dynamic key-value memory network knowledge tracking system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, The dynamic memory update module updates the knowledge state based on a key-value separation memory structure, where key-slot... A fixed set of knowledge points, value slots This indicates the student's current mastery level; the module first integrates attention weights. The read and write locations of the memory slots are determined, followed by erase and write operations; the erase operation is performed by constructing an erase vector. The old value slot content is proportionally decayed; the write operation generates a write vector by combining the current interaction representation with the capability vector. The time was calculated by combining the time interval with the absorbed energy. Adjusting the writing intensity; the above mechanism can effectively simulate the human forgetting process and realize personalized memory updates, which is the core link in realizing dynamic knowledge tracking.

7. The dynamic key-value memory network knowledge tracking system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, After completing the memory update, the prediction output module will retrieve the memory read vector at the current moment. fusion representation vector The final representation is formed by concatenating the predictive and regulatory capability vector. The probability of answering is predicted by using a feedforward neural network; The prediction function is in the form of: During training, to mitigate prediction bias caused by sample imbalance, the system uses Focal Loss as the loss function, which is defined as: , where γ is a hyperparameter; this module combines dynamic memory content with individual ability adjustment to achieve high accuracy and high interpretability in knowledge mastery estimation.

8. The dynamic key-value memory network knowledge tracking system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, Multiple optimization strategies are introduced during system training to improve model robustness and generalization ability. First, a label smoothing mechanism is introduced into the loss function to add small perturbations to the target distribution of the true labels to alleviate overfitting. Second, an L2 regularization term is introduced to limit the parameter size and prevent the model from becoming too complex. Third, a hybrid optimizer combining AdamW and SGD is used, with the former used for fast convergence and the latter to avoid getting trapped in local optima. During training, a cosine annealing learning rate scheduler is used to dynamically adjust the learning rate, combined with an early stopping strategy to prevent overfitting. The system also supports stable training of deep networks based on batch normalization and residual connections, making information transfer between modules smoother and contributing to stable operation under large-scale data.

9. The dynamic key-value memory network knowledge tracing system with fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, The system supports multi-task joint learning extensions, enabling the simultaneous training of auxiliary tasks such as next question prediction, answer time regression, and question difficulty assessment while completing the main task (i.e., knowledge mastery prediction). The network structure designs an independent decoding head for each task, while sharing the front-end encoder module and capability vector representation. Multi-objective training is performed using a weighted joint loss function, taking into account the performance of each sub-task. This multi-task setup not only enhances the discriminative ability of the main task but also improves the model's generalization performance, making it more valuable for practical teaching applications and serving as a foundational module for a personalized learning path recommendation engine.

10. The dynamic key-value memory network knowledge tracking method based on the fusion graph guidance and behavior adaptation mechanism according to claim 1, characterized in that, This is accomplished using the system described in any one of claims 1-9.

Images