Knowledge state evolution graph-based English learning path prediction model construction method

Abstract

The invention relates to the technical field of graph neural networks, and discloses a knowledge state evolution graph-based English learning path prediction model construction method, which comprises the following steps of: constructing a sparse adjacent matrix, configuring an inertia and activity orthogonal bimodal hidden state for a node, calculating an interactive input prediction residual error, comparing the interactive input prediction residual error with a dynamic gating threshold value, and determining an English learning path prediction model. When the residual error is lower than a threshold value, only the target active component is updated, and full icon quantity attenuation is executed; according to the method, a conditional sparse activation architecture of a computational graph is established through a dynamic gating mechanism driven by the residual, the problem of high-frequency computational power redundancy of full-graph convolution is solved, noise diffusion is blocked by using state decoupling, and the robustness of the computational graph is improved. And the prediction precision and stability of the model on the non-stationary sequence are improved.

Description

Technical Field

[0001] This invention relates to a method for constructing an English learning path prediction model based on a knowledge state evolution graph, belonging to the field of graph neural network technology. Background Technology

[0002] When processing data modeling tasks with complex topological dependencies, graph neural networks combined with recurrent neural network architectures use adjacency matrices to represent the node association structure and map discrete interaction events into continuous high-dimensional hidden state vectors through a time-step full-graph message passing mechanism. This captures the evolution of node states over time and performs predictions. End-to-end deep learning frameworks extract data features and fit nonlinear relationships in an ideal computing environment. However, in edge-side inference scenarios with large-scale, high-concurrency requests and limited computing resources, existing full-update strategies face a contradiction between computational cost and state fidelity. Changes in the local state of a single node trigger feature aggregation and parameter updates across the entire graph. Dense computing modes for long sequences or large-scale graphs cause inference latency to increase nonlinearly with the number of nodes, exceeding the computing power load of edge devices.

[0003] To avoid high computational costs, existing technologies attempt to simplify learning behavior using statistical models or shallow machine learning algorithms. For example, Chinese invention patent application CN120387545A discloses an intelligent prediction system for English learning behavior based on big data. This system collects learning time and interactive behavior data, uses linear regression and K-means clustering algorithms to build personalized models, and dynamically adjusts model parameters based on feedback data to recommend learning paths. However, while such shallow logic-based prediction models reduce computational dimensionality, they essentially ignore the deep topological dependencies that objectively exist between knowledge points. They simplify complex knowledge systems into discrete feature vectors, making it difficult for the model to capture the interconnected evolution of knowledge. More importantly, such solutions lack mechanisms for identifying and filtering random noise in non-stationary time series. The parameter update logic is equally sensitive to all input data, failing to distinguish between accidental user errors and substantial cognitive changes. When faced with real teaching data containing a large amount of pseudo-mastery or sudden forgetting, the prediction path is prone to oscillations due to overfitting to short-term fluctuations, lacking the stability required for industrial applications.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a dynamic sparse computing architecture with event-driven characteristics, decouple long and short-term state features and introduce a statistical stability gating mechanism to reduce the real-time inference computing load while improving the model's noise resistance and prediction stability against non-stationary sequences. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of this invention is as follows: A method for constructing an English learning path prediction model based on a knowledge state evolution graph, comprising the following steps: Construct a sparse adjacency matrix of the target domain knowledge graph to define the graph topology, and configure a bimodal hidden state vector for each node in the knowledge graph. The bimodal hidden state vector is composed of mutually orthogonal inertial component vectors and active component vectors. Obtain real-time interactive input features and interactive result labels for the target node, calculate the prediction probability using the bimodal latent state vector of the target node, and calculate the numerical deviation between the prediction probability and the interactive result label. Define the numerical deviation as the prediction residual value. The predicted residual value is compared with a preset dynamic gating threshold, and the first update strategy or the second update strategy is executed according to the comparison result to evolve the state of the graph. If the predicted residual value is less than the dynamic gating threshold, the first update strategy is executed: only the interactive input features are used to update the active component vector of the target node, and the active component vectors of other nodes in the knowledge graph other than the target node are calculated based on the time interval, while keeping the values ​​of the inertial component vectors of other nodes unchanged. If the predicted residual value is greater than or equal to the dynamic gating threshold, execute the second update strategy: lock the set of multi-hop neighbor nodes of the target node in the sparse adjacency matrix, perform graph convolution operation based on the predicted residual value to generate the structure correction gradient, and use the structure correction gradient to correct the inertial component vector of each node in the set of multi-hop neighbor nodes.

[0006] Preferably, the second update strategy further includes performing a local topology self-calibration step based on gradient correlation: after determining the multi-hop neighbor node set, for each neighbor node in the set, calculate the cosine similarity between the bimodal hidden state vector of the neighbor node and the gradient direction vector of the predicted residual value; based on the cosine similarity, dynamically adjust the weights of the connection edges between the neighbor node and the target node in the sparse adjacency matrix; when the cosine similarity indicates that the current state of the neighbor node is positively correlated with the correction direction of the predicted residual value, increase the weights of the connection edges to strengthen the gradient propagation to the source of error; when the cosine similarity indicates a negative correlation, decrease the weights of the connection edges to block the error propagation path in the same direction as the predicted residual value.

[0007] Preferably, after updating the active component vector of the target node, the process further includes performing an inertial fixation step based on second-order statistics: maintaining a sliding time window, recording the magnitude sequence of the active component vector of the target node within the sliding time window; calculating the statistical mean and statistical variance of the magnitude sequence, and calculating the fixation gate coefficient using the following formula: ,in, To solidify the gate coefficient, It is the Sigmoid activation function. To calculate the statistical mean, To statistically analyze the variance, and The adjustment factor with positive weights. The bias constant is To prevent minute amounts with a denominator of zero, a fixed gating coefficient is used as a weighting factor to control the amount of information injected from the active component vector to the inertial component vector. Large updates to the inertial component vector are only allowed when the active component vector satisfies the stability condition of high mean and low variance.

[0008] Preferably, in the first update strategy, the decay factor used in the scalar decay calculation is determined based on the topological depth of the node; the hierarchical depth value of each node in the knowledge graph is calculated under the directed acyclic graph structure; according to the hierarchical depth value, an independent decay rate constant is assigned to each node, and the decay rate constant is configured to be negatively correlated with the hierarchical depth value, so that nodes in the deeper layers of the graph have a slower decay rate of active components than shallower nodes.

[0009] Preferably, the graph convolution operation in the second update strategy further includes: calculating the weights of the connection edges between the target node and each node in the multi-hop neighbor node set and the weighted sum of the corresponding bimodal hidden state vectors of the neighbor nodes; mapping the interactive input features into feature vectors and fusing them with the weighted sum; processing the fused features through a nonlinear activation function to generate an inertial correction gradient for the multi-hop neighbor node set; and using the inertial correction gradient to inversely update the inertial component vectors of each node in the multi-hop neighbor node set, with the update magnitude being proportional to the connection edge weights and the magnitude of the predicted residual value.

[0010] Preferably, the local topology self-calibration step further includes performing a weight-based sparsity truncation operation: after adjusting the weights of the connecting edges, it is determined whether the adjusted weights of the connecting edges are lower than a preset sparsity threshold; if they are lower than the sparsity threshold, the weights of the connecting edges are reset to zero, and the connecting edges are removed from the sparse adjacency matrix to disconnect the information transmission channel between neighboring nodes and the target node, thereby maintaining the sparsity of the computation graph.

[0011] Preferably, the method further includes performing a graph dynamic coarsening step based on state saturation: real-time monitoring of the magnitude saturation of the inertial component vectors of each node in a preset substructure region of the knowledge graph; when the magnitude saturation of all nodes in the substructure region exceeds a preset stability threshold, triggering a graph coarsening operation to aggregate all nodes in the substructure region into a virtual supernode; reconstructing the sparse adjacency matrix, mapping the original connecting edges pointing to each node in the substructure region to connecting edges pointing to the virtual supernode, and pausing the independent state updates of the original nodes in the region until the prediction residual value for the virtual supernode exceeds the restoration threshold.

[0012] Preferably, the inertial component vector in the bimodal hidden state vector represents the long-term structural stability of the node state, and its update depends on the accumulated gradient momentum; the active component vector represents the instantaneous response of the node state to the current input, and its update depends only on the current interactive input features; the two components of the bimodal hidden state vector remain orthogonal in the vector space to achieve independent evolution and non-interference between long-term and short-term features.

[0013] Preferably, the method further includes the step of generating a predicted path based on the updated bimodal latent state vector: traversing candidate nodes in the knowledge graph and calculating the saturation of the inertial component vector of the preceding dependent nodes of each candidate node; using the saturation as a confidence factor to calculate the recommendation potential value of each candidate node in weighted manner; and generating a subsequent sequence of English learning path nodes in descending order of recommendation potential value.

[0014] Preferably, the steps of obtaining real-time interactive input features and interactive result labels for the target node specifically include: receiving interactive data packets from the client, parsing the data packets to extract the unique identifier of the target English knowledge point, the user's answer feature vector, and the binarized correct / incorrect labels; mapping the answer feature vector to an input vector with the same dimension as the bimodal hidden state vector; and converting the correct / incorrect labels into a one-hot encoding format as the benchmark true value for calculating the prediction residual value.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the graph neural network inference process, a prediction bias-based triggering logic is introduced to construct an event-driven dynamic computing architecture. The architecture uses the numerical difference between the real-time interaction results and the forward prediction values ​​as a gating signal. Under normal circumstances where the deviation value is below a preset threshold, the graph topology is locked and low-overhead scalar decay updates are performed only on the current node, bypassing the expensive adjacency matrix multiplication operation. Only when the deviation value exceeds the limit is the local subgraph centered on the target node dynamically activated and multi-hop graph convolution operations are performed. The non-uniform computing resource allocation strategy reduces the average time complexity from linear or quadratic at the full graph scale to local constant level at the algorithm level, reducing the real-time inference load on the end devices. In terms of signal propagation mechanism, a physical isolation barrier is established to block the invalid diffusion of low-confidence small fluctuation signals in the network topology and global parameter perturbations, ensuring that the model has a keen ability to capture abrupt features while improving the overall long-term convergence stability of the network.

[0016] 2. Construct a dual-modal orthogonal state space containing inertial and active components. The active component is configured to have high sensitivity to the input gradient at the current moment and to track rapid fluctuations in short-term contextual features in real time. The inertial component is configured as a long-term memory carrier with high damping characteristics, and its evolution is controlled by independent slow dynamic rules. The high-frequency transient signal and the low-frequency steady-state structure are physically separated. The data structure design solves the contradiction between catastrophic forgetting and plasticity in the continuous learning process of neural networks. This enables the model to accurately fit the current local input changes within a unified computational framework, while strictly maintaining the stability of the global knowledge structure, and avoiding long-term model parameter drift caused by short-term input noise.

[0017] 3. An embedded time-sliding window volatility gating mechanism is used in the state update loop to dynamically adjust the information flow to inject gain into the inertial component by utilizing the second-order statistical characteristics of the variance of the active component in the time dimension. The mechanism constructs a nonlinear transfer function that is negatively correlated with the variance. Based on the network's endogenous ability to distinguish between stable trends and random oscillations, when the input signal is detected to have high mean but high variance pulse characteristics, the gating unit automatically contracts and blocks gradient propagation to prevent unstable transient features from contaminating the long-term state. When the signal meets the stability conditions of high mean and low variance, substantial correction of the inertial component is allowed. The backpropagation gradient weighted modulation processing method is used to improve the model's ability to automatically filter out sudden interference and accurately solidify effective pattern signals in non-stationary input streams, thereby improving the confidence of the path prediction model in complex dynamic environments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the bimodal state evolution and path prediction process of residual gating in this invention; Figure 2 This is a comparison chart of the inference time and CPU resource usage of this invention and mainstream models; Figure 3 This is a timing diagram of the interaction between the first update strategy for topology depth and scalar decay in this invention. Figure 4 This is a schematic diagram of the English learning path prediction model architecture of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] A method for constructing an English learning path prediction model based on a knowledge state evolution graph is presented. This method runs on a server or edge computing terminal with tensor computation capabilities. The implementation relies on a layered, decoupled, event-driven system architecture. This architecture couples various functional entities through an internal high-dimensional tensor data bus, and is divided from bottom to top into a graph base layer, a state evolution layer, and a decision application layer. The graph base layer is responsible for the core architecture, including graph topology construction and state initialization modules, to establish a high-fidelity knowledge graph base, addressing the engineering problems of complex knowledge point dependencies and pseudo-mastery noise in user interaction data in online English learning scenarios. The state evolution layer integrates a bimodal state evolution engine, a cognitive dissonance discriminator, and a dynamic topology calibrator, responsible for conditional updates and structural corrections of hidden states over time. The decision application layer deploys a path potential generator, responsible for mapping the underlying states into executable learning sequences. Under this architecture, a sparse adjacency matrix of the target domain knowledge graph is first constructed. To define the graph topology, where The total number of knowledge points, and the non-zero elements in the matrix. Representation of knowledge points With knowledge points The strength of the directed dependency between nodes, used to distinguish between short-term memory and long-term mastery, is represented by the strength of the dependency between each node in the graph. Configure a bimodal hidden state vector This vector is formed by concatenating two mutually orthogonal subvectors, that is... ,in It is an inertial component vector used to store long-term structured knowledge that evolves slowly over time; The active component vector stores short-term memory features triggered by the current interaction context; both vectors have a dimension of 1. And maintain orthogonality to block the direct penetration of high-frequency noise from the active region to the inertial region.

[0021] The system processes real-time data streams through the following deterministic procedure: receiving interactive data packets from the client and parsing to extract target English knowledge points. Unique identifier, user's answer feature vector and binarized interactive result labels , Answer feature vector Includes normalized response time, key pressure, and number of retries, as well as interaction result tags. In this context, 1 represents correct and 0 represents incorrect, and the answer feature vector is... Mapped to an input embedding vector with the same dimension as the bimodal hidden state vector. Using the target node's current bimodal hidden state vector Calculate the prediction probability ,in Using the Sigmoid activation function, calculate the predicted probability. Labels related to actual interaction results The numerical deviation between them is defined as the predicted residual value. A residual-driven conditional update strategy is adopted to cope with the computational load of full-graph convolution computation on mobile devices, and a dynamic gating threshold is set. The initial value is set to 0.3, and it is dynamically adjusted based on the user's historical average residual, using the predicted residual value calculated in real time. and Compare the results and execute either the first or second update strategy accordingly; when... At this time, the first update strategy is executed. The computation graph is locked, and message passing in the form of matrix multiplication is not performed; only the current interactive input features are utilized. Update target node Active component vector The updated formula is as follows For the graph excluding the target node All other nodes Execution based on time interval Scalar attenuation calculation, i.e. During this process, the inertial component vectors of all nodes Keeping the numerical values ​​constant, the computational complexity is controlled within... Level, attenuation factor Based on node-based topology depth determination, the directed acyclic graph structure is pre-traversed to calculate the depth value of each node level. Assign an independent decay rate constant to each node. ,in and As a preset hyperparameter, this configuration ensures that the core grammatical architecture deep within the graph has a slower rate of active component decay than shallow lexical nodes.

[0022] when At that time, execute the second update strategy to target the node. Centered on the sparse adjacency matrix Locking in Multi-hop neighbor node set within the hop range , For example, if the value is 2, graph convolution is performed only on the subgraph range, calculating the target node and each node in the set. Weighted messages between Embedding interactive input features into vectors By fusing with this weighted message, an inertial correction gradient for the local region is generated through a nonlinear activation function. The gradient is used to inversely update the inertial component vectors of each node in the set, i.e. ,in As the learning rate, this mechanism utilizes the predicted residual value. As a gain factor, the long-term structural state is corrected only when the prediction deviation is large; while executing the second update strategy, a local topological self-calibration based on gradient correlation is performed in parallel for each neighbor node in the activated subgraph. Calculate the hidden state vector of the two modes. gradient direction vector of the predicted residual value Cosine similarity between ,based on Dynamically adjust the weight of the connecting edges ,when When, execute To strengthen dependency paths that contribute to accurate predictions; when When, execute To block the spread of misleading information, a weighted sparsity truncation operation is further performed, if the adjusted weights If the value is below the preset sparsity threshold of 0.01, the element in the sparse adjacency matrix is ​​set to zero, and the connection is broken to maintain the sparsity of the computation graph.

[0023] By introducing an inertial solidification mechanism based on second-order statistics to address the problem of inflated mastery caused by cramming, a length of [value missing] is maintained for each node after the active component is updated. A sliding time window is used to record the magnitude sequence of the active component vector. For example, if the value is 5, calculate the statistical mean of the sequence. With statistical variance The solidification gate coefficient is calculated using the following formula. : ,in For the Sigmoid function, and It is a positive weighting adjustment factor. For bias, To prevent small quantities with a denominator of zero, this formula constructs a nonlinear filter. When the active component exhibits steady-state characteristics of high mean and low variance, Approaching 1 allows for the injection of activity information into the inertial component; if the variance Too large Approaching 0, writing to long-term memory is blocked; a dynamic graph coarsening process based on state saturation is run to cope with the graph expansion caused by deeper learning. The inertial component vector magnitude saturation of each node within a preset substructure region is monitored in real time. When the magnitude saturation of all nodes within the region consistently exceeds a preset stability threshold of 0.85, a coarsening operation is triggered, aggregating all nodes within the region into a virtual supernode. Its state vector is taken as the average of the atomic nodes. The sparse adjacency matrix is ​​reconstructed, and the external connection edges originally pointing to nodes within the region are remapped to this virtual supernode. Supernodes are updated only during subsequent inference until the predicted residual value for that supernode exceeds the restoration threshold, triggering deconstruction restoration. Based on the updated full-graph bimodal hidden state vector, a predicted path is generated. Candidate nodes in the knowledge graph are traversed, and the saturation of the inertial component vector of the preceding dependent nodes of each candidate node is calculated. This saturation is used as a confidence factor, combined with the urgency calculated by the forgetting curve, to obtain the recommended potential value of each candidate node. The nodes are sorted from high to low according to their potential values, and a path sequence containing specific learning content identifiers is generated and output.

[0024] Example 1: In a deployment instance of an online English learning service system handling tens of millions of concurrent requests, the solution of this invention runs on a mobile edge computing node with limited computing resources. It is used to process path prediction and state tracking tasks under non-stationary learning sequences. When faced with complex situations where a large number of users engage in intensive interactions on specific high-difficulty knowledge points, such as subjunctive mood, accompanied by sudden forgetting, the solution parses the received interaction data packets and extracts the target node. Unique identifier, user's answer feature vector and interaction result tags In this high-frequency interaction scenario, the system does not perform full-graph convolution operation for every input, but instead utilizes the current bimodal hidden state vector of the target node. Calculate the prediction probability And obtain the predicted residual value. In most typical practice scenarios, the calculated prediction residual value... Below the dynamic gating threshold At this time, the first update strategy is executed, and the sparse adjacency matrix is... The defined graph topology is locked, blocking cross-node matrix message passing, and only the input embedding vector is used. For the target node Active component vector Perform local updates based on gated cyclic units, while simultaneously updating other nodes in the graph. Perform scalar decay calculation on the active component vector During this process, the decay rate constant With node level depth value The negative correlation configuration causes the active components of core grammatical architecture nodes in the deep layers of the graph to decay more slowly than those of shallow vocabulary nodes. This differential scalar decay mechanism based on topological depth works in synergy with the local update strategy to simulate the non-uniform forgetting pattern of knowledge at different levels through mathematical transformations without requiring high graph convolution computing power. This keeps the computational complexity of a single inference at a constant level, thereby solving the computing power bottleneck problem when edge devices process large-scale graphs.

[0025] When a user makes a series of mistakes on a foundation of knowledge that they previously had a solid grasp of, resulting in incorrect calculations... Greater than or equal to the dynamic gating threshold When cognitive dissonance is detected, a second update strategy is triggered, removing the target node from the list. and skip neighbor node set The locking mechanism allows graph convolution operations to be performed only within the local subgraph; in this case, the predicted residual value... It is converted into a gain factor, and compared with the structure-corrected gradient generated by graph convolution. Multiplication directly affects the inertial component vector. The reverse update process uses the predicted residual value as the trigger condition and gain coefficient for structural correction, achieving dynamic matching of computational resources and information. That is, computational power is only used to correct the long-term memory structure when the model's expectations conflict with reality. For scenarios where users frequently practice the same knowledge point in a short period (i.e., cramming), an inertial solidification mechanism based on second-order statistics runs in parallel during the update process. This mechanism calculates the statistical mean of the active component vector magnitude sequence by maintaining a sliding time window in real time. With statistical variance Using the formula The generated solidification gate coefficient As a one-way valve, it controls the information injection from the active component to the inertial component. When the user performs sudden training, causing drastic fluctuations in the active component, i.e., statistical variance... When it is large, the fixed gate coefficient Approaching zero, physically blocking the high-frequency noise from spreading to the inertial component vector. The system allows information to be written into the inertial component only when the active component exhibits a steady-state characteristic with high mean and low variance. This mechanism uses volatility as a negative feedback term to resolve the contradiction between long-term structural stability and short-term memory plasticity faced by neural networks in continuous learning, ensuring the stability of the prediction model under non-stationary input flow. Based on the updated full-graph bimodal hidden state vector, the system generates and outputs a path sequence containing specific learning content identifiers, achieving accurate tracking and prediction of complex cognitive states with low computing power consumption.

[0026] Example 2: To objectively verify the effectiveness and technical advantages of the proposed dual-modal state evolution and conditional update mechanism in handling real-world non-stationary learning sequences, this example constructs a simulated experimental environment based on real online education data streams. The experiment aims to compare the performance of the proposed solution with existing mainstream knowledge tracing models under complex conditions such as noise interference, sudden forgetting, and computational constraints. The experimental platform uses the publicly available EdNet dataset as the basic data source. To simulate the resource-constrained environment of mobile edge computing, the experiment runs on simulation nodes configured with single-core ARM architecture processors and low-power memory. To reproduce data noise and transmission jitter in the real network environment, 15% of random accidental touch labels are injected into the original interaction sequence, simulating user errors caused by operational mistakes rather than knowledge deficiencies. By introducing a non-uniform time interval distribution, the intermittent learning behavior of users is simulated to test the model's ability to fit long-term forgetting patterns.

[0027] The experimental design includes a multi-dimensional comparison system between the experimental group and the control group. The control group uses a standard deep knowledge tracing model with a single long short-term memory network state vector and performs a full update at each time step. The experimental group of this invention fully deploys a dual-modal hidden state vector, a residual-driven gating update strategy, and an inertial solidification mechanism based on second-order statistics. The core monitoring indicators of the experiment include prediction accuracy, root mean square error, average time consumption per inference, and CPU utilization. After the experiment starts, interactive data streams are fed to each model in time stamp order. In the early stage of processing, when the user is in a stable learning phase, the experimental group of this invention utilizes the prediction residual values. With dynamic gating threshold The comparison mechanism mostly operates in fast-path update mode. During this stage, the CPU utilization of the sample group in this invention remains around 30% of that in the control group, and the predicted AUC value is on par with the control group. When the simulated user suddenly makes consecutive mistakes on knowledge points they have already mastered, i.e., cognitive dissonance occurs, the model in the control group, lacking a mechanism to filter short-term noise, experiences rapid and significant oscillations in its single state vector, leading to overfitting in subsequent estimations of mastery of that knowledge point. In contrast, the sample group in this invention detects... This triggers slow path updates, but the inertial fixation mechanism's fixation gating coefficient... Due to the variance of active components The sharp increase in noise approaches zero, blocking the penetration of this high-frequency noise into the inertial component; see Table 1, under the condition of injecting 15% noise, the AUC value of the sample group of the present invention remains at 0.785, while that of the control group decreases to 0.712.

[0028] Table 1: Performance Comparison Data of Different Models under Noise Interference Environment ; Further experimental data shows that as the proportion of noise, a core variable, increases, the performance decline of the sample group in this invention is significantly slower than that of the control group, exhibiting stronger stability. This nonlinear performance retention capability is not accidental, but stems from the inherent mechanism of this invention, which uses the prediction residual as a closed-loop control signal. This ensures that the model only corrects long-term memory when it has sufficient confidence, thereby maximizing the shielding of random disturbances while maintaining sensitivity to real changes. The final experimental results confirm that the solution of this invention can achieve higher prediction accuracy and lower resource consumption than existing technologies in real-world application scenarios with limited computing power and high noise levels, thus achieving the design goals.

[0029] Example 3: This example combines Figures 1 to 3 This paper describes a method for constructing an English learning path prediction model based on a knowledge state evolution graph. Figure 1 As shown, the overall operation of this English learning path prediction model begins with the monitoring of high-saturation regions and aggregation of virtual supernodes by the dynamic coarsening module of the graph. It then enters the interactive data acquisition stage, parsing unique identifiers and answer feature vectors, and performing bimodal latent state acquisition to extract orthogonal inertial and active components. The core of the process then enters the probability prediction and residual calculation stage, calculating the deviation between the predicted probability and the true value, i.e., the residual value. Next, it enters the residual threshold discrimination step, comparing the residual value with the dynamic gating threshold: if the residual is less than the threshold, the system executes the first update strategy along the left branch, updating only the active component of the target node. The system calculates the residuals of non-target nodes based on topological depth and performs scalar decay calculations. If the residuals are greater than or equal to the threshold, the system executes a second update strategy along the right branch, locks and activates the multi-hop neighbor node subgraph, performs graph convolution structure correction to generate gradient-corrected neighbor node inertial components, and performs local topological self-calibration based on gradient correlation to dynamically adjust edge weights. The left and right branches eventually converge on the inertial solidification mechanism. Based on second-order statistics, namely mean and variance, the injection amount is controlled, and the path potential energy generation step is entered to calculate the recommended potential energy value of candidate nodes. Finally, the path node sequence is output to generate a personalized English learning path.

[0030] like Figure 2 As shown in the figure, the horizontal axis lists four comparison models, and the vertical axis simultaneously represents the time consumed in milliseconds and the percentage of CPU utilization. The legend clearly indicates that the horizontal bars represent the average inference time, and the diagonal bars represent the CPU utilization. The data shows that the traditional GNN model has the highest inference time and CPU utilization, followed by the DKT model and the LSTM model, while the horizontal and diagonal bars corresponding to this invention are shorter than the other three models. Figure 3As shown, the process begins with the interactive input sending the answer feature vector to the target node. The target node reads the current bimodal hidden state and separates the active component vector and the inertial component. It sends the active component and the input feature to the GRU update unit for gated recurrent unit calculation and returns the updated active component. At the same time, the inertial component remains unchanged. The target node triggers a full-map decay command to the decay calculator. The decay calculator queries the topological depth table for the depth of each node level and obtains the returned depth value sequence. Then it traverses other nodes, calculates the differential decay rate based on the depth, so that the decay of deeper nodes is slower, and performs scalar decay update of the active component for other nodes while keeping their inertial components unchanged. Finally, after the decay is completed, it notifies the target node to reassemble the bimodal hidden state vector.

[0031] like Figure 4 As shown, the graph topology construction and state initialization module completes the tasks of constructing a sparse adjacency matrix, configuring the dual-modal hidden state vector, and achieving orthogonal splicing of the inertial and active components, thereby providing initialization configuration for subsequent units. Simultaneously, the system utilizes the interactive perception and parsing module to receive data, parse unique identifiers, answer features, and result labels, and map them into input embedding vectors. These input embedding vectors then enter the cognitive dissonance discriminator, where the predicted residual value is calculated and compared with a dynamic gating threshold. Based on the comparison result, a first control signal or a second control signal is generated. These signals are transmitted to the dual-modal state evolution engine (core computing unit). If the engine receives the first control signal... If the engine receives a second control signal, it triggers the fast path update unit, which responds to the first control signal (low residual) and performs only the update of the active component, while also performing full graph weight decay. If the engine receives a second control signal, it triggers the slow path graph convolution unit, which responds to the second control signal (high residual) and locks multi-hop neighbors, generating structure-corrected gradient-corrected inertial components. During slow path execution, the dynamic topology calibrator collaborates in parallel with the core unit, responsible for calculating gradient correlation cosine similarity and dynamically adjusting edge weights and performing sparsity truncation. After the state update is completed, the data flows to the path potential energy generator, which calculates the recommended potential energy value based on the saturation of the inertial components, and finally outputs the learned path sequence in order.

[0032] Example 4: In a continuous learning tracking scenario spanning several months, as users deepen their mastery of the English knowledge system, the statistical characteristics of the interaction data will experience baseline drift, facing the dual challenges of static hyperparameter failure and computational graph expansion. To address this challenge, a baseline topology is established during the initialization phase. By parsing the logical inclusion relationships defined in the pre-set expert knowledge base, knowledge points are... Points of knowledge Preorder dependency strength is mapped to initial weights in the sparse adjacency matrix. Strong dependencies are assigned a value of 0.8, and weak dependencies are assigned a value of 0.3, constructing an initial directed acyclic graph structure. As interactive data continues to flow in, dynamic gating thresholds are activated. An adaptive calibration process is implemented to prevent the slow path update mechanism from being permanently bypassed due to a general decrease in prediction residuals caused by improved user capabilities. The maintenance capacity is [missing information]. The historical residual observation queue is used to calculate the most recent data in the queue in real time. Arithmetic mean of the predicted residuals generated by each interaction Based on this, the momentum update formula is used. The gating threshold is adjusted periodically, among which... With a preset sensitivity margin, this negative feedback adjustment mechanism ensures that the gating threshold always hovers above the user's current average error level, triggering expensive graph convolution operations only when abnormal interactions deviating from the current capability baseline are captured, thereby maintaining constant sensitivity to cognitive dissonance events throughout the entire lifecycle.

[0033] When it is detected that all nodes in a specific substructure region, such as a relative clause block, satisfy the crystallization conditions of inertial component saturation greater than 0.85 and state fluctuation rate less than 0.05, a graph coarsening operation is immediately performed to compress the computational scale and lock the set of nodes within that region. Calculate the arithmetic mean of the current bimodal hidden state vectors of all nodes in the set, and define it as the newly generated virtual supernode. The state vector, i.e. Remap the connecting edges, for any edge located in the set External nodes The connection weights pointing to virtual supernodes are calculated as the original set that the external node originally pointed to. The average weight of each node in the process, i.e. Through this centroid aggregation and edge weight normalization process, the originally complex local subgraph network is collapsed into an equivalent single-node functional unit, which reduces the matrix dimension in the subsequent reasoning process. Before the user's ability rollback triggers the deconstruction and restoration, the supernode participates in the signal propagation and state evolution of the whole graph as a whole, and the computational load decreases inversely as the user's ability increases.

[0034] Example 5: In the engineering initialization phase of deploying a completely new English learning path prediction system from scratch, due to the lack of historical interaction data, the sparse adjacency matrix... The initial state construction directly determines the performance of the system during the cold start phase. To address the issue of uncertain initial state, this embodiment employs an offline calibration and data filling procedure based on expert priors. It imports a standardized English teaching syllabus and expert knowledge base, defines the logical inclusion and preorder dependencies between knowledge points, and addresses each pair of knowledge points... and If the expert database is defined for If there is a strong preorder dependency, then in the adjacency matrix... Initialize to 0.8; if defined as weakly correlated, initialize to 0.3; if there is no direct correlation, set to 0. This rule-based initialization strategy provides a physically meaningful starting point, ensuring that in the early stages when there is a lack of data-driven gradients, graph convolution operations can still propagate information along a path that conforms to the teaching logic.

[0035] Before deployment is completed and actual traffic is connected, to ensure dynamic gating thresholds With curing gate coefficient Key hyperparameters can be adapted to specific hardware environments and user group characteristics. A set of on-site pre-deployment calibration procedures are executed. In a controlled grayscale testing environment, a representative sample user data stream is introduced, and the gating threshold is set. Set a relatively wide initial range, such as 0.1 to 0.5, and enable full logging mode. Analyze the data under different conditions. Under certain values, the frequency of slow path updates and the corresponding prediction accuracy change curves are used to determine the baseline threshold for balancing computational load and model performance, and to adjust the factors in the inertia fixation mechanism. and By monitoring the variance of active components in the sample data stream By considering the distribution characteristics, these factors are adjusted to ensure that the solidified gating can effectively distinguish between normal learning fluctuations and high-frequency noise caused by cramming, and to ensure that the core control logic is at the optimal operating point adapted to the current environment when it is officially launched.

[0036] Example 6: To verify the engineering reproducibility of the present invention in different application environments, this example establishes a standardized system deployment pre-deployment calibration and model building procedure. The aim is to determine the optimal operating point of the initial dependency weights and dynamic gating threshold of the sparse adjacency matrix through quantitative experiments, thereby eliminating the black box in engineering implementation. Addressing the lack of objective basis for the initial dependency weights, this procedure adopts an automated calibration method based on curriculum structure analysis for knowledge points. and By analyzing the tree structure of the standard teaching syllabus, the path distance between the two in the knowledge tree is calculated. Using the inverse attenuation formula Calculate the initial dependency weights, where Based on the fundamental correlation constant, to determine To find the optimal value, perform a parameter scan on a small test set and set... The scanning range is 0.1 to 1.0, with a step size of 0.1, and in each... Under the given values, run the model and calculate its prediction accuracy for the first 5 interactions of users on the test set. Select the value at which the prediction accuracy reaches its peak. The value serves as the basic association constant during formal deployment, thereby ensuring that the initialization of the adjacency matrix is ​​based on objective structural quantification indicators.

[0037] For dynamic gating thresholds To address the initial setup issues, this procedure employs a baseline calibration process based on historical residual distribution. This involves collecting historical interaction data from the target user group without gating mechanisms enabled, and then calculating the predicted residuals. The probability density function, based on the normal distribution assumption, is used to calculate the mean of the statistical residuals. and standard deviation , set the initial gating threshold Set as ,in The sensitivity coefficient is used to determine... The optimal value is obtained by simulating different scenarios in an offline environment. The system behavior at this threshold is monitored, including the proportion of slow path updates triggered at this threshold and the corresponding improvement in prediction performance. Selection is made based on the ability to improve prediction performance while keeping the computational load within an acceptable range. The value is used to calculate the final initial gating threshold. This procedure transforms fuzzy empirical settings into quantitative calculations based on data statistical characteristics, ensuring adaptability under different data distributions. Simultaneously, to further adapt to the computing characteristics of edge devices, this invention maps the storage and update logic of the dual-modal hidden state vectors to a hierarchical resource scheduling mechanism for the processor. Specifically, in memory space allocation, the system performs hierarchical storage based on the update frequency differences of vector components, mapping inertial component vectors with longer update cycles to low-power cold data storage areas, while residing active component vectors requiring real-time response in the processor's cache or high-frequency hot data areas, thereby reducing the bandwidth overhead of bus read / write operations. At the computing power scheduling level, the logic level output by the cognitive dissonance discriminator directly serves as the trigger signal for instruction set scheduling. When the output level is low, the processor only calls the basic arithmetic logic unit to execute scalar decay instructions, maintaining low-power standby operation. Only when the output level is high, indicating that the residual exceeds the limit, is the neural network processing unit or graphics computing core awakened to execute parallel graph convolution operations, thus achieving adaptive allocation of computing resources at the hardware level according to the fluctuations in cognitive state.

[0038] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for constructing an English learning path prediction model based on a knowledge state evolution graph, characterized in that, Includes the following steps: Construct a sparse adjacency matrix of the target domain knowledge graph to define the graph topology, and configure a bimodal hidden state vector for each node in the knowledge graph. The bimodal hidden state vector is composed of mutually orthogonal inertial component vectors and active component vectors. Obtain real-time interactive input features and interactive result labels for the target node, calculate the prediction probability using the bimodal latent state vector of the target node, and calculate the numerical deviation between the prediction probability and the interactive result label. Define the numerical deviation as the prediction residual value. The predicted residual value is compared with a preset dynamic gating threshold, and the first update strategy or the second update strategy is executed according to the comparison result to evolve the state of the graph. If the predicted residual value is less than the dynamic gating threshold, the first update strategy is executed: only the interactive input features are used to update the active component vector of the target node, and the active component vectors of other nodes in the knowledge graph other than the target node are calculated based on the time interval, while keeping the values ​​of the inertial component vectors of other nodes unchanged. If the predicted residual value is greater than or equal to the dynamic gating threshold, execute the second update strategy: lock the set of multi-hop neighbor nodes of the target node in the sparse adjacency matrix, perform graph convolution operation based on the predicted residual value to generate the structure correction gradient, and use the structure correction gradient to correct the inertial component vector of each node in the set of multi-hop neighbor nodes.

2. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The second update strategy also includes performing a local topology self-calibration step based on gradient correlation: after determining the set of multi-hop neighbor nodes, for each neighbor node in the set, the cosine similarity between the bimodal hidden state vector of the neighbor node and the gradient direction vector of the predicted residual value is calculated. Based on cosine similarity, the weights of the connection edges between neighboring nodes and the target node in the sparse adjacency matrix are dynamically adjusted. When the cosine similarity indicates that the current state of the neighboring node is positively correlated with the correction direction of the predicted residual value, the weight of the connecting edge is increased to strengthen the gradient propagation to the source of error; when the cosine similarity indicates a negative correlation, the weight of the connecting edge is decreased to block the error propagation path in the same direction as the predicted residual value.

3. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, After updating the active component vector of the target node, the process also includes performing an inertial fixation step based on second-order statistics: maintaining a sliding time window, recording the magnitude sequence of the active component vector of the target node within the sliding time window; calculating the statistical mean and statistical variance of the magnitude sequence, and calculating the fixation gate coefficient using the following formula: ,in, To solidify the gate coefficient, It is the Sigmoid activation function. To calculate the statistical mean, To statistically analyze the variance, and The adjustment factor with positive weights. The bias constant is To prevent minute amounts with a denominator of zero, a fixed gating coefficient is used as a weighting factor to control the amount of information injected from the active component vector to the inertial component vector. Large updates to the inertial component vector are only allowed when the active component vector satisfies the stability condition of high mean and low variance.

4. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The decay factor used in the scalar decay calculation in the first update strategy is determined based on the topological depth of the nodes; the hierarchical depth value of each node in the knowledge graph is calculated under the directed acyclic graph structure. Each node is assigned an independent decay rate constant based on its hierarchy depth value. The decay rate constant is configured to be negatively correlated with the hierarchy depth value, with nodes in deeper layers of the graph having a slower decay rate of active components than nodes in shallower layers.

5. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The graph convolution operation in the second update strategy further includes: calculating the weights of the connection edges between the target node and each node in the multi-hop neighbor node set, and the weighted sum of the corresponding bimodal hidden state vectors of the neighbor nodes; mapping the interactive input features into feature vectors, and fusing them with the weighted sum; processing the fused features through a nonlinear activation function to generate an inertial correction gradient for the multi-hop neighbor node set; and using the inertial correction gradient to inversely update the inertial component vectors of each node in the multi-hop neighbor node set, with the update magnitude being proportional to the connection edge weights and the magnitude of the predicted residual value.

6. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 2, characterized in that, The local topology self-calibration step also includes performing a weight-based sparsity truncation operation: after adjusting the weights of the connecting edges, it is determined whether the adjusted weights of the connecting edges are lower than a preset sparsity threshold. If the value is below the sparsity threshold, the weight of the connecting edge is reset to zero, and the connecting edge is removed from the sparse adjacency matrix to disconnect the information transmission channel between the neighboring node and the target node, thus maintaining the sparsity of the computation graph.

7. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The method also includes performing a graph dynamic coarsening step based on state saturation: real-time monitoring of the magnitude saturation of the inertial component vectors of each node in a preset substructure region of the knowledge graph; when the magnitude saturation of all nodes in the substructure region exceeds a preset stability threshold, a graph coarsening operation is triggered to aggregate all nodes in the substructure region into a virtual supernode. Reconstruct the sparse adjacency matrix, map the original connecting edges pointing to each node in the substructure region to connecting edges pointing to the virtual supernode, and pause the independent state updates of the original nodes in the region until the prediction residual value for the virtual supernode exceeds the restoration threshold.

8. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The inertial component vector in the bimodal hidden state vector characterizes the long-term structural stability of the node state, and its update depends on the accumulated gradient momentum. The active component vector represents the instantaneous response of a node's state to the current input, and its update depends only on the current interactive input features; The two components of the bimodal hidden state vector remain orthogonal in the vector space.

9. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The method also includes the step of generating a predicted path based on the updated bimodal hidden state vector: traversing the candidate nodes in the knowledge graph and calculating the saturation of the inertial component vector of the preceding dependent nodes of each candidate node; using the saturation as a confidence factor to calculate the recommendation potential value of each candidate node in weighted manner; and generating the subsequent English learning path node sequence in descending order of recommendation potential value.

10. The method for constructing an English learning path prediction model based on a knowledge state evolution graph according to claim 1, characterized in that, The steps for obtaining real-time interactive input features and interactive result labels for the target node specifically include: receiving interactive data packets from the client, parsing the data packets to extract the unique identifier of the target English knowledge point, the user's answer feature vector, and the binarized correct / incorrect labels; mapping the answer feature vector to an input vector with the same dimension as the bimodal hidden state vector; and converting the correct / incorrect labels into a one-hot encoding format as the benchmark true value for calculating the prediction residual value.

Images