Knowledge tracking method based on hybrid convolution

Abstract

The invention relates to the technical field of knowledge tracking, in particular to a knowledge tracking method based on hybrid convolution, and the method comprises the steps: obtaining the interaction information of a learner and education content; the interaction information is input into a preset knowledge tracking model, the answer correct probability of a target question is predicted, the knowledge tracking model is used for constructing an input sequence based on the interaction information, behavior characteristics in different time ranges in the sequence are extracted through multi-scale causal convolution, and the answer correct probability of the target question is predicted; modeling is carried out in combination with an attention mechanism with a distance penalty term, and finally knowledge tracking is completed in combination with question answering correctness probability prediction of the target question. According to the invention, knowledge tracking is carried out in combination with multi-scale causal convolution and a distance decay attention mechanism, and the ability of modeling the cognitive state of a learner is improved.

Description

Technical Field

[0001] This invention relates to the field of knowledge tracing technology, and in particular to a knowledge tracing method based on hybrid convolution. Background Technology

[0002] Knowledge tracing (KT) is a key technology in intelligent education systems. Its core objective is to dynamically estimate learners' mastery of knowledge components (KCs) based on their historical interaction data with teaching content. This technology is widely used in adaptive learning systems, intelligent tutoring systems, and online education platforms, providing support for personalized teaching recommendations, knowledge diagnosis, and learning path optimization. Currently, mainstream knowledge tracing methods in academia and industry are mainly based on deep learning architectures, specifically including sequence models based on recurrent neural networks (RNNs), such as the DKT (Deep Knowledge Tracing) model, and its subsequent improved versions such as DKT+ and DKT-F; attention models based on Transformer structures, such as SAKT, SAINT, and AKT; and methods such as DKVMN, LPKT, HawkesKT, and DTransformer.

[0003] While the above technologies have achieved certain results, they still have significant limitations when facing challenges such as the high dimensionality, non-stationarity, and individual heterogeneity of learning behavior sequences: (1) They fail to adequately model the heterogeneity of learners' cognitive behavior: Most current knowledge tracing models assume that learners with similar historical behavior sequences should also have similar future performance. This homogeneous modeling assumption ignores the individual differences among learners in terms of cognitive level, behavioral style, and learning pace, which can easily lead to the accumulation of prediction errors. (2) It is difficult to distinguish between long-term trends and short-term fluctuations in cognitive states: Learning behavior sequences usually include stable knowledge acquisition processes (such as cognitive accumulation brought about by repeated practice) and sudden short-term fluctuations (such as fatigue, guessing, and emotional fluctuations). Existing technologies often model both together, lacking the ability to distinguish between them, which leads to the model over-responding to transient abnormal behaviors, thereby reducing the accuracy and robustness of cognitive state estimation. (3) The attention mechanism relies too heavily on distant information: Although the KT model based on Transformer has strong global modeling capabilities, its standard attention mechanism assigns similar attention to all positions in the historical sequence, which easily introduces long-distance but irrelevant information noise, resulting in a decline in prediction performance. Especially when dealing with long sequences, it lacks a "distance-sensitive" modeling mechanism, making it difficult to effectively capture key local behavioral changes. (4) It cannot effectively deal with the phenomenon of "cognitive mirror illusion": In practice, there are a large number of "cognitive mirror illusion" samples, that is, learners with the same historical behavioral trajectory show significant differentiation in subsequent performance. This phenomenon reflects deep-seated cognitive heterogeneity, and existing models have not designed mechanisms to identify or mitigate such samples, leading to biased prediction results and affecting the practicality of personalized learning systems. (5) It lacks adaptive modeling capabilities for sequences of different lengths: The sequence lengths of interaction data from different learners vary greatly. Long sequences usually contain richer cognitive progress information, while short sequences are more volatile. Existing technologies often use a uniform model structure to process all samples without modeling and differentiating the statistical characteristics of sequence length, which can easily lead to overfitting of short sequences and underfitting of long sequences.

[0004] Students' actual knowledge mastery is often influenced by multiple factors, such as fatigue, guesswork, and motivation. These short-term or random fluctuations often mask their true cognitive level, thus affecting the accuracy of predictions. Therefore, there is an urgent need for a knowledge tracking method that can model cognitive changes at different time scales, distinguish between stable cognitive progress and temporary behavioral fluctuations, and adaptively model individual differences, in order to further improve the effectiveness and robustness of personalized learning systems. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge tracking method based on hybrid convolution, which combines multi-scale causal convolution with distance decay attention mechanism to track knowledge and improve the ability to model learners' cognitive states in the process of personalized learning.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] Knowledge tracing methods based on hybrid convolutions include:

[0008] Obtain information about learners' interactions with educational content;

[0009] The interactive information is input into a preset knowledge tracking model to predict the correct answer probability of the target question. The knowledge tracking model is used to construct an input sequence based on the interactive information, extract behavioral features of different time ranges in the sequence using multi-scale causal convolution, and then combine it with an attention mechanism with a distance penalty term for modeling. Finally, the knowledge tracking is completed by combining the correct answer probability prediction of the target question.

[0010] Optionally, the knowledge tracing model includes:

[0011] The context embedding encoding module is used to perform context embedding encoding based on the interaction information and construct a context vector.

[0012] A multi-scale causal convolution module is used to extract features and decouple the frequency domain from the context vector to obtain the decomposed features;

[0013] The attention module is used to process the decomposed features using an attention mechanism with a distance penalty term to obtain attention output;

[0014] The prediction module is used to concatenate the attention output with the target question to make a prediction and output the probability of answering the target question correctly.

[0015] Optionally, the interactive information includes: question identifier. The set of knowledge points Correctness mark .

[0016] Optionally, the context embedding encoding module performs context embedding encoding based on the interaction information to construct a context vector, including:

[0017] ;

[0018] ;

[0019] in, and For context vectors, Embedded with the difficulty level of the question. , For the change in knowledge points, , Embedded for one-hot encoding, , These represent element-wise multiplication and addition, respectively.

[0020] Optionally, the multi-scale causal convolution module performs feature extraction and frequency domain decoupling on the context vector to obtain the decomposed features, including:

[0021] The context vector is divided according to its length, and convolution operation is performed on the divided context vector based on different time scales. Frequency domain decoupling is also performed to obtain low-frequency stable cognitive features and high-frequency fluctuation features.

[0022] The final feature is obtained by fusing the low-frequency stable cognitive features and high-frequency fluctuating features based on a lightweight attention mechanism.

[0023] Optionally, obtaining the low-frequency stable cognitive features and high-frequency fluctuation features includes:

[0024] ;

[0025] ;

[0026] in, This is a low-frequency, stable cognitive characteristic. It exhibits high-frequency fluctuation characteristics. For the first Layer convolution kernel, For the first kernel size of each layer Given the input sequence of interactive features, This represents the current position at the current time step. This represents the time offset steps in the convolution operation.

[0027] Optionally, the attention mechanism with a distance penalty term in the attention module is as follows:

[0028] ;

[0029] ;

[0030] in, For position Query vectors and positions Attention weights between key vectors on the key vectors For adjustable temperature parameters, For position The query vector on For position The key vector on, For position With position Distance penalty between , For learnable parameters, Indicates the distance between positions in the sequence. This represents the number of positions in the sequence.

[0031] Optionally, the prediction module concatenates the attention output with the target question and then calculates the probability of correctly answering the target question through a two-layer fully connected network, including:

[0032] ;

[0033] ;

[0034] ;

[0035] in, This is the output of the first layer fully connected network. This is the output of the second-layer fully connected network. For attention output, Embedded for the target question, , This is the weight matrix of the fully connected layer. , For the bias term of the fully connected layer, For the Sigmoid function, This is the predicted result.

[0036] Optionally, the objective function of the knowledge tracing model adopts a binary cross-entropy loss, which is:

[0037] ;

[0038] in, This represents the binary cross-entropy loss value. This is the actual result.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention combines multi-scale causal convolution with distance-decaying attention mechanisms for knowledge tracking, which improves the modeling accuracy of learners' cognitive states in personalized learning, enhances the model's adaptability to heterogeneous learning behaviors, suppresses long-range noise interference from the attention mechanism, and constructs a knowledge tracking model with good generalization ability and parallel efficiency. Through structural modeling, this invention demonstrates significantly superior performance compared to existing technologies in multiple dimensions, including modeling accuracy, behavioral adaptability, interference suppression, and deployment performance, and possesses promising prospects for practical application and widespread application value. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a schematic diagram illustrating the structure and workflow of the knowledge tracking model in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] This embodiment provides a knowledge tracing method based on hybrid convolution, such as... Figure 1 As shown, it includes:

[0046] Obtain information about learners' interactions with educational content;

[0047] The interactive information is input into a preset knowledge tracking model (MoC-KT) to predict the correct answer probability of the target question. The knowledge tracking model is used to construct an input sequence based on the interactive information, extract behavioral features of different time ranges in the sequence using multi-scale causal convolution, and then combine it with an attention mechanism with a distance penalty term for modeling. Finally, the knowledge tracking is completed by combining the correct answer probability prediction of the target question.

[0048] Furthermore, the knowledge tracing model includes:

[0049] The context embedding encoding module is used to perform context embedding encoding based on the interaction information and construct a context vector.

[0050] The multi-scale causal convolution module (MoC) is used to extract features from the context vector and decouple it in the frequency domain to obtain the decomposed features.

[0051] The attention module is used to process the decomposed features using an attention mechanism with a distance penalty (Kerple-EnhancedAttention) to obtain attention output;

[0052] The prediction module is used to concatenate the attention output with the target question to make a prediction and output the probability of answering the target question correctly.

[0053] Furthermore, the interactive information includes: question identifier. The set of knowledge points Correctness mark .

[0054] Specifically, this embodiment constructs an input sequence based on the learner's interaction records with the educational content. The input at each time step consists of a triple: a question identifier... The set of knowledge points Correctness mark .

[0055] Furthermore, the context embedding encoding module performs context embedding encoding based on the interaction information to construct a context vector, including:

[0056] To enhance representation capabilities, a context vector is constructed that integrates elements such as question content, knowledge points, question difficulty, and the correctness of the answer, defined as:

[0057] ;

[0058] ;

[0059] in, Embedded with the difficulty level of the question. , This represents the change in the amount of knowledge points. , Embedded for one-hot encoding, , These represent element-wise multiplication and addition, respectively.

[0060] This encoding method improves the model's ability to express differences in question types and answer behaviors.

[0061] Furthermore, the multi-scale causal convolution module performs feature extraction and frequency domain decoupling on the context vector to obtain the decomposed features, including:

[0062] The context vector is divided according to its length, and convolution operation is performed on the divided context vector based on different time scales. Frequency domain decoupling is also performed to obtain low-frequency stable cognitive features and high-frequency fluctuation features.

[0063] The final feature is obtained by fusing the low-frequency stable cognitive features and high-frequency fluctuating features based on a lightweight attention mechanism.

[0064] Specifically, to address the aliasing of long-term progress and short-term fluctuations in cognitive states, a multi-scale causal convolutional structure is introduced to decouple the input sequence in the frequency domain, resulting in low-frequency stable cognitive features (long-term trends) and high-frequency fluctuating features (short-term fluctuations).

[0065] ;

[0066] ;

[0067] ;

[0068] in, This is a low-frequency, stable cognitive characteristic. It exhibits high-frequency fluctuation characteristics. For the first Layer convolution kernel, For the first kernel size of each layer Given the input sequence of interactive features, This represents the current position at the current time step. This represents the time offset steps in the convolution operation. It is a learnable high-frequency control factor.

[0069] Low-frequency stable cognitive features (long-term trends) refer to stable cognitive patterns extracted from continuous learning interaction information over long time series, reflecting the long-term accumulation trend of learners' overall knowledge mastery level. This feature is extracted from large receptive field kernels in multi-scale convolutional structures, and can effectively capture stable cognitive signals across topics and time periods, such as the trend of continuously mastering or forgetting a certain knowledge point.

[0070] High-frequency fluctuation features (short-term fluctuations) refer to the rapid changes in learners' response states during short-term interactions, such as fluctuations in attention, guessing behavior, or short-term memory effects. Extracted from small receptive field nuclei, this feature emphasizes rapid response changes within local timeframes and is used to reveal the perturbation or correction effect of short-term cognitive states on prediction results.

[0071] In addition, to adapt to interaction sequences of different lengths L, a segmented modeling strategy is adopted. The boundary threshold is set according to the distribution statistics of the sequence length vector, and three categories of short, medium and long intervals are defined. The samples are divided into three categories of short / medium / long according to the length of the interaction sequence, and are processed by different convolution channels to capture behavioral patterns at different time scales. A lightweight attention mechanism is used to fuse low-frequency and high-frequency features.

[0072] Furthermore, the attention mechanism with a distance penalty term in the attention module includes:

[0073] To overcome the problem of long-distance noise interference in traditional attention mechanisms, an attention mechanism with a distance decay factor (Kerple) is introduced. By adding a distance penalty term to the attention score, attention is guided to focus on local areas.

[0074] ;

[0075] ;

[0076] in, For position Query vectors and positions Attention weights between key vectors on the key vectors For adjustable temperature parameters, For position The query vector on For position The key vector on, For position With position Distance penalty between , For learnable parameters, Indicates the distance between positions in the sequence. This represents the number of positions in the sequence.

[0077] This mechanism can suppress attentional shifts to distant, irrelevant behaviors and improve the model's ability to respond to sudden changes in local behaviors.

[0078] Furthermore, the prediction module concatenates the attention output with the target question and then calculates the probability of correctly answering the target question through a two-layer fully connected network, including:

[0079] Specifically, the model's final output integrates the attention result with the current question embedding. The predicted probability is calculated using a two-layer fully connected network:

[0080] ;

[0081] ;

[0082] ;

[0083] in, This is the output of the first layer fully connected network. This is the output of the second-layer fully connected network. For attention output, Embedded for the target question, , This is the weight matrix of the fully connected layer. , For the bias term of the fully connected layer, For the predicted results, This is the Sigmoid function, used to output the probability of a correct answer.

[0084] Furthermore, the objective function of the knowledge tracing model employs binary cross-entropy loss to supervise and optimize the actual and predicted responses. The binary cross-entropy loss function is as follows:

[0085] ;

[0086] in, This represents the binary cross-entropy loss value. This is the actual result.

[0087] The knowledge tracing model structure constructed in this embodiment has strong parallel computing capabilities and interpretability, and is suitable for:

[0088] (1) Real-time monitoring of knowledge status in online education platforms;

[0089] (2) Dynamic feedback and recommendation of personalized intelligent teaching system;

[0090] (3) Cognitive modeling and answer behavior diagnosis in educational measurement scenarios.

[0091] The method proposed in this embodiment can be implemented in existing deep learning frameworks (such as PyTorch and TensorFlow), supports batch sequence processing, and is suitable for medium-length interactive sequence data (approximately 50-200 characters in length).

[0092] The method proposed in this embodiment has significant advantages in several aspects, including:

[0093] (1) Significantly improves the modeling accuracy of learners' cognitive states;

[0094] By using a multi-scale causal convolution module to decouple learner interaction sequences in the frequency domain, the model can effectively distinguish between long-term cognitive progress and short-term behavioral fluctuations, avoid misjudging occasional behaviors (such as guessing or fatigue), and thus improve the accuracy and stability of knowledge state representation.

[0095] Technical Performance: Without introducing additional features, MoC-KT significantly outperforms 24 existing mainstream knowledge tracking models on multiple real-world datasets. For example, on the NIPS34 dataset, MoC-KT achieves an AUC of 0.8048, higher than the current best model ReKT's 0.8025.

[0096] (2) Enhance the model's adaptability to heterogeneous learning behaviors;

[0097] A segmented modeling mechanism that is aware of sequence length is introduced. Based on the length distribution of interaction sequences of different learners, a dedicated convolutional structure is configured for modeling. This mechanism has the ability to adapt to differences in learning behavior at different time scales and solves the problem that a unified structure cannot take into account both short and long sequences.

[0098] Technical results: Compared with the uniform model that does not distinguish between sample types, the segmented modeling strategy improved the AUC by about 0.8% on the BD2006 dataset (from 0.8189 to 0.8264), verifying the effectiveness of modeling behavioral heterogeneity.

[0099] (3) Significantly suppresses long-distance noise interference in attention mechanisms;

[0100] By introducing a distance-sensitive attention decay mechanism (Kerple), we can actively suppress semantically irrelevant behavioral segments that are far away from the current position during long sequence modeling, thereby enhancing the model's ability to respond to locally important behaviors and avoiding attention dilution.

[0101] Technical results: In the Static2011 dataset (longer sequences, skewed data), MoC-KT achieved an AUC of 0.8311, which is higher than extraKT (0.8307) and AKT (0.8309), demonstrating its robust advantage in long sequence modeling.

[0102] (4) It has good generalization ability and parallel efficiency;

[0103] MoC-KT is built using pure convolution and attention mechanisms, naturally supports parallel computing, and has a compact model structure with reasonable parameter control. It does not require complex graph modeling or comparative learning processes, making it easy to deploy in online real-time inference systems on educational platforms.

[0104] Technical results: Experimental results show that MoC-KT significantly outperforms contrastive learning methods (such as DTransformer) in terms of training time, improving training efficiency by approximately 28.5% while maintaining accuracy.

[0105] In summary, this embodiment, through structural modeling, demonstrates significantly superior performance compared to existing technologies in multiple dimensions, including modeling accuracy, behavioral adaptability, interference suppression, and deployment performance, and possesses promising prospects for practical application and promotional value.

[0106] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A knowledge tracing method based on hybrid convolution, characterized in that, include: Obtain information about learners' interactions with educational content; The interactive information is input into a preset knowledge tracking model to predict the correct answer probability of the target question. The knowledge tracking model is used to construct an input sequence based on the interactive information, extract behavioral features of different time ranges in the sequence using multi-scale causal convolution, and then combine it with an attention mechanism with a distance penalty term for modeling. Finally, the knowledge tracking is completed by combining the correct answer probability prediction of the target question.

2. The knowledge tracking method based on hybrid convolution as described in claim 1, characterized in that, The knowledge tracing model includes: The context embedding encoding module is used to perform context embedding encoding based on the interaction information and construct a context vector. A multi-scale causal convolution module is used to extract features and decouple the frequency domain from the context vector to obtain the decomposed features; The attention module is used to process the decomposed features using an attention mechanism with a distance penalty term to obtain attention output; The prediction module is used to concatenate the attention output with the target question to make a prediction and output the probability of answering the target question correctly.

3. The knowledge tracking method based on hybrid convolution as described in claim 2, characterized in that, The interactive information includes: question identifier. The set of knowledge points Correctness mark .

4. The knowledge tracking method based on hybrid convolution as described in claim 3, characterized in that, The context embedding encoding module performs context embedding encoding based on the interaction information, and constructs a context vector including: ； ； in, and For context vectors, Embedded with the difficulty level of the question. , For the change in knowledge points, , Embedded for one-hot encoding, , These represent element-wise multiplication and addition, respectively.

5. The knowledge tracking method based on hybrid convolution as described in claim 2, characterized in that, The multi-scale causal convolution module performs feature extraction and frequency domain decoupling on the context vector to obtain the decomposed features, including: The context vector is divided according to its length, and convolution operation is performed on the divided context vector based on different time scales. Frequency domain decoupling is also performed to obtain low-frequency stable cognitive features and high-frequency fluctuation features. The final feature is obtained by fusing the low-frequency stable cognitive features and high-frequency fluctuating features based on a lightweight attention mechanism.

6. The knowledge tracking method based on hybrid convolution as described in claim 5, characterized in that, Obtaining the low-frequency stable cognitive features and high-frequency fluctuation features includes: ； ； in, This is a low-frequency, stable cognitive characteristic. It exhibits high-frequency fluctuation characteristics. For the first Layer convolution kernel, For the first kernel size of each layer Given the input sequence of interactive features, This represents the current position at the current time step. This represents the time offset steps in the convolution operation.

7. The knowledge tracking method based on hybrid convolution as described in claim 2, characterized in that, The attention mechanism with distance penalty in the attention module is as follows: ； ； in, For position Query vectors and positions Attention weights between key vectors on the key vectors For adjustable temperature parameters, For position The query vector on For position The key vector on, For position With position Distance penalty between , For learnable parameters, Indicates the distance between positions in the sequence. This represents the number of positions in the sequence.

8. The knowledge tracing method based on hybrid convolution as described in claim 2, characterized in that, The prediction module concatenates the attention output with the target question and then calculates the probability of correctly answering the target question through a two-layer fully connected network, including: ； ； ； in, This is the output of the first layer fully connected network. This is the output of the second-layer fully connected network. For attention output, Embedded for the target question, , This is the weight matrix of the fully connected layer. , For the bias term of the fully connected layer, For the Sigmoid function, This is the predicted result.

9. The knowledge tracing method based on hybrid convolution according to any one of claims 1-8, characterized in that, The objective function of the knowledge tracing model uses binary cross-entropy loss, which is: ； in, This represents the binary cross-entropy loss value. This is the actual result.

Images