A multi-feature knowledge tracking method based on an attention mechanism

Abstract

The present application relates to a kind of multi-feature knowledge tracking methods based on attention mechanism, comprising: the previous knowledge state of student is obtained;The previous knowledge state is input into multi-feature knowledge tracking model, outputs current knowledge state, and predicts the answering performance of student next time based on the current knowledge state;Wherein, the multi-feature knowledge tracking model is based on attention mechanism and fuses multi-dimensional information to calculate attention score, and introduces learning gate mechanism to control the difference between the knowledge state of student between two consecutive interactions.The method of the present application has rationality and effectiveness on knowledge tracking task, can accurately predict the knowledge state of student and provide personalized learning guidance.

Description

Technical Field

[0001] This invention relates to the field of knowledge tracking technology, and in particular to a multi-feature knowledge tracking method based on an attention mechanism. Background Technology

[0002] With the rise of digital and smart education, the demand for personalized learning is gradually increasing. Therefore, accurately analyzing students' knowledge status is becoming increasingly important. Knowledge tracking, as an important component of smart education systems, can dynamically track changes in students' knowledge status over time. However, existing knowledge tracking models still have some problems that need to be addressed: (1) They do not consider changes in students' understanding and absorption of different exercises and knowledge concepts, as well as changes in their forgetting levels. (2) They neglect the identification of individual student differences. Most existing models only consider the accuracy of responses and do not consider that the difficulty effect of questions varies for different students. (3) The use of feature information is not yet sufficient, and the use of multi-feature information needs to be strengthened. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-feature knowledge tracking method based on attention mechanisms, which can accurately predict students' knowledge status and provide personalized learning guidance.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A multi-feature knowledge tracking method based on an attention mechanism includes:

[0006] To obtain the student's prior knowledge status;

[0007] The previous knowledge state is input into the multi-feature knowledge tracking model, the current knowledge state is output, and the student's answer performance at the next moment is predicted based on the current knowledge state; wherein, the multi-feature knowledge tracking model calculates attention score by integrating multi-dimensional information based on attention mechanism, and introduces learning gate mechanism to control the difference in knowledge state between two consecutive interactions of the student.

[0008] Optionally, the multi-feature knowledge tracking model includes an attention mechanism module, a learning and absorption module, a target knowledge concept module, and a prediction module;

[0009] The attention mechanism module is used to calculate an attention score by integrating multi-dimensional information such as questions, answer time, and number of prompts through the attention mechanism.

[0010] The learning absorption module is used to obtain learning absorption based on the previous knowledge state and the knowledge concept vector of the current exercise, and to obtain actual learning absorption by adding the learning gate;

[0011] The target knowledge concept module is used to construct a relationship matrix between exercises and knowledge concepts, and to obtain the knowledge concept vector of the relationship matrix;

[0012] The prediction module is used to combine the attention score, the actual learning absorption, and the knowledge concept vector to obtain the current knowledge state, and to concatenate the question embedding, interval time, answer accuracy, and the current knowledge state to obtain the answer performance at the next moment.

[0013] Optionally, calculating an attention score by integrating multi-dimensional information such as questions, response time, and number of prompts through attention mechanisms includes:

[0014] Constructing learned embeddings with learned feature information:

[0015]

[0016] in, This is a vector concatenation operation, where W1 is the first weight matrix, b1 is the first bias term, and e t For the exercises, at is the embedding matrix of response times, hc is the embedding matrix of the number of prompts given by the student, the subscript t represents time t, and L... t For learning embeddings with learning feature information;

[0017] The learned answer is added to the learning embedding containing the learned feature information to obtain a new value label:

[0018] Based on the new value label, the attention score is calculated using a multi-head attention mechanism based on triangular masking to obtain the attention score.

[0019] Optionally, the method for calculating the attention score is as follows:

[0020]

[0021] Among them, AS t Let Q, K, and V be the attention score, representing the query, key, and value, respectively. For queries with learned embeddings, k τ For key-value pairs with learned embeddings, VL τ For value labels with learned embeddings, τ refers to the time before time t, and d k It is the embedded dimension.

[0022] Optionally, acquiring learning absorption, and acquiring actual learning absorption by adding learning gates, includes:

[0023] Multiply the previous knowledge state by the knowledge concept vector of the current exercise to obtain the previous relevant knowledge state;

[0024] The learning absorption is obtained based on the previously relevant knowledge state, the learning embedding with learning feature information, and the adjacent interaction interval time embedding.

[0025] The actual learning absorption is obtained by projecting the learning absorption range onto a preset range and then multiplying it by the learning gate.

[0026] Optionally, the calculation method for the learned absorption is as follows:

[0027]

[0028] Among them, la t For learning and absorption, W2 is the second weight matrix, b2 is the second bias term, and L... t-1 and L t These are the learned embeddings containing learned feature information from the previous and current time steps, respectively, and 'it' is the temporal embedding of the adjacent interaction interval. This refers to the state of prior related knowledge;

[0029] The calculation method for the learning gate is as follows:

[0030]

[0031] Among them, Γ t σ is the learning gate, σ is the non-linear sigmoid activation function, W3 is the third weight matrix, and b3 is the third bias term.

[0032] Optionally, the knowledge concept vector of the relation matrix is:

[0033]

[0034] Among them, KC t+1 Here, W4 is the knowledge concept vector, W4 is the fourth weight matrix, and b4 is the fourth bias term. To predict e t+1 When solving exercises, extract a specific row from the relation matrix by the exercise number.

[0035] Optionally, the method for predicting a student's performance at the next moment based on the current knowledge state is as follows:

[0036]

[0037] Among them, y t+1 For the student's performance at time t+1, W5 is the fifth weight matrix, b5 is the fifth bias term, and e t+1 For the problem embedding at time t+1, et t+1 Embed the time interval of the same exercise at time t+1, ct t+1 For the time interval embedding of the same knowledge concept at time t+1, ect+1 Embed the accuracy of the exercise answers at time t+1, cc t+1 The embedding of the knowledge concept answer accuracy at time t+1, where σ is the nonlinear sigmoid activation function, and KS t This represents the current state of knowledge.

[0038] Optionally, the objective function of the multi-feature knowledge tracking model is:

[0039]

[0040] Where L is the objective function, a t For the answer, y t This is a prediction of the candidate's performance on the test.

[0041] The beneficial effects of this invention are as follows: This invention comprehensively utilizes attention mechanisms and learning features to effectively model the student's learning process. First, by constructing a learning embedding with learning feature information, it is input into an attention mechanism with an upper triangular mask to obtain attention scores for different exercises. This allows for weighted processing of different exercises according to their importance during the student's learning process, highlighting key learning content. Next, to model the student's learning absorption capacity, the concept of learning absorption rate is introduced into the model. By comparing the differences between continuous interactions, the model can capture the changing trends in students' knowledge absorption. Simultaneously, the model also introduces a learning gate mechanism to distinguish differences in students' knowledge absorption capacity, thereby more accurately predicting the student's knowledge state. In the prediction phase, the student's current knowledge state is fully utilized to improve prediction accuracy. The method of this invention is reasonable and effective in knowledge tracking tasks, accurately predicting students' knowledge state and providing personalized learning guidance. Attached Figure Description

[0042] 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.

[0043] Figure 1 This is a schematic diagram of the interval time in an embodiment of the present invention;

[0044] Figure 2 This is a diagram illustrating the overall architecture of the attention-based multi-feature knowledge tracking model according to an embodiment of the present invention.

[0045] Figure 3 This is a detailed architecture diagram of the attention-based multi-feature knowledge tracking model according to an embodiment of the present invention. Detailed Implementation

[0046] 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.

[0047] 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.

[0048] This embodiment provides a multi-feature knowledge tracing method based on an attention mechanism. This method utilizes a Multi-feature Attentive Knowledge Tracing (MFAKT) model to comprehensively model the learning process using students' learning characteristics. Specifically, the model fully leverages the efficiency of the attention mechanism and the memory capacity of recurrent neural networks when processing time-series data. A new learning embedding is introduced into the attention mechanism, including the question, answer time, and number of prompts, to highlight the impact of students' historical learning behavior on their attention scores. This allows the model to more accurately focus on students' performance and needs at different learning stages. To better model learning absorption rate, a learning gate is designed to control the difference in knowledge state between two consecutive interactions. The introduction of the learning gate allows the model to better capture students' learning progress and changes, thereby improving the accuracy and reliability of knowledge tracing. In the prediction phase, individual student differences are considered, and a series of metrics are employed. First, this embodiment uses each student's answer accuracy to measure the impact of question difficulty. Second, the interval time, i.e., the time interval between two answers, is considered to measure the forgetting effect. Finally, combined with the student's current knowledge state, the model can accurately predict whether the student's answer will be correct or incorrect at the next moment. To verify the rationality and effectiveness of the model, comparative experiments were conducted in this embodiment. The results show that the model has superior performance in knowledge tracing tasks. Furthermore, this embodiment also conducted further ablation experiments to verify the effects and impacts of each component of the model.

[0049] In the knowledge tracing task, suppose there is a student set S = {s1, s2, ..., s...} I The set of exercises is E = {e1, e2, ..., e}. J}, and the set of knowledge concepts is C = {c1, c2, ..., c M Each exercise in the Q matrix contains a specific knowledge concept. It is a matrix composed of 0s and 1s representing the relationship between exercises and knowledge concepts. If exercise e j Includes knowledge concept c m Then Q jm =1, otherwise Q jm =0. Each student's performance on each exercise is unique because they spend different amounts of time and choose whether to need hints or help based on their cognitive level. The entire learning process unfolds through repeated responses to various exercises. During this process, there is continuous interaction between the student and the exercises, with a certain time interval between each interaction. This learning sequence constitutes the student's accumulated experience in the learning process. Therefore, a student's learning sequence in the learning process can be represented as X = (e1, at1, hc1, a1), it1, (e2, at2, hc2, a2), it2, ..., (e t ,at t ,hc t ,a t ),it t , where tuple (e t ,at t ,hc t ,a t ) represents a learning interaction, e t For exercises, at t hc t a represents the number of hints a student receives on this exercise. t This indicates whether a student answered the question correctly or incorrectly (1 indicates a correct answer, 0 indicates an incorrect answer). t This indicates the time interval between adjacent interactions.

[0050] The goal of the knowledge tracking task is to track changes in students' knowledge mastery over time. By monitoring and analyzing students' learning process, it is possible to predict how students will perform in their next attempt at a problem (denoted as e). t+1 The probability of answering correctly at any given time is used to optimize teaching models and strategies, thereby achieving personalized learning guidance. By modeling and analyzing students' learning behaviors and answer performance, it is possible to understand students' mastery of different knowledge concepts and their learning progress. Simultaneously, continuously tracking students' knowledge status helps predict potential difficulties or challenges they may encounter in future learning, thus providing more targeted guidance and support, and promoting students' continuous progress and development. The mathematical symbols and descriptions used in this embodiment are shown in Table 1.

[0051] Table 1

[0052]

[0053] To better understand the overall architecture of the attention-based multi-feature knowledge tracking model, we first briefly introduce the model's four categories of feature embeddings. These feature embeddings include temporal embedding, accuracy embedding, target knowledge concept embedding, and cue count embedding.

[0054] 1. Time embedding:

[0055] In the attention-based multi-feature knowledge tracking model, there are four different temporal embeddings: adjacent interaction interval time, interval between identical exercises, interval between identical knowledge concepts, and response time. These temporal features can influence students' forgetting patterns to some extent. According to the Ebbinghaus forgetting curve theory, the retention of memory gradually decreases over time, and the amount of memory gradually decays; knowledge learned after more than a month will be almost completely forgotten. Therefore, in this model, all time intervals exceeding one month are set to one month. This standardizes the time intervals and takes into account the impact of forgetting. The adjacent interaction interval time is represented by the embedding matrix as follows: it represents the embedding of the time interval between adjacent interactions, d it The number of types representing the time interval between adjacent interactions is obtained by subtracting the start times of adjacent interactions, expressed in minutes and rounded to the nearest integer. Similarly, the embedding matrix for the time interval of the same exercise is represented as follows. The time interval between the same knowledge concepts is represented in the matrix as follows: Where d et It is the number of different types of time intervals between exercises, d ct This represents the number of different knowledge concept intervals. Since response times are typically short, seconds are used as the unit. The embedding matrix for response time is represented as follows: d at It is the number of times the response takes place, d k It is the dimension, which is set to 128 in this embodiment.

[0056] like Figure 1The diagram illustrates the time intervals between student interactions, where 'e' represents an exercise, 'c' represents a concept, 'et' indicates the interval between identical exercises, 'ct' indicates the interval between identical concepts, and 'it' indicates the interval between adjacent interactions. Identical shapes indicate identical exercises; identical borders indicate identical concepts. Without considering time intervals, predictions would be identical regardless of when they are made. However, once time intervals are considered, predictions after ten days and one month show significant differences. This is because students' memories gradually fade over time, leading to decreased learning effectiveness. Therefore, considering time intervals is crucial for accurately predicting students' knowledge status and learning progress.

[0057] 2. Accuracy embedding:

[0058] A student's level of knowledge mastery can be measured by assessing the accuracy of their past answers. A high accuracy rate suggests a comprehensive understanding and mastery of the relevant exercises or concepts; conversely, a low accuracy rate indicates a potentially insufficient understanding. Furthermore, accuracy indirectly reflects the difficulty of the exercises or concepts for the student. If a question is particularly difficult for a particular student, their probability of answering correctly will be relatively low. Therefore, accuracy can serve as an indicator of the difficulty of exercises or concepts. Analyzing student accuracy yields valuable information about their knowledge mastery and the difficulty of the exercises. This helps educators and learning system designers better understand student learning and implement appropriate teaching and personalized learning strategies.

[0059] Students' accuracy rate in answering exercises and their accuracy rate in answering knowledge concepts can be represented by the following definitions:

[0060]

[0061] For a given student, from time 0 to t-1, the problem e j The accuracy of the response is represented by ec(j). t Knowledge concept c m The accuracy of the response is expressed as cc(m). t , where j∈(1,J), represents the set of all exercises, and J is the total number of exercises; m∈(1,M), represents the set of all knowledge concepts, and M is the total number of knowledge concepts. count(e j ) indicates exercise e j The number of times a response was received. Exercise e j Similarly, the number of times the answer is correct can be counted (c).m ) represents the knowledge concept c m The number of times a response was received. Representing the knowledge concept c m The number of times the answer was correct.

[0062] After calculating the two accuracy rates using formulas (1) and (2), the obtained values ​​are then embedded to obtain an embedding matrix. The embedding matrix for the accuracy of the exercise answers is represented as follows: Where d ec The number of categories representing the accuracy of answers to exercises, d k The embedding dimension is represented by the concept response accuracy embedding matrix. Where d cc This represents the number of categories for concept answer accuracy, with accuracy rounded to the nearest integer. Through these definitions and embedding matrices, we can model and analyze students' exercise answer accuracy and concept answer accuracy, thereby gaining a deeper understanding of students' learning progress and knowledge mastery.

[0063] 3. Target knowledge concept embedding:

[0064] Target knowledge concepts play a crucial role in knowledge tracing, representing the specific knowledge concepts students need to learn. Embedding target knowledge concepts into the model helps it gain a deeper understanding of students' learning objectives and needs. In attention-based multi-feature knowledge tracing models, knowledge concepts are represented using an embedding matrix, which is shown as follows: Where d kc d represents the total number of knowledge concepts. k This is the embedding dimension. By assigning a vector representation to each target knowledge concept, the model can associate and learn the target knowledge concept with other features.

[0065] The advantage of this embedding representation is that, during model training, the embedding matrix is ​​iteratively optimized, encoding the relationships and similarities between different knowledge concepts as distance and direction relationships in a vector space. This allows the model to better understand students' learning objectives and the relevance between knowledge concepts, providing more accurate predictions and suggestions for students' learning objectives, thereby offering personalized learning advice and feedback.

[0066] 4. Number of prompts embedded:

[0067] Typically, students only seek hints after answering incorrectly on their first attempt, suggesting they may have some difficulty mastering the knowledge. In other words, students usually need hints only when they are not familiar with the material, so the number of hints can be seen as a reflection of a student's level of understanding. The hint frequency embedding matrix is ​​as follows: dhc It is the number of prompts, d k This is the embedding dimension. By embedding the number of prompts into the model, the model can use this information to better understand the student's learning status and needs.

[0068] The embedding representation of cue counts allows the model to consider student performance and learning progress under different cue counts during the learning process. The model can capture the relationships between different cue counts and their impact on student learning by learning the cue count embedding matrix. This information is invaluable for the model to further understand students' learning progress and difficulties. Therefore, by considering cue counts, the model can provide corresponding support and guidance based on students' different learning needs and levels to help them overcome difficulties and improve learning outcomes. This further enhances the model's effectiveness and capability in knowledge tracking and personalized education.

[0069] This embodiment provides a multi-feature knowledge tracking method based on an attention mechanism, including: obtaining the student's previous knowledge state; inputting the previous knowledge state into a multi-feature knowledge tracking model, outputting the current knowledge state, and predicting the student's answer performance at the next moment based on the current knowledge state; wherein, the multi-feature knowledge tracking model calculates an attention score by fusing multi-dimensional information based on an attention mechanism, and introduces a learning gate mechanism to control the difference in knowledge state between two consecutive interactions of the student.

[0070] Furthermore, the multi-feature knowledge tracking model includes an attention mechanism module, a learning and absorption module, a target knowledge concept module, and a prediction module;

[0071] The attention mechanism module is used to calculate the attention score by integrating multi-dimensional information such as the questions, answer time, and number of prompts. The learning absorption module is used to obtain the learning absorption based on the previous knowledge state and the knowledge concept vector of the current question, and to obtain the actual learning absorption by adding learning gates. The target knowledge concept module is used to construct the relationship matrix between the questions and knowledge concepts, and to obtain the knowledge concept vector of the relationship matrix. The prediction module is used to obtain the current knowledge state by combining the attention score, actual learning absorption, and knowledge concept vector, and to concatenate the question embedding, interval time, answer accuracy, and current knowledge state to obtain the answer performance at the next moment.

[0072] like Figure 2As shown, the model is meticulously designed with four core modules: Attention Mechanism (AMM), Learning Absorption (LAM), Target Knowledge Concept (TCM), and Prediction. These modules work together to undertake the important tasks of student knowledge tracking and prediction. During model operation, the input x is composed of AMM, LAM, and TCM, ensuring the comprehensiveness and accuracy of the information. The student's knowledge state KS plays a crucial role in the entire model. (Previous knowledge state KS) t-1 This will serve as the input for the next time step, and after processing by AMM, LAM, and TCM, the current knowledge state KS will be obtained. t And used to predict students' performance in the next moment. t+1 .

[0073] Figure 3 This paper presents the detailed architecture of a multi-feature knowledge tracking model based on an attention mechanism. In this architecture, the AMM module is an attention mechanism module with an upper triangular mask. This upper triangular masking refers to masking the current and subsequent time steps to prevent the model from seeing the student's answers at those times, aiming to more effectively utilize the student's learning history. This module calculates a specific attention score for each student interaction, enabling the model to better understand and focus on the student's personalized learning patterns. Furthermore, the AMM module introduces a novel learning embedding that integrates multi-dimensional information such as questions, response times, and the number of prompts, further emphasizing the impact of the student's historical learning behavior on the attention score.

[0074] LAM (Learning Absorption Module) utilizes a learning gate to control learning gains and measures a student's ability to digest and absorb knowledge. This gate controls the difference in knowledge state between two consecutive interactions. By introducing the learning absorption gate, the model can better capture students' learning progress and changes, thereby achieving dynamic updates to knowledge state and improving the accuracy and reliability of knowledge tracking.

[0075] The Target Knowledge Concepts (TCM) module selects concept codes from the Q-matrix using exercise numbers, embedding information about the target knowledge concepts to help students better understand the learning objectives. This module provides contextual information about the target knowledge concepts, enabling students to focus more intently and comprehend the content during the learning process.

[0076] The Prediction module is designed to predict performance based on individual student learning differences. It uses a student's current knowledge level and prior learning characteristics to predict their performance in the next exercise. This module comprehensively considers individual student differences and incorporates factors such as question difficulty, time intervals, and the student's current knowledge level for prediction.

[0077] Attention scores are calculated by integrating multi-dimensional information such as questions, response time, and number of prompts through attention mechanisms.

[0078] To eliminate the influence of subsequent learning interactions, this embodiment employs a multi-head attention mechanism with an upper triangular mask. In the input portion of the attention mechanism, an embedding matrix is ​​first used... Let J represent the set of exercises, where J is the number of exercises and d is the number of exercises. k It is the dimension, and by combining the response time, the number of prompts, and the exercises, a learning embedding with learning feature information is constructed:

[0079]

[0080] Where the subscript t represents time t, This is a vector concatenation operation, where W1 is the first weight matrix and b1 is the first bias term. Adding the learned answers to the question embedding with learned features can yield new value labels:

[0081] VL t =L t +a t #(4)

[0082] Where + denotes vector addition. For the answer a... t Expand it into a vector consisting entirely of 0s or 1s. d a It is also the dimension, which is set to 50 in this embodiment. Therefore, the attention calculation formula can be obtained as follows:

[0083]

[0084] Among them, AS t Let Q, K, and V be the attention score, representing the query, key, and value, respectively. For queries with learned embeddings, k τ For key-value pairs with learned embeddings, VL τ For value labels with learned embeddings, τ refers to the time before time t, and d k It is the embedded dimension.

[0085] Because students forget during the learning process, their past performance often has a smaller impact on the prediction of the current question. Therefore, when predicting students' answers to the current question, the model uses an attention mechanism with a multiplicative decay term to calculate the attention score. This attention mechanism can assign different weights to the answers at different points in time based on the students' historical answers. As time goes by, the multiplicative decay term will gradually reduce the weight of past answers, thus paying more attention to recent learning performance. Such a mechanism can better capture the changes and dynamics of students' learning. In formula (5), the range of τ is [1, t-1]. If any value of τ is chosen, the attention score calculation formula is as follows:

[0086]

[0087] in

[0088] Where θ is a trainable decay rate parameter greater than 0, AS t,τ To compute the attention score with multiplicative exponential decay, d(t,τ) is the distance between time steps t and τ. This means that when calculating similarity, in addition to focusing on the similarity between the query and the key, the context-aware time-step distance is also considered:

[0089]

[0090] in

[0091] Acquiring learning absorption, and actually acquiring learning absorption by adding learning gates, includes:

[0092] Because students differ in their ability to absorb and digest knowledge during the learning process, modeling needs to consider learning absorption to simulate students' mastery of knowledge. Learning absorption can be seen as the knowledge progress students acquire during the learning process, reflecting their understanding and mastery of knowledge. Based on past experience, learning absorption can also be called learning gain, similar to "distance traveled," reflecting the difference in students' performance at two time points, that is, the degree to which students digest and absorb knowledge content within this period. By comparing students' learning status or answer performance at different time points, the model can calculate learning absorption, that is, the knowledge content that students have truly absorbed and mastered during this period. To model learning absorption more accurately, the modeling process needs to consider the difference in students' performance in two consecutive learning interactions. This can be achieved by observing students' learning outcomes, comprehension levels, and problem-solving abilities at different time points. In this way, information about students' learning absorption can be obtained, thereby better understanding students' progress and changes in the learning process.

[0093] In attention-based multi-feature knowledge tracking models, differences between two learning processes can be captured by concatenating a student's previous learning embeddings with their current learning embeddings. However, due to individual differences among students, including variations in knowledge reserves and problem-solving time, it is necessary to concatenate the time interval with the student's prior knowledge state. Generally, the shorter the time interval, the more compact and continuous the learning, and the better the ability to digest and absorb knowledge. Furthermore, a student's prior knowledge state also affects learning outcomes. For students with weaker foundations, there is greater potential for improvement; through a series of exercises, they can absorb a large amount of new knowledge and benefit greatly. However, for high-achieving students who have already mastered most of the knowledge content, the opportunity to learn new knowledge is less. To focus on the knowledge state of concepts related to the current exercise, it is necessary to concatenate the prior knowledge state KS. t-1 Knowledge concept vector of the current exercise Multiply to obtain the previously related knowledge state. (Knowledge States):

[0094]

[0095] Where · represents the element-wise product between vectors. Therefore, the formula for learning absorption would be as follows:

[0096]

[0097] in, It is the second weight matrix. This is the second bias term, and tanh is a non-linear activation function. Considering that students have different abilities to digest and absorb information, and not all learning translates into knowledge growth, a learning gate is designed to control absorption capacity.

[0098]

[0099] in It is the third weight matrix. This is the third bias term, where σ is the nonlinear sigmoid activation function. Then, the learning gate Γ... t Multiply by la t To achieve actual learning and absorption:

[0100] LA t =Γ t ·((la t +1) / 2)#(11)

[0101] Since the output value of the tanh function is in (-1, 1), and the learning absorption is positive, the learning absorption range needs to be projected to (0, 1). Therefore, the formula is designed as (la... t +1) / 2.

[0102] The knowledge concept vector obtained from the relation matrix by the target knowledge concept module includes:

[0103] The relationship between exercises and knowledge concepts is usually represented by a Q-matrix. The Q-matrix used in this embodiment is a 0-1 matrix given by an expert. If exercise e... j Includes knowledge concept c m Then Q jm =1; otherwise Q jm =0. When predicting e t+1 When solving exercises, extract a specific row of the Q matrix by the exercise number. Suppose that each knowledge concept is assigned a length of d k From the embedding vector, we can obtain a knowledge concept vector related to the Q matrix:

[0104]

[0105] in It is the fourth weight matrix. It is the fourth bias term, and the knowledge concept vector contains the target knowledge concept encoding.

[0106] The current knowledge state, obtained by combining attention score, actual learning absorption, and knowledge concept vector, can be represented as:

[0107] KS t =AS t +LA t +KC t+1 #(13)

[0108] By combining the embedded exercises, time intervals, answer accuracy, and current knowledge status, the answer performance at the next moment can be obtained, including:

[0109] To predict a student's performance at time t+1, the prediction function concatenates the question embedding, interval time, answer accuracy, and student knowledge state. This concatenated set of question embedding, interval time, answer accuracy, and student knowledge state is then projected into a fully connected layer and activated using the sigmoid function. Thus, the student's expected performance can be represented as:

[0110]

[0111] in It is a weight matrix. It is the bias term, σ is the nonlinear sigmoid activation function, and e t+1 For the problem embedding at time t+1, et t+1 Embed the time interval of the same exercise at time t+1, ct t+1 For the time interval embedding of the same knowledge concept at time t+1, ec t+1 Embed the accuracy of the exercise answers at time t+1, cc t+1 The knowledge concept answer accuracy embedding is used at time t+1. To learn all parameters and vectors in MFAKT, this embodiment selects cross-entropy log loss as the objective function to predict the difference between y and the actual answer a, and uses the Adam optimizer to minimize the objective function on mini-batches.

[0112]

[0113] Experimental Results and Analysis:

[0114] A series of experiments were conducted on four real-world datasets. These experiments aimed to evaluate an attention-based multi-feature knowledge tracking model by predicting future student responses and to compare it with the most representative knowledge tracking methods. The experiments assessed the model's performance on different datasets to understand its effectiveness in predicting future student responses. Ablation experiments were also conducted to understand the advantages of different modules in terms of accuracy and efficiency.

[0115] 1. Experimental Dataset:

[0116] In this embodiment, four classic benchmark educational datasets were used: Statist2011, ASSISTments 2012, ASSISTments Challenge, and Junyi. These datasets were collected from online education platforms and were used to evaluate the effectiveness of the model.

[0117] 2. Comparison Model:

[0118] To verify the effectiveness of the attention-based multi-feature knowledge tracking model, it was compared with seven benchmark knowledge tracking models. The experimental models selected are as follows:

[0119] BKT: BKT is a classic machine learning knowledge tracking model based on Hidden Markov Models. It uses Bayesian rules to update the state of each concept, which is considered a binary variable.

[0120] DKT: DKT obtains the student's knowledge state by providing the student's personal question-and-answer records to an LSTM, and outputs predictions based on the knowledge state.

[0121] DKVMN: Dynamic Key-Value Memory Network Knowledge Tracking Model.

[0122] SAKT: SAKT is a knowledge tracking model based on a self-attention mechanism. It uses a Transformer to model the student's interaction history and predicts the student's performance on the current exercise by considering the relevant exercises in the student's past interactions.

[0123] AKT: Context-Aware Attention Knowledge Tracking Model. AKT defines two self-attention encoders, uses monotonic self-attention to model forgetting, and leverages question difficulty to improve the question embedding of IRT.

[0124] SAINT+: SAINT+ is an extended version of SAINT, a knowledge tracing model based on Transformer.

[0125] LPKT: LPKT takes into account the effects of learning and forgetting, and constrains the evolution of students' states by measuring the learning gains between current and previous practices.

[0126] 3. Comparative Experiment:

[0127] The performance of the attention-based multi-feature knowledge tracking model (MFAKT) and other models in predicting student responses was compared. The experiment used the area under the curve (AUC) and accuracy (ACC), the most common metrics in the knowledge tracking field, as performance evaluation indicators. Bold text indicates the best performance on the dataset, and underlined text indicates the second-best performance. Table 2 shows the performance comparison of each model method across four datasets.

[0128] Table 2

[0129]

[0130] As shown in Table 2, different types of baseline models exhibit significant performance gaps. The attention-based multi-feature knowledge tracing model outperforms other knowledge tracing methods in both AUC and ACC across all datasets, demonstrating the effectiveness of incorporating learned feature information into the attention mechanism for modeling.

[0131] 4. Ablation experiment:

[0132] Fifteen variants of the model were selected, each removing specific modules or features from the original MFAKT model. Details of each variant are described below:

[0133] NoAM: Does not consider attention mechanisms. NoAS: Does not incorporate response time, number of prompts, and interval time into the attention mechanism; only uses exercises and answers as input. NoLA: Does not consider learning absorption rate. NoKC: Does not consider target knowledge concept vectors. NoPM: The prediction layer does not include time intervals and answer accuracy; only exercise encoding is used. Nohc: Does not consider the number of prompts. Noec: Does not consider exercise accuracy. Nocc: Ignores the accuracy of knowledge concepts. NoAcc: Removes answer accuracy, including the accuracy of exercises and knowledge concepts. Noat: Does not consider response time. Noit: Ignores the time interval between adjacent responses. Noet: Omits the time interval between the same exercises. Noct: Removes the time interval between the same knowledge concepts. Noetct: Discards the time interval between the same exercises and knowledge concepts. NoTime: Removes any time information, including response time and time interval.

[0134] This embodiment compares the model prediction performance of MFAKT with other variants on four datasets. All methods use AUC and ACC (higher is better) as performance evaluation metrics. Standard 5-fold cross-validation was used in the experiments. Bold black text indicates the best results on the given dataset, while underlined text indicates modules and learned features with significant impact on that dataset. Table 3 shows the results of the ablation experiments on the attention-based multi-feature knowledge tracking model.

[0135] Table 3

[0136]

[0137] Table 3 summarizes the experimental results of different variants of the attention-based multi-feature knowledge tracking model. By comparing these variant models, we can understand the impact of each module or feature on the MFAKT model. First, it can be observed that the attention mechanism module has the greatest impact on the prediction results. The AUC values ​​of the NoAM model on the four datasets are 78.07%, 76.03%, 80.76%, and 78.52%, respectively, which are nearly 9.24%, 14.72%, 7.81%, and 13.57% lower than the MFAKT model. To further verify the impact of learned embeddings on the attention mechanism, only exercises and answers were used as inputs to the attention mechanism in the NoAS model. It can be observed that the performance of the NoAS model on the four datasets still significantly lags behind the MFAKT model, indicating that learned embeddings play a crucial role. The AUC values ​​of the NoAS model on these four datasets are 80.70%, 77.17%, 77.70%, and 78.99%, respectively. These values ​​are 6.63%, 13.58%, 10.87%, and 13.10% lower than those of the MFAKT model, respectively.

[0138] Furthermore, time information, especially response time, plays a crucial role in model performance. It has been observed that longer response times indicate greater difficulty for students in answering questions. Therefore, considering response time is essential for accurate predictions in the MFAKT model. On the Statics2011 dataset, the Noat model's AUC was 82.11%, nearly 5.22% lower than the MFAKT model. On the ASSIST2012 dataset, the Noat model's AUC was 87.67%, 3.08% lower than the MFAKT model's AUC. On the Junyi dataset, the Noat model's AUC was 82.99%, nearly 9.10% lower than the MFAKT model. However, on the ASSIST2017 dataset, the effect of time was not significant.

[0139] In summary, this method utilizes rich learning feature information to model the student's learning process. Specifically, it combines the efficiency of parallel training of the attention mechanism with the time-series memory capability of recurrent neural networks. In the attention mechanism, a novel learning embedding is constructed, including the question, response time, and number of prompts, to highlight the impact of the student's historical learning behavior on their attention score. To model the learning absorption rate, a learning gate is designed to control for the difference in knowledge state between two consecutive interactions. In the prediction phase, considering individual student differences, the accuracy of each student's answers is used to measure the effect of question difficulty, and the interval time is used to measure forgetting. These factors, combined with the student's current knowledge state, are used to predict the student's performance at the next moment. Experiments were conducted on the Static2011, ASSIST2012, ASSIST2017, and Junyi datasets. The results show that this model outperforms existing knowledge tracking methods. The largest improvement is achieved on the ASSIST2017 dataset, with an AUC value 8.71% higher than the best baseline model, validating the model's rationality and effectiveness.

[0140] 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 multi-feature knowledge tracking method based on an attention mechanism, characterized in that, include: To obtain the student's prior knowledge status; The previous knowledge state is input into the multi-feature knowledge tracking model, the current knowledge state is output, and the student's answer performance at the next moment is predicted based on the current knowledge state; wherein, the multi-feature knowledge tracking model calculates attention score by integrating multi-dimensional information based on attention mechanism, and introduces learning gate mechanism to control the difference in knowledge state between two consecutive interactions of the student. The multi-feature knowledge tracking model includes an attention mechanism module, a learning and absorption module, a target knowledge concept module, and a prediction module. The attention mechanism module is used to calculate an attention score by integrating multi-dimensional information such as questions, answer time, and number of prompts through the attention mechanism. The learning absorption module is used to obtain learning absorption based on the previous knowledge state and the knowledge concept vector of the current exercise, and to obtain actual learning absorption by adding the learning gate; The target knowledge concept module is used to construct a relationship matrix between exercises and knowledge concepts, and to obtain the knowledge concept vector of the relationship matrix; The prediction module is used to combine the attention score, the actual learning absorption, and the knowledge concept vector to obtain the current knowledge state, and to concatenate the question embedding, interval time, answer accuracy, and the current knowledge state to obtain the answer performance at the next moment. Attention scores are calculated by integrating multi-dimensional information such as questions, response time, and number of prompts through attention mechanisms. Constructing learned embeddings with learned feature information: in, This is a vector concatenation operation. This is the first weight matrix. For the first bias term, For exercises, To answer the question of the embedding matrix of time, Embed the number of prompts for students into a matrix, with subscripts t for t time, For learning embeddings with learning feature information; The learned answer is added to the learning embedding containing the learned feature information to obtain a new value label: Based on the new value label, the attention score is calculated using a multi-head attention mechanism based on triangular masking to obtain the attention score; Acquiring learning absorption, and actually acquiring learning absorption by adding learning gates, includes: Multiply the previous knowledge state by the knowledge concept vector of the current exercise to obtain the previous relevant knowledge state; The learning absorption is obtained based on the previously relevant knowledge state, the learning embedding with learning feature information, and the adjacent interaction interval time embedding. The actual learning absorption is obtained by projecting the learning absorption range onto a preset range and then multiplying it by the learning gate.

2. The multi-feature knowledge tracking method based on attention mechanism according to claim 1, characterized in that, The method for calculating the attention score is as follows: in, For attention score, These are queries, keys, and values, respectively. This is a query with learned embeddings. These are key-value pairs with learned embeddings. For value labels with learned embeddings, It refers to the time before time t. It is the embedded dimension.

3. The multi-feature knowledge tracking method based on attention mechanism according to claim 1, characterized in that, The calculation method for the learning absorption is as follows: in, For learning and absorption, This is the second weight matrix. It is the second bias term. and These are the learning embeddings containing learning feature information from the previous and current time steps, respectively. Embed the time interval between adjacent interactions. This refers to the state of prior related knowledge; The calculation method for the learning gate is as follows: in, For learning purposes, It is a non-linear sigmoid activation function. This is the third weight matrix. It is the third bias term.

4. The multi-feature knowledge tracking method based on attention mechanism according to claim 1, characterized in that, The knowledge concept vector of the relation matrix is: in, For knowledge concept vectors, This is the fourth weight matrix. This is the fourth bias term. For prediction When solving exercises, extract a specific row from the relation matrix by the exercise number.

5. The multi-feature knowledge tracking method based on attention mechanism according to claim 1, characterized in that, The method for predicting a student's performance in the next moment based on the current knowledge state is as follows: in, For students' performance at time t+1, This is the fifth weight matrix. This is the fifth bias term. Embedding of exercises at time t+1 Embed the time interval of the same exercise at time t+1. The time interval embedding of the same knowledge concept at time t+1. Embed the accuracy of the exercise answers at time t+1. Embed the accuracy of knowledge concept answers at time t+1. It is a non-linear sigmoid activation function. This represents the current state of knowledge.

6. The multi-feature knowledge tracking method based on attention mechanism according to claim 1, characterized in that, The objective function of the multi-feature knowledge tracking model is: Where L is the objective function, For the answer, This is a prediction of the candidate's performance on the test.

Images