Design of Learning Progress Tracking and Feedback Mechanism Based on Data Visualisation Technology in Music Teaching

With the acceleration of China’s modernization process, music education in colleges and universities has received increasing attention. Music education in colleges and universities is an important content of humanistic and artistic quality education, which can help students establish the correct three views, cultivate creative thinking, and improve the level of students’ aesthetic ability [1–2]. After the reform and opening up, music education in colleges and universities has developed rapidly, and the opinions on strengthening music education in colleges and universities put forward to strengthen the unified planning and leadership of music education, actively carry out extracurricular cultural activities to promote the construction of campus culture, and improve the teaching conditions to ensure that the work of music education is carried out normally [3–4]. In the context of the era of rapid development of the Internet industry and information technology, people’s way of life and learning under the influence of rapidly changing information technology has undergone great changes, the most significant changes in the field of education. Since the 20th century, the state has paid more and more attention to information technology as a means to promote the innovation of education and teaching, and one of the data visualization technologies that has been applied in the field of education it will data in the form of graphic images.

Data visualization technology has been applied in the field of education, which presents data to people in the form of graphic images, which makes people understand the information generated by the data more thoroughly, and makes it easy for decision makers to make decisions efficiently [5–7]. Data visualization technology converts data into graphic or image displays and integrates human-computer interaction theory and technology with it, which uses the technology of computer graphics and image processing [8]. The basic idea of data visualization technology is to obtain the data from the data source after the use of a graphical element to represent each data item, a larger number of data using data images presented. At the same time, the data has multiple attribute values with multi-dimensional data in the form of a display. You can observe the data in many ways so as to have more in-depth observation and analysis of the data [9–11]. Through data visualization technology, the learning progress data generated in music education are processed to feedback the results in a more flexible way, so that deeper information in music education data can be discovered [12]. Visualization technology allows teachers to be able to make more accurate judgments about students’ attendance records, academic performance and homework completion, and other learning situations so as to guide students more effectively. But also depicts students’ learning records in music education, allowing students to carry out more guided opinions on independent learning methods so that students can achieve more efficient learning [13–15].

Data visualization technology is a very important and popular research field, which is widely used in software development, education and scientific research. Literature [16] proposes a related computational method based on scientific computing visualization - the image segmentation method and applies it to music teaching in basic education, aiming to create a good classroom atmosphere and increase students’ interest in music courses so that students can better experience music, enjoy music, and like music. Literature [17] synthesized music graphic images and mathematical and statistical methods such as K-mean clustering and fusion decision tree to construct a music visualization model, which was verified to be effective and accurate through actual cases and performance tests and could present teaching information to students in an intuitive way through graphic images. Literature [18] used the Wilcoxon sign Rank test and Spearman’s correlation to examine the mechanism of the effect of synchronous and asynchronous music lessons based on visualization strategy on students’ music playing skills, and it was found that this strategy allows students to understand the concepts of music playing skills by viewing visual materials, which in turn improves their music playing skills. Literature [19] explored whether machine learning techniques to develop an online music teaching and practicing platform could reduce the burden on teachers and provide students with online real-time performance feedback, and demonstrated the feasibility of the initiative through an evaluation analysis, which allowed for finding or developing music lessons based on student’s learning styles, musical backgrounds, or preferences.

The rapid development of data visualization technology has brought unprecedented opportunities for change in music education. Literature [20] used image segmentation and video tracking technology to design an image recognition algorithm for hand and face, constructed a personalized teaching system for learning resources on this basis, and verified the superior performance of the system through experiments, which can track the learning progress of students in real time, provide timely feedback of teaching information to the teacher, and provide personalized teaching resources services for students. Literature [21] designed a combination of digital media technology and information processing technology singing teaching systems in colleges and universities. The simulation test and analysis found that the system involved has a certain degree of practicality. It is a more graphic, intuitive, convenient way to present the teaching content to achieve the purpose of teaching, and its unique intuition and use of interactivity, so that the teaching of singing is more systematic, professional and modernized. Literature [22] developed a Learning Management System (LMS) and verified through a music assessment information system based on key performance indicators that the LMS can accurately track the progress of students’ vocal practice and provide timely feedback, which serves to simplify learning.

In this paper, based on the deep recurrent neural network model, the deep knowledge tracking technique is introduced, using a large number of neurons to represent the potential learning state and time dynamics of the students, and based on which the self-attention mechanism is added to solve the problem of data sparsity. A strategy that satisfies the normal distribution is used to process the students’ answers as pre-input, and then the retained data is used as formal input. The first embedding layer is constructed based on four aspects: different questions, different test points, students’ learning level of the test points, question features, and test point features. The above features are transformed into an embedding matrix, which is used as the data for the second embedding layer. Using three different cosine similarity functions, the predicted values are calculated and represented using matrix multiplication. Starting from both individual and overall perspectives, a feedback mechanism for music teaching that includes timeliness and phasing is designed, and feedback teaching supports decision-making. Controlled experiments are set up to visualize the progress data of music learning through empirical analysis.

2

Relevant technical theories

2.1

Introduction to Neural Networks

Deep learning is a new research direction in the field of machine learning that has been introduced to bring it closer to artificial intelligence. Deep learning will learn the intrinsic patterns and levels of representation of sample data, and the information gained from these learning processes can be of great help in interpreting data such as text, images, and sounds. The ultimate goal of deep learning is to enable machines to be able to analyse and learn like a human beings and to be able to recognise data such as text, images and sound. Neural networks are the underlying model in deep learning, an algorithmic mathematical model for distributed parallel information processing that mimics the behavioral characteristics of animal neural networks. Such networks rely on the complexity of the system to process information by adapting the relationships between the large number of nodes interconnected within it. Neural networks are widely parallel interconnected networks of simple units with adaptive properties that are capable of modelling the interactive responses made by the biological nervous system to real-world objects.

Single-layer neuronal structures suffer from the problem of non-linear segmentation, hence the need to use multi-layer neuronal structures. A two-layer perceptron is used to solve the heterodyne problem. There is a layer of neurons between the output layer and the input layer called the hidden layer. The neurons in both the hidden layer and the output layer are functional neurons with an activation function. A feedforward neural network with a hierarchical structure that has neurons in each layer fully connected to the neurons in the next layer is a more commonly known entity. Neurons in the input layer of a multilayer feedforward neural network are responsible for receiving input from the outside world, neurons in the hidden and output layers are responsible for processing the input signals, and finally, the processed signals are output through the output layer.

2.2

Deep knowledge tracking model

2.2.1

Knowledge tracking model

The task of knowledge tracking is to model a student’s knowledge state in order to accurately predict how the student will perform in future question-answering interactions [23]. Having data on a student’s history of question-answering sequences, the model is able to derive the student’s current knowledge state from the results of the student’s interactions with the questions and can then predict how the student will perform on the questions.

Most of the previous models rely on manually defined interaction functions, such as the IRT model, but these models cannot dynamically track the students’ knowledge state. The Deep Knowledge Tracking (DKT) model applies temporally deep Recurrent Neural Networks (RNNs) to the knowledge tracking task, using a large number of neurons to represent potential knowledge states and their temporal dynamics, and allows for the learning of potentially variable representations of students’ knowledge from the data. Students’ learning process is affected by many factors, such as learning materials, contextual information, and the order in which they solve the problems, but many of these attributes are difficult to quantify. The DKT model uses two recurrent neural networks: a regular Sigmoid-based RNN and a Long Short-Term Memory model (LSTM).

Figure 1 shows the Sigmoid-based ordinary RNN network. Sigmoid-based ordinary RNN networks map a series of hidden states h₁,h₂,…, h_t obtained computationally from the input sequence x₁,…,x_t to the output sequence y₁,…,y_t. These hidden states are treated as encodings of relevant information from past observations for prediction of the future.

The structure of the network is the same each time, hence the name recurrent neural network. Input x_t is a uniquely hot coded or compressed representation of a student’s performance in answering questions, and output y_t is a vector of probabilities predicting that the student answered the question correctly, defined by equations (1) and (2): 1 $h_{t} = \tanh (W_{h x} x_{t} + W_{h h} x_{t - 1} + b_{h})$ 2 $y_{t} = σ (W_{y h} h_{t} + b_{y})$

LSTM is a more complex variant of RNN. In LSTMs, hidden layers retain their values and do not remove them until they pass through a forgetting gate. As a result, they more naturally retain information over many time steps, which is thought to make them easier to train. Hidden units are updated using multiplicative interactions so that they can perform more complex transformations for the same number of potential units. The DKT model pioneered deep knowledge tracking by applying deep learning techniques to the field of knowledge tracking for the first time. Also, for the first time, DKT considers the influence of students’ answers and uses RNN networks to consider the influence of the time factor of students’ answers.

2.2.2

Knowledge tracking model based on self-attention mechanism

Although the DKT model based on the RNN approach achieves good results, it does not handle sparse data well. In the real world, some students only interact with a very small portion of the knowledge, or there is only a small amount of learning data available for these students. The problem of sparse data can be effectively handled by a knowledge tracking model (SAKT) that is based on the self-attention mechanism [24].

SAKT identifies knowledge information related to a given knowledge from a student’s past performance in answering questions and then predicts their learning based on the selected knowledge information. Since the prediction is based on a relatively small number of past question-answering records, it handles the data sparsity problem better than RNN-based methods. Meanwhile, the SAKT model applies the encoder framework in the transformer to the knowledge tracking domain for the first time and achieves an improvement in model effectiveness.

3

Knowledge tracking model design and implementation

3.1

Pre-input layer

A strategy was used for the response time data that satisfied a normal distribution, i.e., time_i ~ n(a_i, s_i), the mean of the response times for question i, and s_i, the standard deviation of the response times for question i extracted from the dataset. If the answered time of the question does not satisfy 3 standard deviations from the mean, the corresponding data are simply cleared, so that the data on the answered time of the students who actually answered the question are retained. The formulae for mean a_i and standard deviation s_i are as follows, where num represents the total number of times a student answered the corresponding question: 3 $a_{i} = \frac{\sum_{i = 1}^{n u m_{i}} t i m e_{i}}{n u m}$ 4 $s_{i} = \sqrt{\frac{\sum_{i = 1}^{n u m_{i}} {(t i m e_{i} - a_{i})}^{2}}{n u m_{i}}}$

3.2

Input layer

After pre-input processing, the retained data is used as a formal input, which is represented before the data is extracted.Knowledge tracking is the prediction of the student’s future answers based on the student’s historical performance, which can be represented as: 5 ${\hat{y}}_{t + 1} = \hat{p} {a_{t + 1} = 1 | (x_{1}, x_{2}, \dots, x_{t})}$

Where x_i = (t_i, a_i) is a binary, t_i. tabulates the question that the student answered at the i moment, a_i = 1 or 0 denotes a correct or incorrect answer, and $\hat{p}$ is the knowledge tracking objective, i.e., the probability that the student answered correctly at the t + 1 moment.

The first step is to code the input data, the thesis uses one-hot coding, the input to the model is a binary set of questions answered by students and the response situation, i.e., x_i = (t_i, a_i) as described above, where t_i and a_i are two vectors respectively, and the model inputs are spliced together after applying the one-hot coding operation on the two features to form a single, independent, one-hot code,. For example, if the total number of questions in a dataset is N, and the responses to N questions are coded, a vector of dimension 2*N is needed, if a student correctly answers a question with sequence number j, the corresponding positions j and 2* j in the vector are indexed with 1, and the remaining bits are 0. If the student incorrectly answers question number j, the corresponding position j in the vector is indexed with a value of 1 and the remaining bits are 0. Suppose that the total number of questions is 3, i.e., N = 3, and that the student answered question 2 with a correct answer. Then the heat coded solo heat vector is denoted as X = (0,1,0,0,1,0). Conversely, if the student answers question 2 incorrectly, the solo heat vector is denoted as X = (0,1,0,0,0,0).

3.3

First embedding layer

After coding the input data, the following is the feature extraction of the features between the data of the selected dataset, a question contains the question itself and the test points of the question, the set of questions is denoted as $t = {t_{i}}_{i = 1}^{| t |}$ and the set of test points is denoted as $s = {s_{j}}_{j = 1}^{| S |}$ , i={1,…,A}, j={1,…,B}, A and B represent the total number of questions and test points respectively.

3.3.1

Correlation between different questions and between different test points

Use the Jaccard similarity formula and use a matrix R_tt = {r_ij} of size num_t × num_t to store it. For example, if there are two questions in the dataset: t₄ for 4 + 5*6 and t₅ for 4÷5*6, then there are a total of three test points for each of these questions: +×÷, and a total of one test point: ÷. The correlation between questions t₄ and t₅ is the number of test points in common divided by the total number of test points, which is denoted 1÷3 = 0.333 and represented as r₄₅ = 0.333 in the matrix.

The procedure for calculating the correlation between any two questions t_i and t_j using Jaccard’s formula in this study is as follows: firstly, based on the values of the elements of Matrix R_ts, find out the sequences of test points S_i and S_j related to questions t_i and t_j, respectively, and then calculate the correlation between questions t_i and t_j according to the following formula: 6 $r_{i j} = \frac{| S_{i} \cap S_{j} |}{| S_{i} \cup S_{j} |}$

Use another Jaccard formula to calculate the correlation between test points s_i and s_j, and use a matrix R_ss = {r_ij} of size num_s × num_s. Firstly, based on the values of the elements of matrix R_ts, the sequences of questions T_i and T_j related to test points s_i and s_j, respectively, are identified: then the correlation between test points s_i and s_j is calculated according to the following formula, T_i and T_j represent the sets of questions related to test points i and j, respectively: 7 $r_{i j} = \frac{| T_{i} \cap T_{j} |}{| T_{i} \cup T_{j} |}$

3.3.2

The extent to which students learnt the test points

Since the correlation between the questions and the test points has been extracted through the matrix Rpc, the degree of learning of each student on the test points can be obtained through the questions answered by the students, and the degree of learning of the students on the test points is also expressed through the percentage of correct answers and is stored using a matrix R_stu–s = {r_stu–s} of size num_stu × num_s, and num_s is the total number of students in the dataset.

3.3.3

Problem characterisation

Extract the percentage of questions answered correctly: Diff_i is the percentage of questions answered correctly for question t_i, True_i represents the number of times question t_i was answered correctly, and Count_i represents the total number of times question t_i was answered: 8 $D i f f_{i} = \frac{T r u e_{i}}{C o u n t_{i}}$ question t_i, time_k is the time at which question t_i was answered for the k th time, Count_i denotes the total number of times question t_i was answered, and soft max denotes the normalisation function: 9 $S p e e d_{i} = s o f t \max (\sum_{k = 1}^{C o u n t_{i}} t i m e_{k} / C o u n t_{i})$

Extracted question characteristics: tf_i is the characteristics of question t_i, Diff_i is the percentage of correct responses to question t_i., and Speed_i is the speed of response to question t_i: 10 $t f_{i} = \frac{D i f f_{i}}{S p e e d_{i}}$

3.3.4

Characteristics of the examination point

The test point characteristics are expressed in terms of the percentage of correct answers for each test point, as shown in the following equation, where sf_i denotes the characteristics of test point s_i, True_i denotes the number of times test point s_i was answered correctly, Count_i denotes the total number of times test point s_i was answered, and sf_i will be stored in a list sf_i of length num_s, sf = {sf_i}: 11 $s f_{i} = \frac{T r u e_{i}}{C o u n t_{i}}$

With the above learning feature extraction approach, six sets of data representing six learning features are obtained.

3.4

Second embedding layer

Firstly, the question features, test features and student features are represented as three embedding matrices, denoted as: E_c ∈ R^num_t×D, E_s ∈ R^num_s×D and E_stu ∈ R^num_stu×D, with D denoting the dimensions of the embedding matrices, and each row of the matrices representing each of the question content, test features and student features, respectively.

The input layer is the topic features in the dataset t_i, t_i is the set of test points and data in a certain topic, t_i = {c_i, s_i}, where c_i is the data of the problem and s_i is the test points of the problem. The hidden layer uses the softmax function to calculate the weight of each test point and perform a weighted summation as the test point feature of the question, and E_s is represented as the embedding matrix of the test points of the question: 12 $s o t f \max_{s_{j}} = \frac{e^{E_{s_{j}}}}{\sum_{i = 1}^{S} e^{E_{s_{j}}}}$ 13 $E_{s} = \sum_{j = 1}^{n u m_{t_{i} - s}} s o t f \max_{s_{j}} \times E_{s_{j}}$ num_{t_i−s} represents the number of test points associated with question i. The data and test points are then combined in a ratio of λ to form a complete matrix of embedded questions E_t ∈ R^num_t×D: 14 $E_{t_{i}} = λ_{i} \times E_{c_{i}} + (1 - λ_{i}) \times E_{t_{i} - s}$ 15 $λ_{i} = \frac{e^{E_{c_{i}}}}{e^{E_{c}} + e^{E_{t_{i} - s}}}$

3.5

Hidden Layers

In this study, in order to ensure that the predicted values are as accurate as possible, three different cosine similarity functions are used here for the calculation as follows [25]: 1)

Correlation of questions and test points: ${\hat{R}}_{t s}$ denotes the predicted correlation of questions and test points, and ${\hat{L}}_{t s}$ denotes the squared difference loss function between the predicted and true values of the correlation of questions and test points: 16 ${\hat{R}}_{t s} = \frac{E_{t} \times E_{s}^{T}}{| E_{t} | \times | E_{s}^{T} |}$ 17 ${\hat{L}}_{t s} = {({\hat{R}}_{t s} - R_{t s})}^{2}$

2)

Correlation between different questions: ${\hat{R}}_{t t}$ denotes the predicted correlation between different questions, and ${\hat{L}}_{t t}$ denotes the squared difference loss function between the predicted and true values of the correlation between different questions: 18 ${\hat{R}}_{t t} = \frac{E_{t} \times E_{t}^{T}}{| E_{t} | \times | E_{t}^{T} |}$ 19 ${\hat{L}}_{t t} = {({\hat{R}}_{t t} - R_{t t})}^{2}$

3)

Correlation between different test points: ${\hat{R}}_{s s}$ denotes the predicted correlation between different test points, and ${\hat{L}}_{s s}$ denotes the squared difference loss function between the predicted and true values of the correlation between different test points: 20 ${\hat{R}}_{s s} = \frac{E_{s} \times E_{s}^{T}}{| E_{s} | \times | E_{s}^{T} |}$ 21 ${\hat{L}}_{s s} = {({\hat{R}}_{s s} - R_{s s})}^{2}$

The student embedding matrix representing the student characteristics and the knowledge embedding matrix representing the knowledge characteristics are multiplied using matrix multiplication to derive the predicted value of each student’s level of knowledge acquisition ${\hat{R}}_{s t u - s}$ . The error between the predicted value and the true value is then calculated and optimised using the squared error function ${\hat{L}}_{s t u - s}$ using the following formula: 22 ${\hat{R}}_{s t u - s} = E_{s t u} \times E_{s} {\hat{L}}_{s t u - s} = {({\hat{R}}_{s t u - s} - R_{s t u - s})}^{2}$

Two new trainable embedding matrices are introduced, called problem difficulty embedding matrix E_dt ∈ R^num_t×D and test point difficulty embedding matrix E_ds ∈ R^num_s×D, which are multiplied by problem embedding matrix E_t and test point embedding matrix E_s respectively, and the result of the multiplication is the computed predicted value of the problem difficulty ${\hat{f}}_{t f}$ and the predicted value of the knowledge difficulty ${\hat{f}}_{s f}$ , and then the problem features and knowledge features are optimised using two different mean-variance error functions ${\hat{L}}_{t f}$ and ${\hat{L}}_{s f}$ respectively The formulae are as follows: 23 ${\hat{f}}_{t f} = E_{d t} \times E_{t} {\hat{L}}_{t f} = {({\hat{f}}_{t f} - t f)}^{2}$ 24 ${\hat{f}}_{s f} = E_{d s} \times E_{s} {\hat{L}}_{s f} = {({\hat{f}}_{s f} - s f)}^{2}$

Finally, the six loss functions are summed to form a joint optimisation function to pre-train the embedding of questions, test points and student data: 25 $\hat{L} = {\hat{L}}_{t s} + {\hat{L}}_{t t} + {\hat{L}}_{s s} + {\hat{L}}_{s t u - s} + {\hat{L}}_{t f} + {\hat{L}}_{s f}$

3.6

Forecasting and optimisation

Here a neuron is used directly to first splice the question embedding matrix and the test point embedding matrix, denoted as E_t–s = concat (E_t, E_s) : then the matrix E_t−s after the splice is multiplied with the student embedding matrix E_stu to get the matrix representing the current student’s overall learning level of the question and test point, called the learning matrix, denoted as: 26 $E_{J} = E_{t - s} \times E_{s t u}$

Next the learning matrix is passed into a linear transformation and a linear model is used to obtain the predicted values, denoted as: 27 $\hat{y} = σ (W \times E_{J} + B)$ Where W and B are trainable parameters and σ denotes the sigmoid activation function with Eq: 28 $σ (x) = \frac{1}{1 + e^{- x}}$

Finally, the prediction $\hat{y}$ is compared to 0.5 and if $\hat{y} \geq 0.5$ , $\hat{y}$ is set to 1 and vice versa to 0. This prediction $\hat{y} = {0, 1}$ is the final goal of the knowledge tracking model, i.e., whether or not the student answered the question correctly; if $\hat{y} = 1$ , it is assumed that the student will answer the next question correctly, and if $\hat{y} = 0$ , it is assumed that the student will answer the next question incorrectly.

After obtaining the real answer situation y and the final predicted value $\hat{y}$ in the data set, because the predicted value is not 0, i.e., 1, the prediction result is a binary classification problem, a cross-entropy loss function $\hat{L}$ is used for optimization, so that the prediction result of the model is more likely to be close to the true value, and the accuracy of the model prediction is improved in turn, and the calculation formula of $\hat{L}$ is as follows, where num represents the total number of all answer records in the data set: 29 $\hat{L} = - \frac{1}{n u m} \sum_{k = 1}^{n u m} (y_{k} \times \log ({\hat{y}}_{k}) + (1 - y_{k}) \times (\log (1 - {\hat{y}}_{k})))$

4

Design of feedback mechanisms for music teaching

4.1

Immediacy Feedback Algorithm

4.1.1

Individual Immediate Feedback

Individual immediacy feedback is the feedback information of a certain teaching subject at a certain point in time. For the teaching implementer, the most concern about the individual immediacy feedback is the learning concentration of the teaching subject at that point in time since the EVA evaluation model has already evaluated the individual’s learning concentration at a certain point in time, so the individual immediacy feedback can be calculated using the EVA evaluation model.

4.1.2

Overall immediacy of feedback

Unlike individual immediacy feedback, aggregate immediacy feedback refers to feedback from all instructional objects at a point in time, i.e., aggregate immediacy feedback is a collection of individual immediacy feedback from all instructional objects at the same moment in time.

Assuming that a certain teaching object is denoted as l_n, the object’s EVA concentration score at a certain moment t_i is denoted as S_{(t_n,t_i)}, the set formed by combining all l_n is denoted as L, i.e., the set L denotes all the teaching objects, and the instantaneous feedback information of L at a moment of t_i can be denoted as the set S_i, then the set S_i can be expressed by the formula (30): 30 $S_{i} = {l_{n} | S_{(l_{n}, t_{i})}, l_{n} \in L}$

Denote the average of set S_i as ${\bar{S}}_{i}$ which represents the average concentration rating of all teaching subjects.

4.2

Staged Feedback Algorithm

4.2.1

Individual stage feedback

For a given subject l_n, an increase in learning attentiveness over the time interval implies positive feedback from teaching, and a decrease in learning attentiveness over the time interval is considered to be negative feedback from teaching. Assuming that the beginning moment is denoted as t_start, the end moment as t_end, and the time interval is a period from t_start to t_end, which is denoted as T_[start,end], and the EVA evaluation model’s assessment values of learning status at moments t_start and t_end are denoted as S_start and S_end, respectively, and the average value of the EVA evaluation in the interval of T_[start,end] is denoted as $\bar{S_{[s t a r t, e n d]}}$ , the quantitative relationships may have the following four cases. 1)

If $S_{e n d} \geq \bar{S_{[s t a r t, e n d]}}$ and S_end < S_start it means that the subject’s concentration is decreasing in interval T_[start,end] which is recorded as Sc1, and the concentration is increasing, which is recorded as Sc2.

2)

If $S_{e n d} \geq \bar{S_{[s t a r t, e n d]}}$ and $\bar{S_{[s t a r t, e n d]}} \geq S_{s t a r t}$ , it means that the subject’s concentration level is on the rise in the T_[start,end] interval, and the scenario is recorded as Sc2.

3)

If $S_{e n d} < \bar{S_{[s t a r t, e n d]}}$ and S_end ≥ S_start, it means that the subject has a high concentration score in a sub-interval T_[a,b]([a,b] Ö [start,end]) in interval T_[start,end], but a low concentration score in t_start and t_end, which proves that the learner’s concentration has improved for a period of time but is declining, and the scenario is recorded as Sc3.

4)

If $S_{e n d} \geq \bar{S_{[s t a r t, e n d]}}$ and S_end < S_start, it means that the teaching subject has a Period of sub-interval T_[a,b] ([a,b]Ö[start,end]) learning concentration score is low in the T_[start,end] interval, but the learning concentration score is high in t_start and t_end, proving that the learner’s concentration score has had a period of decline but is improving.

Categorising the 4 scenarios it can be concluded that Scenario Sc1 and Scenario Sc3 belong to negative pedagogical feedback, while Scenario Sc2 and Scenario Sc4 belong to positive pedagogical feedback, then the pedagogical feedback score S_Sc for a feedback Scenario Sc can be calculated by f(Sc) using the method (31): 31 $f (S c) = {\begin{matrix} 1, S c \in {S c 2, S c 4} \\ - 1, S c \in {S c 1, S c 3} \end{matrix}$

Denote the standard deviation of the evaluated value of EVA in interval T_[start,end] as σ_[start,end], which is calculated as in Equation (32), where S_i is the evaluated value of EVA at a moment t_i in interval T_[start,end] and N is the total number of moments in interval T_[start,end]. 32 $σ_{[s t a r t, e n d]} = \sqrt{\frac{\sum {(s_{i} - {\bar{S}}_{[s t a r t, e n d]})}^{2}}{N}}$

If σ_[start,end] is larger, it means that the learner’s concentration is more variable and the instructional feedback is stronger in interval T_[start,end], and vice versa, the learner’s concentration performance is smoother and the instructional feedback is weaker. To sum up, the individual stage of instructional feedback F_stage in the T_[start,end] interval is calculated by Equation (33): 33 $F_{s t a g e} = S_{S c} * σ_{[s t a r t, e n d]}$

4.2.2

Overall stage feedback

Overall stage feedback is the feedback information of all teaching objects in a certain time interval. Similar to individual stage feedback, the increase of learning concentration of all teaching objects in the time interval means positive feedback of teaching, and the decrease of learning concentration of all teaching objects in the time interval is regarded as negative feedback of teaching, so the overall stage feedback can be calculated by using the average of the learning concentration of all teaching objects in each moment. Concentration, using a similar approach to individual phase feedback to calculate the score.

4.3

Feedback-based instructional decision support strategies

As per the design of the individual decision support form, there are two situations in teaching where individual reminders need to be sent, one is when the learner is in a state of low concentration for a long time, and the other is when the learner has experienced a sudden drop in concentration over a period of time. For a certain moment t_i, the time interval T_[start,end] can only be from t_start to t_i at most, while from the point of view of the length of the time interval, t_start can be taken from the beginning of the teaching activity moment t₀ to t_i any moment between the previous moment t_i−1, if t_start starts from t₀, which indicates a long interval, the evaluation is the individual from the beginning of the teaching activity to the learning state of the stage of the moment t_i, which is recorded as T_[0,i] if t_start starts from t_n (n < i, n ∈ N*), it means a short interval, which evaluates the learning state of an individual from the moment t_n to the moment t_i. The time interval is denoted as T_[n,i] the two intervals T_[0,i] and T_[n,i] are used to design the individual decision support strategy.

1)

Persistent low concentration state

The time interval on which the persistent low concentration state focuses should be T_[0,i]. The situation is represented by expression (34), where $\bar{S_{[0, i]}}$ is the average concentration score of learner l_n in interval T_[0,i] and md_0.5 is the median score of learner l_n in interval T_[0,i]. To complement the situation where the average score is pulled down by the few times it is scored, 0.6 is the threshold that distinguishes between the concentration levels of less concentration and concentration: 34 $(\bar{S_{[0, t]}} < 0.6) \land (m d_{0.5} < 0.6)$

This expression can be interpreted as learner l_n has hardly been in a state of concentration since the beginning of the teaching activity, therefore, the situation requires a reminder to be sent to learner l_n, as well as to the teacher, denoted as function f 1, as shown in (35): 35 $f 1 = {\begin{array}{l} 1, (\bar{S_{[0, 2]}} < 0.6) \land (m d_{0.5} < 0.6) \\ 0, (\bar{S_{[0, l]}} \geq 0.6) \lor (m d_{0.5} \geq 0.6) \end{array}$

2)

Staged improvement of concentration

Learner l_n focus on concentration stage improvement should focus on the time interval is T_[n,i], then concentration stage improvement can be expressed as F_[n,i] ∈ P, where F_[n,i] is the learner l_n in the T_[n,i] interval of the teaching feedback scores, P is the set of intervals of positive feedback, in this condition it is necessary to abolish the persistent low concentration state of the teaching decision-making, and therefore the function f can be optimised as shown in (36): 36 $f 1 = {\begin{array}{l} 1, (\bar{S_{[0, l]}} < 0.6) \land (m d_{0.5} < 0.6) \land (F_{[n, i]} \notin P) \\ 0, (\bar{S_{[0, l]}} \geq 0.6) \lor (m d_{0.5} \geq 0.6) \lor (F_{[n, i]} \in P) \end{array}$

3)

Plunge in concentration

Learner l_n ’s plummeting concentration is the opposite trend to the phase increase in concentration, again taking time interval T_[n,i]. The situation can be represented by expression (37), where $\bar{S_{[n, i]}}$ is the average concentration score of learner l_n in interval T_[n,i], 0.6 is the threshold that distinguishes between less concentrated and concentrated concentration ratings, F_[n,i] is the pedagogical feedback score of learner l_n in interval T_[n,i], and SN is the set of intervals with strong negative feedback: 37 $(\bar{S_{[n, i]}} < 0.6) \land (F_{[n, i]} \in S N)$

This expression can be interpreted as a sharp drop of learner l_n from the above-focused state to the below-less-focused state in interval T_[n,i]. Therefore, the situation requires a reminder for learner l_n, denoted as function f2, as shown in (38): 38 $f 2 = {\begin{array}{l} 1, (\bar{S_{[n, i]}} < 0.6) \land (F_{[n, i]} \in S N) \\ 0, (\bar{S_{[n, i]}} \geq 0.6) \lor (F_{[n, i]} \notin S N) \end{array}$

The goal of overall decision support is to identify situations in which the overall learning status of the learner is poor, in which case the teacher is alerted to change his/her teaching style, and similar to the individual decision support strategy approach, the consideration of only the overall immediate feedback at a given moment t_i appears to be too one-sided for the decision support. Therefore, the overall decision support has to integrate the overall immediate feedback with the overall stage feedback.

From the perspective of overall immediacy feedback, when the overall learner engagement is low, it is unfavourable for the teacher to explain new knowledge and the teacher needs to be reminded to pay attention to the learner status or to change the way of teaching and learning, and for moment t_i this situation can be represented by the expression (39), where R_inattention denotes the rate of learners whose learning status is below the attentive level, B denotes the threshold of the need for decision support, and G_i denotes the moment t_i. of the instructional feedback level: 39 $(R_{i n a t t e n t i o n} \geq B) \land (G_{i} \in {C -, D +, D -})$

This expression can be interpreted as the ratio of learners whose current learning status is below the concentration level has exceeded the reminder threshold and the instructional feedback level is low, in which case a reminder needs to be sent to the instructor, denoted as function f3, as shown in (40): 40 $f 3 = {\begin{array}{l} 1, (R_{i n a t t e n t i o n} \geq B) \land (G_{i} \in {C -, D +, D -}) \\ 0, (R_{i n a t t e n t i o n} < B) \lor (G_{i} \notin {C -, D +, D -}) \end{array}$

From the perspective of overall stage feedback, when the overall participation of learners plummets, it is also unfavourable for teachers to continue their original teaching activities, and teachers need to be reminded to change their teaching style and pay more attention to the learners’ status or other feedback information. For time interval T_[n,i], the overall stage feedback score Ft_[n,i] on this interval is in the strong negative feedback interval SN, the reminder of the situation is expressed as a function of f3, as (41) shown: 41 $f 4 = {\begin{array}{l} 1, F t_{[n, i]} \in S N \\ 0, F t_{[n, i]} \notin S N \end{array}$

5

Empirical studies on learning progress tracking and feedback

5.1

Study design

5.1.1

Selection of experimental and control groups

In order to verify the effectiveness of the application of the Learning Progress Tracking and Feedback Model (LPTFM), this study adopts the research method of comparing the experimental and control classes, where students in the experimental group use the LPTFM for music learning, and students in the control group use the traditional music learning method. These 2 classes were taught by 2 teachers respectively, and one of the classes led by each teacher was the experimental class, and the other was the control class. Comparisons were made at the end of the semester using final exam grades, overall final grades, and e-learning grades, while questionnaires on independent learning abilities were administered to these 2 classes in order to compare the differences in their motivation and strategies for learning. Finally, one class was randomly selected from each grade from freshman to junior year to conduct an ANOVA before and after teaching.

5.1.2

Teaching practice environment

Enhancing the quality of music teaching has always been the concern of music teachers and researchers, in which the language and cultural environment, teaching mode and methodology have an important impact on the teaching of university music directly or indirectly. The rapid development of information technology provides many means and methods to reform the teaching mode and create a learning culture environment for music, which not only embodies the advantages of individuality, spontaneity, informality, context sensitivity and authenticity but also possesses the characteristics of social interactivity, connectivity and immediacy, and is able to realise the transmission of information and data across time and space in the true sense of pictures, text, sound and video. It also provides students with an authentic, seamless, context-rich, and natural learning environment, which is a perfect fit for music teaching and learning. As the learning management system records a large amount of student learning behaviour data, the mining, analysis and visual presentation of student learning behaviour data using learning analytics provides timely and efficient feedback for teachers, students, administrators and researchers, enabling them to observe and understand the learning process and state of students from a more objective perspective and at a micro level, promoting changes in teaching and learning behaviours of teachers and students and eventually Improve the effect and quality of learning.

5.2

Analysis of the results of tracking progress in learning data visualisation

5.2.1

Analysis of learning strategies

The visualisation in this study mainly provided students with feedback on detailed information about their learning goals and grades in the hope of facilitating the students to be able to reflect and summarise their learning plans and processes and adjust their learning strategies during the learning process. In order to examine whether the visualisation monitoring and feedback had a positive effect on whether the students in the experimental class were self-regulated in their learning, this study analysed the differences between the experimental class and the control class in terms of their learning methods and strategies, as well as the correlation analyses of these methods and strategies with their grades through questionnaires, in order to explore the experimental class students’ self-regulated learning.

The Learning Strategies Scale consists of 39 items divided into six dimensions: general approach (12 questions), learning help (9 questions), learning plan (6 questions), learning summary (5 questions), learning evaluation (3 questions), and learning management (4 questions). The internal consistency coefficient Cronbach’s alpha of the Learning Strategies Scale is 0.95, which indicates high reliability.

The question items from each of the six dimensions of the questionnaire were combined and averaged to generate new variables. Using independent sample t-test (95% confidence interval, two-tailed), the total score of learning strategies and related dimensions of the experimental and control classes were tested for comparison, and Table 1 shows the independent sample t-test of learning strategy dimensions of the experimental-control classes.

Table 1.

Independent sample t test for learning strategy dimensions

/	Group	N	Mean	Std	S.E. of mean
The total score of learning strategies	Experimental class	232	26.4136	2.8565	0.1985
The total score of learning strategies	Control class	243	19.229	2.9496	0.0356
General method	Experimental class	232	4.4986	0.5632	0.0485
General method	Control class	243	3.1659	0.5496	0.0485
Learning help	Experimental class	232	4.5263	0.6297	0.0496
Learning help	Control class	243	3.0896	0.6655	0.0486
Learning plan	Experimental class	232	4.2656	0.6868	0.0422
Learning plan	Control class	243	3.1269	0.6148	0.0493
Learning summary	Experimental class	232	4.5359	0.7586	0.0489
Learning summary	Control class	243	3.2675	0.6418	0.0385
Learning evaluation	Experimental class	232	4.4886	0.7998	0.0364
Learning evaluation	Control class	243	3.4896	0.6145	0.0487
Learning management	Experimental class	232	4.0986	0.5987	0.0386
Learning management	Control class	243	3.0895	0.6493	0.0485
/		Levene’s test
/		F	P	T	df
The total score of learning strategies	Assumed equal variance	0.2655	0.0154	0.0348	472
The total score of learning strategies	Equivariance is not assumed	0.2655	0.0154	0.0348	471.5846
General method	Assumed equal variance	0.4866	0.0486	0.7856	472
General method	Equivariance is not assumed	0.4866	0.0486	0.7855	471.9896
Learning help	Assumed equal variance	0.1985	0.0165	0.4652	472
Learning help	Equivariance is not assumed	0.1985	0.0165	0.4652	471.5966
Learning plan	Assumed equal variance	0.8462	0.0346	0.3482	472
Learning plan	Equivariance is not assumed	0.8462	0.0346	0.3482	471.9645
Learning summary	Assumed equal variance	0.2486	0.0165	-0.5278	472
Learning summary	Equivariance is not assumed	0.2486	0.0165	-0.5278	471.2975
Learning evaluation	Assumed equal variance	0.4098	0.0284	-0.5297	472
Learning evaluation	Equivariance is not assumed	0.4098	0.0284	-0.5297	471.9663
Learning management	Assumed equal variance	1.9486	0.0387	0.3785	472
Learning management	Equivariance is not assumed	1.9486	0.0387	0.3785	471.6869
/		t test for mean identify
/		P(2-tail)	MD	SE	/
The total score of learning strategies	Assumed equal variance	0.0156	0.0145	0.2855
The total score of learning strategies	Equivariance is not assumed	0.0156	0.0145	0.2855
General method	Assumed equal variance	0.0248	0.0498	0.0563
General method	Equivariance is not assumed	0.0248	0.0498	0.0563
Learning help	Assumed equal variance	0.0396	0.0263	0.0587
Learning help	Equivariance is not assumed	0.0396	0.0263	0.0587
Learning plan	Assumed equal variance	0.0189	0.0187	0.0486
Learning plan	Equivariance is not assumed	0.0189	0.0187	0.0486
Learning summary	Assumed equal variance	0.0265	-0.0385	0.0636
Learning summary	Equivariance is not assumed	0.0265	-0.0385	0.0636
Learning evaluation	Assumed equal variance	0.0478	-0.0348	0.0572
Learning evaluation	Equivariance is not assumed	0.0478	-0.0348	0.0572
Learning management	Assumed equal variance	0.0158	0.0289	0.0628
Learning management	Equivariance is not assumed	0.0158	0.0289	0.0628

As can be seen from Table 1, there is a significant difference between the experimental and control classes in the use of learning strategies, and the mean values of the total music learning strategies scores of the experimental and control classes are 26.4136 and 19.229, respectively. Continue to explore the relationship between the students’ overall grades, final exam grades, e-learning grades, and the dimensions of learning strategies. In the independent samples t-test, the hypothetical equal variances of the two-tailed p-values with 95% confidence intervals for the total score of learning strategies, general approach, learning help, learning plan, learning summary, learning evaluation, and learning management were 0.0156, 0.0248, 0.0396, 0.0189, 0.0265, 0.0478, and 0.0158, respectively, which were less than 0.05, which showed that there was a significant relationship between the experimental class and the control class have significant differences in learning strategies.

Through the analysis of the questionnaire items of general approach strategies, it is found that the factors of general approach strategies mainly include the management of time and environment, such as “I will arrange the time to do the most important things first”, and the common strategies such as “I will choose the environment suitable for my self-study”, and the common strategies of effort. Strategies such as “I will try my best to catch up with the subjects that I am not doing well in” and “I will skip the difficult questions and do the easy ones first in the exams”. The general methods were correlated with the overall grade, final exam grade, and e-learning grade, suggesting that this study stimulated the experimental class to apply these methods by visualising the feedback to the students in terms of goals, study hours, and effort values.

5.2.2

Visual tracking and feedback for learning and academic performance analysis

In order to analyse the students’ use of the Learning Tracking and Feedback Model, the students’ usage was recorded. Out of the 232 people who participated in the experiment, 232 (100%) used the Learning Progress Tracking and Feedback Model at least 1 time, and the results of the descriptive statistics of students’ usage and academic performance are shown in Table 2. From Table 2, it can be seen that student use was very uneven, with an overall average of 7.1898 reviews (mean=7.1898, std=4.7652), a total average of 15.6524 minutes of chapter viewing (mean=15.6524, std=14.2625), and a mean number of visits to the chapters of 15.4983 (mean=15.4983, std=14.0625).

Table 2.

Descriptive statistics on usage and academic performance

/	N	Minimum	Maximum	Mean	Std	Skewness	Kurtosis
Final test results	232	19	95.4536	63.4875	15.2089	-0.4862	-0.5136
Musical learning	232	14.4868	29.3157	26.1069	1.7636	-3.4064	15.0546
Total score	232	54.4856	93.4845	76.8615	8.2654	-0.2456	-0.4555
Review number	232	1	35	7.1898	4.7652	1.5682	3.5486
Chapter view length (minute)	232	1	115.8596	15.6524	14.2625	2.7264	12.0855
Access to various chapters	232	1	118.7814	15.4983	14.0625	2.8669	12.6558
Usual test results	232	32.4588	95.4525	85.4886	7.6285	-3.4569	16.5878
Learning hours (minutes)	232	678.1958	19728.4852	4266.0899	2348.5496	1.7855	6.5747
Study effort value	232	1498.9876	8312.4886	4765.0865	1186.4848	0.3758	0.4258

5.2.3

Comparative Analysis of Overall Learning Behaviour of Control Classes in Experimental Classes

Figure 2 shows the transformation of learning behaviour paths in the experimental class, and Figure 3 shows the transformation of learning behaviour paths in the control class. A-K are music learning units 1-10, and L-N is music theory knowledge test, vocal test, and music comprehensive test.

By analysing the significant meaning behavioural paths of the experimental and control classes, it can be found that: 1)

The significant behavioural conversion paths of the experimental class and the control class are 40 and 38, respectively, and there are differences in the specific paths.

2)

Overall, the paths of the students in the experimental and control classes are in accordance with the order set by the platform modules, which is the order of AB, BC, CD, DE, EF, FG, GH, HI, IJ, JK in the paths (with the Z-score value > 50), which is in line with the regular learning process and mode of university music teaching.

3)

The retrospective paths of DC, CB, and LC were clearly found in the experimental class, with Z-score values of 4.9175, 3.7651, and 4.3555, respectively. At the same time, the control class did not have these paths. By analysing the course modules, these paths happen to be the key contents related to basic music theory knowledge and music appreciation in the learning process, indicating that the students in the experimental class have a strong learning strategy, have a clear purpose, and are able to find out the key points and difficult points of learning.

4)

In the learning of the module, the students in the experimental class and the control class repeated the learning. However, from the Z-score scores on repeated learning in Figures 2 and 3, it can be seen that the students in the experimental class invested more learning effort than those in the control class. For example, in the repeated learning of chapter A, the Z-Score of the experimental class and the control class were respectively 97.5445 and 61.6545.

5.3

Learning feedback

5.3.1

Attention analysis

From the trend of the overall attention change in the course, Figure 4 shows the change in the learning attention of the students in the control group, and Figure 5 shows the change in attention of the students in the experimental group, with no obvious learning feedback from the students in the control group, and feedback from the students in the experimental group on their learning progress. The overall attention of the control group fluctuates more, while the experimental group is more consistent overall. The attention change curve of the experimental group shows that when attention is below the threshold (below 40), it usually returns to its normal state quickly. Statistical analysis of the average attention of the two groups, the average attention values of the experimental group and the control group are 81.1853 ± 19.3454 and 67.4544 ± 21.7108, respectively, which shows that the use of learning methods with learning feedback mechanism makes students’ attention more focused.

5.3.2

Learning behaviour analysis

Students in higher education are prone to engaging in irrelevant learning behaviors and unstable learning behaviors during online learning. In order to verify that feedback on learning progress in music learning can reduce learners’ irrelevant learning behaviours, irrelevant learning behaviours during music courses were analysed. Irrelevant learning behaviours refer to behaviours that are not related to learning except for keeping eyes on the screen and necessary note-taking during music learning, including drinking water, stretching, yawning, scratching the head, picking the nails, touching the face, touching the nose, and other common small actions. Learning behaviours were marked by human marking, and the invalid learning behaviours of the video-recorded actors were marked every minute, and the invalid learning behaviours of each group were counted and plotted as an area curve every minute, and the time curve of invalid learning behaviours is shown in Figure 6.

From the graph of data analysis, the irrelevant learning behaviours appeared more in the first 7 minutes, but in a declining trend, probably because the music class began, students need to slowly enter the learning state and gradually reduce the irrelevant learning behaviours. 8-19 minutes of irrelevant learning behaviors were in a steady state, but began to rise slightly in the late stage. 20 minutes before the end of the lesson, the total number of irrelevant learning behaviors rose significantly, and students had more small actions. In the subgroups, the control group’s irrelevant learning behaviors fluctuated more and were on the rise, while the experimental group’s were smoother overall. The number of irrelevant learning behaviors in the experimental group (95) was significantly lower than the number in the control group (137) overall. Among them, the number of irrelevant times in the experimental group was more than that of the control group in the first 12 minutes, and the situation was reversed after 12 minutes, and after 20 minutes, there was a large difference between the two groups. Overall feedback on students’ learning progress was found to reduce the number of invalid behaviors.

5.4

Effectiveness of music teaching

The author used SPSS 21.0 to analyse the data from the questionnaire before (pre-teaching) and after (post-teaching) the implementation of the author’s teaching practice with learning progress tracking and feedback teaching model (post-teaching) using ANOVA and paired analyses. The statistical results for the data before and after teaching music are as follows:

Figure 7 shows the results of ANOVA before and after teaching. Figure (a) is before teaching, and Figure (b) is after teaching.

As can be seen from Figure (a), ANOVA (all known as one-way ANOVA) was used to investigate the variability of the four dimensions of satisfaction with teaching, interest in learning, ability to learn, and effective communication among the three grades before the instructional practice of tracking and feedback on the progress of music teaching. From the figure, it can be seen that all three grade samples will not show significance for teaching satisfaction before teaching, learning interest before teaching, learning ability before teaching, and emotional communication before teaching, and the p-value is greater than 0.05, which means that the three grade samples show consistency for teaching satisfaction before teaching, learning interest before teaching, learning ability before teaching, and emotional communication before teaching, and there is no significant difference.

From Figure (b), it can be seen that the samples of different grades show significance (p<0.05) for teaching satisfaction, learning interest, learning ability and emotional communication after teaching practice, which means that the samples of the three grades have significant differences for teaching satisfaction, learning interest, learning ability and emotional communication after teaching practice. Specific analyses show that the three grades showed a 0.01 level of significance (F=5.5646, p=0.0020**) for teaching satisfaction after teaching practice. A specific comparison of the differences shows that there is a more significant difference in teaching effectiveness between the three grades. The results of the comparison of group mean scores show that after first year teaching (4.0505) > after third year teaching (3.9468) and after first year teaching (4.0505) > after second year teaching (3.6502).

6

Conclusion

This paper applies a temporal deep recurrent neural network to the knowledge tracking task, creating a deep knowledge tracking model. Meanwhile, the self-attention mechanism is used to address the issue of sparse data. The knowledge tracking model is built and implemented from six parts, including the input layer, first embedding layer, second embedding layer, hidden layer, and optimization layer. Design a feedback mechanism for music teaching by combining timely feedback and stage feedback. The control experimental group is selected to visualize and analyze the learning progress tracking data through empirical research. 1)

The mean values of the total music learning strategy scores of the experimental and control classes are 26.4136 and 19.229, respectively. T-tests of independent samples were conducted on the dimensions of learning strategies, and the test results showed that the hypothetical variance values of the dimensions had p-values less than 0.05 and that there was a significant difference between the learning strategies of the experimental and control classes.

2)

The students’ progress was tracked, and overall, the average number of times students reviewed the music subject was 7.1898, and the average total time spent on chapter viewing was 15.6524 minutes. The average number of times chapters were visited was 15.4983.

3)

In module learning, both the experimental and control classes repeated their studies, with the experimental class putting more effort into their studies than the control class. The Z-Score values for the experimental and control classes were 97.5445 and 61.6545, respectively.

Langue:: Anglais

Périodicité:: 1 fois par an
Sujets de la revue:: Sciences de la vie, Sciences de la vie, autres, Mathématiques, Mathématiques appliquées, Mathématiques générales, Physique, Physique, autres

RSS Feed de la revue

Design of Learning Progress Tracking and Feedback Mechanism Based on Data Visualisation Technology in Music Teaching

Peihan Lin

Publié en ligne: 05 févr. 2025

Reçu: 19 sept. 2024

Accepté: 01 janv. 2025

DOI: https://doi.org/10.2478/amns-2025-0047

Mots clésKnowledge tracking, Self-attention mechanism, EVA evaluation, Visual tracking of learning progress

© 2025 Peihan Lin, published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Mots clés
Knowledge tracking, Self-attention mechanism, EVA evaluation, Visual tracking of learning progress