A learning effect prediction method based on students' facial expressions

Abstract

The present application relates to the field of education, and specifically provides a learning effect prediction method based on students' facial expressions, which includes the following steps: S1, collection and processing of expression data; S2, construction of a learning effect value-added evaluation; S3, construction of a learning effect prediction model and training it; S4, performance evaluation of the trained learning effect prediction model. The learning effect prediction model is a support vector machine binary classification model. The trained support vector machine binary classification model is evaluated using the indicators F1‑score and ROC curve. When the micro-expression frequency domain signal is input and the kernel function is a Gaussian function, the prediction accuracy is 89.7%; the value of the indicator F1‑score is 0.87, which is greater than 0.7; the area AUC between the ROC curve and the horizontal axis is 0.92, which is close to 1. The prediction accuracy of the method of the present invention is high, and the prediction results are reliable.

Description

Technical Field

[0001] The present application relates to the field of education, and in particular, to a method for predicting learning effects based on students' facial expressions. Background Art

[0002] In the teaching process, student learning outcomes are a crucial indicator. On the one hand, they can monitor student progress, and on the other hand, teachers can adjust the teaching pace in a timely manner based on student outcomes, ensuring that as many students as possible achieve better results. This ultimately boosts student confidence and increases motivation to learn. However, it's difficult for teachers to accurately monitor all students' facial expressions and assess their learning outcomes in real time.

[0003] In traditional teaching, students' learning outcomes are often judged by assigning grades based on a specific test. Since test results reflect more than just their learning achievements over a period of time—their own knowledge and experience also influence the final test results—this traditional evaluation method fails to effectively measure a student's true learning progress over a period of time and is often lagging. Furthermore, judging learning outcomes solely based on whether or not students can answer relevant questions ignores students' emotional changes and feelings during the classroom learning process.

[0004] Facial expressions can truly reflect students' emotions in class, and emotions can accurately reflect students' learning outcomes. Experienced teachers can roughly judge students' learning outcomes in class and adjust the teaching schedule in a timely manner. However, teachers have limited energy and cannot observe every student. Moreover, different teachers have different experiences and their judgments on the same student may be different. This makes teachers' judgments on students' learning outcomes inaccurate and prevents them from observing every student in a timely manner. Computer technology can help teachers predict students' learning outcomes, which can significantly improve teachers' classroom effectiveness. It allows teachers to pay attention not only to individual outstanding students, but also to the learning progress of each student, and to provide targeted reminders and guidance to students in class. However, existing methods for predicting learning outcomes usually use large movements such as head raising rate to predict learning outcomes, which is not accurate enough.

[0005] Existing learning effect prediction methods do not closely link facial expressions and learning effects, resulting in low accuracy in learning effect prediction. Summary of the Invention

[0006] The purpose of the present invention is to provide a learning effect prediction method based on students' facial expressions to address the deficiencies in the above-mentioned prior art.

[0007] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows: This application provides a learning effect prediction method based on students' facial expressions, which includes the following steps: S1, collection and processing of expression data; S2, construction of a value-added evaluation of learning effects; S3, construction of a learning effect prediction model and training it; S4, evaluation of the performance of the trained learning effect prediction model.

[0008] The data of the embodiment of the present invention comes from the class videos of 103 students in two classes of a middle school, which were obtained through 6 classroom practices, with a total of 2178277 sets of data. Based on the obtained class videos, expression values ​​are extracted. Specifically, it is divided into four processes: face recognition, face tracking, photo wall construction, and expression value extraction. The extracted expression value data is denoised and normalized to obtain micro-expression time domain signals and micro-expression frequency domain signals. The denoising process is first performed using a bandpass filter and then a moving average method. The normalization process is performed using the relative strength index (RSI). When the RSI is between 30-70, it can be considered that the student's expression has not changed significantly. Less than 30 or greater than 70 means that the student's expression has changed significantly.

[0009] The value-added evaluation of learning outcomes is calculated by subtracting the pre-test student ability from the post-test student ability. The median of the value-added is taken. Values ​​greater than the median are considered "good" learning outcomes, represented by "1" and a positive sample. Values ​​less than the median are considered "poor" learning outcomes, represented by "-1" and a negative sample. The pre-test and post-test student ability values ​​are derived using the Rasch model, which takes into account the difficulty of the questions and the varying difficulty levels of the same question for students with different foundations.

[0010] The learning effect prediction model is a support vector machine binary classification model. The micro-expression time-domain and frequency-domain signals obtained in step S1 and the learning effect value-added evaluation results obtained in step S2 are input into the support vector machine binary classification model for training, and a functional relationship between the two is obtained. The kernel function of the support vector machine binary classification model is one of a linear function, a quadratic polynomial function, a cubic polynomial function, and a Gaussian function. The prediction accuracy is the highest when the kernel function is a Gaussian function, reaching 89.7%, indicating that the prediction accuracy of the present method is high.

[0011] The learning effect prediction model trained in step S3 was evaluated using the F1-score and ROC curve. An F1-score value exceeding 0.7 indicates that the corresponding model can be used for prediction and has high reliability. When the micro-expression frequency domain signal is input and the kernel function is a Gaussian function, the F1-score value is 0.87, which is greater than 0.7, indicating that the model obtained in step S3 of the present invention is reliable and the prediction results of the present method are reliable.

[0012] Compared with the prior art, the present invention has the following beneficial effects: the expression value of the acquired expression data is extracted, and denoising and normalization are performed simultaneously, so that the noise that does not carry micro-expression signals is removed, and the expression values ​​of students with different expression levels are significantly changed through normalization, thereby improving the prediction accuracy of learning effects. The use of value-added evaluation of learning effects allows the labels of learning effects to reflect the actual learning effects, so that the error between the predicted learning effects and the actual learning effects is small. Compared with other machine learning methods, the support vector machine binary classification model has the advantage of being very convenient for solving complex problems, such as high-dimensional, nonlinear, and multi-classification problems. The evaluation results show that the prediction accuracy of the method of the present invention is as high as 89.7%, which is a high prediction accuracy; the F1-score value is 0.87, which exceeds 0.7; the AUC is 0.92, which is close to 1, and the prediction results of the method of the present invention are highly reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of a learning effect prediction method based on students' facial expressions provided by the present invention;

[0014] Figure 2 The normalized result of a student's "happy" expression in step S1 of the learning effect prediction method based on student facial expressions provided by the present invention;

[0015] Figure 3 The frequency domain signal of a student's "happy" expression in step S1 of the learning effect prediction method based on student facial expressions provided by the present invention;

[0016] Figure 4 This is an example diagram of the classification results in step S3 of the learning effect prediction method based on students' facial expressions provided by the present invention;

[0017] Figure 5 The classification results of different kernel functions on time domain signals and frequency domain signals in step S3 of the learning effect prediction method based on student facial expressions provided by the present invention;

[0018] Figure 6 The confusion matrix diagram in step S4 of the learning effect prediction method based on student facial expressions provided by the present invention;

[0019] Figure 7 This is the ROC curve in step S4 of the learning effect prediction method based on students' facial expressions provided by the present invention. DETAILED DESCRIPTION

[0020] In order to make the implementation process of the present invention clearer, it will be described in detail below with reference to the accompanying drawings.

[0021] The present invention provides a learning effect prediction method based on students' facial expressions. Figure 1 As shown, the method includes the following steps:

[0022] S1, collection and processing of expression data;

[0023] The expression data of the present invention can be classroom videos of students or data in an existing database. The expression data in this embodiment comes from classroom videos of 103 students in two classes of a middle school, obtained through 6 classroom practices, with a total of 2178277 sets of data. Based on the obtained class videos, expression values ​​are extracted. Specifically, the process is divided into four steps: face recognition, face tracking, photo wall construction, and expression value extraction. In the present invention, a face detection algorithm based on Haar features is used for face recognition, but the identified students are mixed. If the expression change data of each student is to be obtained, the face tracking needs to be classified. Based on speed and complexity considerations, the present invention uses Kalman filtering for face tracking to obtain the face tracking results of each student. The resulting photos of the face tracking are classified and a photo wall is constructed to facilitate uploading to the Megvii Face++ platform dedicated to expression classification and expression value calculation. The photo wall is uploaded to the Megvii Face++ platform dedicated to expression classification and expression value calculation, and the expression data shown in Table 1 is obtained. At each moment in the data, the sum of the expression values ​​of any student is 100. This means that the obtained expression values ​​represent relative sizes, that is, the proportion of the expression, rather than absolute values. This makes the characteristics of the expression more prominent and also makes the numerical differences of the expression values ​​at different moments between 0 and 100. The changes in the expression values ​​can more clearly show the changes in expression at different times. Specifically, this invention exploits students' micro-expression data. Micro-expressions are typically short-lived and difficult to observe. They are determined by physiological laws and are unconscious manifestations of a person's inner emotions. Therefore, using micro-expressions to predict learning outcomes is more accurate. More specifically, because students' expressions often remain largely unchanged, calmness was added to the six expressions of happiness, fear, surprise, sadness, anger, and disgust, for a total of seven.

[0024] Because the video capture was conducted in the classroom, the resulting profile pictures obtained through image processing were blurry. Furthermore, the accuracy of expression recognition and extraction when uploaded to the Megvii Face++ platform was not 100%. Therefore, the noise generated appeared as white noise. White noise is a random signal or process with a constant power-wavelength spectral density. The generation of these random signals is unavoidable in a teaching environment and requires denoising. The extracted expression values ​​were denoised using a bandpass filter and a moving average method to obtain micro-expression signals in the time domain. Fast Fourier transforms were then used to obtain micro-expression signals in the frequency domain, and the moving average length was then determined. Specifically, a band-pass filter is first used for denoising to obtain a micro-expression signal in the time domain, and a fast Fourier transform is used to obtain a micro-expression signal in the frequency domain. The moving average method is then used to obtain the length of the moving average, and the expression signal obtained by the band-pass filter is moved averaged according to the length of the moving average. More specifically, the band-pass filter can retain signals within a certain frequency band, and signals above the upper limit of the range and below the lower limit of the range will be removed. The time interval of micro-expression changes is 169-503ms, and the corresponding frequency is 2-6Hz. The band-pass filter can remove signals outside the 2-6Hz range. These signals are irrelevant to micro-expressions and are noise, that is, the band-pass filter can remove noise that is obviously not a micro-expression signal. The moving average method is then used to further reduce the noise of the signal output by the band-pass filter to remove random signals. More specifically, the expression of the moving average method is: Among them, n is the expression change cycle obtained by fast Fourier transform, and n is an odd number, m represents the frame number, H represents the actual expression value of the frame, and F represents the expression value after moving average. More specifically, any continuously measured time series signal can be expressed as an infinite superposition of sinusoidal signals of different frequencies. Through the above fast Fourier transform, the frequency domain signal is obtained, and the main frequency of the student's expression change can be obtained to calculate the period of expression change. Figure 3 In the frequency domain signal shown, the dominant frequency of the student's expression "happy" is 3.4286 Hz. Since the length n of the moving average can only be an odd number, the period of his expression change is n = fs / 3.3817 ≈ 7. This period can be used as the length n of the moving average.

[0025] Since each student's facial expressions vary, for example, some students laugh out loud when they are happy, while others just smile. The same level of happiness can have very different facial expression values. Therefore, the relative strength index (RSI) is used to normalize the denoised signal to avoid the inaccurate reflection of student emotional changes due to differences in facial expressions, thereby affecting the accuracy of learning effect prediction. Specifically, the relative strength index (RSI) is expressed as: Among them, A represents the increase in the upward fluctuation of the expression value within a cycle, and B represents the increase in the downward fluctuation of the expression value within a cycle; A+B represents the total fluctuation amplitude, such as Figure 2 As shown in the figure, the RSI corresponding to the students' expression values ​​is mostly between 30-70. When the RSI is between 30-70, it can be considered that the students' expressions have not changed significantly. When it is less than 30 or greater than 70, it means that the students' expressions have changed greatly. This is consistent with the actual situation. Most of the time in class, the changes in students' expressions are relatively weak and the changes in expressions are not frequent.

[0026] Table 1: Expression value extraction results

[0027]

[0028] S2, constructing value-added evaluation of learning effects;

[0029] It is necessary to quantify students' learning outcomes. This is usually done by comparing the performance of students in a class or school on standardized tests with the overall (province-wide or even national) performance to evaluate the learning outcomes of students in that class or school or the effectiveness of teachers' teaching. This evaluation method cannot effectively measure students' actual learning over a period of time. Value-added evaluation refers to obtaining students' test scores at different time points through research design for comparison. It evaluates students' learning status and outcomes over a period of time by observing changes in their knowledge level, thinking ability, and other aspects over different time periods. Compared to the original evaluation method that only considers the results of a single test, value-added evaluation better reflects students' true learning outcomes.

[0030] Specifically, as shown in Table 2, students' learning value-added is derived from their pre- and post-class test scores. The pre-test represents the pre-class competence value, the post-test represents the post-class competence value, and the value-added is the post-test minus the pre-test. More specifically, using a dichotomous model, correct answers are marked as 1 and incorrect answers are marked as 0, which are then edited into a text file format and entered into the WINSTEPS software. This allows the students' pre- and post-class competence values ​​to be derived, giving the specific values ​​for the pre- and post-tests, and thus the value-added. More specifically, the Rasch model is used to derive the competence value from the test results. To evaluate students' learning outcomes over a period of time (e.g., a class), the median value-added across all students is taken. A score greater than the median is defined as good, represented by "1," while a score less than the median is defined as poor, represented by "-1." This transforms the abstract learning outcome into "1" and "-1." The pre- and post-class tests are largely identical in terms of content, ensuring that differences in difficulty do not lead to inaccurate learning outcomes. Specifically, the same score improvement will have different learning effects for students with different academic performance. In other words, simply using the score improvement value cannot accurately reflect the student's learning effect. For example, if a student with a better foundation improves from 80 points to 90 points and a student with a weaker foundation improves from 20 points to 30 points, the former's ability improvement is greater. Therefore, it is necessary to calculate based on each student's own situation. Specifically, the Rasch model is used to obtain the student's ability, which is expressed as: Among them B n Indicates the student's ability, that is, the value of the pre-test or post-test, D i Indicates the difficulty of question i, which is related to the student's answer. n and D i The unit is logit, P ni The probability that the nth person answers the i-th question correctly is 1-P ni represents the probability that the nth person answers the i-th question incorrectly; B n Smaller than D i , that is, the most difficult question in the whole test paper should be such that no one can do it. Specifically, B n and D i The value range of -5 to 5 is a reasonable range, indicating that the difficulty is moderate, which can make the test more reliable. Otherwise, the reference value is not very significant. The value-added evaluation of the present invention takes into account the difficulty differences of the questions and the different difficulty levels of the same questions for different students. Therefore, the value-added evaluation of the present invention can more realistically reflect the students' learning effects.

[0031] Table 2: Example of value-added evaluation

[0032]

[0033] S3, building a learning effect prediction model and training it;

[0034] For teachers and educational researchers, using facial expressions to explore student learning outcomes is a tedious and complex task. Therefore, a new, scientific, and rational technology or method is needed to carry out this task. Support vector machines (SVMs) fit the bill perfectly. SVMs are a new machine learning method based on statistical learning theory and belong to the supervised learning model category. That is, their training data consists of a set of inputs and associated output values. Compared to other machine learning methods, SVMs are particularly adept at solving complex problems, such as those involving high dimensions, nonlinearity, and multi-classification.

[0035] The present invention constructs a support vector machine binary classification model to obtain the functional relationship between the processed micro-expression signal obtained in step S1 as a classification sample set (X) and the corresponding learning effect value-added evaluation result obtained in step S2 as a classification label (Y). Specifically, the support vector machine binary classification model is constructed using MATLAB programming software. The constructed support vector machine binary classification model can obtain the functional relationship between the classification sample set (X) and the classification label (Y) through training. Specifically, the support vector machine binary classification model is similar to a black box. The support vector machine is a new machine learning method based on statistical learning theory and belongs to a supervised learning model in machine learning. Its mechanism is to find an optimal classification hyperplane that meets the classification requirements, so that the hyperplane can maximize the blank area on both sides of the hyperplane while ensuring the classification accuracy. That is, its training data consists of an input set and an associated output value.

[0036] The processed micro-expression signal obtained in step S1 is used as the classification sample set (X), and the corresponding learning effect value-added evaluation result obtained in step S2 is used as the classification label (Y). "1" indicates a good learning effect, and "-1" indicates a poor learning effect. X and Y are input into MATLAB, and the fitckernel function is called for training. In order to improve the accuracy of the training, the corresponding X and Y are divided into a test set and a training set according to a ratio of 2:8, wherein the X and Y in the test set and the training set correspond to each other. Specifically, the training set and the test set samples are used for training, and both the time domain micro-expression signals and the frequency domain micro-expression signals are trained. More specifically, the learning effect prediction model corresponding to the four kernel functions of linear, quadratic polynomial, cubic polynomial, and Gaussian is used for training to obtain the classification accuracy. The training is performed at least once. Since the classification accuracy obtained by training is consistent when the ratio of the training set and the test set of the same sample set is the same, the present invention does not need to repeat the training process multiple times. The larger the sample size in the sample set, the more accurate the prediction result obtained.

[0037] More specifically, the average of 33 sample sets is taken, and the classification results are shown in the figure below, taking "happy" and "sad" as examples. Figure 4 As shown, the horizontal axis is "sadness", the vertical axis is "happy", and the straight line is the classification line. Corresponding to the trained model, the data points on the left side of the classification line indicate poor learning effects, and the data points on the right side of the classification line indicate good learning effects. More specifically, the input micro-expression signal can be the time domain signal obtained in step S1, or it can be the frequency domain signal obtained in step S1. Preferably, the input micro-expression signal is the frequency domain signal obtained in step S1. Compared with the time domain signal, the change of the micro-expression signal in the frequency domain signal is more obvious, the classification accuracy is higher, and a frequency can be obtained in a few minutes. The machine learning process is easier to find patterns, the training process is faster, and the output of the prediction results is also faster. The kernel function can be a linear function, a quadratic polynomial function, a cubic polynomial function, or a Gaussian function; preferably, the kernel function is a Gaussian function, and the expression of the Gaussian function is: where x i and x j represents the sample vector, ||x i -x j || represents the norm of the sample vector, 2σ 2 represents the hyperparameters of the Gaussian kernel function, specifically, the hyperparameter 2σ 2 The value is 1; the Gaussian kernel function is a kernel function with strong locality, which can map a sample into a higher-dimensional space. Compared with the polynomial kernel function, it has fewer parameters. Due to the smaller number of parameters and stronger mapping ability, the Gaussian function is better when used as the kernel function of the learning effect prediction model of the present invention, with higher prediction accuracy and faster calculation speed.

[0038] When the kernel function is a Gaussian function, the classification accuracy is higher, that is, the prediction accuracy of the learning effect is higher. Figure 5 is the classification result of the model with different kernel functions on time domain signals and frequency domain signals, such as Figure 5 As shown in the figure, different kernel functions have different classification accuracy rates. When the kernel function is a Gaussian function and the model input is a frequency domain signal, the classification accuracy of the corresponding model is higher, and the accuracy of the learning effect prediction is higher, reaching 89.7%.

[0039] S4, evaluate the performance of the trained learning effect prediction model.

[0040] When making a prediction, it is only necessary to perform step S1 on the signal of the expression to be tested, and then directly input the learning effect prediction model obtained in step S3 to output the prediction result. Since the classification accuracy is not the only factor in evaluating the learning effect prediction model, based on the comparison of the prediction result and the actual result of the learning effect prediction model obtained in step S3, the present invention further evaluates the classification performance of the trained model using frequency domain micro-expression signals and selecting Gaussian function as the kernel function. The use of the indicator F1-score can comprehensively consider the precision and recall rate. The precision rate refers to how many of the samples predicted to be positive are truly positive samples; the recall rate refers to how many positive examples in the sample are predicted correctly. The specific value can be obtained from the confusion matrix diagram. Figure 6 As shown in the figure, the prediction of the classification model in each category can be obtained from the confusion matrix diagram. This figure is obtained when the model is trained using frequency domain micro-expression signals and the Gaussian function is selected as the kernel function. That is, when the model is trained using MATLAB, the corresponding confusion matrix diagram is obtained. The horizontal axis is the prediction situation, the vertical axis is the actual result, and the cells represent the arrangement of the predicted results and the actual results. The numbers on the cells represent the prediction accuracy of the cells. The 74% area represents how many samples with the actual result of "-1" are correctly predicted as "-1", that is, correctly predicted as negative samples (True Negative or TN), 97% represents how many samples with the actual category of "1" are correctly predicted as "1", that is, correctly predicted as positive samples (True Positive or TP), 26% represents how many samples with the actual category of "-1" are incorrectly predicted as "1", that is, incorrectly predicted as positive samples (False Positive or FP), and 3% represents how many samples with the actual category of "1" are incorrectly predicted as "-1", that is, incorrectly predicted as negative samples (False Negative or FN). Specifically, the larger the TN and TP of the learning effect prediction model, the smaller the FN and FP, the better the classification performance of the model and the higher the prediction accuracy. The F1-score is calculated as follows:

[0041]

[0042]

[0043]

[0044] Generally, if the value of the parameter F1-score exceeds 0.7, it means that the corresponding model can be used for prediction and has high reliability. Therefore, the learning effect prediction model obtained in step S3 is highly reliable, and the method of the present invention can be used to predict students' learning effects.

[0045] ROC curve is also one of the important bases for evaluating the performance of learning effect prediction models. When the model is trained, the corresponding ROC curve will be generated. The ROC curve can show the true positive rate (TPR) and false positive rate (FPR) of the model. Figure 7 As shown, the false positive rate is 0.26 and the true positive rate is 0.97, indicating that the current model incorrectly predicts 26% of observations as class "1" and correctly predicts 97% of observations as class "1". Furthermore, the closer the area between the ROC curve and the horizontal axis (AUC) is to 1, the better the model's classification performance. The results show that the AUC value is 0.92, indicating that the learning effect prediction model obtained in step S3 is highly reliable and that the method of the present invention can be used to predict student learning outcomes.

[0046] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A learning effect prediction method based on students' facial expressions, characterized in that: The method comprises the following steps: S1, collection and processing of expression data; S2, constructing value-added evaluation of learning effects; S3, building a learning effect prediction model and training it; S4, evaluate the performance of the trained learning effect prediction model; The expression in step S1 is a micro-expression, and the processing includes denoising and normalization to obtain a micro-expression time domain signal and a micro-expression frequency domain signal; The value-added of the learning effect value-added evaluation in step S2 is obtained by subtracting the pre-test student ability from the post-test student ability, and taking the median of the value-added. If the value-added is greater than the median, it is defined as a "good" learning effect, and if the value-added is less than the median, it is defined as a "poor" learning effect; the values ​​of the pre-test student ability and the post-test student ability are expressed by Get, where B n Indicates the pre-test student ability or the post-test student ability. When calculating the pre-test student ability, D in the expression i and P ni When the test results before learning are used to calculate the post-test student ability, the D in the expression i and P ni Bring in the test results after students learn, D i Indicates the difficulty of the i-th question, P ni represents the probability of answering the i-th question correctly; The learning effect prediction model in step S3 is a support vector machine binary classification model. The micro-expression time domain signal and micro-expression frequency domain signal obtained in step S1 are used as classification sample sets, and the learning effect value-added evaluation result obtained in step S2 is used as a classification label to input into the support vector machine binary classification model for training.

2. The learning effect prediction method based on student facial expressions according to claim 1, characterized in that: The denoising process is first performed using a bandpass filter and then a moving average method. The frequency band passed by the bandpass filter is 2-6 Hz. The expression of the moving average method is: , where n is the expression change cycle obtained by fast Fourier transform, and n is an odd number, m represents the frame number, Indicates the actual expression value of the frame, and F indicates the expression value after moving average.

3. The learning effect prediction method based on student facial expressions according to claim 2, characterized in that: The normalization process is performed using the relative strength index (RSI), which is expressed as: , where A represents the increase in the upward fluctuation of the expression value within a cycle, B represents the increase in the downward fluctuation of the expression value within a cycle, and A+B represents the total fluctuation amplitude. RSI between 30-70 means that the student's expression has not changed significantly, and less than 30 or greater than 70 means that the student's expression has changed significantly.

4. The learning effect prediction method based on student facial expressions according to claim 3 is characterized in that: The kernel function of the support vector machine binary classification model is one of a linear function, a quadratic polynomial function, a cubic polynomial function, and a Gaussian function.

5. The learning effect prediction method based on student facial expressions according to claim 4 is characterized in that: The step S4 uses the indicator F1-score and ROC curve to evaluate the learning effect prediction model trained in the step S3.

6. The learning effect prediction method based on student facial expressions according to claim 5 is characterized in that: The calculation method of the indicator F1-score is: Among them, TP indicates how many samples with the true category as positive samples are correctly predicted as positive samples, that is, correctly predicted as positive samples; FP indicates how many samples with the true category as negative samples are incorrectly predicted as positive samples, that is, incorrectly predicted as positive samples; FN indicates how many samples with the true category as positive samples are incorrectly predicted as negative samples, that is, incorrectly predicted as negative samples.

Images