The role of the O2O blended teaching model in improving the teaching effectiveness of physical education classes

3.1 Design of a hybrid teaching condition evaluation model based on an improved K-modes algorithm

O2O teaching mode refers to a teaching mode that combines online education and offline physical education. O2O provides course content, teaching resources, and learning support through online platforms while combining with offline physical education venues for practical teaching and tutoring services. O2O teaching mode makes full use of the advantages of Internet technology to make up for the limitations of traditional offline education and, at the same time, provides a more flexible and personalized learning mode. O2O teaching mode makes full use of the advantages of Internet technology to make up for the limitations of traditional offline education and, at the same time, provides a more flexible and personalized learning mode.

In view of the advantages of the O2O teaching mode, the study integrates the O2O teaching mode into physical education teaching. Although physical education belongs to a discipline with strong practicality, physical education still involves the teaching of certain theoretical knowledge, and the online teaching part of the program can provide sports-related video lectures, online reading materials, and so on. In the post-epidemic era, facing some special circumstances, such as during the epidemic or when students are physically unable to come to school, online teaching can be used as a way of distance learning to ensure that students do not interrupt their learning due to external factors.

In the O2O blended teaching model, the interaction between teachers and students leaves a large amount of data on online teaching activities, and valid teaching data can provide reliable support for teaching evaluation. Assessment mechanisms occupy an important position in the design of the blended teaching model, which requires more scientific and quantitative evaluation methods, and the teaching quality and teaching evaluation system needs to be improved in parallel [11]. Therefore, the study introduces clustering algorithms and attention mechanisms into the hybrid teaching model to build a more scientific and reasonable teaching evaluation and learning prediction model and improves the hybrid teaching model in the reverse direction based on the evaluation feedback.

The study constructs a teaching evaluation model for the O2O blended teaching model based on the K-modes clustering analysis technique. The K-modes clustering algorithm is a typical division-based clustering algorithm, an unsupervised learning algorithm that uses simple matching distance (SMD) to classify variables [12,13].

In the K-modes algorithm, X = { x 1 , x 2 , … , x g } denotes the set of sample points, the attribute of the sample point x i = { x i 1 , x i 2 , … , x i h } is denoted as { A 1 , A 2 , … , A h } , and the value of the attribute A i is defined in equation (1). In equation (1), f denotes the number of objects in the attribute.

The K-modes algorithm uses the Hamming distance as the basis for sample classification. The distance of any sample point x i l , x j l is calculated in equation (2).

For a given initial number of clusters and an initial cluster centre, the sample set is assigned to different data classes based on the Hamming distance size. Then, the average distance from each sample data point to its corresponding centre is calculated, and the position of the cluster centre is continuously updated. If each cluster centre and sample agree with the last clustering result, the clustering is completed; if there are differences in the results, the steps are repeated until the end of the iteration, and a schematic illustration of the K-modes clustering analysis process is shown in Figure 1.

Figure 1

K-modes cluster analysis process diagram.

Most of the traditional K-mode algorithms are clustering for the known number of clusters K . However, the value of the number of clusters K is not certain in practical application scenarios, and a priori knowledge or experience is needed to determine the value of the number of clusters K , which leads to difficulty in selecting the K value. At the same time, in the face of fuzzy data or high-dimensional data, in order to ensure the clustering effect, the determination of the number of clusters K needs to be more cautious. At present, most studies use software packages to determine the number of clusters K , but the software packages have a large amount of computation, computational efficiency is low, and there are even some problems that cannot be calculated. In this regard, the study used sum of squared error (SSE) to determine K . The SSE is calculated in equation (3), where k indicates the number of clusters, Z l indicates the cluster centre of the l rd. cluster L l , and Dist ( x , Z l ) indicates the similarity of the samples x and Z l . The size of the SSE indicates the good or bad clustering result.

The number of occurrences of each value in the sample data is calculated in equation (4). In equation (4), x i j is the sample attribute value, where i = 1 , 2 , 3 , … , s and s are the number of sample data; j = 1 , 2 , 3 , … , m and m are the number of sample data dimensions; f ( x i j ) is the frequency of x i j ; C is the set of allowed values of x i j , where C 1 is the number of times the first element appears in the sample data set, 5 is “excellent,” 4 is “good,” 3 is “moderate,” 2 is “pass,” and 1 is “fail.”

Traditional K-modes use a frequency-based anomaly detection algorithm for similarity calculation in equation (5), where l is the sample data. After the similarity is obtained, K-modes are used to perform pre-clustering operation, and the optimal number of clusters is obtained by comparing the size of K value with the change of SSE size. The optimal K value is related to the amount of sample data and the dimensionality of the data.

The traditional K-modes clustering algorithm usually selects the cluster centres randomly, but this increases the probability of isolated points being selected and also increases the probability of different cluster centres being close together, making the clustering results less stable and accurate. The study improves on this by assuming that all data points are likely to be cluster centres and calculating the SSE value, with the minimum point taken as the initial cluster centre for l clustering. This operation is repeated for all remaining data points until the completed clusters are bigger than the known classifications. The minimum error square algorithm solves for the initial cluster centres, as shown in Figure 2 [14,15].

Figure 2

Small error squared algorithm for solving initial cluster centres.

The clustering centres are calculated in equation (6). In equation (6), Z is the initial set of clustering centres, z l represents the sample data points that make the SEE fall fastest and z l varies, s represents the number of sample data, and x represents the sample data.

The K-modes clustering algorithm measures the clustering distance by the similarity between samples x and Z l , which is not suitable for dealing with complex situations where different attributes of a sample or the same attribute are linked, and the clustering results can be too different from the actual situation. The study uses the co-occurrence method to improve the traditional clustering centre distance measure.

The distance between two different samples of data under the attribute A i for the attribute A i is defined in equation (7). In equation (7), x A i is the set of values of under x A i , is the set of values of another attribute when u A j x A i P is taken; is the conditional probability, all values of the conditional probability are in the range [0,1]. u ' is the complement of the set of values of another attribute taken at the time of A j x A i with respect to u . When different samples of data take the same value under the A i attribute, the distance to A j under A i is 0 and the distance is minimal.

The distance metric between two sample data is defined in equation (8).

3.2 Design of a model for predicting achievement in a blended teaching model based on improved self-attention

O2O is a hybrid teaching model that combines online and offline teaching and learning and is guided by computer-based Internet technology that allows for the integration of multiple teaching resources for innovative teaching and learning. Various online learning behaviour data can be used to predict learning outcomes, provide timely feedback to teachers, and monitor student learning [16].

Learning behaviours include online teaching attendance, online assignment completion rate, online learning hours, number of questions and posts, and platform visits. A Self-Attention-based performance prediction model was constructed. Each feature that affects learning performance is modelled, and feature weights are calculated by the Self-Attention mechanism to improve the accuracy of the algorithm. The attention mechanism is a form of deep learning. With limited computer power, the attention mechanism allocates computational resources to more important tasks, reduces the attention of non-critical information to solve the algorithm information overload, and improves the processing efficiency and accuracy of the algorithm, which is widely used in image processing, language processing, and speech recognition. The network schematic of the attention mechanism is shown in Figure 3 [17,18].

Figure 3

Schematic diagram of attention mechanism network.

The calculation process of the attention mechanism is shown in equation (9), where Q , K , V denotes the matrix consisting of the vectors Query, Key, and Value, respectively; softmax denotes the normalization function; d k denotes the dimension of Key; and the weight matrix W Q , W K , W V of Q , K , V is also included in the self-attentive mechanism.

The input coding layer of the model needs to be preceded by pre-processing of the raw data for a uniform data coding transformation. Data with large values of attributes are normalized using the min–max normalization method. The coding layer first performs discrete-value vectorization, which will linearly vary the input feature vectors, as shown in equation (10), where W Q , W K , W V denote the high-dimensional weight matrix, W Q , W K , W V ∈ ℝ k × n , x i denote the feature vectors, n denote the number of attribute features, and k denote the attribute feature dimensions.

The self-attentive mechanism mines the degree of influence of different generic features with greater accuracy, and the weights are calculated in equation (11). q i , k i , v i represents a mapping vector.

Following this, a vector of attribute features of student behaviour can be obtained from f i , which is calculated in equation (12). In equation (12), β i represents the attention weight score of the feature attribute.

However, students’ historical grades are also critical to the prediction of grades. Students’ final grades are strongly correlated with their historical grades in the first two stages. The study designs an improved two-way attention (TWA) mechanism to further explore the correlation information between each feature data and grades, calculate the attention scores of the two-stage grades using two attentions, and predict grades after further feature fusion. The TWA model framework is shown in Figure 4 [19,20].

Figure 4

Two-way attention model framework.

The coding layer of TWA is similar to the single attention mechanism, using a high-dimensional parameter matrix to transform discrete attribute values and historical grades into a feature matrix A and grade vectors g 1 and g 2 , as defined in equation (13). In equation (13), A , g 1 , g 2 ∈ ℝ k , n denotes the number of attribute features, and k denotes the dimensionality of the attribute features and vectors; T denotes the vector transpose symbol.

A multi-layer perceptron (MLP) with two fully connected layers is used for attention weighting, and the MLP structure is shown in Figure 5.

Figure 5

Structure diagram of multi-layer perceptron.

The weight calculation process is shown in equation (14), where A i represents the feature vector at the position corresponding to the column vector of the feature matrix A , A i = ( a 1 i , a 2 i , … , a k i ) T ; u i and v i represent the attention weights based on the first and second stage historical performance, respectively; [ ; ] represents the feature stitching operation.

The attention scores for each of the two historical stages were obtained by renormalizing ownership to β i , γ i and the calculation process is shown in equation (15).

The two attention scores are weighted and summed with A i to obtain the feature vectors f 1 and f 2 , which are calculated in equation (16).

Finally, the features of f 1 and f 2 are fused by Maxpooling, and the maximum values of the corresponding positions of f 1 and f 2 are selected as input; the expressions are given in equation (17).

Ultimately, the label prediction layer predicts student achievement categories based on attribute features and calculates scores for each category using the softmax function, see equation (18).

The model was trained using a backpropagation algorithm with a classification loss of cross entropy [21]. The L2 regular term is used to constrain the function to avoid overfitting, and the loss function is given in equation (19). In equation (19), λ ∥ θ ∥ 2 denotes the L2 regular term, N denotes the total amount of data in the training set, p i denotes the predicted probability, θ denotes the set of parameters, and y i denotes the label of the first i student.

4.1 Performance analysis of the teaching evaluation model of the O2O blended teaching model

The evaluation data came from the real data in the academic affairs management system of a university, and the specific teaching evaluation indicators included teaching quality, teaching attitude, teaching skills, and extra-curricular links. The trend of clusters 1, 2, and 3 is shown in Figure 6. From Figure 6, most of the teaching evaluations are distributed in three grades: “excellent,” “good,” and “moderate.” With the use of the blended teaching mode, the first and second clusters show an increasing trend over the semester, while the third cluster shows a decreasing trend, with the percentage of “excellent” rating increasing from 15.7 to 27.9% and the percentage of “good” rating increasing from 24.2% to 28.8. The blended teaching mode has achieved better teaching results, and the teaching evaluation has gradually increased.

Figure 6

Proportion of different cluster evaluation indicators.

Publicly available datasets from different domains, including Iris flower dataset (Iris), Wine Quality Dataset (Wine Quality Dataset), Anomalous Network Traffic Dataset, Twitter user dataset, and Facial Expression Recognition Dataset, are used to divide the datasets required for the experiments into the training set and the testing set at a ratio of 9:1. The clustering centre, number of clustered samples, minimum error sum of squares, accuracy, and recall were introduced as evaluation metrics. The research constructed model was compared with the algorithms of random determination of clustering centres + simple ratio distance and SSE method to determine clustering centres + AVF, and the outcomes are denoted in Table 1.

As seen in Table 1, the improved K-modes performed optimally in all metrics. The minimum sum of squared errors of the improved K-modes algorithm is 58.18 and 36.25% lower than the other two algorithms, respectively. The SSE is smaller when the samples within each type of cluster are more similar, and the clusters are tighter. Therefore, the improved K-modes algorithm clustering is optimal, with the data points within each cluster being closest to the centre of mass.

The core idea of the K-modes algorithm is to make the clusters themselves as compact as possible and the different clusters as separated as possible; thus, the optimal goal of clustering is to make the average intra-class distance small and the average inter-class distance large. However, the three algorithms do not agree on the measure of distance, so the proportion of the average intra-class distance to the average inter-class distance is used as the rating index, and the experimental results are expressed in Table 2. In Table 2, the improved K-modes clustering algorithm has the smallest distance, and the smaller ratio indicates better clustering and the stronger the relevance of the clustering results, and the algorithm proves the effectiveness of the blended teaching model.

The statistical results of the Jaccard coefficient (JC) and Fowlkes and Mallows Indicator (FMI) are shown in Figure 7. As can be seen from Figure 7(a), the JC value of the research design model is the highest on all datasets, and the clustering results are similar to the real situation. As can be seen from Figure 7(b), the FMI value of the research design model is higher than the other two models, and the results take values above 0.75, which comprehensively evaluates that the SSE + Co-occurrence model clustering has a high accuracy rate.

Figure 7

Comparison of clustering performance of different models. (a) JC and (b) FMI.

4.2 Performance testing of the O2O blended learning model performance prediction model

The study conducted performance tests on the TWA algorithm. The dataset contained 1,784 valid data, including 30 feature attributes such as students’ basic information, learning behaviour information, and students’ historical grades in the previous two stages. The rate of the training set to the test set was 7:3, and the prediction results of the TWA algorithm for the final grades are shown in Figure 8. The results predicted by the model are basically the same as the actual grades, with only minor differences, which can be considered to be within the normal error range. The model predicts with high accuracy and is able to make accurate predictions about the teaching performance of the blended teaching mode of PE. The model can be used in conjunction with the blended teaching approach to provide timely feedback and early warning of learning, prompting teachers to change their teaching strategies and urge students to adjust their learning styles and attitudes.

Figure 8

Comparison between predicted values and actual values of predicted data.

Both historical and final predicted scores were divided into four grades, A for “excellent,” B for “good,” C for “moderate,” and D for “fail.” The attention weighting results for all attributes are shown in Figure 9. The TWA model visualises the weighting of factors and allows teachers to develop more precise and detailed blended personalized instruction for different students.

Figure 9

Attribute probability distribution. (a) Level A, (b) level B, (c) level C, and (d) level D.

TWA was compared with five common prediction methods, namely Self-Attention, support vector machine, parsimonious Bayes, decision tree, and logistic regression. The model training results are expressed in Figure 10, in which the TWA algorithm performed best in all four evaluation metrics, accuracy, precision, recall, and F1-value, with 98.23% prediction accuracy and 98.42% precision, which were 57.45 and 135.57% better than the worst-performing plain Bayes, respectively, and better than the Self-Attention algorithm which did not consider 6.62% improvement in accuracy and 4.45% improvement in precision compared to the Self-Attention algorithm which did not consider the fusion of historical performance features. For the contradictory indicator of accuracy, recall, the TWA algorithm still showed the highest value of 92.45%, and the algorithm had a low miss rate; the overall evaluation indicator F1 was 91.78%, which was a good balance between accuracy and recall, and the algorithm had the best overall learning performance.

Figure 10

Performance indicators of different algorithms.