Virtual Learning Environment and Intelligent Interaction Design for the Construction of Online Platform for Teaching Aesthetic Education in Colleges and Universities under Three-Taiwan Fusion Environment

Cloud computing provides powerful computing resources and storage capacity for the online platform, enabling it to easily cope with large-scale concurrent student access and complex data processing needs. Through virtualization technology, the resources of physical servers are integrated and distributed so that users do not need to care about the construction and maintenance of the underlying infrastructure, and only need to obtain the required computing resources on demand, which greatly reduces the platform construction and operation costs, and at the same time improves the flexibility and expandability of the system.

Internet of Things (IoT) technology connects various devices (e.g., sensors, intelligent terminals, etc.) to the Internet, realizing the information interaction between people and things, and between things and things. In the online platform for teaching aesthetic education in colleges and universities, IoT technology can be used to collect students’ learning behavior data, environmental data, etc., providing data support for personalized teaching and intelligent interaction. For example, the physiological state of students during the process of art creation is monitored through smart wearable devices, so as to adjust the teaching rhythm and content difficulty to achieve the best learning effect.

Big data technology is capable of storing, managing, and analyzing massive and diverse data. In the platform, big data can be used to analyze students’ learning habits, interest preferences, knowledge mastery, etc., to dig out the potential education laws and students’ needs, and to provide a basis for accurate teaching, resource recommendation and teaching evaluation. Through in-depth analysis of historical data, teachers can identify problems in the teaching process in a timely manner and improve teaching methods and strategies in a targeted manner.

According to the architecture of 3D virtual learning environment, this paper designs a simulation prototype, i.e. cloud computing platform, whose architecture is shown in Fig. 1. High performance and 3D virtual learning environment can be constructed through this platform.

Figure 1.

Prototype design

Among them, the servers mainly include login servers, space servers, simulators, scenario file servers, file servers, database servers, and web servers, which are distributed in storage and application nodes, respectively. The servers in the cluster system install RedHat 9.0 Linux operating system, and the parallel environment uses MPICH 2.0 to present.

The leftmost is the cloud client, the user through the embedded JSP and JAVA and other plug-ins of the IE browser to interact with the prototype system in the method of service packaging, functionality and business logic is mainly realized by middleware technology. The following describes the operation process of the simulation prototype.

First of all, when the user performs an operation, the cloud client will make a judgment and different types of operations will be submitted to the main server for processing in different ways. When the user logs in, registers, manages users, and orders business, the main server compares, enters, or takes out the information in the database according to the type of request, and returns the information the user wants. Secondly, when the user carries out the business application, since it is not a simple operation, the master server will perform load balancing according to the condition of each node, and assign the application nodes and computing nodes that are the most suitable for the user’s application to the user, so that the user’s operation can achieve the highest efficiency. The following is a concrete introduction to the functional design of the 3D virtual learning environment through the designed simulation prototype, i.e., the cloud computing platform.

The concept of three-dimensional virtual learning environment is very broad, can be provided by text, graphics, images, sound, video and other multimedia forms (such as electronic courseware PPT, tutorial lesson plan documents, teaching case video, MP3 voice resources, etc.), can also be presented by way of three-dimensional physical objects that can provide real-time interaction (such as virtual experiments, modeling demonstrations, virtual campus, etc.). The three-dimensional virtual learning environment for teaching aesthetic education in colleges and universities discussed in this paper mainly provides the following functions: 1)

User registration and login

The user registers through a unique account, and at the same time gives the user’s nickname and gender in the learning environment, and chooses a favorite virtual image as an avatar. After the system validation, the user is received as a legitimate user. After successful registration, through the login page, you can login to enter the three-dimensional virtual learning environment for learning.

Collect data such as students’ basic information, learning history, interests, learning styles, etc., and transform the data into a format that Python can handle, so as to obtain user labels to construct user profiles. In the process of extracting course labels, it is necessary to firstly perform word splitting on the course text, and use the TFIDF algorithm to count the results of word splitting [20], locate the course label keywords, and then construct the user portrait. After the course is subworded, the label keywords of the course can be obtained, such as “Principles of Aesthetic Education in Colleges and Universities” can be obtained after subwording [“colleges and universities”, “principles of aesthetic education”], and these two These two keywords constitute the vector space of the course, which represents the characteristics of the learning resource. By counting the labeled keywords of all the courses that the user has studied, the user label and weight can be obtained, and the weight ω_{i, t} of label t in user i can be calculated by equation (1): (1)ωi,t=Ni,tNi

where N_{i, t} is the number of times tag t appears after binning all the course text of the course taken by user i, and N_i is the total number of tags obtained after binning the course taken by the user.

As the number of courses a user takes increases, the number of course tags also increases, but the percentage of course tags in the user profile should not increase linearly with the increase of tags. For example, a label i appeared 5 times, label j appeared 10 times, although the number of times j appeared is twice as much as i, but it does not mean that the user’s interest in j is also twice as much as i. Therefore, formula (1) is further optimized by using the log function to normalize the number of times the label appeared, and formula (2) can be obtained, and the label weights obtained by this formula can more accurately reflect the user’s knowledge structure as well as interest preferences: (2)ωi,t=logNiNi,t

Through the above method, the user profile can be constructed by utilizing the user attributes and historical learning information, and the user profile is represented using the representation method of vector space model. Next, the relevant recommendation algorithms are implemented and improved to generate a recommendation candidate set, after which the fit between the user profile and the items in the candidate set is calculated, and the most suitable learning resources for the user are filtered out.

In this paper, user-based collaborative filtering algorithm is selected to realize collaborative filtering. In the calculation method of user similarity, several more classical similarity calculation methods such as Pearson correlation coefficient and cosine similarity are usually used. In the dataset used in this paper, only the overall score data of the course is included, and from the formula, Pearson similarity is not applicable due to the need for user ratings of individual items to participate in the calculation process, so this paper adopts the cosine similarity and Jekyll and Hyde similarity to conduct experiments and make targeted improvements. After the data is processed, the cosine similarity [21] can be calculated using equation (3): (3)sin(u,v)=u→⋅v→|u→|*|v→|=u1*v1+u2*v2+⋯+un*vnu12+u22+⋯+un2*v12+v22+⋯+vn2=unionlen(u,v)len(u)*len(v)

The u→ in the formula can be interpreted as a vector of all items, denoted as [u₁, u₂, ..., u_n]. Since only the courses that the user has taken are used as dictionary values when constructing the dictionary, the courses that the user has taken can be denoted as 1 in vector u→ and 0 otherwise. The denominator in the formula denotes the product of the modes of the two vectors, but the courses that the user hasn’t taken aren’t saved in the dictionary, so u12+u22+⋯+un2 is the value in the constructed dictionary length of the constructed dictionary. After the optimization of the formula, it can be calculated by using the & (with) operation and the len() function in Python, which reduces the amount of computation.

Equation (3) simplifies the calculation, but it is not optimized for the characteristics of the user’s study courses, although it can achieve better calculation results in some data backgrounds, but the stability is not high, therefore, this paper adjusts the weights by introducing the method of course category percentage, as shown in Equation (4): (4)sin(u,v)=∑i=1n(ωi*li)*unionlen(u,v)len(u)*len(v)

Eq. (4), compared with Eq. (3), introduces the ∑i=1n(ωi*li) parameter, which is equivalent to adding a reconciliation parameter to the original similarity, where ω_i indicates the proportion of the course categories i in the courses taken by the target user to the total number of courses taken by the user, and l_i indicates the proportion of the same course categories i in the courses taken by other users to the total number of courses taken by the user, and by this parameter, the effect of the main course categories on the similarity is amplified, and the finding of the more uniformity in subject preferences between the similar users and the target users, and thus better recommendation results can be achieved in the education domain.

For Jaccard similarity [22], which itself is used to measure the degree of association between two sets consisting of Boolean values, computationally, it is only necessary to compute the ratio of the intersection of the two sets to the concatenation of the two sets, and again, to amplify the effect of the major course disciplines on the similarity, it is optimized by using Eq. (5): (5)sin(u,v)=∑i=1n(ωi*li)*|N(U)∩N(V)N(U)∪N(V)|

After experiments using cosine similarity and Jekyll and Hyde similarity to obtain similar recommendation results, both of which are in line with the user’s original learning preferences. However, after repeated experiments, it is found that the percentage of each subject category in the results obtained by using the improved cosine similarity is more similar to the percentage of course categories reflected in the user’s studied courses, i.e., the recommended results are more in line with the user’s original interests, so in the similarity calculation of this algorithm, this paper selects the optimized cosine similarity as the calculation method.

Learning Resource Recommendation Algorithm with Fusion of User Profiles By fusing labeled user profiles with recommendation algorithms and learning resources, the fit between the user and the learning resources in the recommendation candidate set is calculated, and the N items with the highest fit are recommended to the user.

When improving and realizing the recommendation algorithm, in order to ensure the effectiveness of the recommendation results, a single algorithm is set to output no more than 10 recommendation results based on the similarity score, so as to reduce redundant and invalid recommendations. Therefore, after mixing the results, there will be at most 30 results in the candidate set, and only need to calculate the fit between these 30 candidate resources and users to get the final recommendation results, which greatly reduces the amount of computation compared to calculating the fit between all resources and users. The fit between resource i and user u can be calculated by equation (6): (6)ψu,i=∑t∈Tωu,t*δi,t

In Eq. (6), ω_{u, t} denotes the weight of label t in the vector space model of user u, δ_{i, t} denotes the weight of label t in the features of this learning resource, i.e., the proportion of label t in all labels of the course after the course i has been subdivided into words, and T denotes the set of all labels. By calculating the fit ψ_{u, i} between the user and all the courses in the coarse candidate set, the N items with the highest fit can be obtained, and the user profile as well as the resource features can be fully utilized by this method to get the fit between them.

Since the raw audio captured is of low quality, audio processing is performed before speech recognition.After Kinect captures the audio, it employs algorithms to improve the audio quality such as echo cancellation algorithms, automatic gain control algorithms, and noise suppression algorithms. When the sound returns from the microphone, it creates an echo, Kinect eliminates the echo by extracting the user’s voice pattern and then extracting specific audio from the received audio based on this pattern. Automatic gain control is to keep the amplitude of the user’s voice consistent with the time, avoiding the impact on speech recognition due to differences in the user’s position. Noise suppression is to remove ambient noise from the audio signal, which allows the user’s voice to be captured more clearly and accurately by Kinect.

Speech recognition can be categorized into continuous speech recognition and isolated word speech recognition according to how the speech is told. Isolated word speech recognition recognizes separated words and phrases, which has a high recognition accuracy although this limits what the speaker says and is not as natural and convenient as continuous speech recognition. Since this paper wants to realize the control of smart character movement through voice commands, only isolated word speech recognition is required.

After starting the Kinect device, the first step is to initialize the audio data stream and register the speech recognition engine, which is the core part of speech recognition and is used to analyze and interpret the audio data stream. Then it is also necessary to create an XML grammar file in which all the required voice commands are written. After writing the grammar XML file, it is loaded into the speech recognition engine, and then the voice commands captured by Kinect are matched with the predefined voice commands in the XML file; if the match is successful, the next step of event processing is performed, and then it continues to recognize new voice commands issued by the user. If the match fails, this voice data is discarded.

The experiments in this paper use a real dataset in the external environment of E-learning, namely the Aesthetic Education 2017-2018 (Ae) dataset, which is proposed from an online platform for teaching aesthetic education in a university. The dataset contains implicit information about learners’ interactions with the online learning platform and learning resources. The dataset contains 359 learners, 1386471 interaction messages, 5237.16 total credit hours, and the sparsity of the dataset is 77.62%. In order to evaluate the performance of the algorithm, the dataset is divided into training set and test set according to 7:3.

The definition of rating sparsity is shown below: (11)Sparsity=(1−∑Ratings∑Users×∑Items)×100%

In E-learning environments, some information about learners, either explicitly or implicitly, is collected from their active sessions by observing their behavior and interaction with the platform. In order to make the data suitable for mining and implementing recommendations, the raw dataset needs to be cleaned and preprocessed. Implicit learning ratings are represented by a value between 0 and 5.

The main purpose of the experiment is to test the prediction accuracy of the method proposed in this paper, and MAE is chosen for the evaluation, defined as follows: (12)MAE=∑(u,j)|R˜u,j−ru,j|M

where M is the total number of ratings for all learners, R˜u,j is the learner u‘s predicted rating of the learning object j and r_{u, j} is the actual rating.

Consider the algorithms of this paper for different similarity metrics to compare the performance: Ours-Pearson, Ours-Tanimoto, Ours-Jaccard, Ours-Cosine.The first experiments were performed to find the best value of the number of nearest neighbors, K, on the static and incremental datasets, and the results of the experiments are shown in Fig. 4. In the figure, in order to determine the optimal value of the nearest neighbor number K for this paper’s method on the Ae dataset, the experimental selection range is controlled between 0 and 300. By increasing the value of K, it is found that there is a significant difference in the MAE values of this paper’s algorithm using different similarity measures. On the Ae dataset, the MAE values of the Pearson and Tanimoto similarity measures are significantly higher than those of the Jaccard and Cosine similarity measures, while the Cosine similarity measure achieves smaller MAE values when the number of nearest neighbors is greater than 180, i.e., the Ours-Cosine method possesses higher prediction accuracy. Overall, by changing the number of nearest neighbors, it can be found that Ours-Cosine performs better than other methods on the Ae dataset.

Figure 4.

Comparison of the MAE values of different similarity metrics

Figure 5 shows the results of the experiments conducted under the four similarity metrics with K values set to 30, 60, 180 and 200 respectively. It can be seen that for the Pearson and Tanimoto similarity metrics, the MAE values are higher under all four K values. It indicates that in the Ae dataset, the use of Pearson and Tanimoto similarity metrics could not achieve the desired prediction results.Jaccard, although more than the Pearson and Tanimoto methods MAE values of the MAE values at K values of 30, 60 and 180, its MAE values are at a lower level. However, when the K value is greater than 180, the MAE value of Jaccard produces a significant increase.The MAE value of Cosine method can be maintained at a lower level of 0.5~0.8 under all four K values, and the corresponding MAE value is the smallest when K=180. The experimental results show that the selection of the Cosine similarity metric can obtain higher prediction accuracy and model performance stability, which demonstrates the reasonableness of choosing it as the similarity metric for learning resources recommendation in this paper.

Figure 5.

Comparison of the MAE values of different similarity metrics

Fig. 6 shows the comparison results of the prediction accuracy of this paper’s learning resource recommendation method incorporating user profiles with the item-based learning resource recommendation method (Item), the deep learning-based learning resource recommendation method (Deep-learning), and the hybrid learning resource recommendation method (Hybrid) under the Ae dataset. From the figure, it can be seen that the method in this paper has a lower MAE value compared to the other three learning resource recommendation methods. Therefore, it has higher prediction accuracy than the other three methods, and can recommend more accurate and personalized learning resources for learners in the online platform for teaching aesthetic education in colleges and universities, which can fully satisfy the diversified learning needs of different students in the process of learning aesthetic education.

Figure 6.

Comparison of prediction accuracy of different methods

Expanding on the poses proposed in related studies, a database containing 60 whole-body poses is established to validate the effectiveness of the pose recognition method in this paper. 1)

Offline experiment

Six subjects (3 males and 3 females) were recruited to conduct the offline experiment.The Kinect device was placed horizontally, 50cm from the ground.The angle of the drive motor was positive 20 degrees. The background was a white wall. Subjects faced the Kinect device with their whole body in the field of view, 250 cm away from it, and performed 60 postures in sequence. For subjects 2~6, 110 samples were collected for each posture, respectively 6600 samples. For subject 1, 220 samples were collected for each posture, totaling 13,200 samples. 60% of the pose data for subject 1 was used for training and the other 40% was used for testing. All of the pose data for subjects 2~6 were used for testing.

To get the number of joint angle features with the highest accuracy, the number of joint angle features is gradually increased from 5 to 30. When the optimal number of joint angle features is obtained, the accuracy of analyzing this paper’s method is gradually increased from 2 to 60 classes.

The number of joint angle features gradually increases from 5 to 30 to analyze the pose recognition accuracy and the results are shown in Figure 7. With the increase in the number of joint angle features, the accuracy rate shows an overall increasing trend. The accuracy rate is highest when the number of joint angle features is 28, which is 97.51%. When the number of joint angle features is greater than 28, the accuracy rate tends to be stable.

Figure 7.

Different number of features and accuracy

The number of pose types increases gradually from 2 to 60 to analyze the pose recognition accuracy and the results are shown in Figure 8. With the increase in the number of pose types, the overall trend of pose recognition accuracy is decreasing. When the number of pose types is 3, 10, 13, 15, 18 and 26, the accuracy reaches 100%. When the number of pose types does not exceed 29, the accuracy maintains a high level, and when it exceeds 29, the accuracy begins to decline.

Figure 8.

Different categories of postures and accuracy

The results of the online experiment are shown in Table 1, with an accuracy rate of 98.43% for subject 1. The accuracy rates of the three newly recruited subjects are all above 95%. It shows that the pose recognition method constructed based on Kinect device in this paper can complete the learner’s pose recognition in the aesthetic education teaching platform with a high accuracy rate, which lays a solid foundation for the intelligent interaction in the teaching platform.

The Kinect device uses the TensorFlow framework to implement speech text recognition. The Conformer encoder uses 14 identical Conformer modules, with a number of attention heads of 5 and an attention dimension of 250 dimensions. The dimension of the feed-forward network is 3042 dimensions, and to prevent overfitting, the model uses a preheat learning strategy with a preheat learning rate of 0.005, a preheat step count of 1000, and an initial learning rate of 0.01. The model training is set with batch_size=125, num_workers=20, and the AISHELL-1 dataset is used, and only 180 rounds of training are performed in this experiment, taking into account the model size, the training duration is about 55h, and the results are shown in Fig. 9, (a)-(d) denote the training set LOSS, the test set LOSS, the training learning rate and the test set word error rate, respectively. The training results show that the best LOSS for 180 rounds is 0.524, the best word error rate is 0.058, and the model recognition correctness can reach over 95%.

Figure 9.

Speech recognition model training results

In this subsection, model training is performed in the experimental platform using the homemade CA-ES dataset, and the results are shown in Fig. 10, where (a)-(d) denote the training set LOSS, the training set model classification accuracy, the test set model classification accuracy, and the model recognition confusion matrix, respectively. The same training 180 rounds, the best LOSS is 0.264, and the best accuracy rate can reach more than 95%. The confusion matrix shows the correspondence between the model’s prediction results on the test set and the real labels, as well as the accuracy and error of the classification, as can be seen from the figure, except for the Surprise emotion, the accuracy of the other five emotions reaches more than 90%, and due to the small number of the Surprise emotion speech, the accuracy rate is lower. Overall, the use of Kinect device in this paper can effectively complete the voice text recognition and voice emotion recognition in the online platform of aesthetic teaching, which significantly improves the experience of intelligent interaction design in this paper.

Figure 10.

Speech emotion recognition model training results

Combining all the experimental results and analyses, it can be concluded that the improved learning resource recommendation method adopted in the virtual learning environment design of this paper can effectively meet the personalized needs of students regarding the learning of aesthetic education. In the intelligent interaction design, the pose recognition and voice interaction methods based on the Kinect device in this paper also demonstrate superior performance. Applying the virtual learning environment and intelligent interaction module designed in this paper to the online teaching platform of aesthetic education in colleges and universities, focusing on the construction of the platform in the “three-station fusion”, can significantly promote the benign reform of the teaching of aesthetic education in colleges and universities, enhance the level of the teaching of aesthetic education in colleges and universities, and promote the development of students’ ability of aesthetic education.