Optimization of Physical Education Course Resource Allocation Model Based on Deep Belief Network

Sun, Lili

doi:https://doi.org/10.1155/2023/8457760

Mathematical Problems in Engineering

On this page

Abstract Introduction Literature Review Conclusion Data Availability Conflicts of Interest Acknowledgments References Copyright Related Articles

Special Issue

Intelligent Target Detection in Complex Scene Image

View this Special Issue

Research Article | Open Access

Volume 2023 | Article ID 8457760 | https://doi.org/10.1155/2023/8457760

Optimization of Physical Education Course Resource Allocation Model Based on Deep Belief Network

Lili Sun¹

Academic Editor: B. Sivakumar

Received20 Aug 2022

Revised23 Sept 2022

Accepted03 Oct 2022

Published29 Apr 2023

Abstract

In order to meet the optimization needs of physical education curriculum resource allocation, the author proposes a deep belief-based physical education curriculum resource allocation technology. The efficient feature abstraction and feature extraction capabilities of deep belief technology fully explore the interests and preferences of learners on course resources. Because deep belief has strong capabilities in feature detection and feature extraction, it has unique and efficient feature abstraction capabilities for different dimensional attributes of input data; the author proposes a DBN-MCPR model optimization method based on deep belief classification in the MOOC environment. Experimental results show that when the number of iterations reaches about 80, the RMSE of DBN-MCPR trained with the training dataset without learner feature vector is 77.94%, while the RMSE of DBN-MCPR trained with the dataset with learner feature vector is 77.01; DBN-MCPR with full eigenvectors tends to converge after about 40 iterations, while DBN-MCPR without learner eigenvectors starts to converge after about 15 iterations; this result is in line with the characteristics of the internal network structure of DBN. Conclusion. This application proves that the technical research based on deep belief can effectively meet the needs of the optimization of physical education curriculum resource allocation.

1. Introduction

With the decline of physical fitness of Chinese students year by year and the continuous deepening of the new curriculum reform, the physical fitness of middle school students has become a basic political task in China, which all shows the importance of Chinese physical education and health courses [1]. In the context of China’s basic education reform, the research and application of physical education curriculum resources play a key role in China’s physical education work; the reform of the physical education curriculum is mainly reflected in the physical education and health curriculum of each school, and students are the participants of the school physical education class and the object of teaching, which is used to enhance the students’ physical fitness. At present, the best way to exercise is to be reflected in the school physical education class, which shows that Chinese physical education teachers should focus on enhancing the physical fitness of students in physical education classes; this way of class not only allows students to exercise but also cultivates students’ love for sports; if the will to persist in exercising is to be greatly improved, it should start from the physical education class, and whether students can persist in exercising independently; it also depends on the role that teachers play in the course of the class. At present, the classrooms of primary and secondary school students still need to rely on the equipment and venue resources in the physical education curriculum resources to complete the teaching goals; this can not only achieve the effect of exercise but also allow students to fully participate in the classroom and then exercise the will quality of students [2].

Accordingly, the problem of the allocation of physical education curriculum resources has been paid more and more attention. The optimal allocation of curriculum resources is a challenge to education from an economy and a learning society and an urgent need for the cultivation and growth of innovative talents. In the course of curriculum reform, the allocation of curriculum resources has become a brand-new research field in education and scientific research. With the continuous advancement of a new round of college education curriculum reform, the optimal allocation of physical education curriculum resources is attracting attention and discussion in the field of physical education theory and practice; the effective allocation of physical education curriculum resources is an important basis and premise for the achievement of physical education curriculum goals, and it has become more and more people’s consensus.

On the basis of limited physical education curriculum resources, optimize the allocation of resources that are conducive to the construction of physical education courses, fully explore the best benefits of physical education curriculum resources with the help of the advantages of college sports resources, and provide a strong and favorable guarantee for improving the comprehensive quality of talents; this is not only an important task of deepening education reform but also the core of improving the utilization rate of China’s educational resources, which is of special significance for promoting the reform and development of China’s higher education.

2. Literature Review

With the vigorous development of Internet technology, the era of big data is quietly coming, emerging concepts such as artificial intelligence, machine learning, and deep learning are surging, “Internet + education” came into being, and always the trend of the information age is followed [3]. The in-depth combination of Internet technology and education has made MOOC (Massive Open Online Courses) platforms represented by XuetangX, NetEase Cloud Classroom, Guoke, Coursera, and so on, which are strongly impacting the ecology of traditional education and rapidly reshaping the way learners learn. With the digitization and network sharing of educational resources, MOOC platform course resources are increasing day-by-day, and they are increasing at the petabyte level every day. Faced with countless educational resources of varying quality, on the one hand, learners can always discover the course resources they really need; on the other hand, due to the differences in learners’ own cognitive ability and knowledge structure, it is difficult for them to correctly identify the content of the resource itself; it is impossible to filter out the learning resources that interest you in a very short period of time, so you lose your direction or choose blindly, wasting more precious time. This leads to the dilemma of rich curriculum resources but difficult-to-select resources, which makes learners have the information trek [4]. Information trek is mainly blamed on the problem of “information overload” caused by too many courses. How to help learners quickly and accurately find suitable learning resources among the rapidly growing MOOC resources is an urgent problem to be solved in the field of educational big data. Therefore, it is very important to study the personalized recommendation system based on course resources.

On the basis of the current research, the author proposes a recommendation method for building a DBN-MCPR personalized model based on deep belief classification in a MOOC environment. The DBN-MCPR recommendation process mainly includes the following steps: (1) first, according to the multifeatures defined by the learner’s interest model, the original feature data that can fully represent the learner’s interest preference is collected; (2) second, through a series of data preprocessing operations such as feature mapping, character data digitization, deletion of data that seriously deviates from the normal value, and feature normalization, a standard data set is formed and divided into training set, validation set, and test set; (3) then, the DBN classification model is used to perform feature learning on the training set, and the unsupervised greedy algorithm is used to pretrain each layer of RBM in the DBNs layer-by-layer; finally, the trained DBNs are used as the feature vector input of BP supervised learning; combined with the class label of the course rating, the feature of BP error back propagation is used to fine-tune the entire network; finally, a well-trained DBN personalized recommendation model is formed; (4) the scores of the model is predicted, and course resources with higher predicted scores are recommended to learners in order of their scores from high to low. Figure 1 shows the data preprocessing module.

3. Research Methods

3.1. Basic Model and Overview of Deep Belief Networks

In recent years, with the continuous maturity of deep learning application research, the use of deep learning for resource recommendation has gradually become the mainstream trend [5]. The deep belief network (DBN) is one of the main implementation methods of deep learning. The DBN is a generative model with several layers of latent variables. Latent variables are usually binary, while visible units can be binary or real [6]. Although it is possible to construct DBNs with relatively sparse connections, in general models, each unit in each layer is connected to each unit in each adjacent layer, and there is no connection within a layer. The DBN can be constructed by sequentially stacking several restricted Boltzmann machines (RBMs); the learning process is divided into two stages, that is, the RBM is firstly pretrained layer-by-layer unsupervised and then the entire network is supervised by the back propagation algorithm [7]. The structures of RBM and DBN are shown in Figure 2, respectively.

Given the model parameters , the joint probability distribution of the visible layer and the hidden layer is defined by the energy function as follows:

Among them, is the normalization factor, and the marginal distribution of the model about is as follows:

For a Bernoulli (visible layer) distribution–Bernoulli (hidden layer) distributed RBM, the energy function is defined as follows:

Among them, is the connection weight of RBM, and and represent the bias of visible layer nodes and hidden layer nodes, respectively. Then, the conditional probability distribution can be expressed as follows:where is a sigmoid function.

Since the visible layer and the hidden layer are Bernoulli binary states, the standard for judging their binary probability is often achieved by setting a threshold.

By calculating the gradient of the log-likelihood function , the RBM weight update formula can be obtained as follows:

In the formula, and represent the number of iterations and learning rate of the RBM, respectively, and and represent the expectation of the observed data in the training set and the expectation on the distribution determined by the model, respectively. In particular, RBMs have an interesting property that when trained with maximum likelihood-based learning rules, the update of specific weights connecting two neurons depends only on the statistics collected by these two neurons under different distributions: and . The rest of the network is involved in shaping these statistics, but the weight parameters can be updated with absolutely no knowledge of the rest of the network or how these statistics were produced. This means that the learning rules are “local,” which makes the RBM’s learning seem to be somewhat biological.

As a deep network model, the DBN has the dual properties of a generative model and a discriminative model. Because the pretraining process of the DBN is mainly used to express the high-order correlation of data or describe the joint statistical distribution of data, it has the characteristics of the generative model [8]. The DBN supervised that tuning process is usually used to classify the intrinsic pattern of the data or describe the posterior distribution of the data, which has the characteristics of a discriminative model. At the same time, as a generative model, the generative adversarial network (GAN) has also received a lot of attention in recent years and has been widely used.

The advantage of the DBN learning model is that by combining many RBMs, taking the feature excitation of the previous layer of RBM as the training data of the next layer, the hidden layer can be learned efficiently. Recurrent neural network’s (RNN) depth can even reach the same length as the input data sequence. In the unsupervised learning mode, the RNN is used to predict future data sequences based on previous data samples, and class information is not used in the learning process [9].

The convolutional neural network (CNN) is another deep learning network with discriminative performance, and each module of it is composed of a convolutional layer and a pooling layer. The convolutional layer shares weights, and the pooling layer downsamples the output of the convolutional layer, reducing the amount of data in the next layer. Studies have found that the application of the CNN is mainly concentrated in the field of computer vision or image recognition, and the effect is relatively good [10]. The applications of the DBN are widely distributed in the fields of computer vision and data modeling and prediction.

In addition, the DBN has many hyperparameters, which can be divided into two following categories: one is the training parameters (such as learning rate and momentum term); the other is the parameters that define the network structure (such as the number of network layers and the number of neurons in each layer). The automatic tuning of the former belongs to the category of hyperparameter optimization (HO), while the automatic tuning of the latter is generally called neural architecture search (NAS) [11].

3.2. DBN-Based Learner Interest Model Construction

3.2.1. Demographic Characteristics

At present, in the online learning platform in the field of education, most personalized resource recommendation systems usually only consider part of the learner’s network behavior information when constructing the learner’s interest model, which is unfavorable for fully expressing the learner’s interest. Descriptions about a series of natural attributes and social conditions of people are called demographic characteristics, such as gender, age, and education level. In addition, for MOOC platform users, things like the school grade of the learner can also be attributed to demographic information. The introduction of demographic information can effectively solve the cold start problem of new users; when a learner registers a new learning platform, the system does not have any behavior information of the learner; at this time, it is possible to mine and predict the interests and hobbies of learners according to the demographic information of learners, and rational use of demographic characteristics can help improve the recommendation efficiency [12].

3.2.2. Learner Behavior Characteristics

Explicit feedback information refers to information such as grades and evaluations that learners actively give to learning resources after browsing or using resources. It is more objective and specific to reflect some natural attributes of learners on the platform. For example, in the MOOC platform, after a learner has finished learning a certain course resource, the displayed rating of the course reflects the learner’s direct preference for the course resource. When learners register on the platform, they manually input some basic information to express the user’s interests and hobbies [13].

For example, learners’ search, click, comment, share, collection of courses, and learners’ behaviors are the image depiction of learners’ inner feelings and reflect learners’ interest preferences [14]. The extraction of implicit feedback information is done without the learner’s awareness and without any additional burden on the learner, and the learner’s interest is extracted by analyzing the learner’s behavior log files.

3.2.3. Content Attributes and Characteristics of Course Resources

This phenomenon leads to various types of course resources on the MOOC platform, such as document courseware, videos, and exercises. In addition, there are structured, semistructured, and unstructured distinctions in course resources, and they are heterogeneous. At the same time, the curriculum resources in different subjects and grades are also different. The definition and coding rules of the curriculum resource classification system are shown in Table 1 [15]. According to the content characteristics of course resources, the division strategy can express the course resources themselves well to a certain extent, and these course resource content characteristics can well complement the expression of learners’ interest and preference for course resources.

3.3. Personalized Model Recommendation Method Based on DBN Classification

3.3.1. Composition of DBN-MCPR

DBN-MCPR is the key to the study of personalized recommendation for learners; it is mainly composed of four modules, namely, data collection, data preprocessing, feature learning, and rating prediction. The overall service architecture diagram of DBN-MCPR is shown in Figure 3.

3.3.2. DBN-MCPR Training

The quality of the recommendation performance depends on the degree of training the DBN classification model, and training the model to optimize its performance is the key to realizing personalized recommendation. The goal of training the model is to make the mapped data fit the original input data as much as possible after the deep feature abstraction of the DBN. The model training process is divided into unsupervised pretraining and supervised parameter fine-tuning. The DBN-MCPR training process is shown in Figure 4 [16].

In the model training process, the most important work is to determine the final parameter set = of each layer of RBM; since the pretraining process of the DBN can be regarded as training several relatively independent RBMs, the training set is firstly used as the input of the first RBM; the first RBM is fully trained using the CD-k algorithm to save the weights and biases of the visible and hidden layers of the trained RBM. Then, the hidden layer of the first RBM is used as the input of the second RBM; after the second RBM is fully trained, the weights and biases of its visible and hidden layers are saved; this process is repeated until all the parameters defined in the model are trained, and the RBM layer completes the unsupervised pretraining process. Finally, the hidden layer of the last RBM trained is used as the input of the visible layer of BP; combined with the course rating class label, the weights and biases of the entire network are fine-tuned using the BP reverse error propagation algorithm in a supervised manner. Finally, when the training error meets a certain set value, the DBN output layer (i.e., the output layer of BP) is classified by a logistic regression classifier, and the score prediction is finally completed [17].

In the DBN-MCPR training process, the training samples come from the learning behavior information and course ratings generated by the learners on the selected course resources. The “learner-course resource” feature vector and the learner’s rating of the course together constitute a sample of the learner; for any course resource that the learner has studied, the corresponding learner sample can be obtained; all samples constitute a huge dataset for training and testing [18, 19]. The number of neurons in the visible layer is set as the attribute dimension of the feature vector of “Learner-Course Resource.” In order to improve the recommendation effect of DBN-MCPR, each parameter needs to be effectively set before the model runs; the settings generally include the initialization of training parameters and the parameter settings of influence factors.

4. Results Analysis

(i) Influence of training data set on the recommendation accuracy of DBN-MCPR: considering the complex structural characteristics of the DBN classification model itself, due to its strong feature expression ability, a large amount of data is required to avoid overfitting; if the model is overfitted, its generalization ability will drop sharply, which will be very detrimental to its training. The authors verify the variation of DBN-MCPR recommendation accuracy with the training dataset by adjusting the size of the training dataset, as shown in Figure 5 [20]. As can be seen from Figure 4, the number of training sets directly affects the recommendation accuracy of DBN-MCPR. Since the smaller the training data set is, the weaker the correlation between the data is, which leads to the easier it is to generate “fragments” of the relationship between the samples [21]. In the case of weak correlation between sample information, DBN-MCPR cannot mine the complex relationship between data well, resulting in low recommendation accuracy. As the number of samples in the training data set increases, the correlation between samples becomes more and more abundant, which is more conducive to DBN-MCPR to mine the hidden relationship between samples, so the recommendation accuracy will be higher. When using the entire training set to train DBN-MCPR, its RMSE is as low as 76.68% [22].(2) The influence of learner feature vector on the recommendation accuracy of DBN-MCPR: learner demographics and course resource content attributes together influence learners’ interest preferences; a comparative experiment is set up here to verify the influence of the learner’s feature vector on DBN-MCPR and to further verify the effectiveness and practicability of the learner’s interest model [23]. Eliminate the relevant demographic feature dimensions and course resource content attribute feature dimensions in each dataset; these attribute dimensions include age, gender, grade, level_of_education, school, course_subject, course_grade, course_knowledge, course_creator, and course_school. Further experiments show that the influence of learner feature vector on the recommendation accuracy of DBN-MCPR is shown in Figure 6 [24].

5. Conclusion

The author proposes a DBN-MCPR model optimization method based on deep belief classification under the MOOC environment, which is mainly based on the efficient feature abstraction and feature extraction capabilities of deep belief technology, fully excavates learners’ interests and preferences for course resources, and builds a DBN-MCPR model to test the optimization effect of course resource allocation. From the test results, it can be seen that when the number of iterations reaches about 80, the RMSE of DBN-MCPR trained with the training dataset without learner feature vector is 77.94%, while the RMSE of DBN-MCPR trained with the dataset with learner feature vector is 77.01, the RMSE of DBN-MCPR with full eigenvectors is 1.21% lower than that of DBN-MCPR without learner eigenvectors; it is fully proved that the learner feature vector has a great influence on the prediction accuracy of DBN-MCPR. DBN-MCPR with full eigenvectors tends to converge after about 40 iterations, while DBN-MCPR without learner eigenvectors converges after about 15 iterations; this result is in line with the characteristics of the internal network structure of the DBN, which effectively verifies that the model in this study has a high optimization effect and further proves that the technical model based on deep belief can effectively meet the needs of the optimization of physical education curriculum resource allocation.

Data Availability

The data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest

The author declares that he has no conflicts of interest.

Acknowledgments

The study was supported by the University-Level Educational Reform Project “Research on the Teaching Mode of 1123N New Integrated Curriculum of Health Education in Yunnan Nationalities University,” project no. 2021JG-077.

References

Y. Luo, “Demand-led course construction and teaching resource development and application of practical Chinese in secondary vocational colleges,” Education Teaching Forum, 2020.
View at: Google Scholar
S. Zheng, “Research on Distance teaching System of English Course Based on Wireless Network technology,” Hindawi Limited, vol. 2021, Article ID 3275340, 2021.
View at: Publisher Site | Google Scholar
J. C. Carey, L. S. Beitelspacher, J. Tosti-Kharas, and E. Swanson, “A resource-efficient modular course design for co-teaching integrated sustainability in higher education: developing the next generation of entrepreneurial leaders,” Entrepreneurship Education and Pedagogy, vol. 4, no. 2, pp. 169–193, 2020.
View at: Publisher Site | Google Scholar
L. N. Dong and J. W. Liu, Design of security sharing system for online teaching course resources of fine arts specialty International Conference on E-Learning, E-Education, and Online Training, Springer, Berlin Germany, 2021.
J. K. Green, “Designing hybrid spaces for learning in higher education health contexts,” Postdigital Science and Education, vol. 4, no. 1, pp. 93–115, 2022.
View at: Publisher Site | Google Scholar
J. Williams, M. J. Davies, B. Suesee, and D. Hunt, “Teachers’ experiences of teaching the australian health and physical education health benefits of physical activity curriculum and the need for greater reality congruence,” Curriculum Perspectives, vol. 42, no. 1, pp. 27–37, 2022.
View at: Publisher Site | Google Scholar
X. Zhu, L. Bao, H. Wang, and N. Zhao, “The train of thought and strategy of Chinese school physical education to cultivate unique minority students’ motor quotient in yunnan,” Advances in Physical Education, vol. 11, no. 04, pp. 481–488, 2021.
View at: Publisher Site | Google Scholar
C. Tan and L. Xue, “Application research of big data mining in personalized teaching of internet education platform,” Journal of Physics: Conference Series, vol. 1992, no. 2, Article ID 022120, 2021.
View at: Publisher Site | Google Scholar
S. Wang, T. Sun, X. Qu, and W. Xu, “Online education of atomic physics based on mooc platform,” Journal of Physics: Conference Series, vol. 1881, no. 3, Article ID 032009, 2021.
View at: Publisher Site | Google Scholar
L. C. Redmond and V. J. Howard, “Open educational resource-enabled pedagogy to improve student outcomes in nutrition course,” Journal of Nutrition Education and Behavior, vol. 53, no. 7, pp. 631–634, 2021.
View at: Publisher Site | Google Scholar
R. D. Sushir and D. G. Wakde, “An improved detection of blind image forgery using hybrid deep belief network and adaptive fuzzy clustering,” Multimedia Tools and Applications, vol. 81, no. 20, pp. 29177–29205, 2022.
View at: Publisher Site | Google Scholar
A. Decelle and C. Furtlehner, “Restricted Boltzmann machine: recent advances and mean-field theory,” Chinese Physics B, vol. 30, no. 4, Article ID 040202, 2021.
View at: Publisher Site | Google Scholar
B. Zhou, H. Duan, Q. Wu et al., “Short-term prediction of wind power and its ramp events based on semi-supervised generative adversarial network,” International Journal of Electrical Power & Energy Systems, vol. 125, no. 2, Article ID 106411, 2021.
View at: Publisher Site | Google Scholar
K. E. Arunkumar, D. V. Kalaga, C. M. S. Kumar, M. Kawaji, and T. M. Brenza, “Forecasting of covid-19 using deep layer recurrent neural networks (rnns) with gated recurrent units (grus) and long short-term memory (lstm) cells,” Chaos, Solitons & Fractals, vol. 146, no. 1, Article ID 110861, 2021.
View at: Publisher Site | Google Scholar
Y. N. Fu’Adah, I. Wijayanto, N. K. C. Pratiwi, F. F. Taliningsih, S. Rizal, and M. A. Pramudito, “Automated classification of alzheimer’s disease based on mri image processing using convolutional neural network (cnn) with alexnet architecture,” Journal of Physics: Conference Series, vol. 1844, no. 1, Article ID 012020, 2021.
View at: Publisher Site | Google Scholar
B. M. Oloulade, J. Gao, J. Chen, T. Lyu, and R. Al-Sabri, “Graph neural architecture search: a survey,” Tsinghua Science and Technology, vol. 27, no. 4, pp. 692–708, 2022.
View at: Publisher Site | Google Scholar
C. L. Dobbs, N. K. Caselli, E. Hartzell, C. Flanagan, and Y. Yan, “Understanding middle graders’ language borrowing: how lexical and demographic characteristics predict similarity,” Reading and Writing, vol. 35, no. 4, pp. 971–994, 2021.
View at: Publisher Site | Google Scholar
G. Lin, F. Liang, W. Pan, and Z. Ming, “Fedrec: federated recommendation with explicit feedback,” Intelligent Systems, IEEE, no. 99, p. 1, 2020.
View at: Google Scholar
B. Wang, J. Han, and Y. Cuan, “Improved metric factorization recommendation algorithm based on social networks and implicit feedback,” Journal of Physics: Conference Series, vol. 1634, no. 1, Article ID 012037, 2020.
View at: Publisher Site | Google Scholar
I. S. Harahap, I. Z. Matondang, E. K. Indah, E. K. Indah, and I. Fitri, “Mapping climate classification of oldeman in agricultural resources management in south tapanuli district,” IOP Conference Series: Materials Science and Engineering, vol. 1156, no. 1, Article ID 012002, 2021.
View at: Publisher Site | Google Scholar
X. Ruhang, “Efficient clustering for aggregate loads: an unsupervised pretraining based method,” Energy, vol. 210, no. 9, Article ID 118617, 2020.
View at: Publisher Site | Google Scholar
Y. Guo, C. Zhang, L. Gao, and X. Kang, “Research on Online Teaching Practice Based on Information Technology: A Case Study of the Course "Human Resource Management"online Teaching during the COIVD-19 Pandemic,” in Proceedings of the 2021 2nd International Conference on Education, Knowledge and Information Management (ICEKIM), Xiamen, China, January 2021.
View at: Google Scholar
V. Shah, S. Murthy, J. Warriem, S. Sahasrabudhe, G. Banerjee, and S. Iyer, “Learner-centric mooc model: a pedagogical design model towards active learner participation and higher completion rates,” Educational Technology Research & Development, vol. 70, no. 1, pp. 263–288, 2022.
View at: Publisher Site | Google Scholar
W. Zhong, K. Xie, Y. Liu, C. Yang, and S. Xie, “Multi-resource allocation of shared energy storage: a distributed combinatorial auction approach,” IEEE Transactions on Smart Grid, vol. 11, no. 5, pp. 4105–4115, 2020.
View at: Publisher Site | Google Scholar

Copyright

Copyright © 2023 Lili Sun. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

PDF Download Citation

Download other formats

Order printed copies

Views

60

Downloads

26

Citations