Research on Interaction Design of Big Data College Short Video Parenting Platform Construction Based on Deep Learning

The continuous development of all-media, ubiquitous and unused information has led to profound changes in the ecology of public opinion, media pattern, and communication methods [1-2]. Short video has attracted a large number of users with its concise and vivid form, and has had a great impact on marketing, local promotion, popularization of science and other aspects [3-6]. As the most sought-after emerging media by youth groups in the mobile Internet era, university educators should change their working attitude in time, seize the main position of short video ideological and political education, integrate the cultural penetration of short video with the education work in depth, and build up momentum and empower college students for their comprehensive development [7]. Short video parenting work involves the integration of different disciplines such as communication, sociology, pedagogy, management, etc. College educators lack of relevant disciplines and professional knowledge base, in the construction and operation of short video platforms tend to lose sight of the other side of the coin, it is difficult to continue [8-9].

At present, short video platforms include Jitterbug, B station, headline, quick hand, small red book and so on. However, due to lax regulation and lack of standardization, there are a large number of vulgar, violent, negative and radical views in short videos, which make viewers suffer from adverse effects [10-11]. These contents not only seriously hurt students’ physical and mental health, but also easily trigger imitation behavior and violence. In addition, the negative phenomena and radical views appearing in short videos can also affect the formation of students’ values and worldviews [12-13]. In the era of big data, how to use short videos to infuse parenting content into all aspects of college students’ learning and life, and help them grow up and become successful, is an important issue that educators in colleges and universities need to explore urgently. Therefore, it is necessary to build a safe and reliable short-video parenting platform that is far away from violence, vulgarity, negativity and radicalization. And deep learning is a new field in machine learning research, which is motivated by building, simulating the human brain for analyzing and learning neural networks, which mimics the mechanism of the human brain to interpret data, such as images, sounds and texts [14]. It can be well combined with big data to build a suitable educational short video platform.

In order to solve the problem of interaction of contemporary education platform, this paper applies Web3D virtual technology to build 3D area, 3D college and 3D studio, which form the Web3D short video learning and education platform. The real learning process is simulated, and the human-computer interaction design is realized through the functions of commenting, private messaging, and liking. The deep learning-based recommendation strategy proposed in this paper is applied to the video recommendation scenario on the short video learning and education platform to screen out the short videos of interest for the users among the massive educational resources. It enhances the learners’ online learning effect and improves their interactive experience in the online learning process. In order to verify whether the various functions of the system can achieve the expected goals of the learning and education platform and the related algorithms to test the platform function and performance.

2

Interaction design for Web3D short video learning and education platforms

2.1

Platform construction

With the booming development of new media, short videos are welcomed and favored by the majority of young students with the advantages of low threshold, wide audience, strong interaction and fast dissemination, and have become the main carrier of information exchange and interaction for higher education students.

Web3D short video learning and education platform builds three basic frameworks: 3D area, 3D academy, and 3D studio. It loads and imports the designed models, and constantly adjusts them to ensure their rationality. Among them, each existing college is distributed on the 3D area, which realistically restores the distribution of the area module, 3Dizes the area module, and realizes the patterned coordinates. 3D college restores the simulated park scene, sets up colorful scenes such as panoramic sky, water, grass, buildings, park boards, etc., and provides map resources of different sizes, so that colleges can choose different themes to form a 3D world with more complete features. 3D studio simulates the real The 3D studio simulates the real classroom learning scene, provides multi-functional tools such as practice table, bookshelf, tutor, display screen, etc. It restores the campus learning scene and fits the real learning atmosphere, and the studio provides special features other than the same functions (e.g., the medical studio sets up the function of simulating animal experiment table, the chemistry studio sets up the function of chemistry experiment table, and the computer class studio places the code practice table). Learners can conduct immersive learning in the 3D studio. The basic components of the platform are depicted in Figure 1.

2.2

Faculty interaction information presentation

Use axiso (request tool) to access the studio information interface to get studio information. Entrance signposts in the map show the list of open studios, and the smooth animation operation of clicking on the signposts to zoom in is realized by tween.js. Use key-value pairs to store building and studio information binding, information drawing presentation. The information is drawn and rendered in two ways: canvas and div. The div module is the presentation of the entrance signpost, and the canvas is the presentation of the text above the building. The canvas will start at the center of the building and shift upward and forward to a certain position to create an eye-catching display. After binding the studio and the building, add the corresponding click event, i.e., when clicking on the building, search for the value of the studio number with the building number as the key value, and then make a route jump to the studio. Route jumping is realized by vue-router.

Its human-computer interaction flow is shown in Figure 2. Taking the teaching of Java course content as an example, the course is divided into five staged tasks: code foundation, JavaWeb, JavaEE, introduction to microservices, development, and maintenance. Each phased task is divided into a number of sub-tasks, and each sub-task list contains a number of task videos. Each task video corresponds to the corresponding task practice questions. The inter-component interaction process simulates the real learning process, starting with the start of receiving tasks, which triggers both the NPC dialog and task receiving components. The task receiving component has a detailed task description. The task description contains five elements: scenario, expectation, tool, verifiable, and skill. The task list is used as an aid to the task description, and it outlines in detail the necessary content items for completing the task. The task list includes the task video and the knowledge point exercise video. Students can submit their answers by completing the video and exercises as outlined in the task list, but if they do not complete them, the NPC tutor will not proceed to the next task in the student’s progress.

2.3

Establishment of course short video resource library

The content of the short video is based on the presentation of the main points of the course teaching content, and guides students to independently expand their access to relevant knowledge and thinking. For each knowledge point, one or more short videos are produced according to the situation, and the short videos should be compact and direct, open to the point, go straight to the topic, and explain the knowledge points clearly in the shortest time.

Before each classroom teaching, through the video platform and the establishment of the video resource library, in accordance with the curriculum teaching plan, release the corresponding course learning task files, as well as typical short video construction examples, so that students can use the fragmented time before the class, many times and in detail on the subsequent course teaching content for pre-study, guiding the students to learn and think about the content of the textbook knowledge points. For example, when teaching deep foundation pit construction, the first release of deep foundation pit descending methods, support construction methods and other learning content, and uploaded in the video platform Hengyang a deep foundation pit construction site video and related information for students to learn and watch.

2.4

Ongoing communication discussions on video-interactive platforms

Through the comments, private messages, likes and other functions of the video platform, teachers and students from inside and outside the school and the construction personnel of various enterprises have extensive discussions on the teaching content, and students have a newer and deeper understanding of the relevant course content through the study and discussion records. For example, in the study of deep foundation pit support, discuss the detailed construction process and application of the support forms such as row pile, diaphragm wall, internal support, etc., and for the case of Hengyang, a deep foundation pit in the actual construction of the problems. The enterprise personnel comment on the discussion between teachers and students through the platform, and add to the shortcomings. They also record a video on the construction site for explanation. This method can more widely mobilize the construction personnel of the enterprise to participate in the course teaching process, so as to avoid a one-sided understanding of a certain construction process method.

3

Deep learning-based hybrid short video recommendation model

This paper takes educational short video resources as the research object to build a short video education platform. The deep learning-based recommendation platform is used to screen out suitable short videos for users among the huge amount of short video resources, so as to provide an interactive platform for learners of various specialties to learn and enjoy.

3.1

Scoring Data Extraction Module

In this section, the rating matrix data is feature extracted using the Hidden Semantic Model proposed by Funk, which is a recommendation algorithm that utilizes existing rating data and trains the model to derive the predicted ratings based on a machine learning approach, which is simple to implement and has superior performance. In a short video recommendation platform, LFM first divides all short videos into several categories according to certain criteria, then calculates the user’s interest in each category and the weight of each short video in each category, the higher the weight means that this short video and the higher the degree of conformity of the category, and finally the user’s interest in the category and the weight of the item in the category do the multiplication to get the user’s favorite degree of the video. Finally, we multiply the user’s interest in the category and the weight of the item in the category to get the user’s favorite degree of the short video, and in most cases we use the user’s rating of the short video to indicate the favorite degree.

The preferences of user u₂ for hidden factors f₁,f₂,f₃ and Lf_k are p₂₁,p₂₂,p₂₃ and Lp_2k, respectively, and the weights of item i₁ in hidden factors f₁,f₂,f₃ and Lf_k are q₁₁,q₁₂,q₁₃ and Lq_1k, respectively, from which the predicted score of user u₂ for item i₁ is obtained as ${\hat{r}}_{21} = p_{21} q_{11} + p_{22} q_{12} + p_{23} q_{13} + L + p_{2 L} q_{1 k}$ . The predicted scores of user u₂ for all items are shown in equation (1): (1) $\begin{matrix} {\hat{r}}_{21} = p_{21} q_{11} + p_{22} q_{12} + p_{23} q_{13} + L + p_{2 k} q_{1 k} \\ {\hat{r}}_{22} = p_{21} q_{21} + p_{22} q_{22} + p_{23} q_{23} + L + p_{2 k} q_{2 k} \\ {\hat{r}}_{23} = p_{21} q_{31} + p_{22} q_{32} + p_{23} q_{33} + L + p_{2 k} q_{3 k} \\ M \\ {\hat{r}}_{2 n} = p_{21} q_{n 1} + p_{22} q_{n 2} + p_{23} q_{n 3} + L + p_{2 k} q_{n k} \end{matrix}$

Eventually you can get a rating for each user u for each item i: (2) ${\hat{r}}_{u i} = \sum_{f = 1}^{l} p_{u f} q_{i f}$

where p_uf represents the degree of preference of user u for hidden factor f, q_if represents the weight of item i in hidden factor f, and l represents the number of hidden factors. The smaller l is, the higher the degree of generalization of the model is, and the larger l is, the more complex the model is, and at the same time, the lower the degree of generalization of the model is.

The information of the original user-item rating matrix can be maximally retained by minimizing the objective function, which is shown in equation (3): (3) $L = \sum_{(u, i) \in T} {(r_{u i} - \sum_{f = 1}^{l} p_{u f} q_{i f})}^{2} + λ ({‖ p_{u} ‖}_{2}^{2} + {‖ q_{i} ‖}_{2}^{2})$

where λ represents the regularization coefficient used to reduce overfitting, and T represents the set of all samples that are iteratively optimized for variable values by stochastic gradient descent algorithm.

Derivation is performed for p_uf and q_if as shown in equation (4): (4) $\begin{array}{l} \frac{\partial L}{\partial p_{u f}} = - 2 Ω (r_{u i} - \sum_{f = 1}^{l} p_{u f} q_{i f}) q_{i f} + 2 λ p_{u f} \\ \frac{\partial L}{\partial q_{i f}} = - 2 Ω (r_{u i} - \sum_{f = 1}^{l} p_{u f} q_{i f}) p_{u f} + 2 λ q_{i f} \end{array}$

The LFM uses the inner product to regress p_u and q_i, and finally predicts the user’s rating of the item. Since the rating prediction module of the Deep-AM model proposed in this paper is handled by the fusion and prediction module, only the user hidden vector Output_{u_Rating} and item hidden vector Output_{i_Ratring} generated by matrix decomposition are simply used in this subsection, i.e., the rating matrix is decomposed into two hidden vector matrices to represent it as shown in Equation (5): (5) $R_{m \times n} = U I^{T}$

Where matrix R_m×n represents the ratings of m users on n items, this rating can be represented using the product of user hidden vector matrix U and item hidden vector matrix I. Where each row of matrix U represents a user and is represented using k dimensional vector Output_{u_Raing_i} as shown in equation (6): (6) $I_{n \times k} = {\begin{matrix} O u t p u t_{i_R a t i n g_1} \\ M \\ O u t p u t_{i_R a t i n g_n} \end{matrix}}$

Each row in matrix I represents an item, represented using a k-dimensional vector Output_{i_Rating_i}, as shown in equation (7): (7) $I_{n \times k} = {\begin{matrix} O u t p u t_{i_R a t i n g_1} \\ M \\ O u t p u t_{i_R a t i n g_n} \end{matrix}}$

3.2

Comment data extraction module

In short video websites, some users are used to giving high ratings for each short video, while others rate harshly, and even the well-performing short videos cannot be favored by such users, which shows that the rating data is highly influenced by users’ subjectivity. Review data can be used to alleviate the sparsity of the rating matrix and help improve the accuracy of recommendations. Extracting features from the comment text is similar to extracting features from the rating matrix by learning the feature vectors of users and items from the original data (rating matrix, comment text), the difference is that extracting features from the comment text requires a feature extractor for the comment data, and the latest deep learning techniques are usually used to feature extract the comment text to learn the personal preferences and emotional tendencies, attribute features of items and other information.

3.3

Integration and Forecasting Module

The two types of data used in this paper, the rating data and the review data, belong to different source data and exist in different feature spaces, so they cannot be simply joined head to tail to form a vector. In this paper, inspired by the factorization machine, Output_{u_reviow}, Output_{i_review}, Output_{u_Rating}, and Output_{i_Rating} extracted by the convolutional neural network and LFM above are used for low-order feature combination and high-order feature extraction. In order to facilitate the better cross-fertilization of feature vectors from different sources, all four feature vectors are k-dimensional vectors.

1)

Low-order features

It has been shown that the most direct and effective way to fuse comment vectors and rating hidden vectors is to directly add the corresponding dimensions of the two vectors. Therefore, this chapter adopts this approach to construct low-order features directly, and the low-order features of users can be expressed as follows: (8) $U_l o w = O u t p u t_{u_R a t i n g} e O u t p u t_{u_r e v i o w}$

where e denotes the corresponding dimensional summation operation, and the same can be obtained for the low-order feature I _ low of the item.

2)

High-order features

Higher-order features can be obtained by DNN, which consists of fully connected multilayer neural networks with strong modeling ability and can be used to extract nonlinear higher-order features. User vectors and item vectors are taken as inputs and fed into the DNN to learn the higher-order features respectively, and the computational process is shown in equation (9): (9) $\begin{matrix} O u t p u t_{u} = O u t p u t_{u_R a t i n g} \oplus O u t p u t_{u_r e v i s w} \\ U_{h i g h}^{1} = σ^{1} (w^{1} O u t p u t_{u} + b^{1}) \\ U_{h i g h}^{2} = σ^{2} (w^{2} O u t p u t_{u} + b^{2}) \\ M \\ U_{h i g h}^{i - 1} = σ^{i - 1} (w^{i - 1} U_{f e a t u r e}^{i - 1} + b^{i - 1}) \\ U_{h i g h}^{i} = σ^{i} (w^{i} U_{f e a t u r e}^{i - 1} + b^{i}) \end{matrix}$

Where $U_{f a c t u r e}^{i}$ is the layer i output vector, bⁱ is the layer i bias, w^j is the layer i weight parameter and σⁱ is the layer i activation function. Similarly, the high-order feature of the item can be obtained as I_high.

The final user depth feature can be obtained by splicing user low-order features and user high-order features: (10) $U_{f e a t u r e} = U_{l o w} \oplus U_{h i g h}$

Similarly, the item depth features can be obtained I_feature.

In order to improve the recommendation accuracy, most of the current recommendation algorithms based on review text or based on rating matrix focus on how to better learn the user features and item features.

The rating prediction part of this paper uses the method that the final user depth feature vector U_feature and item depth feature vector I_feature obtained are directly spliced, and then fed into the full connectivity layer, so that the user depth features and item depth features are linearly combined with different weights, and finally the user’s rating of the item is regressed, and thus the final predicted rating is shown in equation (11): (11) ${\hat{r}}_{u i} = W \times m u l t i p l y (U_{f e a t i r e}, I_{f e a t i r e})$

where 1 represents the final predictive score of the model, W is the weight, and multiply(a,b) represents an operation that can multiply the corresponding elements of two equal dimensional vectors. When W is ${\hat{r}}_{u i}$ , it is a standard LFM.

Since this section is based on the task of score prediction, which is essentially a regression problem, the objective function can be obtained as shown in equation (12): (12) $L = \sum_{u, i \in T} {({\hat{r}}_{u i} - r_{u i})}^{2}$

where r_ui represents the true rating.

4

Performance evaluation of the short-video teaching platform

After the functional test to verify the normal operation of the various functions of the platform, it is also necessary to verify the performance of the platform. Therefore, the platform also needs to carry out the relevant performance test, through the performance test to verify whether the platform can still maintain a stable and good running state in complex situations. Good performance can make users have a good experience and improve their satisfaction. In this subsection, the performance test of Web3D short video learning and education platform will be conducted, and the specific test cases are shown in Table 1. According to the implementation and testing situation all aspects of the system have achieved the expected results.

Table 1.

Systematic test

Use case number	Use case description	Concrete operation	Response time	Conclusion
1	The overall response time of the user	Access all pages of the client	<2s	Pass
2	The overall time of the teacher’s end	Access the teachers’ all pages	<3s	Pass
3	The administrator side is the same time	Access the administrator all the pages	<2s	Pass
4	Users play the course waiting time	Click the play course	<3s	Pass
5	User search courses wait for time	User input keyword search course	<1s	Pass
6	The teacher uploaded the course waiting time	Teachers select course resources to click upload	<3s	Pass
7	User concurrent testing	Write the test script to simulate 1,000 users for a short time	<1s	Pass
8	Teacher concurrent test	Write test scripts to simulate 100 teachers	<2s	Pass
9	Administrator concurrent testing	Write test scripts to simulate 100 administrators for a short time	<2s	Pass

The functional tests of the recommendation page are explained in detail as shown in Table 2. The platform has the ability to suggest popular courses to the user on the home page, as well as course videos that they may find appealing. After testing, the platform’s popular course recommendation and personalized recommendation functions are perfect and error-free.

Table 2.

Recommended page function test case table

Use case number	Use case name	Use case description	Expected result	Actual result	Conclusion
1	Popular courses recommend	Click the search course and recommend a popular course recommendation below the search results	Show popular courses	Congruity	Pass
2	Personalized course recommendation	The user enters the home page automatically to display	Show recommended courses	Congruity	Pass

For the traditional recommendation algorithm ItemCF, UserCF in the precision rate, recall rate, F1 value of the three aspects of the comparison test respectively, here in this paper set Top-N is 4. The comparison results are shown in Table 3, which shows that the hybrid short video recommendation model based on deep learning proposed in this paper is significantly better than the two traditional algorithms, UserCF and ItemCF, in all the indexes, and its precision rate, recall rate, and F1 value are 60.845, 56.943, and 59.349, respectively. After analysis, it is due to the fact that traditional algorithms cannot accurately model the user’s interest preferences, and therefore the recommendation effect is not as good as the recommendation algorithm proposed in this paper. Moreover, considering the dimensions of algorithm performance, UserCF and ItemCF have high dimensional and extremely sparse user-item matrices, and therefore require a large amount of resources to maintain them, which results in a greatly limited recommendation performance and recommendation effect.

Table 3.

Traditional algorithms compare experimental results

Method	Precision	Recall	F1_score
UserCF	43.198	43.764	42.917
ItemCF	44.813	45.419	45.076
Our method	60.845	56.943	59.349

Further, this paper also conducts comparison experiments with deep neural network DIN model, DeepFM model, AFM model, and Video2Vec in the dimensions of Precision, Recall, and F1 value under different Top-N, as shown in Fig. 3, respectively.

The horizontal coordinate in the figure represents the size of the recommendation list, and the vertical coordinate corresponds to the corresponding Precision, Recall, and F1 values when the length of the recommendation list is 1, 2, 3, 4, and 5, respectively. It can be seen that the Precision value of the five algorithms gradually decreases with the increase of Top-N, and both Recall and F1 values increase with the increase of Top-N. On the whole, the model proposed in this paper outperforms the other four models in terms of performance.

Among them, the Video2Vec model performs poorly overall because it fails to dig deeper into the user’s interests and cannot grasp the user’s personalized needs. In contrast, since this paper is a hybrid short video recommendation algorithm based on deep learning, it can effectively solve the limitation of structural sequential processing of videos and can parallelize the processing of data, so it is better than DeepFM in all assessments. Although the attention mechanism is also introduced in DIN and AFM, time consideration is missing. Similarly, the platform in this paper demonstrates better recommendation performance in course learning.

To summarize, the method in this paper achieves better performance relative to other methods because this paper takes into account the temporal factor of user behavior and the interests of similar users in the model, while other methods only recommend from the similarity of the users or the course itself. This approach reinforces the user’s recent interest, which is often more important in online learning platforms. This is because the courses that the user is recently interested in are the ones that the user is going to take recently, whereas a long ago interest represents that the user has either learned it or the user has passed that learning stage and does not need to take that course anymore. Thus, the model proposed in this paper achieves better results from all perspectives.

5

Analysis of the results of the experience test

5.1

User test design

In order to compare the sense of experience of using this paper’s platform with traditional teaching methods, a questionnaire was designed in this paper, and the detailed questions are shown in Table 4.The questionnaire consists of eight questions, Q1-Q4 is the feeling of the user’s learning with the platform, and Q5-Q8 is the user’s feeling of the platform’s design. The answers were set according to a seven-point Likert scale, with seven options for each question: strongly agree, somewhat agree, fair, somewhat disagree, disagree, and strongly disagree. In order to test the experience of students of different age groups on the use of the platform, the testers chose 10 university students (5 males, 5 females) versus 6 compulsory school students (3 males, 3 females).

Table 4.

User experience questionnaire

Serial number	Index	Problem description
1	Concentration	I focus on what I study
2	Learning atmosphere	I feel good learning atmosphere
3	Fatigue	I feel tired after learning
4	Learning motivation	I feel the dynamics of the course
5	Interface quality	The interface in the system makes me comfortable
6	Comprehensibility	I can understand the tips and guidelines in the learning process
7	Interactivity	I like how we interact in the learning process
8	Personal preference	I like this learning style

5.2

Analysis of test results

The results of the user testing of this paper’s platform and traditional teaching methods are shown in Figure 4. For the seven options in the Likert scale of strongly agree, agree, relatively agree, average, relatively disagree, disagree, and strongly disagree, respectively, assigned a value of 7, 6, 5, 4, 3, 2, and 1 points, and the average of the scores of all the tested persons was taken as the score of the platform on this index. In the two indicators of concentration and learning atmosphere, the platform scores higher than the traditional teaching methods, indicating that the platform can create a learning atmosphere for students and improve their concentration. In the indicator of fatigue, the platform scores lower than traditional teaching methods, indicating that in the way of answering questions between two sessions, compared with using the mouse to answer practice questions, the platform in this paper’s motion answering can to a certain extent alleviate students’ fatigue. In terms of learning motivation index, the score of this platform is slightly higher than that of traditional teaching methods, which indicates that the visual and audio real-time feedback of this platform can motivate students to learn and improve their learning motivation when the test subjects answer the questions correctly. In terms of interface quality indicators, the score of this platform (4.7) is lower than that of traditional teaching methods (5), indicating that there is still room for improvement in the interface design of this platform. In terms of comprehensibility, this platform scores lower than traditional teaching methods, which indicates that this platform needs more hints and guidelines for users to operate and use the platform. In terms of interactivity, this platform scores higher than traditional teaching methods, which means that the interaction method adopted by this platform is more attractive to users than traditional interaction methods. In the indicator of personal preference, the platform scores slightly higher than traditional teaching methods, indicating that the platform has a better learning experience for users.

In order to further compare the differences in experience metrics between this paper’s platform and traditional teaching methods, a paired-sample t-test was conducted on the collected questionnaire data. T-test is mainly used for sample content is small overall standard deviation of the unknown normal distribution, for two samples of the T-test obtained P-value represents the level of significance of the difference, P-value the smaller the more reason to think that the two overall mean difference is significant, when P < 0.05 indicates that the difference is significant. The results of the paired-sample t-test of the two platforms are shown in Table 5, in which the difference between the two platforms in the indicators of learning atmosphere and interactive experience is significant (P<0.05), indicating that the platform of this paper can provide students with a better learning atmosphere and a sense of interactive experience compared with traditional teaching methods.

Table 5.

Test results of the experience index

Experience indicator	Teaching method	M	SD	T	P
Concentration	Teaching platform	5.51	1.36	1.375	0.120
Concentration	Traditional method	5	0.98	1.375	0.120
Learning atmosphere	Teaching platform	5.42	1.22	2.154	0.047
Learning atmosphere	Traditional method	4.6	0.97	2.154	0.047
Fatigue	Teaching platform	3.84	1.53	-1.132	0.283
Fatigue	Traditional method	4.47	1.27	-1.132	0.283
Learning motivation	Teaching platform	5.22	1.42	1.472	0.163
Learning motivation	Traditional method	4.48	1.20	1.472	0.163
Interface quality	Teaching platform	4.75	1.38	-0.885	0.401
Interface quality	Traditional method	5.18	1.27	-0.885	0.401
Comprehensibility	Teaching platform	4.84	1.64	-1.214	0.257
Comprehensibility	Traditional method	5.56	1.35	-1.214	0.257
Interactivity	Teaching platform	5.26	1.35	0.842	0.022
Interactivity	Traditional method	4.87	1.13	0.842	0.022
Personal preference	Teaching platform	5.36	1.43	1.123	0.296
Personal preference	Traditional method	4.95	0.93	1.123	0.296

6

Conclusion

In this paper, a short-video learning and education platform is established by using Web3D technology, and 3D areas, 3D colleges, and 3D studios are built through virtualization technology, so that online learners can have an immersive experience, which improves the fun of learning and the interactive experience. The deep learning hybrid short video recommendation model is applied to the online learning platform to personalize course recommendations for users, which helps to create a learning atmosphere and improve students’ learning concentration and learning effect. In this paper, the educational platform maintains a stable and good operating state even in complex situations, recommending popular educational short videos with good performance. Compared with traditional teaching methods, the mean value of the interactivity index of this paper’s platform is higher, indicating that the interaction method adopted by this paper’s platform is more attractive to users than the traditional interaction method.

Funding:

This research was supported by the Special task project of Humanities and Social Science Research of the Ministry of Education: "Research on the Construction of Short Video Education Platform in Universities under the perspective of Big Data" (22JDSZ3174).

Langue:: Anglais

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

RSS Feed de la revue

Research on Interaction Design of Big Data College Short Video Parenting Platform Construction Based on Deep Learning

Yan Yang

Publié en ligne: 21 mars 2025

Reçu: 17 oct. 2024

Accepté: 14 févr. 2025

DOI: https://doi.org/10.2478/amns-2025-0627

Mots clés<kwd>Web3D virtual technology</kwd>, <kwd>Recommendation model</kwd>, <kwd>Deep learning</kwd>, <kwd>Education platform</kwd>

© 2025 Yan Yang, published by Sciendo

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

Mots clés
<kwd>Web3D virtual technology</kwd>, <kwd>Recommendation model</kwd>, <kwd>Deep learning</kwd>, <kwd>Education platform</kwd>