Construction and Research of Personalized University English Learning Platform Based on Recommendation Algorithm

English is the most common language in the world and is most widely used in globalized communication, and the importance of students’ English proficiency to their future development is very prominent. It is the focus of foreign language learning to cultivate students with good foreign language communication skills, broad knowledge, cross-cultural communication, conscious awareness of knowledge innovation, scientific and rational critical thinking, and good humanistic cultivation [1–4].

College students’ English learning is a long-term and continuous process. College students need to master a certain level of English skills in order to adapt to future social development and vocational needs. According to the large variability in the distribution of students’ intelligence, their different personalities, different interests, different abilities, and different directions of professional application, personalized English learning is a means that must be used in the process of improving students’ English skills [5–6]. However, most university English courses adopt uniform syllabus and teaching materials, and students receive the same teaching content at the same time and place, which cannot satisfy students’ individualized learning needs, resulting in a lack of motivation and interest in learning; the traditional teaching mode requires teachers to invest a great deal of energy and time, and it is difficult to provide personalized guidance and counseling to each student in a timely and effective manner [7–10]. And today’s society is the information age, information technology penetrates into every aspect of modern people’s lives and changes people’s traditional way of life. Nowadays, students can conveniently use network technology to connect with the outside world and obtain knowledge and information from anywhere. Therefore, the establishment of personalized university English learning platform has become a trend. This kind of platform makes full use of modern information technology to meet students’ personalized learning needs, improve the autonomy and self-discipline of learning, reduce the teaching pressure of teachers, and promote the overall development of college students’ English learning [11–14]. Taking English vocabulary as an example, there are existing software on the market, such as Baiwu Chop, Multi Neighboring Countries, Fanbei Word, etc., each of which has its own characteristics, and learners can choose according to their own preferences. However, these platforms have a fixed mode and do not provide personalized services according to learners’ existing vocabulary level and learning interests, which leads to learners’ low efficiency in remembering words [15–17]. For this reason, such learning platforms need to be optimized. As for personalized learning platforms, due to the richness of learning materials and the demand for efficient learning, suitable personalized recommendations are needed to achieve real-time tracking of user needs and timely adjustments, and to proactively provide users with learning resources with the required content [18–20].

With the expanding scope of foreign exchange, both the socialization of college students and their English learning are affected, and various college English learning platforms have been launched one after another. First, this paper builds an English learning platform and introduces the functions of each subsystem. Secondly, k-Means clustering algorithm is used to construct the student portrait. The student portraits obtained based on the clustering results are introduced into the user-based collaborative filtering algorithm, and then the improved version of Pearson’s similarity algorithm is used to calculate the student similarity, and finally the improved version of collaborative filtering algorithm is obtained. Finally, HR and NDCG evaluation metrics are used to evaluate the performance of the model. In this paper, how to effectively use network resources in the context of network education technology, based on the recommendation algorithm to study the university English learning platform, the optimization of personalized recommendation of the English learning platform has certain theoretical significance.

2

Overview

The rise of personalized learning has led most scholars to develop and design various learning platforms and systems to provide personalized learning for learners. For example, platforms or tools such as One Learning Squirrel [21], Beelingua [22], ChatGPT [23], GamesHUB [24], and Massive Open Online Course (MOOC) [25]. In addition, the personalized e-learning platforms designed in the literature [26] combine visual, auditory, and reading with kinesthetic presentations and games to increase students’ learning interest and efficiency. Similarly, literature [27] study showed that Moodle plug-in can promote students’ ability to personalize their English grammar learning. Whereas, literature [28] study enhanced Moodle’s personalized services to promote personalized learning and improve learning efficiency from students’ learning materials, paths, and other processes. Some studies have designed platforms or systems only for accurate recommendation of block learning in English learning. For example, personalized English reading system [29], personalized vocabulary model for student characteristics [30], etc. The latter is designed by collaborative filtering based on students’ needs and interests. Similarly, literature [31] utilizes knowledge graph and user’s daily platform learning behavior characteristics to collaboratively filter and recommend in parallel to personalize and push students’ English learning materials. And literature [32] proposes a framework for a personalized learning recommendation system, intending that students can find their desired learning materials in a learning management system that brings together course resources and web resources for personalized learning. It can be understood that these personalized recommendations are designed based on students’ interests or user platform trajectories, but personalized learning also requires a gradual process. Therefore, literature [33] proposes an ordered learning path recommendation system based on learners’ course sequences and learning preferences to achieve personalized learning also based on a step-by-step learning process to improve learning efficiency.

3

Introduction to the Functional Modules of Each Subsystem of the English Learning Platform

3.1

User management subsystem

In order to facilitate management, the background management of all the columns is unified on the management page, teachers and administrators are able to manage the columns within their authority after logging in, such as adding, deleting, modifying and other operations. Based on its functional modules, draw the functional module structure diagram, the user management module is shown in Figure 1.

The user management subsystem involves the design of 2 modules, the design of the user login module and the design of the user management module. Initial users (teachers, students) must first register for the application, providing the system user name and password. The administrator logs in at the backstage, audits the users who have applied for registration and loads the records into the Users table of the database, and the administrator can also delete the existing user information. 1)

Administrator user

(1)

The administrator can log in to the front page and change his/her password after successful login, but cannot change the user name.

(2)

The administrator enters the website backstage login page through the website management, and the page needs to realize that the administrator user enters his/her own user name and password and passes the verification, and then logs in to the backstage management in order to carry out the corresponding operation.

(3)

After logging in, the administrator has the authority to view teachers, students, classes and messages.

(4)

After login, the administrator has the authority to add teachers, students and classes.

(5)

After logging in, the administrator has the authority to review messages and register users.

2)

Teacher User

(1)

Teachers can log in to the system through the front-end user login function; user registration can also be realized; after logging in to the system, you can also change your own password, but not the user name.

(2)

After logging into the system, teachers can see the assignments showing the latest learning releases, the discussion topics showing the latest releases, the resources showing the latest learning uploads, and they can realize the uploading and editing functions for the assignments, topics and resources.

3)

Student Users

(1)

Students can log in to the system through the front-end user login function; user registration can also be realized; after logging in to the system, they can also change their own passwords, but they cannot change their user names at will.

(2)

After logging into the system, students can see the assignments showing the latest learning releases, the discussion topics showing the latest postings, and the resources showing the latest learning uploads, and can download the resources or post comments.

3.2

Operations management subsystem

The assignment management subsystem mainly involves the design of the assignment display module and the assignment management module. The assignment management module is used by teachers and students. The operations of students include: submitting assignments online to the teacher of the class, and the submitted assignments support a variety of file formats, including text, pictures, animations and tables. Submitted assignments support various file formats, including text, pictures, animations and tables. Submitted assignment attachments support various file formats, including the commonly used RAR compression format. In order to standardize the assignment title format, the system automatically records the user’s name and adds it to the end of the assignment title. Based on its functional modules and drawing the functional module structure diagram, the homework management module is shown in Figure 2. 1)

Teachers upload the assignments for the class, select the classification according to the course name when adding, and the data will be synchronized into the homework table of the database after adding, and displayed in the latest learning published assignments.

2)

Teachers can also edit or delete the uploaded assignments, and the data in the homework table of the database will be operated accordingly.

3)

The system displays the most recently uploaded 30-question homework in the webpage according to the time when the teacher uploaded the homework, sorted by the most recent upload time.

4

Construction of Student Portrait Based on K-Means Algorithm for English Learning Platform

4.1

Introduction to the K-means clustering algorithm

As a common and effective division clustering algorithm, K-means algorithm (K-means clustering algorithm) is widely used because of the advantages of ease of use and high algorithmic efficiency, for a given set of data points, in order to achieve high-quality clustering effect, K clustering algorithm needs to be based on a certain distance function, will be repeatedly subdivided into the K clusters (specified by the user in advance). Connected regions located in higher density will form clusters of clusters (a kind of multidimensional space of point sets), and the aggregation of points in the space can be tested by completing the testing process using clusters of classes, separating different clusters of classes based on the density of the point set region between clusters of classes. Clustering method in the unknown operation premise, can for continuous unknown data clusters to the corresponding center of mass or characteristics based on the completion of the calculation of the distinction process. Assuming that the number of data objects is n, the K-means clustering algorithm first divides them and selects K clustering centers from them, and then the closest distance of the remaining objects from the centers of mass as the object of categorization, and recalculates the center of the clusters until all the clusters converge, and the process of the K-means clustering algorithm can be described as the completion of the first clustering clusters and the set of data objects (denoted by K and D, respectively) as an input, and on the basis of this, complete K clusters. On this basis to complete the output of K clusters (to meet the minimum criteria), the specific algorithm process is as follows:

Complete the random selection of K initial centers (denoted by μ⁽¹⁾, μ⁽²⁾, ⋯, μ⁽ⁿ⁾, n = 1, 2, ⋯, k) from D;

Repeat for i = 1, 2, ⋯, m;

Calculate the mean value of the samples of class cluster objects and assign them to similar class clusters;

End for;

for j = 1, 2, ⋯, n;

Calculate the mean of the object samples in the class clusters to obtain the center of mass;

If the maximum number of iterations is not reached, the current center of mass is updated;

If the maximum number of iterations is reached, the current clustering target is kept unchanged;

End for

By using the function of evaluating the clustering performance (criterion function) to achieve the acquisition of tight values to ensure that the iterative process of the K-means clustering algorithm to achieve the optimal results, the algorithm is suitable for the analysis of online learning behavior in the process of continuous attribute value clustering, K-means clustering algorithm in the face of dense clustering of data (data differences between the clusters is large) to obtain good results. The calculation effect.

4.2

Application of K-Means algorithm for student profiling

4.2.1

Deep Neural Network Based Student Feature Extraction

In order to enhance the expressive ability of students’ features and effectively mine students’ learning characteristics, this paper adopts deep neural networks to extract fine-grained features from students’ raw features. In order to enhance the expressive ability of the extracted student features, this paper constructs the positive and negative samples of the optimization objective with the key of whether there is a common learning difficulty between students and students to optimize the DNN feature extractor. Specifically, for a student, other students who have chosen the same difficulty point as the student are taken as the positive samples of the student, and students who have chosen different difficulty points are taken as the negative samples of the student, and the positive and negative samples are combined in order to achieve the optimization of the DNN algorithm. The detailed process of feature extraction is described below:

First, the original features of student users are mapped to the hidden space in the input layer, as shown in Equation (1): 1 $h_{0} = W^{i n p u t} x + b^{i n p u t}$

Where h₀ is the hidden features after processing by the neural network in the input layer, W^input is the neuron weights in the input layer, x is the original features of the student user after label encoding, and b_input is the bias parameter of the neurons in the input layer.

After the processing of the raw features by the neurons in the input layer, the user features can be further extracted and used to comprehensively portray the student learning profile. In order to increase the nonlinear fitting ability of the DNN model as well as mine out the deep features of the data, this paper sets two hidden layers in the number of neural network layers and uses relu activation functions. The details are shown in Equation (2): 2 $h_{l} = r (W^{l} h_{l - 1} + b^{l})$

Where h_l is the output of the l nd hidden layer, W^l is the parameter of the l th hidden layer neuron, b^l is the bias parameter of the l th hidden layer neuron and r(·) is the relu activation function.

After the processing of input layer and hidden layer, the student’s academic features are further extracted. In order to get the final academic features of students, this paper sets an output layer to obtain the final features of users. The details are shown in Equation (3): 3 $h_{o u t} = W^{o u t p u t} h_{l} + b^{o u t p u t}$ where h_out is the final extracted features of student users and h_l is the output of the hidden layer.

In order to effectively optimize the student’s academic features, this paper constructs the optimization objective of similar users and uses the constructed positive and negative samples to optimize the DNN. For the constructed positive and negative samples, a classifier is used to judge them, as shown in Equation (4): 4 $p_{i j} = s i g m o i d (f ([h_{o u t}^{i}; h_{o u t}^{j}]))$ where p_ij is the probability that user i and user j are common interest users, f (·) is the classifier, and $[h_{o u t}^{i}; h_{o u t}^{j}]$ is the vector after the feature splicing operation for user i and user j. For the classifier, a one-layer neural network is used to obtain the output values of its spliced vectors, and a sigmoid activation function is used to calculate the probability that student i and student j are students with the same difficulty.

After obtaining the probability, the model is optimized by applying the cross-entropy loss function for binary classification, as shown in Equation (5): 5 $l o s s = - p_{i j} \log p_{i j} - (1 - p_{i j}) \log (1 - p_{i j})$

By establishing the optimization objective to optimize the user’s features, the abstract features of student learning can be obtained, so as to achieve an accurate portrait of student learning.

Considering the dimensional size of the data, in this study, the number of layers of the deep neural network hidden layer is set to 2 layers, the number of neurons is set to 6, and the output layer is set to one neuron. The final neural network abstract feature output results of students’ learning attitudes and learning habits were obtained as shown in Table 1:

Table 1.

Neural network abstract characteristics of learning attitudes and habits

Student number	Learning attitude	Learning habit
1	-0.02601	-0.03124
2	-0.01827	-0.01996
3	0.13212	-0.02431
…	…	…
110	0.46121	0.05428
111	0.089412	0.124109
112	0.301224	0.097611

The absolute magnitude of the values in the table reflects the strength of the characteristics, with negative values indicating negative feedback, i.e., relatively poor attitudes toward learning, and positive values indicating positive feedback, i.e., relatively good attitudes toward learning. The same applies to study habits.

4.2.2

K-Means based cluster analysis of student population

Abstract features of students were extracted using deep neural networks in the previous section. In order to realize the portrayal of students’ learning characteristics, the sampled features extracted by deep neural network are used as labels for K-Means clustering portrait in this section. The K-Means clustering results constructed using the sampled features about students’ learning attitudes and learning habits obtained by deep neural network processing obtained in the previous section are used, and the k-Means clustering results of deep neural network features are shown in Figure 3.

The clustering result divides students into 3 categories: 28 students with poor learning engagement, accounting for 25%; 34 students with good learning engagement, accounting for 30%; and 50 students with excellent learning engagement, accounting for 45%. The clustering result reveals that students are more motivated to learn English at university.

The student profiles obtained through k-Means clustering algorithm provide the basic data for the next step of calculating the similarity between all students based on the student, learning resource matrix.

5

Research on recommendation algorithm modeling for English learning platforms

This section mainly introduces the implementation of collaborative filtering algorithm based on student portrait, user-based collaborative filtering algorithm adding student portrait can improve the accuracy of recommendation, the specific process steps are as follows: 1)

Construct the user - item scoring matrix, the user stands for the student, and the item stands for the learning resource which is the course, so the first step is to construct the student - learning resource matrix.

2)

Calculate the similarity between all students according to the student - learning resources matrix, find the set of students with similar interests to the current students, this is an important step in the user-based collaborative filtering algorithm, this paper will focus on the calculation of student similarity next.

3)

Based on the student similarity obtained in the second step, predict the students’ ratings (degree of interest) for other courses, and generate a recommendation list in accordance with the ratings, fusing the student profiles to calculate the degree of fit between the learning resources and the students, and generating the final list of recommendations to the target students.

5.1

Calculation of student similarity

5.1.1

Improvement of Pearson’s similarity algorithm

In this paper, the final method of calculating student similarity is formed by integrating Pearson similarity, learning resource weights, and scoring differences, and the formula is shown in (6). 6 $s i m (u, v) = s i m {(u, v)}^{P e a r s a n} \times w e i g h t (u, v) \times \frac{1}{d i f (u, v)}$

We get the k nearest neighbors with the highest similarity to student u by Pearson similarity, and then predict student u’s rating of learning resource i by the ratings of these k similar students on learning resource i, and then recommend the top N course resources with high predicted ratings to the students, as shown in (7). 7 $P_{u i} = {\bar{s}}_{i} + \frac{\sum_{v \in V} s i m (u, v) \times (s_{v i} - {\bar{s}}_{i})}{\sum_{v \in V} s i m (u, v)}$

In Equation (7), P_ui denotes the predicted value of student u on learning resource i, ${\bar{s}}_{i}$ denotes the mean value of student’s rating on resource i, V is the set of the top K students who have the highest similarity with student u, sim(u,v) is the similarity between student u and student v, and s_vi denotes student v ‘s rating on learning resource i.

Since the similarity between different students and student u is not the same, we use sim(u,v) as a weight to indicate the degree of influence on student u, and the mean score is introduced in Eq. (7) in order to eliminate the effect of the difference in ratings between different students.

We calculate the fit between the learning resources in the obtained initial source recommendation list and the students to get the final recommendation results, and the fit is calculated as shown in Equation (8). 8 $ψ_{u i} = \sum_{t \in T} r_{u, t} \times δ_{i, t}$

In Eq. (8), r_u,t, denotes the behavioral weight accumulated by student u on top of tag t, and δ_i,t denotes the feature weight of tag t in learning resource i, by which we get the final recommendation list.

5.1.2

Effectiveness of the improved version of Pearson’s similarity algorithm

In order to prove the optimization of the improved Pearson similarity algorithm for the student similarity method, this paper calculates the trend of RMSE and MAE with the number of neighbors k by using different similarity calculation methods, and the variation of RMSE and MAE with k under different similarity degrees is shown in Fig. 4, and we compare the improved Pearson similarity (Person2) with original Pearson similarity (Person), cosine similarity and Jekyll and Hyde similarity, from the figure, we can see that with the increase of k, the RMSE and MAE gradually decrease, which represents the accuracy of recommendation is getting higher and higher, and the improved Pearson is better than the other computation methods with the same value of k.

5.2

Effectiveness of Improved Version of Collaborative Filtering Algorithm

In order to verify the effectiveness of the improved version of collaborative filtering algorithm model for learning resource recommendation model, this paper adopts Canvas Network open dataset to conduct experiments, and the training set, validation set, and test set in the dataset are in the ratio of 14:3:3, and the performance of the model is evaluated by using HR and NDCG evaluation indexes.

For the improved collaborative filtering algorithm model (UserCF2), it is compared with the original collaborative filtering algorithm model (UserCF) and the deep neural network model which is widely used in the learning resource recommendation system at present, respectively. On the other hand, in order to seek the influence of the number of prioritized recommendations N on the experimental results, the experiments are carried out respectively by setting the value of N to 20, 30, 40, 50, and the performance comparison graphs for different models’ performance comparison graphs are shown in Fig. 5.

As the number of prioritized recommendations N increases, all three models increase in both HR and NDCG evaluation metrics, this is because the number of recommendations increases, then the higher the probability that the student will get the learning resources he/she wants. When the number of recommendations N takes the value of 20, the improved version of collaborative filtering algorithm, compared with collaborative filtering algorithm and deep neural network, improves 0.06 and 0.08 in HR evaluation indexes, and improves 0.003 and 0.01 in NDCG evaluation indexes, respectively, so that it has more advantages in processing user sequential data than both collaborative filtering algorithms and deep neural networks, and so the improved version of collaborative filtering algorithm has a better performance in the two evaluation metrics, the improved version of collaborative filtering algorithm has better performance.

6

Conclusion

In this paper, in view of the phenomenon of the sharp increase of university English learning platform, based on the original model of English teaching, a university English learning platform system is built. The K-Means algorithm is applied to the student portrait, the student group is clustered and analyzed, and the recommendation algorithm module of the English learning platform is studied in depth, and its optimizable and improved directions are proposed.

In the case that the number of neighbors k takes the same value, the improved Pearson similarity RMSE and MAE are lower than other calculation methods, indicating that the improved Pearson similarity is better than other calculation methods.

When the number of recommendations N takes the value of 20, the improved collaborative filtering algorithm, compared with collaborative filtering algorithm and deep neural network, improves 0.06 and 0.08 in HR evaluation indexes, and improves 0.003 and 0.01 in NDCG evaluation indexes, respectively, which compares that the improved collaborative filtering algorithm has more advantages in processing user sequential data, and has a better performance compared with collaborative filtering algorithm and deep neural network with better performance.

Sprache:: Englisch

Zeitrahmen der Veröffentlichung:: 1 Hefte pro Jahr
Fachgebiete der Zeitschrift:: Biologie, Biologie, andere, Mathematik, Angewandte Mathematik, Mathematik, Allgemeines, Physik, Physik, andere

Zeitschrift RSS Feed

Construction and Research of Personalized University English Learning Platform Based on Recommendation Algorithm

Ping Chen

Online veröffentlicht: 19. März 2025

Eingereicht: 30. Okt. 2024

Akzeptiert: 17. Feb. 2025

DOI: https://doi.org/10.2478/amns-2025-0502

SchlüsselwörterEnglish learning platform, User profiling, Pearson similarity, Collaborative filtering recommendation algorithm, Personalized recommendation

© 2025 Ping Chen, published by Sciendo

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

Schlüsselwörter
English learning platform, User profiling, Pearson similarity, Collaborative filtering recommendation algorithm, Personalized recommendation