Research on the Innovation of Ideological and Political Education Methods for College Students under the Background of “Internet+”

Obtaining a dataset rich in learning data dimensions is the first task to carry out the problem of predicting learners’ online learning performance, because only by obtaining comprehensive data on learners’ online learning can we more fully and accurately reflect the learners’ online learning status, and better provide a strong data support for the next steps, such as the construction of data mining models. In this paper, Moodle [14] is used as the platform for learners’ online learning, and the online learning data of college students’ ideological and political education is collected as the analyzed data set for this experiment.

The collected data cannot be used directly because some of the data may be incomplete and there may be some “dirty data”, which will affect the accuracy of the knowledge model extracted by data mining. At the same time, the multi-algorithm fusion prediction process is dynamic, complex and variable, and is usually affected by singular values, improper selection of influencing factors and other elements. In order to reduce the impact of the differences between different prediction algorithms on the prediction results, appropriate and scientific data processing and learning performance influence factor selection methods should be sought in the multi-algorithm fusion prediction process. Data preprocessing is one of the main data operations before model construction, which performs normalization operations on the data to make the initial prediction dataset into the standardized data format required by data mining methods.

In the process of building a prediction model, the algorithms chosen differ according to the research tasks and can generally be divided into two categories: one is classification tasks and the other is prediction tasks. Regarding the prediction aspect of learning behavior in students’ ideological and political education, prediction tasks are usually used to make predictions about specific learning behaviors of students, as well as the use of classification tasks to predict the level of students’ learning behaviors. In real-world settings, students’ individualization varies greatly and the specific results are taken too broadly; therefore, past research has concluded that in the case of using students’ learning behavior prediction as a prediction task, its prediction results are usually more accurate. Accordingly, in classification tasks, where student scores are categorized into a limited number of classes, ideal prediction results are often obtained. In addition, when the prediction of student learning behaviors was used to identify academically at-risk students, more attention was paid to the student’s behavioral class than to specific behavioral scores, so the study examined the prediction of student learning behaviors as a classification task.

The XGBOOST algorithm [16] is similar to the GBDT algorithm in its basic idea of continuous feature splitting to grow into a tree, learning one tree per round, which, in fact, is designed to fit the residuals of the predicted and actual values of the model from the previous round. After obtaining k trees at the end of the training, we will predict the scores for the samples, based on the nature of the samples, each tree will be dropped to the corresponding leaf node, each leaf node corresponds to a score, and finally, just add the corresponding scores of each tree, that is, the predicted value of this sample.

Given the dataset D={(x₁,y₁)}, XGBoost is trained to learn the K tree and predict the samples using the following method: (2) y˜=ϕ(xi)=∑k=1Kfk(xi)⋅fk∈Γ

Each round of training corresponds to a tree, all with the goal of minimizing the loss function, i.e., what each tree is trained to achieve. The loss function of XGBOOST differs from GBDT in that it is not just a measure of the model fit error; regularization terms are also added, i.e., tree complexity penalty terms for each tree, to constrain the complexity of the tree and avoid overfitting.

The objective function in the parameter space: (3) Obj(θ)=L(θ)+Ω(θ)

Inside the formula L(θ) denotes the error function: the error between the model prediction and the actual value, and Ω(θ) denotes the regularization term: the complexity of the tree with penalty. The objective function in the function space is of the form: (4) L(ϕ)=∑it(y˜v,yi)+∑kΩ(fk)

where the regularization term: (5) Ω(f)=γT+12λ‖ w ‖2

T is the number of nodes in the leaf and w is the score of the leaf node. The second-order Taylor expansion of the error function, after the t th iteration, the model prediction is equivalent to the last t–1 model prediction and the t th tree prediction cumulatively: (6) y^zt=y^t(t−1)+fi(xi)

Objective function: (7) L(t)=∑i=1nl(yi,y^t(t−1)+ft(xt))+Ω(ft)

Eqs. yi,y^t(t−1) , are known, and the model only needs to study the t-tree f_t. The above function undergoes a second-order Taylor expansion at y^t(t−1) : (8) L(t)□∑i=1n[ L(yi,y^t(t−1))+gift(xi)+12hift2(xi) ]+Ω(ft) (9) gi=∂y^i(t−1)l(yi,y^i(t−1)) (10) hi=∂y^i22(t−1)l(yi,y^i(t−1))

Removing the constant term inside the formula gives: (11) L(t)=∑i=1n[ gift(xi)+12hift2(xi) ]+Ω(ft)

Write (f_t), Ω(f_t) in the form of a tree structure, i.e., bring the following equation into the objective function: (12) f(x)=wq(x)Ω(f)=γT+12λ‖ w ‖2 (13) L(t)=∑j=1T[ (∑i∈Ijgi)wj+12(∑i∈Ijhi+λ)wj2 ]+γT

Minimum loss is: (14) L˜*=−12∑j=1TGj2Hj+λ+γT

XGBoost scoring function when the structure of the regression tree is determined: (15) L˜*=−12∑j=1TGj2Hj+λ+γT

Split a leaf node and define the gain before and after the split [17]: (16) Gain=GL2HL+λ+GR2HR+λ−(GL+GR)2HL+HR+λ−λ

The higher the Gain, the faster the L drops after splitting. When a leaf node is segmented, the Gain corresponding to all candidate features is calculated and the one with the largest Gain is selected for segmentation.

The exact algorithm is to traverse all possible segmentation points of all features, calculate the gain value, and select the feature with the largest value to segment. Approximate algorithm for each feature, only examine the split point, reduce the computational complexity. Globa1 for learning each tree before, propose candidate cut points, Local: each time before disaggregation, re-propose candidate cut points.

XGBoost has several regularization methods to prevent overfitting [18], and limits the complexity of the tree by including in the objective function not only the fitting error function of the model, but also by adding a penalty term on the complexity of each tree, i.e., the number of leaf nodes as well as a squared term on the fraction of leaf nodes. Like GBDT, it is possible to multiply each model by a step a, a∈(0,1], which is used to descend the contribution of each model to the prediction. Row sampling can be done with column sampling, similar to random forests.

This paper takes the third-year computer science major (1) and (2) students of university A as the object of analysis, and selects the learning behavior data related to the ideological and political education courses in the online learning platform of the students in the two classes. The learning behavior characteristics include chapter, check-in interval, classroom interaction score, answer completion rate, answer correct rate and score rate In order to make the connection between behavioral characteristics and learning status more intuitive, and to simplify the output of the subsequent prediction model, this paper further recombines the behavioral characteristics of the classroom stage (excluding the characteristic “chapter”) to obtain the three types of classroom performance, namely, “class motivation”, “class participation”, “knowledge mastery”, “class participation”, “class participation”, and “knowledge mastery”, “Class participation” and ‘knowledge mastery’. Among them, “classroom interaction score” and “question completion rate” reflect students’ active participation in the whole class, so they are combined to get “classroom participation”, and “correct answer rate” is the most important one. The “correct answer rate” and “score rate” imply the mastery of the knowledge in this class, so they are combined to get the “knowledge mastery”, and finally, the “check-in interval” is a measure of students’ participation in the class. Finally, “check-in interval” is renamed as “class motivation”. Through data segmentation and integration, a classroom performance dataset containing 2312 valid data was obtained, and the information of each category is shown in Table 1. The sample sizes of those who performed positively in class motivation and class participation were 365 and 1124, respectively.

The student learning behavior prediction experiments involve three types of models, which are traditional classification algorithm SVM, BP neural network and LSTM neural network based, so the above three models are selected as comparison models in this paper. In order to enhance the robustness of the models, all the experiments in this subsection use 5-fold cross-validation method. In addition, 80% of the data is selected as the training set and 20% of the data is selected as the test data set. The following three types of experiments are designed correspondingly.

In order to further verify the effect of the learning behavior prediction model proposed in this paper, the MA-LSTM model with better prediction performance and the XGBOOST learning behavior prediction model proposed in this paper are selected to analyze the process of F1 value enhancement of the two models, and the analysis results of the model’s F1 value change with the number of iterations are shown in Figure 2. It can be seen that the MA-LSTM model is really not enough to find the optimal ability in the early stage, while it also falls into the local optimum in the later stage. In addition, the F1 value of the XGBOOST model is higher than that of the MA-LSTM model in both the early and late stages, and the final improvement of the model performance is very considerable. When the number of model iterations is the first, the F1 value of XGBOOST model in predicting the behaviors of “class motivation”, “class participation” and “knowledge mastery” has reached 0.6559, 0.6477 and 0.6719. This indicates that using the split node algorithm in the XGBOOST model to find the parameters of the attention layer is more advantageous than the traditional way. The MFO algorithm in the MA-LSTM model can find relatively better parameters of the attention layer, but its defects lead to falling into the local optimum at a later stage, which ultimately does not improve the performance of the prediction model much. However, the experimental results of the XGBOOST model show that the split node algorithm in the model can make it jump out of the local optimum, and further improve the performance of the prediction model and training efficiency.

Figure 2.

The improvement curve of F1 values

In this paper, a practical experiment on ideological and political education was conducted in (1) class (72 students) and (2) class (75 students) of computer science majors in the third year of college A. The students in (1) class were taught using a semester-long precision intervention teaching method based on the results of the prediction of learning behaviors proposed in this paper. At the same time, the traditional lecture method was used in the ideological and political education course of the (2) class of computer science majors. After one semester of study, the learning behaviors of students in each class in the ideological and political education course were analyzed. Meanwhile, in order to objectively assess the effectiveness of the practice, examination papers were designed at the beginning and the end of the semester, and the data collected on the performance of the two classes were compiled and analyzed in order to visually compare the two teaching methods. A total of 72 and 75 test papers were distributed to the two classes, and the recovery rate of valid test papers was 100%, and the effective rate also reached 100%. The details of the question papers were test questions on Ideological and Political Education in Colleges and Universities. The paper will have a total of 100 points and will consist of multiple-choice, fill-in-the-blank, short-answer, and synthesis questions covering all of the highlights of the semester’s ideological and political education course.

(The results of the analysis of the learning behaviors in the ideological and political education course at the end of the semester and the beginning of the semester for class (1) and class (2) are shown in Table 3. The analysis of the actual and predicted results of class (1) shows that the learning behavior prediction model is extremely accurate in predicting the learning behavior of college students in ideological and political education, which further confirms the feasibility of the model. From the comparative analysis of the learning behavior results of the (1) class under the precise intervention education model and the (2) class taught by the lecture method, it can be seen that at the beginning of the semester, the class motivation, class participation, and knowledge mastery of college students’ ideological and political education in the (1) class and the (2) class were not high, and the number of students at the motivated level ranged from 11 to 18. After a semester of educational practice, the ideological and political education learning behaviors of the students in class (1) were significantly improved, in which the number of students with positive behavior levels in class motivation and class participation reached 41 and 45 respectively, accounting for 56.94% and 62.50% of the total number of students in the class. The number of students in the level of inactive behavior also decreased significantly, with only 2 students in the level of inactive class. On the other hand, the learning behavior of class (2) did not improve significantly under the traditional education model, and the number of students with inactive behavior in the level of class motivation even increased from the initial 14 to 16. This shows the accuracy of predicting learning behavior in ideological and political education based on XGBOOST, and also finds that the implementation of the innovative educational method of precise intervention in ideological and political education in the era of “Internet+” has obvious effects.

In this paper, the performance of ideological and political education courses is divided into three grades, which are Grade A (80-100 points), Grade B (60-79 points) and Grade C (<60 points). The analysis results of the learning achievement of ideological and political education courses of the students in class (1) before and after the educational practice are shown in Fig. 3, and the analysis results of the learning achievement of ideological and political education courses of the students in class (2) are shown in Fig. 4, in which (a) and (b) both represent the analysis results before and after the educational practice is carried out. In terms of the average value of learning achievement in ideological and political education courses, the average scores of classes (1) and (2) before the implementation of educational practice were 64.92 and 67.27 respectively, and the average learning achievement of the two classes increased to 86.52 and 78.05 after the practice was completed, with the achievement of class (1) being significantly higher than that of class (2). It is often the intermediate level, i.e., Level B, that has the highest percentage of students in a class, which means that these students may be neglected in terms of their academic performance. Students at this level already have a foundation to build on, and with enough effort, they can expect to reach the A level. However, they may also fall to Level C if they are not careful enough, which means that they will lose more opportunities and advantages and need to work harder to get back to their original level. Therefore more attention and support should be given to these students to motivate them to pursue higher learning goals and challenge higher levels of learning content. This will help develop their self-confidence and sense of self-challenge so that they can better cope with their future learning and career development. Under the innovative education model of precise intervention, teachers make precise interventions for each student’s learning status, and each student can obtain the joy of success through the education method suitable for him/her. After the implementation of the innovative education model, the number of A-level students in the class of (1) reached 55, accounting for 76.39% of the total number of students in the class, which indicates that the implementation of the innovative education model has achieved significant results, and it can effectively improve the students’ academic performance and learning interest in ideological and political education.

Figure 3.

Analysis results of academic achievement (Class 1)

Figure 4.

Analysis results of academic achievement (Class 2)