A student performance prediction method based on multi-model fusion in hybrid teaching environment
By adopting a multi-model fusion method in a hybrid teaching environment, combining polynomial naive Bayes, AdaBoost and Gradient boosting with logistic regression models, the problem of low accuracy in predicting grades among students in the existing technology is solved, and a higher level of achievement is achieved. Prediction accuracy and early risk discovery enhance the accuracy and timeliness of teaching management.
Patent Information
- Application Number
- CN202210329720.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2025-01-03
- Estimated Expiration
- 2042-03-31
AI Technical Summary
The existing technology is difficult to accurately and timely analyze students' learning status in a hybrid teaching environment, which leads to the widespread failure phenomenon, the traditional evaluation method has low accuracy, and it is impossible to effectively predict students' grades. Especially in the online learning environment, the data is complex and changeable, and a single prediction algorithm There are limitations in accuracy.
Multi-model fusion method, including polynomial naive Bayes, AdaBoost and Gradient boosting as primary classification models, combined with logistic regression as secondary classification models, build basic learning characteristics by obtaining students' learning behavior data and grade data, and conducting Data preprocessing and model training, use 50% cross-validation and grid search method to adjust parameters, and establish a student achievement prediction system that integrates two-layer models.
It improves the accuracy of student achievement prediction, can detect academic risks in advance, intervene and guide in a timely manner, and enhances learners' enthusiasm. It performs better than a single model, and the data source is simple and accurate.
Smart Images
Figure CN115099295B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a student performance prediction method using multi-model fusion in a hybrid teaching environment, and belongs to the technical field of model fusion. Background Art
[0002] With the development of information technology, online teaching has become widely popularized, and the mode of hybrid teaching formed by combining online and offline teaching is developing rapidly. Improving teaching effects by combining online and offline teaching data with technologies such as big data and cloud computing has gradually become a new trend in current teaching development. At present, the phenomenon of failing courses is relatively common among college students. However, failing courses may cause a series of bad consequences such as students being unable to graduate on time. If schools can detect students' abnormal learning behaviors at an early stage and intervene and guide them in time, it is possible to prevent these unfortunate events from happening. Therefore, by using data mining and analysis technology to mine and analyze students' learning behavior data, find out the rules hidden in the learning of different students, and predict their grades based on students' learning behavior data, it is possible to warn students whose academic performance is in danger, intervene and guide them in time, and improve students' grades and teaching quality. At present, most schools use the learning evaluation method and status feedback of students based on the subjective cognition of teachers to infer students' learning status. The inference method based on the subjective cognition of teachers has low accuracy, and with the increase in the number of online learning students, the traditional evaluation method will be unable to cope with it. With the development of big data technology, some scholars have conducted relevant research on methods for predicting students' online learning performance. However, the prediction algorithms used are relatively simple. Compared with single online teaching, students in blended teaching have more learning behavior characteristics, the data are relatively complex and changeable, and there are more outliers. The existing single prediction algorithm has certain limitations in accuracy. Summary of the invention
[0003] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a student performance prediction method in a hybrid teaching environment that can fully utilize the data generated by the platform to analyze the multi-dimensional characteristics of learning behavior and select effective features, accurately and timely analyze the students' learning status and intervene.
[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0005] A student performance prediction method based on multi-model fusion in a hybrid teaching environment includes the following steps: obtaining students' mid-term exam performance data and learning behavior data that affect final exam performance to construct basic learning feature data;
[0006] Inputting the basic learning feature data into the trained fusion algorithm model to obtain the student performance prediction result;
[0007] The fusion algorithm model includes a primary classification model constructed based on three classification models: polynomial naive Bayes, AdaBoost and Gradientboosting, and a secondary classification model constructed based on a logistic regression model; the secondary classification model predicts the final grade according to the output results of the primary classification model;
[0008] The training method of the fusion algorithm model includes:
[0009] Collect students' historical midterm exam scores, historical final exam scores, and historical learning behavior data that affect their historical final exam scores;
[0010] The basic learning characteristic data are constructed based on the students' history midterm exam scores and history learning behavior data, and the history final exam scores are used as the target characteristic variables;
[0011] Preprocess the basic learning feature data and target feature variables respectively;
[0012] The linear correlation between the preprocessed basic learning feature data and the target feature variable is calculated by the Pierce correlation coefficient, and the basic learning feature data that is not linearly correlated with the target feature variable is deleted to obtain a processed student feature data set;
[0013] Divide the processed student feature data set into a training set and a test set;
[0014] The training set is used to train the primary classification model and the secondary classification model in the fusion algorithm model to obtain a trained fusion algorithm model.
[0015] Methods for preprocessing the basic learning feature data and the target feature variables include:
[0016] The interval scaling method is used to scale the basic learning characteristic data of students in different classes to the same interval and then perform binning processing;
[0017] The target feature variable is classified with 60 points as the boundary: those with a score greater than or equal to 60 are success, and those with a score less than 60 are failure.
[0018] Before using the training set to train the primary classification model and the secondary classification model in the fusion algorithm model, the training set is divided into 5 training subsets, and the primary classification model and the secondary classification model are trained using the 5 divided training subsets.
[0019] Students’ learning behavior data include chapter study times, video viewing duration, video viewing rumination ratio, class attendance, and homework completion.
[0020] The basic learning feature data scaling formula is:
[0021]
[0022] In the formula, x is a feature set with a total of m feature variables, x j is the jth feature variable in the feature set, x j ' is x j The scaled value, max(x) is the maximum value of the characteristic variable of the set, and min(x) is the minimum value of the characteristic variable of the set.
[0023] The formula for the Pierce correlation coefficient is:
[0024]
[0025] Among them, g is the total number of data in the entire data set, Z i is the data corresponding to a certain learning feature in the data set, is the mean of all data corresponding to a certain learning feature in the data set, Y i The data set is different from Z i The data corresponding to a certain learning feature of To distinguish Z from i The mean of all data corresponding to a certain learning feature.
[0026] The training set D accounts for 70% of the student feature dataset, and the test set B accounts for 30% of the student feature dataset.
[0027] The three primary classification models, Multinomial Naive Bayes, AdaBoost and Gradient boosting, were trained using 5-fold cross validation.
[0028] Each cross validation consists of the following two steps:
[0029] First, randomly select 5 data sets D divided from the training set D i , 4 data sets in i∈{1,2,3,4,5} are used as the primary classification model training set for training, and the remaining 1 is used as the test set for the primary classification model prediction, and the prediction data set a is obtained. i ,i∈{1,2,3,4,5}, integrate them into one column and record it as A1;
[0030] Then, the primary classification model trained based on the four data sets in the previous step predicts the test set B and obtains the predicted data set p i ,i∈{1,2,3,4,5}, and then the results of the 5 predictions p i, i∈{1,2,3,4,5} are added row by row and the mean is taken to obtain the data set P1; after the three primary classification models complete training and prediction, two new matrix data sets (A1,A2,A3) and (P1,P2,P3) will be obtained; in the second layer, the secondary learner logistic regression uses the (A1,A2,A3) matrix as the training set and the (P1,P2,P3) matrix as the test set, and trains on them to obtain the final prediction result, that is, whether the student can pass the final test.
[0031] Each primary classification model and secondary classification model are trained and adjusted individually using the grid search method to obtain the optimal parameters of each model.
[0032] The beneficial effects of the present invention are as follows: the present invention provides a method for predicting student performance by integrating multiple models in a hybrid teaching environment, which describes the student's learning status through the student's online learning behavior characteristics and the student's basic information, and the sample data can be obtained from the school's teaching management system and online learning platform. The data source is simple and accurate, and is convenient for promotion in colleges and universities. The method can effectively predict students who are about to have academic problems for colleges and universities, so that teachers can grasp the learners' learning dynamics in time, discover the learners' learning risks in advance, intervene in a targeted manner, and improve the learners' learning enthusiasm. The method integrates three classification models, namely, polynomial naive Bayes, AdaBoost and Gradient boosting, as the first-level primary classification model of the fusion model, and adopts the logistic regression model as the second-level classification model of the fusion model, to establish a two-layer multi-model fusion student performance prediction model, and provide early warning for students' academic performance. Compared with a single prediction model, the multi-model fusion prediction model performs better in terms of prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a flow chart of a method for predicting student scores by integrating multiple models in a hybrid teaching environment of the present invention;
[0034] Figure 2 This is a 5-fold cross-validation flow chart of the fusion model in the present invention;
[0035] Figure 3 This is a graph showing the results of the correlation analysis of the learning behavior characteristics of middle school students in the present invention;
[0036] Figure 4 This is a comparison chart of the accuracy of each model in the present invention. DETAILED DESCRIPTION
[0037] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the protection scope of the present invention.
[0038] like Figure 1 As shown, the present invention discloses a student performance prediction method using multi-model fusion in a hybrid teaching environment, comprising the following steps:
[0039] Step 1: Obtain students’ learning behavior data from the online teaching system, mainly including chapter learning times, video viewing time, video viewing rumination ratio, class attendance, homework completion, etc.; obtain students’ personal information data and academic performance data from the school’s academic affairs system, mainly including midterm exam scores and final exam scores, etc.
[0040] The data information of various types of student behavior is shown in Table 1.
[0041] Table 1
[0042]
[0043] Step 2: Use the interval scaling method to scale the basic learning feature data of students in different classes, such as chapter learning times, video viewing time, video viewing rumination ratio, class attendance, homework completion, midterm exam scores, etc., to the same interval and perform binning. The feature data scaling formula is:
[0044]
[0045] In the formula, x is the same feature set, with a total of m feature variables, x i is the i-th feature variable in the feature set, x i ' is x i The scaled value, max(x) is the maximum value of the characteristic variable of the set, and min(x) is the minimum value of the characteristic variable of the set.
[0046] The target feature variable final exam score is divided into two categories with 60 points as the boundary, that is, scores greater than or equal to 60 are success, and scores less than 60 are fail.
[0047] Step 3: The linear correlation between the basic features such as chapter learning times, video viewing time, video viewing rumination ratio, class attendance, homework completion, midterm exam scores and the target feature final exam scores is calculated using the Pierce correlation coefficient. The results are as follows: Figure 3 As shown, the basic features such as the department where the students are located whose linear correlation with the target feature is 0 are deleted. The formula for the Pierce correlation coefficient is:
[0048]
[0049] Among them, g is the total number of data in the entire data set, Z i is the data corresponding to a certain learning feature in the data set, Y is the mean of all data corresponding to a certain learning feature in the data set. i The data set is different from Z i The data corresponding to a certain learning feature of To distinguish Z from i The mean of all data corresponding to a certain learning feature.
[0050] Step 4: Divide the student feature data set generated by the blended teaching after steps 2 and 3 into 70% as the training set D of the prediction model, and the remaining 30% as the test set B of the model, and divide the processed training set into 5 data sets D on average. i (i∈{1,2,3,4,5}).
[0051] Step 5: Select three classification models, namely, Multinomial Naive Bayes, AdaBoost and Gradient boosting, as the first-level primary classification model of the fusion model, and use the logistic regression model as the second-level classification model of the fusion model. The second-level classification model uses the results predicted by the first-level primary classification model as features to predict the final results. The model prediction flow chart is shown in the figure below: Figure 2 shown.
[0052] In the first-layer prediction model, three primary classification models, namely, multinomial naive Bayes, AdaBoost and Gradient boosting, were trained using 5-fold cross validation, respectively. Figure 2 shown.
[0053] Specifically, taking the Adaboost primary classification model as an example, each cross-validation includes the following two processes.
[0054] First, randomly select 5 data sets D from the training set divided in step 4. i The four data sets in (i∈{1,2,3,4,5}) are used as the training sets for the Adaboost primary classification model, and the remaining one is used as the test set for Adaboost model prediction to obtain the predicted data set a i (i∈{1,2,3,4,5}), integrate them into one column and record it as A1.
[0055] Then, the Adaboost model trained based on the four data sets in the previous step predicts the test set B and obtains the predicted data set p i (i∈{1,2,3,4,5}), and then the results of the 5 predictions p i(i∈{1,2,3,4,5}) is added row by row and the mean is taken to obtain the data set P1. After the three primary classification models complete training and prediction, two new matrix data sets (A1,A2,A3) and (P1,P2,P3) are obtained. In the second layer, the secondary classification model logistic regression uses the (A1,A2,A3) matrix as the training set and the (P1,P2,P3) matrix as the test set, and trains on them to obtain the final prediction result, that is, whether the student can pass the final test (fail, success).
[0056] Furthermore, each primary classification model and secondary classification model used in step 5 are individually trained and parameter adjusted using the grid search method to obtain the optimal parameters of each model. The optimal parameters of each model are shown in Table 2.
[0057] Table 2
[0058]
[0059]
[0060] Figure 4 The comparative experimental results of the classification accuracy of multinomial naive Bayes, AdaBoost, Gradient boosting, logistic regression and multi-model fusion prediction model on the data set are shown. Among the five methods, the classification accuracy of the multi-model fusion prediction model on the data set is higher than that of the other four models, indicating that the multi-model fusion prediction model can achieve good results when applying performance prediction to classification problems, further demonstrating the advantages and potential of the method of the present invention.
[0061] The present invention adopts a fusion model to give full play to the advantages of each learner. For the individual model with poor prediction in the test set, a discard strategy is adopted. The fusion model adopts multiple and different individual learners. The second-level classification model can make up for the prediction results of multiple primary classification models in the first layer, and can effectively reduce the risk of poor generalization performance of a single classification model to a certain extent. On the other hand, from the perspective of model optimization, in the training process of a single classification model, the model often has the risk of falling into a local minimum point, and the generalization performance of the model corresponding to some local minimum points may be poor. After the prediction by multiple primary classification models, the results are combined through the secondary classification model, which can effectively reduce the risk of falling into a local minimum point.
[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A method for predicting student scores by integrating multiple models in a hybrid teaching environment, characterized by: The following steps are involved: Obtain students' midterm exam score data and learning behavior data that affect final exam scores to construct basic learning feature data; Inputting the basic learning feature data into the trained fusion algorithm model to obtain the student performance prediction result; The fusion algorithm model includes a primary classification model constructed based on three classification models: polynomial naive Bayes, AdaBoost and Gradient boosting, and a secondary classification model constructed based on a logistic regression model; the secondary classification model predicts the final grade according to the output results of the primary classification model; The training method of the fusion algorithm model includes: Collect students' historical midterm exam scores, historical final exam scores, and historical learning behavior data that affect their historical final exam scores; The basic learning characteristic data are constructed based on the students' history midterm exam scores and history learning behavior data, and the history final exam scores are used as the target characteristic variables; Preprocess the basic learning feature data and target feature variables respectively; The linear correlation between the preprocessed basic learning feature data and the target feature variable is calculated by the Pierce correlation coefficient, and the basic learning feature data that is not linearly correlated with the target feature variable is deleted to obtain a processed student feature data set; Divide the processed student feature data set into a training set and a test set; The training set is used to train the primary classification model and the secondary classification model in the fusion algorithm model to obtain a trained fusion algorithm model.
2. The method for predicting student scores by integrating multiple models in a hybrid teaching environment according to claim 1 is characterized by: Methods for preprocessing the basic learning feature data and the target feature variables include: The interval scaling method is used to scale the basic learning characteristic data of students in different classes to the same interval and then perform binning processing; The target feature variable is classified with 60 points as the boundary: those with a score greater than or equal to 60 are success, and those with a score less than 60 are failure.
3. According to claim 1, the method for predicting student scores by integrating multiple models in a hybrid teaching environment is characterized by: Before using the training set to train the primary classification model and the secondary classification model in the fusion algorithm model, the training set is divided into 5 training subsets, and the primary classification model and the secondary classification model are trained using the 5 divided training subsets.
4. According to claim 2, the method for predicting student scores by integrating multiple models in a hybrid teaching environment is characterized by: Students’ learning behavior data include chapter study times, video viewing duration, video viewing rumination ratio, class attendance, and homework completion.
5. According to claim 1, the method for predicting student scores by integrating multiple models in a hybrid teaching environment is characterized by: The basic learning feature data scaling formula is: In the formula, x is a feature set with a total of m feature variables, x j is the jth feature variable in the feature set, x j ' is x j The scaled value, max(x) is the maximum value of the characteristic variable of the set, and min(x) is the minimum value of the characteristic variable of the set.
6. The method for predicting student scores by integrating multiple models in a hybrid teaching environment according to claim 1 is characterized by: The formula for the Pierce correlation coefficient is: Among them, g is the total number of data in the entire data set, Z i is the data corresponding to a certain learning feature in the data set, is the mean of all data corresponding to a certain learning feature in the data set, Y i The data set is different from Z i The data corresponding to a certain learning feature of To distinguish Z from i The mean of all data corresponding to a certain learning feature.
7. The method for predicting student scores by integrating multiple models in a hybrid teaching environment according to claim 1 is characterized by: The training set D accounts for 70% of the student feature dataset, and the test set B accounts for 30% of the student feature dataset.
8. The method for predicting student scores by integrating multiple models in a hybrid teaching environment according to claim 1 is characterized by: The three primary classification models, Multinomial Naive Bayes, AdaBoost and Gradient boosting, were trained using 5-fold cross validation.
9. The method for predicting student scores by integrating multiple models in a hybrid teaching environment according to claim 8, characterized in that: Each cross validation consists of the following two steps: First, randomly select 5 data sets D divided from the training set D i , 4 data sets in i∈{1,2,3,4,5} are used as the primary classification model training set for training, and the remaining 1 is used as the test set for the primary classification model prediction, and the prediction data set a is obtained. i ,i∈{1,2,3,4,5}, integrate them into one column and record it as A1; Then, the primary classification model trained based on the four data sets in the previous step predicts the test set B and obtains the predicted data set p i ,i∈{1,2,3,4,5}, and then the results of the 5 predictions p i , i∈{1,2,3,4,5} are added row by row and the mean is taken to obtain the data set P1; after the three primary classification models complete training and prediction, two new matrix data sets (A1,A2,A3) and (P1,P2,P3) will be obtained; in the second layer, the secondary learner logistic regression uses the (A1,A2,A3) matrix as the training set and the (P1,P2,P3) matrix as the test set, and trains on them to obtain the final prediction result, that is, whether the student can pass the final test.
10. The method for predicting student scores by integrating multiple models in a hybrid teaching environment according to claim 1, characterized in that: Each primary classification model and secondary classification model are trained and adjusted individually using the grid search method to obtain the optimal parameters of each model.