Predicting Next-Level Grade For Students In English Language Institute

The use of machine learning,data mining,and statistics in the educational field is called educational data mining.It turns the extensive educational databases’ raw data into beneficial information that can be utilized for decision-making in educational systems and a better knowledge of students and their learning progress.EDM helps to measure the current situation in any field despite the variety of data sources and goal orientation.By understanding the correlation between available input variables and output(i.e.,label),they can be concentrated and given more attention to give insights in any way to help education development.Predicting student performance is a technique to find variables that best describe students’ conditions or learning results.It is also among the most crucial topics for learning environments like schools and universities since it aids in designing efficient processes that,among other things,promote academic performance and prevent students’ dropout.Historical data on students’ achievements can be a handy way to foretell the results of upcoming courses or terms students haven’t studied yet.Students can use this information to choose the majors that are the best fit for them and to schedule classes for various difficulties effectively.Retention rates of students can be increased in case next-term grade prediction incorporates into an early-warning system.Also,effective prediction models would provide important insights into the variables affecting students’ achievement across different subpopulations.A special kind of educational institution called English language institute where non-English speakers can take special,intensive classes to improve their English proficiency.This thesis aims to apply the next-term grade prediction approach to an English language institute since current studies were implemented only in universities and schools.The program structure for English language institutes is more flexible than universities and schools,for it is built on a chain of levels instead of courses.Students’ English proficiency determines which level they should begin at,and there is no mandate for them to finish all along levels to graduation.Therefore,this study can provide information that might be used to deliver knowledge to advisors and teachers when students require additional assistance.Early acknowledgment of at-risk students is crucial for avoiding their discouragement and dropout.The result of multiclass classification can be utilized to cluster students with similar English proficiency in same class to not make limit the diversity between students in class.This thesis conducted four experiments to achieve the goal of this research as follows1-The extracted dataset for final grades of levels was reshaped into two subsets.The first subset was initialized to build models to predict students’ nextlevel grades in a specific level,and it’s named level-specific prediction.Thus,these models focus more on levels to understand their difficulty and how they associate with each other.The second model predicted the grades of levels’ ordinal numbers for students and named it student-specific prediction.This way attended to tested students’ knowledge in predicting scores of the next level despite the name of levels.The number of prior levels was parametrized as a number in a set of(4,6,8)for both approaches to notice the change in prediction performance when more prior levels were provided as input variables to the model.The results showed no major difference between the two approaches,yet some algorithms gave level-specific prediction advantages over student-specific prediction.Also,most algorithms performed well with four prior levels due to the massive number of examples compared to other sets.2-Three main machine learning techniques were investigated to see how next-level grade prediction in English language institutes would differ.Regression algorithms: Linear Regression,Decision Tree Regressor,and Random Forest Regressor took a role in predicting grades in continuous values.Classification algorithms were also added to find out the grade in discrete values.Classification algorithms for this experiment are K-Nearest Neighbors,Decision Tree Classifier,Random Forest Classifier,and Na(?)ve Bayes.The last technique was taken from a recommender system approach named Collaborative Filtering which finds similarities between students to make a prediction.This experiment’s output displayed that Linear Regression achieved results with the lowest RMSE and MAE values among regression algorithms.In the classification technique,the highest precision values went to k-nearest neighbor,random forest,Na(?)ve Bayes,and Decision Tree sequentially.Collaborative Filtering results obtained higher errors(i.e.,RMSE and MAE)than linear regression.3-This experiment was conducted to see the variety of algorithms’ performance when the final grades of levels are 100-scale and 4.0-scale.All mentioned algorithms were taken grades in 100-scale to build prediction models;then grades were transformed to 4.0-scale to create prediction models.The outcome of this part presented a considerable difference between metrics values in regression and Collaborative Filtering with a preference for 4.0-scale.In classification,however,the contrast wasn’t major diversity with preference to 4.0-scale as well.4-The last experiment split the classification task into multiclass classification and binary classification.Similar algorithms were tested for both to detect the contrast in metric evaluation values.Binary classification results were higher and better than multiclass ones.This research showed the possibility of predicting students grades in the following level in English language institutes,and it is a base for further advanced researches.The nature of available data that only provides final grades for students caused limitations in building and evaluating prediction models.