1 Introduction

Education plays an important role in both personal growth as well as social change around the world today. The value of education cannot be overstated as it equips individuals with the skills necessary for succeeding in their respective careers and concurrently serving their communities’ needs. Correspondingly, beyond individual gains lies the underlying national benefit that comes with an educated populace that drives social advancement alongside economic empowerment. In fact, education is critical towards fostering global economic innovation by generating productive citizens who compete favourably within dynamic markets.

A significant parameter used in determining a student’s academic performance is the Cumulative Grade Point Average (CGPA) scorecard as it shows a student’s overall progress throughout their study years. CGPA has a significant impact on a student’s future, as their future prospects and employment opportunities depend on it (Gamboa et al. 2018). CGPA is calculated by averaging the performance of a student in each semester, also known as the Semester Grade Point Average (SGPA). Predicting SGPA has numerous benefits for both faculties and students alike. The prediction of this metric allows faculties to identify students lacking the necessary support they need from time-to-time or those at risk of substantially falling behind others without tailored interventions unique to their learning needs at an early stage. Furthermore, predicting SGPA allows students to set reasonable future goals and make informed decisions based on it.

Machine Learning (ML) and Deep Learning (DL) have shown significant enhancement in various domains such as healthcare (Pathan et al. 2020), education (Shila Pawar 2020), agriculture (Goel et al. 2020), and intelligent transportation systems (Rai et al. 2023). In literature, various methods and techniques for predicting CGPA have been proposed ranging from ML algorithms including Decision Trees (DT) (Alyahyan and Dusteaor 2020) and Neural Networks (NN) (Saifuzzaman et al. 2021) (Verma et al. 2022) to statistical techniques such as regression analysis (Akuma and Abakpa 2021a) or Data Mining (DM) approaches like association rule mining (Bhaskaran 2021) etc. Despite varied effectiveness amongst these methods, the onset of COVID-19 has amplified needs for precise predictive modelling.

The data for this study was collected from undergraduate students (B.Tech CSE and IT) studying at DIT University, India through a Microsoft questionnaire designed to capture various influential factors. These factors incorporated several aspects such as academic performance, daily physical activity, social media interaction and study habits like help from outside sources like YouTube and other online platforms, help from friends, SGPA of semesters 1–4 etc. On the collected data, various classification and regression techniques were applied to determine the most effective techniques in predicting SGPA. This work has the potential to aid faculties, students, and policymakers by providing them with more accurate information about student performance.

This paper is structured as follows: Sect. 2 comprises review of related literature. Section 3 provides description of considered ML and DL models. In Sect. 4, the proposed work is discussed. Section 5 presents the findings of this study. Also, a comparison between existing works and the proposed work is depicted in Sect. 6. Finally, Sect. 7 outlines the conclusions and future directions for this research.

2 Background

The literature review related to ML and DL algorithms for predicting student academic performance is explained in this section. In (Baashar et al. 2022), several ML algorithms were used to predict the academic performance of postgraduate students. The study used a dataset of 635 master’s students from a well-known private university in Malaysia. The best results were obtained using an Artificial Neural Network (ANN) model, which achieved an error rate of only 11% when predicting the students’ CGPA. A Gaussian Process Regression (GPR) model also performed well, achieving an accuracy of 71%.

The authors of (Yakubu and Abubakar 2022) aimed to predict students’ performance in academics using various early detection indicators (i.e., gender, CGPA, age etc.) for enabling early intervention. It was done using logistic regression, a predictive analysis model appropriate for two dependent variables like CGPA. The dataset was divided into 30% testing data and 70% training data for supervised ML methodology. It achieved training accuracy of 84.7% and testing accuracy of 83.5%. Furthermore, Support Vector Machine (SVM) has also been used for predicting academic success in online learning and achieved an accuracy of 83.5% (Samsudin et al. 2022).

Several studies have been conducted to analyze students’ academic performance using ML algorithms. Sharma and Aggarwal (2021) have used Linear Regression (LR) to analyze the effect of parental influence on academic performance. The study found that family size, parents’ education, educational support from the family, internet access at home, paid classes and semester-wise marks were significant predictors. The accuracy was measured using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) which were 3.155 and 3.76, respectively. In (Gajwani and Chakraborty 2021), feature selection along with ML algorithm was used to analyze students’ performance based on behavioral and academic parameters such as nationality, gender, place of birth, student participation in discussion groups, raising a hand in classes, using external resources, grade along with semester marks. It was found that gradient boosting achieved an accuracy of 75%, followed by the Random Forest (RF) classifier achieving an accuracy of 74.31%.

In (Rai et al. 2021), various ML algorithms were applied to a dataset of university students and found that the RF classifier was the most accurate, with an accuracy of 94%. The study helped faculty to take early action and assist students which were categorized as poor and average. Other studies have taken different approaches to predict students’ academic performance. For example, in (Shetu et al. 2021) students’ results based on both academic status and environmental attributes were evaluated. It achieved an accuracy of 77% and MAE of 0.1087. Prime data mining methodologies like association, classification, clustering and DT were studied to analyze factors affecting a student’s academic performance (Ganorkar et al. 2021).

Also, in (Akuma and Abakpa 2021b) a model using LR Algorithm to predict student performance in academics based on their CGPA and prior academic records was developed. It accurately forecasted the fourth year CGPA of 70 undergraduate computer science students and achieved an accuracy of 87.84% and a correlation of 0.9338. Alangari and Alturki (2020) have explored the use of Educational Data Mining (EDM) to predict students’ performance. They tested 15 classification algorithms and found that the Naïve Bayes and Hoeffding tree models could predict students’ final GPA with 91% accuracy. These models outperformed others on multiple datasets, achieving an average classifier accuracy of 71%.

In (Yulianto et al. 2020), a comparison between data mining methods such as DT and K-Nearest Neighbors (KNN) was studied. It was found that DT performed worse than KNN, with an accuracy of 54.80%, while the latter with 59.32% accuracy. Furthermore, Lau et al. (2019) have used an ANN technique to model input variables, achieving an accuracy of 84.8%. Also, Islam Rifat et al. (2019) have predicted students’ CGPA using a DT algorithm as the base model along with Deep Neural Network (DNN) approach. The results obtained for Mean Square Error (MSE) with DT and DNN were 0.0226 and 0.008, respectively. Also, Table 1 summarizes the existing literature review.

Table 1 Literature review
Full size table

3 Description of models

ML and DL models can be broadly categorized into two types i.e., regression and classification. Following describes the models used in this work:

3.1 Regression models

Regression is a type of supervised ML algorithm which utilizes labeled data as input. It helps in determining the relation between dependent and independent variables.

  • Linear regression (Su et al. 2012): A statistical method that is used to determine linear relation between dependent and independent variables using a linear equation.

  • Ridge regression (McDonald 2009): This type of LR includes a regularization term in the loss function. It helps to prevent overfitting by reducing the magnitude of the coefficients towards zero.

  • Lasso regression (Ranstam & Cook 2018): It is a type of LR that adds regularization term to the loss function and uses the L1-norm instead of the L2-norm. It can result in some coefficients which are shrunk to zero, effectively performing the variable selection.

  • Elastic net regression (Hans 2012): A hybrid method combining Ridge and Lasso Regression by including L1 and L2 regularization terms in the loss function.

  • RF regression (Liu et al. 2012): An ensemble learning method that utilizes multiple DTs to make predictions.

  • Bayesian ridge regression: It is a type of regression analysis that estimates a probabilistic model of the regression problem.

3.2 Classification models

This supervised learning algorithm divides the dataset into classes based on various parameters. It is trained on a training set and then uses this information to classify the data into different classes.

  • RNN (GRU) (Dey & Salemt 2017): GRU is a form of RNN that effectively handles sequential data by utilizing GRUs. These units are a modified version of LSTM units and they possess fewer parameters, thus, allowing for faster training.

  • RNN (LSTM) (Staudemeyer & Morris 2019): It is a type of RNN that uses LSTM units to process sequential data. These are designed to retain information for longer periods, making it suitable for tasks such as language modelling and speech recognition

  • RF classification (Liu et al. 2012): It is a type of ensemble learning technique that utilizes multiple DTs to make predictions. It can be used for both classifications and regression.

  • Convolutional neural networks (CNN) (O’Shea & Nash 2015): CNN constitutes a neural network architecture that exploits convolutional layers in order to extract key features from input data. It is typically employed in image recognition and processing domains

  • Logistic regression classification (LaValley 2008): It is a statistical technique used for binary classification. It calculates the likelihood of an instance being in a specific class using a logistic function.

  • DT classification (Myles et al. 2004): A tree-like model is generated for prediction. The data is split into subsets based on the values of the input features.

  • Naïve bayes classification (Webb 2016): A probabilistic algorithm based on Bayes’ theorem that supposes independence between the features. It can be used for both binary and multi-class classification tasks.

  • Hoeffding tree algorithm (Kirkby 2007): It is a way to classify stream data using DTs. It employs Hoeffding trees and the Hoeffding bound to select the best splitting attribute from a small data sample.

4 Methodology

The proposed work predicts the SGPA of students using various regression and classification techniques by utilizing data collected through a survey form consisting of questionnaires related to academics and personal routines. The setup used for this experiment included INTEL I7-9750H CPU, NVIDIA GTX 1660ti GPU with 16 GB ram and 1 TB storage with Windows 11, python (3.9.16), VS Code (1.78.2), NumPy (1.24.3), Pandas (1.5.3), TensorFlow (2.10.1), and scikit-learn (1.2.1).

4.1 Dataset collection

For this study, the dataset was collected from undergraduate students (B.Tech CSE and IT) studying at DIT University, India through a survey form (Jain 2021). The details of the form are discussed in (Sect. 4.2) and is accessible at https://forms.office.com/r/pVUWHZ9gWJ.

4.2 Dataset description

The survey form consisted of following attributes:

  • Daily social media interaction: “less than 1 h”, “between 1 and 2 h”, “more than 2 h”.

  • Physical activity frequency: rated on a scale of 1 to 10.

  • Programming language knowledge: rated on a scale of 1 to 10.

  • Class Attendance: “85–95%”, “75–85%”, “65–75%”, “55–65%”, “Below 55%”.

  • Attention during lectures: rated on a scale of 1–10.

  • Daily study duration: “less than 1 h”, “between 1 and 2 h”, “between 2 and 4 h”, “more than 4 h”.

  • Study duration during exams: “between 2 and 4 h”, “between 4 and 6 h”, “between 6 and 8 h”, “more than 8 h”.

  • Frequency of taking help from external sources (YouTube, Google, Library): rated on a scale of 1–10.

  • Frequency of how often student clears doubts of other friends while studying in a group: rated on a scale of 1–10.

  • SGPA for semester 1, 2, 3 and 4.

4.3 Implementation

Initially, the data was imported to Python, and a label encoder (Bisong 2019) was used to convert the non-numeric values in the dataset to numeric classes. Label encoder transforms the non-numeric values to numeric classes ranging from 0 to n-1 classes. Afterwards, to fill in the missing values mean imputation technique was utilized (Jadhav et al. 2019). Furthermore, a split ratio of 0.2 was used for all the experiments (Roshan et al. 2022). The following techniques were used in this study:

4.3.1 Regression

In this study, a relationship was established between students’ personal daily routines, previous SGPA scores, and Semester 4 SGPA using regression models for prediction purposes. The comparison between true and predicted values for various regression models is shown in Fig. 1.

Fig. 1
figure 1

Output of regression models

Full size image

4.3.2 Classification

In this study, the SGPA of students was divided into 3 classes i.e., 0, 1 and 2 for SGPA in range “0–6.5”, “6.5–8.5” and “8.5 above”, respectively, using these the model was trained to predict class category for 4th semester SGPA of students. The comparison between true and predicted class values for various classification models is shown in in Fig. 2.

Fig. 2
figure 2

Output of classification models

Full size image

5 Result

The results obtained have been discussed in this section. This study used different classification and regression models to predict students’ 4th semester SGPA using the collected data. These models were evaluated and compared based on their MAE, MSE, RMSE along with accuracy values.

The regression models were evaluated using MAE, MSE, RMSE and R2 score as evaluation metrics (Hodson 2022). The results obtained after applying different regression techniques are summarized in Table 2, and a graphical comparison is shown in Fig. 3.

Table 2 Results of the Regression Models
Full size table
Fig. 3
figure 3

Comparison of Regression models

Full size image

Figure 3 concludes that Elastic Net regression has achieved the lowest MAE, MSE, RMSE and highest R2 score among all, proving to be the best among the regression techniques at predicting the SGPA of students.

The performance of the classification models was evaluated based on accuracy, F1 score and Recall. The results obtained after applying various classification techniques are summarized in Table 3, and a graphical comparison is shown in Fig. 4.

Table 3 Results of the classification models (in %)
Full size table
Fig. 4
figure 4

Accuracy comparison of classification models

Full size image

Figure 4 concludes that Naïve Bayes and Logistic regression classification have achieved the highest accuracy, F1-score and recall values among all applied classification models.

6 Comparative analysis

This section compares existing studies, namely (Gajwani and Chakraborty 2021) and (Shetu et al. 2021), with the proposed work as shown in Table 4. The comparison was based on dataset, techniques applied, and various performance metrics. The proposed study has utilized 10 features in the considered dataset whereas (Gajwani and Chakraborty 2021) and (Shetu et al. 2021) have used 7 and 9, respectively. Also, Gajwani and Chakraborty (2021) and Shetu et al. (2021) have applied 5 and 1 ML models, respectively, in their work while the presented work has applied 14 ML and DL models. Furthermore, the proposed study has achieved better accuracy i.e., 94.4% than other studies. The highest values of F1 Score and Recall were achieved by Logistic Regression Classification and Naïve Bayes Classification of 94 and 94 respectively, however, these metrics were not computed by Gajwani and Chakraborty (2021) and Shetu et al. (2021).

Table 4 Comparative analysis
Full size table

7 Conclusion and future scope

The proposed work predicts student’s SGPA using various regression and classification techniques. The data used for this study was collected from undergraduate students (B.Tech CSE and IT) studying at DIT University through a Microsoft survey form consisting of questionnaires based on academic parameters and daily personal routine. Using the collected data, regression and classification techniques were applied and the results were evaluated based on MAE, MSE, RMSE and R2 score for the regression models and accuracy, F1 score and recall for the classification models. Among regression models, Elastic Net Regression yielded the lowest MAE, MSE, RMSE and highest R2 score values and proved to be the best regression technique for predicting student SGPA, while Naïve Bayes and Logistic Regression yielded the highest accuracies among all other classification models. This work is helpful for both faculties as well as students as it can predict SGPA (of students) using ML and DL models with low error rate and high accuracy.

In future scope, this research could be expanded by collecting more data and using more algorithms to improve the accuracy of predictions. Other factors that may affect a student’s performance in academics can be considered and included in the analysis.