Interactive Online Learning Method for Students Based on Artificial Intelligence

Synthesis: Li and Yin (2025) propose a hybrid DMO-GRU framework for AI-powered interactive online learning, combining the Dwarf Mongoose Optimization (DMO) algorithm with a Gated Recurrent Unit (GRU) neural network. DMO performs feature selection and automatic hyperparameter tuning, while GRU captures temporal patterns in sequential student data. Evaluated on classification and regression tasks against Linear Regression, Random Forest, SVR, and XGBoost, the model reports superior accuracy (91.2%), F1-score (90.8%), precision (90.1%), recall (92.5%), and the lowest error (MAE 2.70, RMSE 3.40, R² 0.92), while also supporting interactive audio, video, and one-to-one learning modes intended to boost engagement.

Key Findings

• Best classification performance: DMO-GRU achieved 91.2% accuracy, 90.1% precision, 92.5% recall (sensitivity), 89.8% specificity, and a 90.8% F1-score, outperforming GRU, LSTM, CNN-GRU, Random Forest, and XGBoost baselines.
• Lowest prediction error: On regression metrics the proposed model posted MAE of 2.70 and RMSE of 3.40 versus XGBoost (3.50/4.70), Random Forest (3.90/5.00), SVR (4.20/5.50), and Linear Regression (4.80/6.20), with the highest R² score (0.92).
• Efficient training: DMO-GRU trained in ~98 s — far faster than LSTM (138 s) and CNN-GRU (150 s) — while remaining more accurate than the faster classical models.
• Automatic tuning and interpretability: DMO auto-optimizes GRU hyperparameters (learning rate, batch size, hidden units, dropout); the paper points to SHAP/LIME and attention visualizations to improve model interpretability for practitioners.
• Engagement supports: The system integrates synchronous interactive modes — two-way audio classrooms, one-to-many video classrooms, and one-to-one video sessions — intended to improve learner engagement and mimic the traditional classroom online.

Study Design & Method

The study is presented as a hybrid modeling effort built on a literature review of AI in online/interactive education (sources from IEEE Xplore, ACM, and Google Scholar, largely the last decade). Data preprocessing removed or imputed missing values, normalized numeric features with Min–Max scaling, one-hot encoded categorical variables, excluded outliers via Z-score (|Z| > 3), and applied SMOTE for class imbalance.

The core contribution is the DMO-GRU framework: DMO (modeled on dwarf mongoose foraging, with alpha/scout/babysitter groups and an added Learning Strategy) performs feature selection and hyperparameter tuning, while a GRU network — with reset and update gates to handle the gradient-vanishing problem in sequence learning — captures temporal patterns in student engagement and performance. The model was evaluated against GRU, LSTM, CNN-GRU, Random Forest, and XGBoost using classification metrics (accuracy, precision, recall, sensitivity, specificity, F1) and regression metrics (MAE, RMSE, R²), with training time reported. Experiments were run in MATLAB R2023a. The authors acknowledge limitations: dependence on Benchmark datasets, limited real-world testing, model interpretability concerns, and limited attention to ethical issues such as data protection and fairness.

What this means for practice

• Designers. Combine a metaheuristic optimizer with a sequence model when building learner-performance prediction, and report regression and classification metrics together — the reported 91.2% accuracy came alongside MAE 2.70 and RMSE 3.40, and accuracy alone would hide which at-risk students are being missed.
• Designers. Ship explainability alongside prediction: the paper pairs the engagement classifications with SHAP/LIME feature attribution and attention visualizations, so instructors can see which inputs drive an at-risk flag before acting on it.
• Designers. Keep the architecture lightweight and auto-tuned for deployment, since DMO-GRU trained in about 98 s against 138 s for LSTM and 150 s for CNN-GRU at higher accuracy than the faster classical baselines.
• Learners. Treat an at-risk or disengaged classification as a trigger to use the platform's interactive modes — two-way audio classrooms, one-to-many video, and one-to-one sessions — rather than as a fixed verdict on ability.

Limitations

• Validation used benchmark datasets with no deployment on a live platform, which the authors name as their first stated shortcoming, so the reported gains are preliminary rather than established.
• The authors also flag that the model is not interpretable and that ethical issues such as data protection and fairness received little attention in the study.
• Comparisons were run only against classical baselines (Linear Regression, Random Forest, SVR, XGBoost) in MATLAB R2023a, and the preprocessing pipeline — Z-score outlier removal at |Z| > 3 and SMOTE balancing — shapes the training data without separate ablation.

Citation

Li, C., & Yin, W. (2025). Interactive online learning method for students based on artificial intelligence. Discover Artificial Intelligence, 5, 201.