Authors: Associate Professor Shilpa K
Abstract: This paper presents a multimodal, data driven framework for forecasting academic risk that was developed utilizing a three phase approach that prioritized model variety, data quality, and open decision making. Phase 1 involves the construction of a big data pipeline that connects the records of over 10,000 students from colleges throughout Karnataka. The analysis revealed thirty traits that are useful for predictive modeling. Among these are behavior, mental health, demographics, and college courses. The researchers studied some classical and neural models in Phase 2. These are Neural Networks (NN), Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), Naive Bayes (NB), K-Nearest Neighbors (KNN) and Decision Trees (DT). Five assessments of pairs of comparison models with different patterns are regularly made: RF frequently commits less mistakes compared with NN and DT, retains more information, and is more precise. SVM has lower false negatives compared to NN and is dependable. Two of the areas that LR outperforms NB are precision and F1 measures. KNN, which has greater classification stability, is a major challenge to NN. These findings indicate that structured ensemble and distance based models are more effective for heterogeneous educational data. SHAP and LIME are examples of explainable AI tools that are part of Phase 3. These tools make it easy to share features, check for fairness and create unique intervention programs for each college. The findings indicate that the combination of multimodal feature engineering, ensemble driven modeling and interpretable AI constitutes a robust and ethical approach to predicting academic risks.
