I’ve built a machine learning project to predict student employability: Trained and evaluated multiple ML models (SVM, Decision Tree, Logistic Regression, KNN) using Python. Choose the best-performing model (SVM) based on F1-score after applying SMOTE to balance the dataset. Saved the trained model and scaler using joblib. Deployed the model using a Streamlit app where users can play around with predict the employability. When I run predictions in the Streamlit app: All predictions are always the same (Less Employable) — even if I adjust the input sliders to extreme values. The predicted probabilities also don’t change much. What I’ve Checked In the notebook: model evaluation works fine. The SVM produces correct predictions & probabilities. .pkl files (model & scaler) are generated and loaded correctly. The input features in Streamlit are named and ordered the same way as during training. I scale the input using the same scaler. I’ve uploaded all my code and data to this GitHub repo. It contains: Training notebook (Train_Employability_Model_Final.ipynb) Streamlit app (student_employability_app_final.py) Dataset (Student_Employability_CP2_2025.xlsx) Trained model & scaler (employability_predictor.pkl, scaler.pkl) requirements.txt Could someone help me review: My training notebook: to confirm that the .pkl files are correct. My app.py Streamlit code: to check if the way I prepare and scale the input before prediction is correct. Continue reading...
