Placement Predictor ML System

1. Problem Statement

Educational institutions often struggle to identify which students are at risk of not securing a placement during campus recruitment drives. Failing to identify these students early prevents targeted intervention, such as extra training or interview preparation. This project solves that problem by providing an end-to-end classification system predicting the likelihood of a student being placed based on their academic and demographic profiles.

2. Dataset

Source: Internal university academic records (anonymized) supplemented by historic placement data.
Dataset Size: Approximately 10,000 student records spanning 5 years.
Features: Includes numerical features (CGPA, high school marks, aptitude test scores) and categorical features (branch of study, internships completed, backlogs).
Preprocessing: Handled missing values using median imputation for continuous variables and mode imputation for categorical ones. Scaled numerical features using standard scaling.

3. Feature Engineering

Polynomial Features: Applied degree-2 polynomial interactions between aptitude scores and CGPA to capture non-linear relationships.
Categorical Encoding: Used target-guided ordinal encoding for categorical features like “branch” strictly based on historical placement rates for that branch.
Feature Selection: Used Recursive Feature Elimination (RFE) to whittle down 24 initial features to the 12 most predictive ones, ensuring the model remains computationally efficient and robust against overfitting.

4. Models Tested

Model	Accuracy	F1 Score
Logistic Regression (Baseline)	0.82	0.78
Decision Tree Classifier	0.88	0.86
Random Forest Classifier	0.94	0.93
XGBoost	0.93	0.92

5. Final Model

The Random Forest Classifier was selected as the final model. Despite XGBoost performing similarly, Random Forest proved slightly more resilient to outliers in our specific academic domain without requiring extensive hyperparameter tuning. It strikes an excellent balance of high precision (avoiding false placement hopes) and high recall (catching at-risk students), which is precisely what university administration requires.

6. Evaluation

Confusion Matrix Details: Showed a very low False Positive Rate (predicting placed when they weren’t), which was critical for ensuring all at-risk students were captured.
ROC Curve: Reached an Area Under Curve (AUC) score of 0.95.
Cross-Validation: Implemented 5-fold stratified cross-validation, resulting in a stable accuracy variance of ±0.012 across folds, proving the model generalizes well to unseen cohorts.

7. System Architecture

graph TD A[Raw Student Data] --> B[Data Imputation Pipeline] B --> C[Feature Engineering & Scaling] C --> D[Random Forest Model] subgraph Model Training and Evaluation D --> E[Stratified K-Fold CV] E --> F[Hyperparameter Optimization] F --> G[Saved Model .pkl] end subgraph Inference System G --> H[FastAPI Endpoint] H --> I[Admin Dashboard Prediction] end

8. Key Learnings

Implementing a robust baseline model early saved weeks of debugging later. I learned the critical importance of preventing data leakage in the context of academic data, specifically ensuring that temporal splits (training on older years, testing on the newest year) were used instead of random splits to simulate true real-world deployment.

9. GitHub Repository

Source Code: Placement Predictor ML

10. Future Improvements

Integrate NLP analysis on student resumes to extract additional features.
Migrate the inference pipeline from FastAPI to a fully containerized Kubernetes deployment to handle large concurrent query bursts during peak recruitment seasons.