Project Description

A project focused on building a robust, reproducible ML workflow to predict student depression risk, balancing model performance (Recall vs. Precision) with real-world impact.

Responsibilities

• Engineered a robust, end-to-end Scikit-learn workflow, including custom transformers and pipelines, to preprocess diverse data types (categorical, numerical, geographical).
• Implemented a custom callable `refit` strategy in `GridSearchCV` to select the best model by maximizing **Recall** first, then using F1-score as a tie-breaker, directly aligning with the business objective of minimizing False Negatives.
• Adopted an iterative development strategy, validating hypotheses with baseline models (Logistic Regression, Linear SVC) before engineering more complex, non-linear models (Random Forest, KNN, SVC).
• Successfully diagnosed and resolved a critical bug related to **Pandas index misalignment** within custom transformers during cross-validation, a deep dive into the internal mechanics of Scikit-learn and Pandas.
• Constructed a final `VotingClassifier` by strategically ensembling diverse base models (linear, tree-based, kernel-based) to leverage their different predictive strengths and improve robustness.
• Demonstrated a clear strategic choice by training and evaluating a suite of models, from a balanced **LinearSVC** (87% F1-score) to a high-recall specialist **SVC** (99.6% Recall) designed to never miss an at-risk student.
• Conducted in-depth EDA to identify key factors correlated with depression risk, such as `suicidal_thoughts`, `academic_pressure`, and `financial_stress`.

Related Links

• GitHub Repository

Technology