CRUP Fall 2025 Final Project

Project Overview

You will design and execute a complete Machine Learning project that answers a real-world question and present your findings through an interactive website. This project integrates both machine learning and software engineering skills.

Core Requirements

  • Formulate a specific, real-world problem that can be solved with Machine Learning
  • Build and train an ML model using a real dataset
  • Create an interactive website that explains your process and allows users to interact with your model

Part A: Machine Learning Component

1. Problem Formulation & Dataset Selection

Your task is to identify a clear, answerable question that requires machine learning to solve.

Dataset Requirements

  • Must come from a real dataset (e.g., Kaggle, Hugging Face)
  • Your project must be one of the following:
    • Classification (e.g., “Will this customer churn?”)
    • Regression (e.g., “What will this house sell for?”)

2. Complete ML Pipeline Implementation

You must implement all stages of a professional ML workflow.

a) Data Acquisition and Cleaning

  • Download and load your dataset
  • Perform Exploratory Data Analysis (EDA) with visualizations and summary statistics
  • Handle missing values (remove, impute, etc.)
  • Encode categorical variables (e.g., one-hot encoding or label encoding)

b) Model Training and Hyperparameter Tuning

  • Try multiple models appropriate for your task
  • Train each model
  • Tune hyperparameters using Grid Search, Random Search, etc.

c) Rigorous Model Evaluation

For Classification

  • F1 Score
  • AUC
  • Accuracy (use carefully)
  • Confusion Matrix

For Regression

  • R-squared (R²)
  • MAE
  • RMSE

d) Artifact Preservation

  • Save your trained model (e.g., using pickle or torch.save)
  • You will load this model into your website

Part B: Software Engineering Component

1. Public-Facing Website

  • Built using React
  • Serves as documentation + interactive demonstration

2. Required Website Content (Documentation)

a) Central Problem & Real-World Impact

Explain:

  • What question you’re answering
  • Why it matters
  • Who benefits
  • What real-world decisions your model could influence

b) Data Source & Nature

Include:

  • Dataset link
  • What each row represents
  • Features included
  • Number of examples
  • Any limitations or biases

c) ML Methodology

Clarify:

  • Which algorithms you tried
  • Which you chose
  • Why you chose it
  • What hyperparameters you tuned

d) Final Performance Metrics

Report:

  • Your final evaluation metrics
  • A direct answer to your core question
  • Limitations + failure modes

3. Interactive Component (MANDATORY)

Your website must include at least one interactive ML-powered element.

Acceptable Options

  • Prediction Form (user enters input → model predicts)
  • Slider-Based Dynamic Prediction
  • Interactive Visualizations
  • Comparative Predictions (What-if Analysis)

Part C: Deadlines & Deliverables

📅 Deadlines

  • Research Proposal — Due: End of Thanksgiving Break
    • One paragraph
    • Includes central question, dataset, and approach
  • Final Project — Due: Before Banquet
    • Full ML pipeline
    • Fully functional website
    • Complete documentation

✅ Deliverables Checklist

  • Research Proposal
  • ML Solution
    • Dataset acquired & cleaned
    • Multiple models compared
    • Best model selected
    • Model evaluated
    • Model saved
  • Website Component
    • React website (public-facing)
    • Full documentation
    • At least one interactive component
    • Accessible, clear design

🌟 Exemplary Projects for Inspiration

  • https://llm-attacks.org
  • https://thinkingmachines.ai/blog/modular-manifolds
  • Distill-style explorations:
    • Feature Visualization
    • Activation Atlas
    • Handwriting with Neural Networks
    • Building Blocks of Interpretability
  • https://distill.pub

Tips for Success

  • Choose a focused question
  • Select a manageable dataset
  • Document continuously
  • Build the interactive component early
  • Make explanations accessible to non-ML audiences

Good luck! 🚀