askarnik / project_madelon

Project 3: Feature Selection + Classification

Domain and Data

You're working as a data scientist with a research firm. You're firm is bidding on a big project that will involve working with thousands or possibly tens of thousands of features. You know it will be impossible to use conventional feature selection techniques. You propose that a way to win the contract is to demonstrate a capacity to identify relevant features using machine learning. Your boss says, "Great idea. Write it up." You figure that working with a synthetic dataset such as Madelon is an excellent way to demonstrate your abilities.

Requirement

This work must be done on AWS.

Problem Statement

Your challenge here is to develop a series of models for two purposes:

1. for the purposes of identifying relevant features.
2. for the purposes of generating predictions from the model.

Solution Statement

Your final product will consist of:

1. A prepared report
2. A series of Jupyter notebooks to be used to control your pipelines

Tasks

Data Manipulation

You should do substantive work on at least six subsets of the data.

• 3 sets of 10% of the data from the UCI Madelon set
• 3 sets of 10% of the data from the Madelon set made available by your instructors

Prepared Report

Your report should:

1. be a pdf
2. include EDA of each subset
   • EDA needs may be different depending upon subset or your approach to a solution
3. present results from Step 1: Benchmarking
4. present results from Step 2: Identify Salient Features
5. present results from Step 3: Feature Importances
6. present results from Step 4: Build Model

Jupyter Notebook, EDA

• perform EDA on each set as you see necessary

Jupyter Notebook, Step 1 - Benchmarking

• build pipeline to perform a naive fit for each of the base model classes:
  • logistic regression
  • decision tree
  • k nearest neighbors
  • support vector classifier
• in order to do this, you will need to set a high C value in order to perform minimal regularization, in the case of logistic regression and support vector classifier.

Jupyter Notebook, Step 2 - Identify Features

• Build feature selection pipelines using at least three different techniques
• NOTE: these pipelines are being used for feature selection not prediction
• Use the results from step 2 to discuss feature importance in the dataset

Jupyter Notebook, Step 3 - Testing Model Pipelines

• Considering these results, develop a strategy for building a final predictive model
• recommended approaches:
  • Use feature selection to reduce the dataset to a manageable size then use conventional methods
  • Use dimension reduction to reduce the dataset to a manageable size then use conventional methods
  • Use an iterative model training method to use the entire dataset
• This notebook should be a "playground" where you try various approaches to solving this problem

Jupyter Notebook, Step 4 - Build Model

• Implement your final model
• (Optionally) use the entire data set

---

Requirements

• Many Jupyter Notebooks
• A written report of your findings that detail the accuracy and assumptions of your model.

---

Suggestions

• Document everything.