Disaster Response Pipeline Project

Project Overview

This project is meant to implement a complete machine learning pipeline to demonstrate Data Engineering skills which is part of Udacity Data Scientist Nanodegree.

In this project, real disaster data from Figure Eight is analyzed to build a model for an API that classifies disaster messages.
The used dataset contains real messages that were sent during disaster events.
A machine learning pipeline is implemented to categorize these messages so that later on, every message can be sent to an appropriate disaster relief agency.

The project includes a web app where an emergency worker can input a new message and get classification results in several categories.

Let's see it working

• To run the web app and try out the model prediction:

  • Run the following command in the app directory

  python run.py

  • Then Go to http://localhost:3001/

• To try out the ETL or ML pipelines, run the following commands in the project's root directory.

  • To run ETL pipeline that cleans data and stores it in database

  python data/process_data.py data/disaster_messages.csv data/disaster_categories.csv data/disaster_messages.db

  • To run ML pipeline that trains classifier and saves it

  python models/train_classifier.py data/disaster_messages.db models/classifier.pkl

Project Components

1. ETL (Extract, Transform and Load) Pipeline.
   Pipline is implemented in data/process_data.py script, where it:
   • Merges data/disaster_messages.csv and data/disaster_categories.csv datasets.
   • Extracts categories and transforms them to binary values.
   • Cleans data and removes duplicates.
   • Loads cleaned data into SQLite database data/disaster_messages.db.
2. ML (Machine Learning) Pipeline.
   Pipline is implemented in models/train_classifier.py script, where it:
   • Loads data from the SQLite database - resulted from ETL pipeline - into feature & target/label values.
   • Splits both features & labels into training and test sets.
   • Builds a text processing and machine learning pipeline that performs multi-output classification on the 36 categories in the dataset.
   • Trains and tunes the model using GridSearchCV.
   • Evaluates the resulted model by running prediction on the test set and displaying a classification report.
   • Exports the final model's best estimator as a pickle file models/classifier.pkl.
3. Flask Web App.
   The app implementaion exists inside app directory.
   • App layout is implemented in app/templates.
   • App logic is implemnted in app/run.py, where it:
     • Uses plotly to compose a couple of graphs that helps explore the dataset.
     • Loads the saved model from the pkl file and uses it to predict the categories of the message entered via the app.
4. Directory notebooks includes two jupyter notebooks that I used during implementing ETL & ML pipelines.

The project repo: https://github.com/merocode-com/Disaster-Response-Pipelines

The Model Performance

Here is the classification report on the final model.

And the accuracy score.

I had to use "10" which is a really low number of estimators for the RandomForestClassifier to reduce the file size of the final model so that Github let me upload it to the repo :)

Locally I trained the model setting n_estimators to "200", and results were greatly improved.

Here is the classification report on the final model.

And the accuracy score.