DisasterResponse

Project Motivation
Data Set
Results
File Description
Libraries and Dependencies
Licensing, Authors, and Acknowledgements

Project Motivation

This project was done within the framework of the Data Science Nanodegree offered by Udacity. The goals of the project are:

To build an ETL pipeline extracting information from a dataset of text messages and loading it into an SQL database
To build a machine learning pipeline to predict to which categories each message belongs
To build a web application where new messages can be entered and classified and some graphics containing information about the data et are displayed.

Data Set

The data set contains text messages related to natural disasters. The file /data/disaster_messages.csv contains 26248 text messages in english (partly translated) as well as an original language. A second file /data/disaster_categories.csv contains binary labels for 36 categories (such as floods or storms). Each text message may belong to multiple categories.

Implementation and Results

The ETL pipeline consists of the following steps:

Extract the english messages from the messages data set
Extract the categories and their labels from the categories data set and create one column for each category
Create a table where each row has one text message and 36 binary labels
Drop duplicate rows
Drop categories with only one distinct value since we cannot train on a single label. This applies to the category column child_alone
Store the table in an SQL database

The ML pipeline consists of the following steps:

Load data from the SQL data base
Create train and test splits
Define a machine learning pipeline including:
- NLP pipeline
- Logistic regression classifier for each of the 36 categories
Train the model using GridsearchCV for tuning parameters and cross validation
Evaluate the model and print out scores.
Save the model using pickle. The ML pipeline (including training) runs in about two minutes on a 2014 MacBook Pro.

In the web app, the user can manually enter a text message that is then classified for each of the 36 categories. Further, there are three graphics displayed:

Distribution of the message genres
Distribution of the message categories
Correlations between different categories

Note: the data set did not include any examples for the category child_alone. While this category is included in the app, the predictions are not reliable for this category.

File Description and Usability

Here is the file structure of the project.

- app
| - templates
| |- master.html  # main page of web app
| |- go.html  # classification result page of web app
|- run.py  # Flask file that runs app

- data
|- DisasterResponse.db   # database with cleaned data
|- disaster_categories.csv  # data to process 
|- disaster_messages.csv  # data to process
|- process_data.py


- models
|- train_classifier.py
|- classifier.pkl  # saved model 


- ETL Pipeline Preparation.ipynb
- ML Pipeline Preparation.ipynb
- README.md

The data merged and cleaned by the ETL pipeline is stored in the database data/DisasterResponse.db. The ETL process can be replicated by running

python process_data.py disaster_messages.csv disaster_categories.csv DisasterResponse.db

from within the data folder. The ML pipeline includes an NLP processing part and a logistic regression classifier for each of the 36 message categories. It can be retrained and replicated by running

python train_classifier.py ../data/DisasterResponse.db classifier.pkl

from within the model folder.

To run the web app locally, run

python run.py

from within the app folder. The web address to access the app should be displayed in the terminal.

The two jupyter notebooks ETL Pipeline Preparation.ipynb and ML Pipeline Preparation.ipynb have been used to build the ETL and ML pipelines and can be used to experiment with changes in the pipelines.

Libraries and Dependencies

The notebook uses the following Python libraries:

Licensing, Authors, and Acknowledgements

The data set was provided by figure eight for the Udacity Data Science Nanodegree. A similar version of the data set (which is already preprocessed) is available here. Parts of the code used here was provided by Udacity.

larpal/DisasterResponse