Data cleansing is a common task in model training. While in most cases, a person can validate the correctness of data reliablely, when it comes to domain specific knowledge, we would like to seek agreement among individuals.

This application provides a platform which allows user to clean data without considering the allcation and summarization of data cleansing tasks.

It provides function including:

• assign different tasks to users
• validate and update data based on responses from users
• generate new dataset for training
• generate working report for management purpose

1. Install necessary packages: pip install -r requirements.txt
   OR
   Activate virtual environment: source datacleansingenv/bin/activate

2. Install psycopg2 manually if the previous step fails to install it:
   For Mac:
   xcode-select --install
env LDFLAGS="-I/usr/local/opt/openssl/include -L/usr/local/opt/openssl/lib"
pip install psycopg2
   For Ubuntu:
   If psycopg2 cannot be installed, install psycopg2-binary instead:
   pip install psycopg2-binary

3. Install PostgresSQL For Mac:
   brew install postgresql
   For Ubuntu:
   sudo apt-get install postgresql
   Create User:
   sudo -u postgres createuser datacleansinguser
alter role datacleansinguser with superuser;
alter role datacleansinguser with password 'L1feI5T0ugh';
   Switch to the created user:
   set role datacleansinguser;
   Create database:
   create database datacleansing;

   If PostgresSQL version doesn't meet the requirement: brew postgresql-upgrade-database

   If you would like to use the other user or database:
   Approach settings.py - DATABASES. Set NAME, USER, PASSWORD as required.

4. Migrate: python manage.py migrate

5. To set up data for testing: run python set_up.py

6. To start Django server: run python manage.py runserver

7. To log into the website:

   Superuser account: superuser@gmail.com password: superuser

   Admin account: admin@gmail.com password: admin

   User account: alice@gmail.com password: alice

8. To generate test report:

   coverage run --source=. manage.py test

   coverage report [-m]

   If .xml file is needed, runcoverage xml

Superuser is a group of people who do NOT require domain specific knowledge. Their role is to manage different groups, accounts and dataset at higher level.

Superuser is expected to be the account that has the most permissions.

Details

• To create/delete group

• To approve/reject user account
• To activate/deactivate user account
• To assign the admin role

• To view/download dataset
• To update data

• To view/download working report
• To view admin log

Admin is a group of people who do NOT require domain specific knowledge. Their role is to manage accounts and dataset within their own group.

Details

• To approve/reject user account
• To activate/deactivate user account

• To view/download dataset
• To update data

• To view/download working report
• To view admin log

• To assign tasks
• To summarize and update individual report

User is a group of people who require domain specific knowledge. Their role is to finish any tasks allocated to them by admin. User should NOT have any permission regarding management.

Details

• To submit response for tasks allocated to the user
• To change password

Validating tasks require users to agree or disagree with question-answer pairs. The users should either approve a pair or provide the correct answer if they disapprove a pair.

The responses to these tasks will be collected each time a response is submitted. When responses from one side are confirmed to be greater than the other side, the data is considered to be cleaned and the majority will be the response for it.

If the majority considers the data as false, the data will be input to voting tasks, and all new answers submitted will be converted to its choices.

If the majority considers the data as true, the data will directly be finalized.

Voting tasks require users to select the best answer from at most 3 choices for a question.

The choice that has the highest votes will be considered as the answer to the question. Whenever a tie occurs, the task will be re-assigned to two more users, until the number of assignees is greater than 15. It will be transferred to Controversial Tasks after more than 15 users have voted for it.

Keyword selection tasks require users to select keywords for both questions and answers from a pair. The keyword does not have to be a single word, it can be any substring of the question and answer.

All responses submitted will be union together and considered as the keywords for a question-answer pair.

Image label validation tasks share the same logic as voting tasks, while instead of question-answer pairs, its target is image-label pairs.

Controversial tasks require users to select the best answer from at most 3 choices for a question.

The choice selected will be considered as the answer to the question.

Controversial tasks are assigned to a user by admins. Since the user's response can directly affect the data, the assigned user is expected to be an expert in the area.