# Fake Data Hackathon

Fake data is every where, but instead of resiting it we will embrace it, use it to our own advantage.
You can use it however you feel like. Only restriction is you use it.


### Problem Statement : 
L&T Finance provides vehicles loans to its customers . However they do want to find out which customers are trustworthy and will not default. You are asked to create a customer default model (binary classification) to predict which customer will default or not.
The dataset and competition details can be found [in this Kaggle competition](https://www.kaggle.com/c/fake-data-hackathonv2/overview)

### Data : 
You have the following datasets:
train.csv - Training data
test.csv - Test data
data_dictionary - Dictionary containing the explanations of various data columns. 

As this is fake data challenge, you are required to use fake/generated data in the model creation process. No feature generation does not count as fake/generated data !

You can use the [CTGAN library](https://github.com/sdv-dev/CTGAN) or any other library of your choice for generating fake tabular data. You should use at-least X rows of fake data while training your models.

Some ideas for including fake data include : 

1. Augment your training data with generated data.

2. Create stacked classifier with fake/ generated data.

3. Use generated data in your test set as [mentioned in this presentation](https://gitlab.com/ing_rpaa/hackathons/fake-data-hackathon/-/blob/master/resources/model_uncertainty_slides.pdf)

4. [Data augmentation using GAN to improve Risk Models] (https://confengine.com/odsc-india-2019/proposal/11560/data-augmentation-using-gan-to-improve-risk-models-for-new-credit-card-customers) by American express. They also [published a paper](https://arxiv.org/pdf/2002.02271.pdf) about it.

5. [Improve class imbalance using GAN](https://www.researchgate.net/publication/334409523_Generative_adversarial_fusion_network_for_class_imbalance_credit_scoring)

Let your creative juices flow.

### Environment

The datasets are small enough to use on your laptops, however if you want you can use the RPAA GCP environment as well. 

**Please shut-down the instance after use to avoid high costs**

### Evaluation Metric : 
Log loss will be used as the evaluation metric . You can read more about it [here](http://wiki.fast.ai/index.php/Log_Loss)

### Rules : 

- You need to use atleast 3000 rows of generated data in your model training process.
- The winner will not only be the one with the best score, but also the one with the most innovative approach to the problem. This will be decided during the presentation.
- 20 submissions max.
- 2-3 people max per team
- deadline 11th Sept 2020, 16:00 CEST

### Submission : 
You need to create a `submission.csv` file with the `uniqueid`'s from `test.csv`, and add a column `loan_default` which as the probabilities.
Here is the [sample notebook](https://gitlab.com/ing_rpaa/hackathons/fake-data-hackathon/notebooks/) to create baseline model and submit the results.

## Heroes

> Heroes: Dan, Tingting, Anil