Oracle High Performance Computing Contest - Fall 2020

Author: Nicolas Racchi, BSc Automation Engineering student at Polimi

This repository contains all the content for my submission to the Graph ML contest.

Installing and running my submission

The easiest way to run my submission is by using the run.py script I prepared, which will run the final submission and output the testing cases predictions.

NOTE: The dataset is pulled from the original repository of the contest, so you don't have to install anything manually. Refer to (utils/load_data) for more info.

# 1. Clone the repo
git clone https://github.com/nicolas-racchi/hpc2020-graphML

# 2. cd in the project folder
cd hpc2020-graphML

# 3. (Optional): activate your virtualenv
virtualenv venv && source venv/bin/activate

# 4. install dependencies
pip install -r requirements.txt

# 5. run the main script
python run.py

Solution overview

The final pipeline is composed of:

1. Loading data
2. Preprocessing data
3. An embedding model: 2 HinSAGE layers that produce node embeddings (32-dimensional tensors), trained with Deep Graph Infomax in a semi-supervised way. This must be repeated for each node type (5 times).
4. A Decision Tree classifier that takes as input the node embeddings and outputs the predicted Case ID.

Please refer to my written report for all the implementation specific details.

Predictions CSV

You can also find the ready-made predictions in output/full_predictions.csv. Please note that, as stated in my report, I predicted the extended case only for the nodes of type Account, Customer, and Derived Entity, which together make more than 80% of all cases.

External Libraries

The main libraries leveraged in this solutions are:

• Stellargraph 1.2.1
• Tensorflow 2.3.1
• Pandas 1.1.5
• Numpy 1.18.5

You can find the complete list of the dependencies as well as their version in requirements.txt.

Please note that some libraries were only used in the development phase and are not relevant to the final solution but might be only used in some notebooks. The ones mentioned above are the ones that I used the most.