A Kaggle competition | Build a recommender system based on real-world e-commerce sessions.
Keywords: Recommender System Graph Neural Network Heterogeneous Graph
Tools/Frameworks: Python Pandas Polars PyTorch PyTorch Geometric
Dataset: OTTO – Multi-Objective Recommender System
The goal of this competition is to predict e-commerce clicks, cart additions, and orders. You'll build a multi-objective recommender system based on previous events in a user session.
Your work will help improve the shopping experience for everyone involved. Customers will receive more tailored recommendations while online retailers may increase their sales.
In this project, I participated in the OTTO Kaggle competition to develop a Multi-Objective Recommender System using a Link Prediction Approach. The dataset used in the competition consisted of 12 million real-world user sessions, 220 million events, and 1.8 million unique articles in the catalog.
I took the initiative to study and work on this project independently, which allowed me to expand my knowledge and skills in the field of machine learning and data processing. I utilized the PyTorch framework and the PyTorch Geometric library to implement Graph Neural Networks. Furthermore, I optimized the code and model to run efficiently on a CUDA GPU and used Polars as the DataFrame library for maximum GPU and memory efficiency.
The main objective of the competition was to build a multi-objective link prediction on a large-scale e-commerce heterogeneous graph. I utilized graph neural networks and link prediction techniques to recommend the most relevant articles to users based on their preferences and past behaviors.
An illustration of the heterogeneous graph (left) which consists of multiple node and edge types and multi-objective recommendations (right) which takes prediction on multiple event types and items.
An illustration of the overall workflow.