Explore the world of recommendation systems using Graph Neural Networks (GNNs) in this 26-minute video from Databricks. Dive into RECKON (RECommendation systems using KnOwledge Networks), a machine learning project designed to enhance entity intelligence. Learn how to represent site interactions as a heterogeneous graph, with nodes representing various entities and edges depicting interactions between them. Discover the GNN-based encoder-decoder architecture used in RECKON to learn entity representations by leveraging individual features and interactions through graph convolutions. Gain insights into personalized recommendation techniques and follow an end-to-end product building process on Databricks. Understand key concepts such as Neural Message Passing, neighborhood-aware embeddings, and the RECKON graph schema. Explore solutions for newsletter recommendations, link prediction, and the cold start problem. Get a comprehensive overview of the RECKON pipeline, model architecture, and scoring process, as presented at the DATA+AI SUMMIT 2022.
Overview
Syllabus
Intro
Graph ML is a branch of ML that deals with graph data
Neural Message Passing Goal is to obtain "neighborhood-aware" embeddings
Message Function
Aggregate/Reduce Function
Update Function
Single GNN Layer
RECKON - Graph Schema
Newsletter Recommendations
Link Prediction
RECKON - Cold Start Problem
RECKON - Pipeline
Model Architecture
RECKON - Scoring
DATA+AI SUMMIT 2022
Taught by
Databricks
