Relational Deep Learning Approach for Predictive Tasks

This document is a technical report that introduces Relational Deep Learning, a novel approach designed to learn from data spread across multiple relational tables without the need for manual feature engineering. The report outlines the challenges associated with traditional machine learning methods that typically require data to be aggregated into a single table, which can lead to loss of information and inefficiencies. The authors present a framework that views relational tables as a heterogeneous graph, allowing for the application of Message Passing Neural Networks to extract meaningful representations directly from relational data. Additionally, the report introduces RELBENCH, a benchmarking tool that facilitates research in this area by providing datasets and an implementation of the proposed method. The document details the four main steps of the model pipeline, including the specification of predictive tasks, extraction of entity-level features, learning of new node representations, and production of predictions. Overall, this work aims to broaden the applicability of graph machine learning techniques to various artificial intelligence use cases.