Machine Learning on Graphs

This project focuses on developing machine learning techniques and its applications over graph-structured data across multiple domains like cybersecurity, compliance, healthcare and recommenders.

Project Details

Machine Learning on Graphs

Machine Learning on Graphs

This project focuses on developing machine learning techniques and its applications over graph-structured data across multiple domains like cybersecurity, compliance, healthcare and recommenders.

Project Overview

Graphs are trending - in academia and industry - due to their ability to model latent relationships as first-class citizens and capture linked information (i.e., entity relationships) that other data models fail to capture. While at the same time, machine learning techniques are continuously being improved upon across multiple domains to uncover “hidden insights” through training from patterns and trends across enormous volume of data. The recent advances in machine learning over graph-structured data has enabled new insights into the latent pattern learning over available data and extract more complex patterns (e.g., multi-hop relations or correlations across neighbors’ attributes). This enhanced capability of machine learning on graphs is becoming increasing popular for discovering relationships (e.g., link prediction for recommender systems), classifying information (e.g., drug categorization), identifying complex anomalies in data (e.g., fraud detection) and many more such applications. 

Most of these capabilities are attributed to research advancements to autonomously learn vector representations of graph-structured data in supervised, unsupervised or semi-supervised manner and thereupon its application for the desired objectives ranging across different verticals, namely, cybersecurity, healthcare, retail. For instance, cybersecurity is a growing concern in any cloud provider and threat detection is increasingly challenging due to heterogeneity in collected data from underlying applications, for e.g., logs. Graph data model enables to connect these heterogeneous data elements and thereupon detect anomalous patterns by leveraging the above-mentioned graph machine learning techniques on such heterogenous graphs. Recommender systems is yet another application domain where graph machine learning stands out due to its ability to correlate user features (e.g., demography, gender, age, budget,…), item features (e.g., price, product hierarchy, category,…) as well as user-item interactions (e.g., preference, dislike, share, tweet,…).

Principal Investigator

Damien Hilloulin

Research Manager

Damien Hilloulin is a Senior Member of Technical Staff at Oracle Labs Zurich.

He joined Oracle Labs in 2016 after completing his master degree in Computer Science at EPFL (Swiss Federal Institute of Technology in Lausanne).

Damien is focusing his work on PGX (Parallel Graph AnalytiX), mostly on its single machine runtime (PGX.SM). There he has been working on several topics, including the partitioned graph support (data model, internal representation, memory optimizations, query and algorithm optimizations) and Graph Machine learning (integration of GraphML models and addition of supporting data-structures such as PgxFrame). He is currently focusing on improving the graph mutability support in PGX.SM to be able to update graphs with billions of vertices and edges at high throughput and low latency while keeping memory consumption at bay.