Ondemand Webinar

Explainability in Graph Learning 

A deep dive into how you can apply best-in-class explainability on state-of-the-art GNN models in production.

Watch Now

What Is It?

Graph Neural Networks (GNNs) have emerged as the most effective approach for machine learning (ML) on graphs. Graphs are useful for describing most real-world applications (including financial graphs, social networks, e-commerce and online marketplaces) where phenomena can be broken down into entities, their relationships, and their dynamic interactions.

When it comes to applying ML for these use cases in production, the ability to understand the reasoning behind the model’s decision-making and the impact on the output are essential for organizations. ML practitioners typically leverage explainability for:

  1. Building trust and transparency in the predictions
  2. Debugging and improving overall performance and quality of the model
  3. Identifying key drivers and important attributes in the data
  4. Meeting regulatory requirements across any industry.
While explainability and interpretability are well-understood and highly effective in traditional supervised ML, these concepts are more challenging and less common when applied to GNNs. In this webinar, we’ll deep dive into how you can apply best-in-class explainability on state-of-the-art GNN models in production. We’ll cover best practices and leading techniques to make the most of GNN model outputs, we’ll demonstrate the latest PyTorch Geometric improvements with the release of a new explainability framework fully integrated with Captum, and finally, we’ll show you how you can leverage Kumo to explore predictions, generate fine-grained column contribution statistics, and deep-dive into entity-level and class-level explainability.

 

Speakers

TinYun_Ho

Tin-Yun Ho

VP, Product

Kumo AI

Prior to Kumo, Tin-Yun was at Google Cloud, where he was a founding PM and co-lead on Vertex, Google Cloud's flagship ML platform. He also founded and led multiple enterprise-facing AI offerings for automated modeling on structured data (incl. AutoML Tables, AutoML Forecasting), personalization (incl. Recommendations AI), and optimization (incl. Cloud Fleet Routing). He also previously covered Cloud AI's video and edge offerings. Prior to that, he was the first student to graduate from Stanford with a joint CS and MBA degree (AI-focused), and worked at McKinsey and the Bill & Melinda Gates Foundation.

RexYing

Rex Ying

Engineer

Kumo AI

Rex is an assistant professor in the Department of Computer Science at Yale University. His research focus includes algorithms for graph neural networks, geometric embeddings, and trustworthy ML on graphs. He is the author of many widely used GNN algorithms such as GraphSAGE, PinSAGE and GNNExplainer. Rex works on a variety of applications of graph learning in physical simulations, social networks, NLP, knowledge graphs and biology. He developed the first billion-scale graph embedding services at Pinterest, and the graph-based anomaly detection algorithm at Amazon. He is the winner of the dissertation award at KDD 2022. Rex served as an area chair for Learning on Graph Conference and the PhD Consortium Chair for KDD Conference, and has organized of 4 workshops on graph learning and simulation in ICML, NeurIPS, ICLR and KDD.

Kumo, 2023