This talk is to get the audience excited about graph ML and hopefully inspire them to think of their data as a potential graph which can open doors for this new paradigm of ML. Graphs are complex, but then again so is most of the data now.
The talk will be supported by code snippets and will be using an open source recommendation dataset.
Outline:
1. Intro to Graph Machine Learning (6 minutes)
- Brief overview of traditional data vs. graph-based data and graph components
- Networks in real-life and relevance
- ML on graphs: recent advancements
2. Graph Neural Networks (GNNs) and DGL (6 minutes)
- What are GNNs and how do they work?
- Overview of the Deep Graph Library (DGL)
- Architecture for building recommendation models using GNNs
- GNN Model Architecture
3. Training a GNN model in DGL (15 minutes)
This section is meant to address the following key questions on training and serving models using DGL:
- How do we transform tabular data into meaningful nodes and edges?
- How do we prepare data for graph modeling?
- Can we split train / test data randomly as is convention?
- How do we represent categorical/numerical features?
- How do we train and make predictions on this model?
- (optional): How do we decide on a model architecture?
- How do we evaluate this model?
- How do we re-train this model?
- What are the data / scalability limitations when training GNNs?
We will also discuss certain common pitfalls and how to avoid them.
4. Q&A and Closing Remarks (3 minutes)
After the talk, the audience should have a good understanding about:
- How to structure their data as graph components and prepare it for modeling
- What graph machine learning is all about, what a GNN is and how do we train and serve GNNs in DGL?
- End-to-end ML tasks on graph data: best practices
