2023-03-09
This is Marco's daily open-notebook.
Today is 2023.03.09
Women in Data Science Geneva event 2023
Buidling a data practice from scratch for cheap
Christelle Marfaing, Head of Data, May
How to build your stack
- define and model your target
- Ingestion and processing frameworks
- technical monitoring
- data quality
- docs
- analysis tools
- build and model a road map
How to build a team
- decentralize
- data mesh
- don't decentralize for nothing or reorganize all the time.
- learn to let people go.
Graph neural networks
Nadya Chernyavskaya, Senior Reasearcher, CERN
What is a graph
- a set of nodes and edges with some features
- directed/undirected, connected/disconnected, sel-loops,
- they offer much more flexibility
What are they for ?
- node level: predict a property of a node
- edge, link level : predict links between two nodes
- subgraph level: detect if nodes form a community (for ex: diagnostics)
- graph level : molecule property prediction for example
How to implement and represent a graph neural network
- create adjacency matrix, or edge list, or adjacency list
- a graph can then be implemented into a feature matrix and adjacency matrix
- each node has to keep its own local information, take new information into account and update it's state based on the previous two.
impact ?
example Google maps:
- nodes: road segments
- edges : connectivity between road segments
- prediction: time of arrival
Conclusion
- When ? Graph neural networks is a very powerful tool for complex and sparse data
- What for ? They can be successfully applied to a variety of different tasks such as node classification, graph classfication, edge prediction, etc...
- How ? We learned how to convert graph into ML read representation and build GNN