Speakers:
• Filippo Minutella, Chapter Lead of AI, LARUS Business Automation
• Valerio Piccioni, AI Engineer, LARUS Business Automation
Session type: Full Length Session
Abstract: Reproducibility – both in machine learning and data science – is an emerging theme because you need to repeatedly run your algorithms on different features that can also be obtained with different graph projections to discover which one performs best on a particular dataset. Feature Store is the right tool for this objective. It can manage different versions of point-in-time features for both training and inference phases. We will present a full pipeline, starting with a graph on Neo4j and repeatedly transforming and loading it on Feast while also using Neo4j Graph Data Science to create new features to train a simple neural network. The talk will be organized with the first part on Feature Store and Neo4j's capabilities, and we will end with a notebook to present the full pipeline.