Predictive modeling is a core factor in trendy programs, and powers capabilities equivalent to fraud detection, mortgage approvals, and suggestion programs. These programs sometimes function on structured, relational knowledge saved in enterprise databases, with rows, columns, and interlinked tables. Whereas pc imaginative and prescient and pure language processing have undergone a neural community revolution, the tabular knowledge layer underpinning predictive modeling nonetheless largely depends on handbook function engineering and task-specific fashions.
Relational deep studying proposes a brand new method. It treats databases as graphs and applies transformer-style consideration mechanisms immediately over structured relational knowledge. Researchers at the moment are constructing basis fashions for tabular knowledge that intention to generalize throughout predictive duties with out painstaking function engineering.
Jure Leskovec is a Professor of Pc Science at Stanford College and he beforehand served as Chief Scientist at Pinterest and was an investigator on the Chan Zuckerberg Biohub. Most just lately, he co-founded the machine studying startup, Kumo.AI.
On this episode, Jure joins Sean Falconer to debate the constraints of conventional predictive modeling, why structured enterprise knowledge requires its personal modality-specific neural architectures, how graph transformers generalize consideration to relational databases, and extra.