MACHINE LEARNING & RESEARCH
Exploring what machine learning can do.
I build and study machine learning systems, with work spanning ad auctions, contextual bandits, and drug discovery.
01 / RESEARCH
Ideas, experiments, and systems
that make it into the real world.
Showing all 8 papers
A production system for setting bid floors in mobile ad auctions with heavy-tailed bids. The accompanying FloorGym simulator studies cascades that pair a primary floor with a safety floor and compares their revenue and fill-rate trade-offs.
An interface that helps domain experts understand and manage contextual bandits. A “value gain” metric explains the impact of individual components, while a qualitative user study explores how to make these systems accessible to non-experts.
Introduces OBSI, an algorithm that adds features to a batched contextual bandit only as evidence of their relevance grows. Synthetic experiments show improvements in regret, feature relevance and computational cost, helping keep irrelevant features out of decisions.
Thousands of experiments examine five embedding models on two drug-discovery knowledge graphs. Training choices, hyperparameters, random seeds and data splits can change both performance and model rankings, highlighting the need for reproducible, carefully reported comparisons.
A tool for cross-referencing disease ontologies and moving between their hierarchies to support knowledge graph data integration.
Combines tensor factorisation with knowledge graph embeddings to predict clinical outcomes for unseen gene–disease pairs. Adding learned gene representations improves prediction accuracy, with a joint neural-network and tensor model outperforming the evaluated baselines.
A guide to public biomedical data for building drug-discovery knowledge graphs. The review categorises data sources, compares existing graphs and examines practical challenges and future research directions for applying graph-based machine learning in this domain.
Explores how multiple imputation handles missing values, censoring and measurement error in early-stage drug discovery. Using simulated drug-discovery datasets, Python and Stan, the work compares imputation approaches and examines their effects on uncertainty estimates and investment decisions.
02 / EXPLORATIONS
A few things I’ve built
and learned along the way.
Fine-tuning Stable Diffusion to bring Simpsons characters to life from text prompts.
Read the write-up ↗MACHINE LEARNING ↗An implementation of the Deterministic Uncertainty Quantification model.
Explore the notebook ↗An experiment in cryptocurrency and smart contracts.
03 / JUST FOR FUN