Sidharth Kumar of Flipkart at Machine Learning Developers Summit 2019
summary
A genetic-algorithm feature selector augmented with a multi-dimensional covariance map and an SVM regressor picks a tiny subset (3–4) from thousands of photometric features, replacing expensive spectroscopy and achieving 99% accuracy on star/galaxy classification and accurate quasar distance regression. The method converges ~100x faster than a vanilla GA (from ~5,000 to ~50 generations) by Bayesian-updating multi-feature covariances; experiments run on ~100k objects, ~3k features, parallelized across ~20 cores in about an hour.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 3:23The research problem
- 4:58Data and methodology
- 11:18Feature selection approaches
- 16:50Deep dive into SVM and GA
- 26:15Genetic algorithm mechanics
- 35:12Covariance map approach
- 40:09Results and Q&A
related talks
36mTwinkle twinkle little star, how I wonder what you are
37mCounting the Stars
21mAnalytics & Science for Customer Experience in E-commerce
38mWhy giving your algorithm ALL THE FEATURES does not always work
29mHow GenAI is transforming e-commerce catalogues
41mExploring Dark Energy & Star Formation w/ ML
30mAI in Retail - Advancing Promo Demand Forecasting
42mAn Algorithm of Style