Machine Learning over Small Data: generating new variables with predictive power
summary
A demand model is built from 72 restaurants’ weekly sales by enriching coordinates with Google Places to create 336 generated variables across 21 categories and 16 radius/nearest schemes, plus a custom economic concentration index that encodes density and distance. Dimensionality is reduced with backward selection (and options like Ridge/Lasso, PCA, CV explored) and a linear regression achieves R² ≈ 72% and RMSE 382 on mean 2,800 weekly transactions, picking the better location about 9/10 times.
AI-written summary · about
chapters (12)
- 0:00Introduction and speaker
- 2:09Why creativity matters in data science
- 4:21Project background with small data
- 6:32Generating new variables from maps
- 10:58Formulating hypotheses and roadmap
- 13:26Designing the economic concentration index
- 22:02Testing and refining variables
- 27:49Reducing dimensionality and model evaluation
- 31:21Final model and results
- 33:53Delivering the heat map solution
- 34:14Six key takeaways
- 36:25Q&A session
related talks
10mPredicting Store Cannibalisation using ML
33mMachine Learning explained to humans
10mIs Bigger Really Better? Predictive Analytics with Fine-grained Behavior Data
22mFrontiers of Machine Learning from the Lens of Retail
56mPredictive Analytics for Practitioners
01h 25mUnlocking the Predictive Power of Relational Data with Automated Feature Engineering
22mPredict the future (or don't) with machine learning in applications
28mUse case in demand forecasting