Zillow Machine Learning Team
- Metadata
- Summary:
- Notes:
- Stock is down 30% and laid off 25% of their staff
- Their home value estimation models were not producing the results as expected
- They started using it for buying and flipping homes
- Zillow was going to buy homes and flip them for profit, but this relied on a very robust model that can accurately predict house prices in 6 months
- It also relied on a linear increase in house values which doesn't happen in short term
- Stability is a key assumption for statistical models, and that's not a given in complex systems like the housing market
- The training data's distributions have distributions
- Topology and Differential Geometry is important
- Distribution of distributions can be modeled as deformations in the inference space
- Why overfitting creates miraculous accuracy
- Models needs to learn about deformations by having training data from disconnected sources to the outcome
- To figure out the relationships between these features and the outcome requires experimentation and a chain of models
- [[Causal Inference]]
- Deep learning models
- Throw all the features at it and let it figure it out
- Large production model architectures are expensive and time consuming to set up
- Feature stores need to recommend features which could improve model accuracy
- Experiment management system need to be able to track different experiments
- Review is difficult
- Tips for Data Scientists
- When revenue starts getting booked against ML projects - model reliability is critical