Problem
Zillow's automated valuation model (AVM), the Zestimate, uses structured public and MLS data such as bedrooms, bathrooms and square footage. Price-relevant attributes like a private backyard, a paver stone patio or a newer roof often appear only in the free-text listing description. Raw text is hard to use: it also holds contact details and showing times, subjective words ("charming"), ambiguous words ("pool" can mean a pool table) and aspirational claims. The valuation team also needs the new inputs to stay explainable.
Example from the article: a description with "Newer Roof", "Trex Deck" and "Paver Stone Patio" holds facts that county and MLS records lack.
Solution in brief
Embed the phrases that an existing tagging product (Home Insights) extracts from each description with a fastText model fine-tuned on listing text. Match each phrase to its nearest term in a real estate taxonomy with k-nearest neighbors. A binary feature (for example, has a swimming pool) is set only if the similarity passes a tuned threshold. A backtest against production then checks the features.
Questions this case answers
- How to add features mined from free text to a tabular prediction model while keeping each feature explainable and avoiding subjective or ambiguous phrases
- How to turn free-text item descriptions into reliable binary features for a valuation model?
- How to add text-derived features to a price model without losing explainability?
- How to choose a similarity threshold when mapping extracted phrases to feature categories?
Tools and methods
- Home Insights tags — phrases extracted from each description
- Real Estate Taxonomy — hierarchy of home attribute terms; chosen subtrees define the features
- fastText — embeddings of terms and tags, fine-tuned on listing text
- k-nearest neighbors — closest taxonomy term for each tag
- Similarity threshold — turns the match into a 0/1 feature
- Kubeflow — runs the Zestimate backtests
Google or a one-time email code. No password.