Prioritizing home attributes based on guest interest

Joy Jing, Jing Xia16 Feb 2023

Problem

A marketplace wants suppliers to add and highlight the attributes buyers care about. Airbnb collects 800+ home attributes, and what matters differs by home: a mountain cabin is judged on other things than an urban apartment. Hosts fill in the list incompletely, so advice must also know what the home already has. Counting keyword mentions per fine-grained home segment fails because segments multiply and most have little data.

Example from the article: for a cabin for six at an average daily price of $50 the system ranks hot tub, fire pit, lake view, mountain view, grill and kayak first. For an urban apartment it ranks parking, restaurants, grocery stores and subway stations.

Solution in brief

Extract attribute mentions from guest text with an NER model plus embedding matching to a fixed attribute list. Normalize mention frequencies across sources, then predict the expected frequency for each home from its characteristics instead of counting inside segments. Combine the importance with the estimated probability that the attribute is already present to choose advice for hosts.

Questions this case answers
  • How to derive item attribute importance from unstructured user text and rank it per item segment when fine-grained segments are too sparse to count directly
  • How to combine attribute importance with the probability that the attribute is already present to produce per-item advice for suppliers
  • How to find out which listing attributes guests care about from messages, reviews and support tickets?
  • How to rank attributes separately for each type of listing when most narrow segments have too little data?

Tools and methods

  • textCNN NER — extracts key phrases in five categories
  • Embedding-based entity mapping — maps each phrase to the nearest attribute label by cosine distance, with a confidence score
  • Frequency inference model — predicts the expected mention frequency of each attribute for a home from its characteristics
  • Bayesian inference model and WiDeText — estimate whether a home already has an attribute
Continue reading — it's free

Google or a one-time email code. No password.