2020-Improving Deep Learning For Airbnb Search

Table of Contents

1. Problem Definition
2. Methods
3. TODO & Questions & Further Reading

Problem Definition

1. capture the majoy improvement after launching the DNN model.
2. Evaluation method
   1. NDCG
   2. increasement NDCG (especially for cold start and position bias)
3. Several ideas
   1. But the highlight of our journey is the realization that to push the boundaries of our DNNs, the inspiration was not going to come from some external source. For that we had to follow the lead of our users.
   2. Using NDCG to quantify the position of the booked listings in search results has been the most reliable gauge of model performance for us.
   3. the price features are easy to learn, while locaton, quality are harder to learner. So the initial versions of model will be biased toward price features

Methods

1. Focusing on three parts:
   1. architecture -> two-tower architecture with pairwise loss
      1. impact 0.6% in booking, -2.3% in average price, 0.75% in revenue; -33% in 99pertentile latency
   2. Cold start
      1. predict engagement features VS baseline (statistics)
      2. impact: In online A/B experiment, we observed an improvement of +14% in bookings of newly created listings
   3. position bias
      1. Methods
         1. position as a feature; set it to 0 while doing inference
         2. use dropout (dropout rate 0.15) for position feature while training, balancing the relevance prediction and position bias prediction while decding the dropout rate
      2. impact:
         1. 0.7% bookings; 1.8% revenue

TODO & Questions & Further Reading

[x] more papers from airbnb

• Previous
  2017-Search Ranking and Personalization at Airbnb
• Next
  2019-Neural Product Retrieval at Walmart.com