SLIDE 1
Applying Deep Learning To Airbnb Search
MALAY HALDAR / April 2019 / QCon
Applying Deep Learning To Airbnb Search Unique Challenges High - - PowerPoint PPT Presentation
Applying Deep Learning To Airbnb Search Unique Challenges High - - PowerPoint PPT Presentation
MALAY HALDAR / April 2019 / QCon Applying Deep Learning To Airbnb Search Unique Challenges High Value Transactions Sparse Listing Data Sparse User Data Users are picky when spending Each listing can be booked at Travel is not a daily use
SLIDE 2
SLIDE 3
SLIDE 4
Unique Challenges
High Value Transactions Sparse Listing Data Sparse User Data Each listing can be booked at most 365 times a year. Travel is not a daily use case. Users are picky when spending hundreds of dollars.
SLIDE 5
How Does Airbnb Search Work?
SLIDE 6
SLIDE 7
Evolution Of Search Ranking At Airbnb
2014 2015 2016 2017 2018 2019 Scoring Functions GBDT DNN
SLIDE 8
SLIDE 9
First Neural Net : April 2017
{ Continuous features with heavy feature engineering } 31 Relus L2 loss
SLIDE 10
Key Learnings
- Keep the first launch as simple as
possible
- Manage expectations, aim for
neutral
SLIDE 11
Second Neural Net: Lambdarank NN June 2017
- sigmoid_cross_entropy_with_logits X lambdarank_weights
{ Booked listing features } { Unbooked listing features }
SLIDE 12
Key Learnings
- Pairwise preference
- Tensorflow
SLIDE 13
Third Neural Net: Hybrid Mar 2018
SLIDE 14
DNN : June 2018
{ Raw query and listing stats } 127 Relus 83 Relus
SLIDE 15
SLIDE 16
Key Learnings
- Transitioning to deep learning is
about scaling the system.
- Complexity is easy to build,
simplicity is hard.
SLIDE 17
Gains
SLIDE 18
Problem Formulation
SLIDE 19
Pointwise
SLIDE 20
Pairwise
SLIDE 21
Feature Engineering
Do Not Cook Your Features
SLIDE 22
Continuous Features
SLIDE 23
Categorical Features
Embedding Categorical Feature (0x3fc90)
- 1.203
- 1.862
1.701
- 2.003
0.192
:
SLIDE 24
System Engineering
- I/O
- I/O
- I/O
SLIDE 25
SLIDE 26
Feature Importance
...This problem is easily solved by showing a rug (indicators for data points on the x-axis) or a histogram. The assumption of independence is the biggest issue with PD plots. It is assumed that the feature(s) for which the partial dependence is computed are not correlated with other features. For example, suppose you want to predict ...
https://christophm.github.io/interpretable-ml-book/pdp.html
SLIDE 27
TopBot Analysis
SLIDE 28
Hyperparameter Tuning Optional, not a necessity.
SLIDE 29
SLIDE 30
Many more interesting topics...
Failed models!
SLIDE 31
Applying Deep Learning To Airbnb Search
Malay Haldar, Mustafa Abdool, Prashant Ramanathan, Tao Xu, Shulin Yang, Huizhong Duan, Qing Zhang, Nick Barrow-Williams, Bradley C. Turnbull, Brendan M. Collins, Thomas Legrand
https://arxiv.org/abs/1810.09591