Ranking hotel images using deep learning
summary
Expedia builds a model that ranks hotel photos using transfer learning from VGG and trains a regressor in Keras with the TensorFlow backend, labeling 100,000 images via Mechanical Turk as ground truth out of a 10 million image corpus, achieving about 0.5 mean absolute error (≈12.5% on a 1–10 scale). Engineering work includes GPU/CPU pipelining optimizations and batching so scoring that corpus drops from a week to about a day. A/B testing runs on one affiliate partner; future work is per-customer personalization and tighter outlier filtering.
AI-written summary · about
chapters (12)
- 0:00Expedia and team intro
- 1:51Expedia Affiliate Network
- 3:17Why ranking hotel images matters
- 5:45Manual ranking challenges
- 7:18Deep learning as solution
- 8:00Dataset with Mechanical Turk
- 10:57Model architecture and transfer learning
- 13:07Training the model
- 14:35Model performance and errors
- 15:45Optimizing training speed
- 16:48Old vs new ranking results
- 18:14Q&A
related talks
33mDeep Learning at Booking.com
30mJourney Towards Deep Learning for Ranking
20mPragmatic Deep Learning for image labelling
29mFrom Trees to Transformers: Our Journey Towards Deep Learning for Ranking
39mInsights About Places with Deep Learning Computer Vision
38mLearning to Rank journey at GetYourGuide: Our Logbook #bbuzz
17mLaying the foundation for deep learning
35mHow Booking.com serves Deep Learning Model Predictions