Deep Learning at Booking.com
summary
Booking.com adapts convolutional nets to its hotel-image corpus and runs 200,000 images per day through custom multilabel classifiers, links image labels to personalization signals and validates impact with A/B tests. Recommendation models use a wide and deep architecture (logistic cross features plus learned embeddings) trained with TensorFlow and hyperparameter search. Production uses containerized Kubernetes + TensorFlow serving, scheduled retraining, model health checks, metadata tracking and inference optimizations like quantization to int8 and lookup tables.
AI-written summary · about
chapters (8)
- 0:00Deep learning intro
- 1:51When to use DL effectively
- 5:01Analyzing hotel images
- 8:28Personalizing with images
- 10:36Deep networks for hotel recommendations
- 13:23Wide and deep networks
- 22:33Scaling DL with Kubernetes
- 32:00Q&A
related talks
35mHow Booking.com serves Deep Learning Model Predictions
17mKeynote: Scaling Deep Learning Models in Production Using Kubernetes
17mLaying the foundation for deep learning
29mRanking hotel images using deep learning
40mMachine Learning
20mPragmatic Deep Learning for image labelling
44mMachine learning @ booking.com
37mDeep Recommender Systems