Machine Learning @ Carousell
2 likes772 views
Machine learning is being used at Carousell in several ways to improve the user experience. Category and title suggestions using convolutional neural networks help users list items faster and improve listing quality. Chat reply suggestions apply dot product ranking and recurrent neural networks to previous messages to suggest relevant responses. These models are trained on Google Cloud and deployed using TensorFlow and Kubernetes to provide fast, efficient predictions at scale.
Machine Learning @ Carousell
- 1. Machine Learning at Carousell Shannon Chan Data Scientist
- 2. THE CAROUSELL APP SNAP, LIST, SELL List your item for sale in 30 seconds IN-APP CHAT Chat directly with sellers without revealing personal information IMAGE-CENTRIC Main focus of the app on images SOCIAL FEATURES Share listings on social media channels & join groups of people with similar interests
- 3. 110 million listings 15+ Avg Mins on the app for active users 43 million Sold 19 Cities 7 Countries Largest and fastest growing mobile classifieds in the region
- 5. CLASSIFIEDS 2.0: DESKTOP WEB
- 6. CLASSIFIEDS 3.0: MOBILE-FIRST WHAT CAROUSELL PIONEERED
- 7. CLASSIFIEDS 4.0: AI-FIRST WHAT CAROUSELL IS BUILDING AND PIONEERING
- 8. Machine Learning for User Experience Buying Experience Chat reply, Recommender system, Search Selling Experience Category, Title suggestions Trust & Safety Bad actor, Spam detections
- 9. Category & Title Suggestions Buying Experience Chat reply, Recommender system, Search Selling Experience Category, Title suggestions Trust & Safety Bad actor, Spam detections
- 10. Category & Title Suggestions
- 11. Category & Title Suggestions
- 12. Title Suggestion: Ranking of whitelisted titles Samsung galaxy edge Red dress Forever 21 skirt PS3 Slim 500GB iPhone X Wine cooler iPhone case Toyota Wish Samsung galaxy edge Red dress Forever 21 skirt PS3 Slim 500GB iPhone X Wine cooler iPhone case Toyota Wish
- 13. Convolutional Neural Network (CNN)
- 14. Dot Product Ranking ImageTitle x y score = x . y N-gram embeddings CNN Image Title
- 15. Dot Product Ranking 0.1 0.9 0.2 ... 0.5 Image vector Forever 21 skirt vector iPhone X vector 0.1 0.9 0.2 ... 0.5 0.1 0.9 0.2 ... 0.5 Image . Forever 21 skirt = -0.7 Image . iPhone X = 0.9
- 16. Dot Product Ranking Image Suggested titles
- 17. Dot Product Ranking Whitelist of titles x scores = x . Y precom puted vectors Y Image CNN
- 18. Category Suggestion ImageTitle x y score = x . y N-gram embeddings CNN category prediction softmax loss
- 19. Visualizing the shared image / title space
- 20. Visualizing the shared image / title space
- 21. Vector Direction red direction France + (Berlin - Germany) = Paris capital direction
- 22. Vector Direction
- 23. Training and Deployment Write ML models using TensorFlow Train models using Google Cloud ML Engine Deploy models on Google Kubernetes Engine
- 24. Category & Title Suggestions Fast training iteration Trained with Google Cloud ML for 2 days No manual labelling User-written titles as labels Performance Train on tens of millions of images Efficient inference Single pass with one matrix operation
- 25. Category & Title Suggestions - Results ~10s faster in time to list on average Improved listing completion rate Less wrongly categorized listings
- 26. Chat Reply Suggestions Buying Experience Chat reply, Recommender system, Search Selling Experience Category, Title suggestions Trust & Safety Bad actor, Spam detections
- 27. General Chat Reply Model
- 29. Chat Reply Suggestions Whitelist of messages x scores = x . Y precom puted vectors Y Previous Messages N-gram embeddings Extension of Dot-Product model with Recurrent Neural Network RNN
- 30. Question Answer Model Whitelist of messages x scores = x . Y precom puted vectors Y Previous Messages - Sequence Examples - Bi-directional RNNs - Hash Embeddings - Attention layers N-gram embeddings RNN
- 31. References (1) Efficient Natural Language Response Suggestion for Smart Reply https://arxiv.org/abs/1705.00652 Authors: Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, Ray Kurzweil (2) Question-Answer Selection in User to User Marketplace Conversations https://arxiv.org/abs/1802.01766 Authors: Girish Kumar, Matthew Henderson, Shannon Chan, Hoang Nguyen, Lucas Ngoo
- 32. Future Directions Personalized Feed ; Pricing Suggestion ; Caroulens ; Carousell Assistant ;
- 33. Were hiring! (1) Data Scientist , SG (Ads Ranking) (2) Business Intelligence Intern , SG Check out careers.carousell.com