"Are You Sure?": blending product comparisons and recommendations with A.I.
summary
Pipeline generates comparison-as-recommendations by first building dense product vectors from session co‑views and KNN retrieval, then refines candidates with a binary neural classifier trained on unsupervised co‑view/co‑purchase labels, cleaned by image-vector cleaning and augmented via substitute‑graph sampling. Properties are chosen by query frequency, page mentions and entropy; products sample three price bins (same, lower, higher). A/B‑style experiments on two shops show image‑based cleaning raises precision ~7–10 points on the PR curve; agreement with humans is higher for popular, non‑technical categories.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:21Why is this important
- 3:13Recommendation systems
- 4:51Challenges
- 6:53Pipeline overview
- 8:24Fast retrieval
- 9:59Machine learning
- 13:22Visual similarity
- 14:36Data augmentation
- 15:59Representation
- 16:42Model
- 17:25Data
- 18:05Property importance
- 21:00Experiments
- 24:13User study
- 25:51Real examples
- 26:30Next steps
- 27:15AI Labs
- 28:02Questions
related talks
25mProduct Management and AI
31mAI Search for Marketplaces: Experiences & Tech
35mNatural and conversational shopping experiences
16mSolve for AI
24mAI: Redefining Customer Experience
26mAI and ML for Product Management
41mAI is here – time to throw away our search engines?
40mAI or the Metaverse - What should brands bet on?