Personalizing search results in real-time
You type a query in a search box and get great search results back. Sounds pretty simple, right? But the same search results ranking can be great for you, but not for someone else.
With widely known Learn-to-Rank process, you need to deduce which search item was relevant for a user in the past to get the better ranking in the future. But relevancy does not depend only on the query itself, it’s also dependant on the current context. For example, for the same “shoes” query:
- during the winter, slippers are not that popular;
- people of different genders will probably prefer different models;
- if someone landed on a specific product from google search before, he may be looking for a specific brand.
In this talk Roman will share a lessons learned experience on building personalized search platform for thousands of merchants at Findify:
- Does better search relevancy improve sales? (spoiler: yes)
- Why does real-time reaction matter?
- Fighting position bias while training the ML models, when people are clicking on the first results because they are just first.
- Solving cold-start problem for fresh merchants with transfer learning.
Presented at MICES 2019, 19 June 2019