Contribute Media
A thank you to everyone who has made this possible: Read More

The Dark Art of Search Relevancy

Summary

Building a search engine is a dark art that is made even more difficult by the nebulous ever-changing concept of search relevancy. When, and to what degree, is a result deemed to be relevant for a given search term? In this talk I will describe how we built a Lyst search relevancy data set using heuristics, crowd-sourcing and Xbox Live matchmaking.

Full details — http://london.pydata.org/schedule/presentation/1/

Description

Search is a hard area to work in. Techniques are not made public due to their value and little academic work is done in the area. Furthermore, Google has made the exceptional an everyday experience so the bar for success is very high from the outset.

Search data sets are also hard to create due to the nebulous ever-changing concept of search relevancy. When, and to what degree, is a result deemed to be relevant for a given search term? The ElasticSearch documentation states it well: " Search relevancy tuning is a rabbit hole that you can easily fall into and never emerge".

In this presentation I'll give a introduction to building a search relevancy data set with python using crowd-sourcing and the Trueskill algorithm from Microsoft. Trueskill is used for matchmaking on XBox Live and it allows us to transform moderated pairwise comparisons into rankings. The rankings can then be used to learn what results best match a given search phrase. I'll briefly cover how we're modeling the moderated rankings at Lyst using deep learning.

References

M. Hadi Kiapour, Kota Yamaguchi, Alexander C. Berg, Tamara L. Berg. Hipster Wars: Discovering Elements of Fashion Styles (2014).

Yelong Shen, Xiaodong He, Jianfeng Gao, Li Deng, and Grégoire Mesnil. Learning semantic representations using convolutional neural networks for web search (2014).

Ralf Herbrich, Tom Minka, and Thore Graepel. TrueSkill(TM): A Bayesian Skill Rating System (2007).

Details

Improve this page