Overcoming position and presentation biases in search and recommender systems

Roman Grebennikov, CTO, Metarank Labs
People's behavior is full of implicit biases. We click on first items because they're first and not because they're relevant: Google has trained us to avoid scrolling. We prefer popular things because they're popular, thus making them even more popular.
In ML tasks, taking these biases into account is a key way to improve the quality of your model. In this talk we'll go over the most typical implicit biases in the data, and discuss different approaches to overcome it and make your model more stable. We will also do a live bias-removal demo with Metarank on an open movielens-based dataset.
Get in touch with us
Join our Slack community: toloka.ai/comm...
Check out more of our events: toloka.ai/events
Read our Medium: / toloka
Follow us on social networks to make sure you won't miss any updates.
Twitter: / tolokaai
Facebook: / globaltoloka
Linkedin: / toloka