Panel: Causality in search and recommendation systems

Опубликовано 25 января 2022, 1:29

Moderator: Emre Kiciman, Senior Principal Researcher, Microsoft Research Redmond
Panelists:
Amit Sharma, Senior Researcher, Microsoft Research India
Dean Eckles, Associate Professor, Massachusetts Institute of Technology
Eren Manavoglu, Vice President, Search Ads, Microsoft
Yixin Wang, LSA Collegiate Fellow, University of Michigan

With the scale of search and recommendation, real-time robust and explainable decision-making is at the heart of search and recommendation systems that work robustly even as the user-base changes, new content appears, and topics rise and fall in popularity. These changes can lead to brittle models that fail to capture knowledge about the domain, for example, how people search or which trends are expected to be related based on external knowledge. In addition, since user feedback is only available for items that were shown by the current system, the training data and labels are skewed towards items that were recommended previously, leading to unintended feedback loops. In this talk, we discuss advances in causality that increase model robustness against anchoring too strongly to past observational data. We also speculate on the future of more robust and explainable models as we advance learning methods that go beyond correlation to causation. 

Learn more about the 2021 Microsoft Research Summit: Aka.ms/researchsummit