Just in time for #JSM2022, our new paper has been published in the Journal of Machine Learning Research! jmlr.org/papers/v23/21-… Elena A. Erosheva and I propose the first joint statistical model for rankings and scores (1/n)
Rankings and scores are two common types of preference data, which occur in contexts like voting, polling data, recommender systems, and peer review. Because it’s difficult to combine ordinal rankings and cardinal scores, they’re almost always modeled separately (2/n)
But rankings and scores provide different, and complementary, information! Rankings make direct comparisons but are coarse. Scores are more granular but make only implicit (and often unreliable) comparisons (3/n)
What if we could model rankings and scores together? Past researchers suggested this could be more reliable and improve inference on preferences, but no models existed. Instead, they had to convert or ignore some of the data (4/n)
The Mallows-Binomial fills this gap. The model has shared parameters between a Mallows ranking model and Binomial rating models to learn group preferences, with uncertainty! (5/n)
The model works in many realistic settings: Partial rankings, missing scores, or individuals who express inconsistent preferences between rankings and scores (6/n)
In our paper, we apply the model to real #peerreview data and show how it efficiently combines information from rankings and scores. We think the model can help make better decisions and understand the inherent uncertainty in them (7/n)
To try out the Mallows-Binomial model for yourself, check out our R package “rankrate” on CRAN. Thanks for reading! (8/8)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
