I will discuss work of Lanckriet-McFee and Dasarathy-Eriksson-Novak-Singh on clustering given partial ordinal data on similarities; it turns out that in this case one can still do a good job of building a heirarchical clustering on S (that is, embedding S in the leaf-set of a tree.) We discuss some questions about how one might go further, including: can we use these methods to efficiently evade calibration problems (“one rater’s “very similar” is another rater’s “somewhat similar”?) How do we choose what kind of target metric space to try to embed S in? What is the non-metric version of the Netflix problem and how might one attack it? Is it correct to talk about “compressed clustering”?
Please be warned that the talk will contain no new theorems or data, and that the speaker is wandering dangerously far from his own area of expertise.
Discovery Building, Orchard View Room