Summary: Approximate classification via earthmover metrics
Aaron Archer # Jittat Fakcharoenphol + Chris Harrelson # Robert Krauthgamer §
Kunal Talwar ¶ ’
Eva Tardos #
Abstract
Given a metric space (X, d), a natural distance measure
on probability distributions over X is the earthmover
metric. We use randomized rounding of earthmover
metrics to devise new approximation algorithms for
two well­known classification problems, namely, metric
labeling and 0­extension.
Our first result is for the 0­extension problem. We
show that if the terminal metric is decomposable with
parameter # (e.g., planar metrics are decomposable
with # = O(1)), then the earthmover based linear
program (for 0­extension) can be rounded to within an
O(#) factor.
Our second result is an O(log n)­approximation for
metric labeling, using probabilistic tree embeddings in
a way very di#erent from the O(log k)­approximation

Source: Archer, Aaron - Algorithms and Optimization Group, AT&T Labs-Research

Collections: Computer Technologies and Information Sciences