Massively Parallel Entity Matching with Linear Classification in Low Dimensional Space.

In entity matching classification, we are given two sets R and S of objects where whether r and s form a match is known for each pair (r, s) in R x S. If R and S are subsets of domains D(R) and D(S) respectively, the goal is to discover a classifier function f: D(R) x D(S) -u003e {0, 1} from a certain class satisfying the property that, for every (r, s) in R x S, f(r, s) = 1 if and only if r and s are a match. Past research is accustomed to running a learning algorithm directly on all the labeled (i.e., match or not) pairs in R times S. This, however, suffers from the drawback that even reading through the input incurs a quadratic cost. We pursue a direction towards removing the quadratic barrier. Denote by T the set of matching pairs in R times S. We propose to accept R, S, and T as the input, and aim to solve the problem with cost proportional to |R|+|S|+|T|, thereby achieving a large performance gain in the (typical) scenario where |T|u003cu003c|R||S|. This paper provides evidence on the feasibility of the new direction, by showing how to accomplish the aforementioned purpose for entity matching with linear classification, where a classifier is a linear multi-dimensional plane separating the matching and non-matching pairs. We actually do so in the MPC model, echoing the trend of deploying massively parallel computing systems for large-scale learning. As a side product, we obtain new MPC algorithms for three geometric problems: linear programming, batched range counting, and dominance join.