Data sampling through collaborative filtering for algorithm selection

Algorithm selection has been studied to specify the best possible algorithm(s) for a given problem instance. One of the major drawbacks of the algorithm selection methods is their need for the performance data. The performance data involves the performance of a set of algorithms on a group of problem instances. Depending on the problem domain, algorithms and the experimental settings, generating such data can be computationally expensive. ALORS [1] as a collaborative filtering based algorithm selection strategy addresses this issue by performing matrix completion. Matrix completion allows to generate algorithm selection models when the performance data is incomplete. Although ALORS is able to deal with varying data incompleteness levels, it ignores the quality and cost of the performance data. The present study offers a collaborative filtering based sampling strategy to designate which algorithm(s) to run on which instance(s). The goal is to provide either computationally cheap or highly informative incomplete performance data for algorithm selection.