Analogical Classification: A Rule-Based View

Analogical proportion-based classification methods have been introduced a few years ago. They look in the training set for suitable triples of examples that are in an analogical proportion with the item to be classified, on a maximal set of attributes. This can be viewed as a lazy classification technique since, like k-nn algorithms, there is no static model built from the set of examples. The amazing results (at least in terms of accuracy) that have been obtained from such techniques are not easy to justify from a theoretical viewpoint. In this paper, we show that there exists an alternative method to build analogical proportionbased learners by statically building a set of inference rules during a preliminary training step. This gives birth to a new classification algorithm that deals with pairs rather than with triples of examples. Experiments on classical benchmarks of the UC Irvine repository are reported, showing that we get comparable results.