Machine learning based on similarity operation

Abstract

The paper describes a machine-learning paradigm that uses binary semi-lattice operation for computing similarities between training examples, with Formal Concept Analysis (FCA) providing a technique for bitset encoding of the objects and similarities between them. Using this encoding, a coupling Markov chain algorithm can generate a random sample of similarities. We provide a technique to accelerate convergence of the main algorithm by truncating its runs that exceed sum of lengths of previous trajectories. The similarities are hypothetical causes (hypotheses) for the target property. The target property of test examples can be predicted using these hypotheses. We provide a lower bound on necessary number of hypotheses to predict all important test examples for a given confidence level.