Incremental Update of Locally Optimal Classification Rules

Incremental learning is a traditional topic that has particularly gained importance in the wake of big data and stream mining. Discrete symbolic representations do not easily allow for gradual refinements of the learned concept. While the problem is less severe for incremental induction of decision trees, it is much harder for incremental rule learning in that there are hardly any incremental rule learning algorithms which are really successful. In this paper, we introduce iLord algorithm, an adaptation of a recently proposed rule learning algorithm Lord, which aims at finding the best rule for each individual example, to an incremental learning setting. After being initialized with a first batch of training examples, iLord relies on efficient data structures to summarize the information contained in the training examples, which can be quickly updated and allows to retrieve the best rule for each incoming example. The behavior of iLord is evaluated with different parameterizations, and compared to other best-known incremental symbolic learning algorithms such as HoeffdingTree and VFDR.