| Abstract |
Inductive Logic Programming (ILP) is a form of symbolic machine learning, where the goal is to learn human- readable rules from examples. Although some state-of-the-art ILP systems are able to guarantee optimality and generalisation of the knowledge learned, the certainty of predictions made using the learned rules strongly depends on the datasets from which they have been learned. So, when making such predictions in real- world applications, being able to give an estimate of the certainty of a prediction is crucial. Current state-of-the-art ILP systems have no way of computing such an estimate of the certainty.
In this paper, we describe how we have extended the FastLAS ILP system with a new predict mode, which takes as input the usual set of training examples, together with one additional example for which it is supposed to make a prediction. FastLAS then computes the scores of the best solutions (w.r.t. the training set) that accept and reject the new example. The intuition is that if there is little difference between these two scores then a small change in the training data would cause FastLAS to reverse its prediction (so there is low certainty), whereas a large difference in the two scores indicates a high certainty in the prediction.
FastLAS’s predict mode provides a first step towards estimat- ing the uncertainty of ILP systems. It is also expected to aid the integration of neural and symbolic approaches to machine learning.
|