Machine Learning: Between Accuracy and Interpretability
Predictive accuracy is the usual measure of success of Machine Learning (ML) applications. However, experience from many ML applications in difficult, domains indicates the importance of interpretability of induced descriptions. Often in such domains, pre
- PDF / 1,631,051 Bytes
- 15 Pages / 481.89 x 691.654 pts Page_size
- 54 Downloads / 277 Views
I. Bratko
Ljubljana University, Ljubljana, Slovenia
ABSTRACT Predictive accuracy is the usual measure of success of Machine Learning (ML) applications. However, experience from many ML applications in difficult domains indicates the importance of interpretability of induced descriptions. Often in such domains, predictive accuracy is hardly of interest to the user. Instead, the users' interest now lies
in the interpretion of the induced descriptions and not in their use for prediction. In such cases, ML is essentially used as a tool for exploring the domain, to generate new, potentially useful ideas about the domain, and thus improve the user's understanding of the domain. The important questions are how to make domain-specific background knowledge usable by the learning system, and how to interpret the results in the light of this background expertise. These questions are discussed and illustrated by relevant example applications of ML, including: medical diagnosis, ecological modelling, and interpreting discrete event simulations. The observations in these applications show that predictive accuracy, the usual measure of success in ML, should be accompanied by a criterion of interpretability of induced descriptions. The formalisation of interpretability is however a completely new challenge' for ML.
G. Della Riccia et al. (eds.), Learning, Networks and Statistics © Springer-Verlag Wien 1997
164
1
I. Bratko
Introduction
In Machine Learning (ML), the usual criterion of success has been considered to the predictive accuracy.
However, observations from many ML applications show that the user's mam interest was in the interpretation of the descriptions induced from examples. These observations come from applications in various domains. In this paper we look at some of such applications that come from ecological modelling [6], predicting the mutagenicity of chemical compounds [12], assessing river water quality [3], modelling the deer population dynamics [13], and analysis of discrete event simulation results [9]. For example, Boris Kompare, a user of ML techniques, studied in his thesis the growth of algae in the Lagoon of Venice and the biodegradability of chemicals [6]. The main contribution of the thesis consisted in discovering, with machine learning, of interesting patterns in the environmental data. Although the predicitve accuracy of the induced descriptions was very limiLed, they were of interest due to their expert interpretations. The interpretations then made the learning results useful as a source of ideas about the domain of investigation. Without the in-depth interpretations of learning results, the thesis would be of little interest. Comments from system modelling and simulation experts are also quite revealing in respect of experts' interests. Simulation results are rarely used really for making for predictions. Instead, models are studied and simulations done to deepen the expert's understanding of the problem. All this experience indicates that in machine learning, the predicitve accuracy cr
