Interpretability in data science refers to those methods and models that make the behavior and predictions of machine learning systems understandable to humans. In other words, it can be defined as the degree to which a human can understand the cause and effect of a decision [Mil19]. This term is essential in those application fields in which an accurate prediction or decision only partially solves the original problems [Mol20].
The research community widely accept decision trees-based and rule-based models as expressive ways of capturing interactions between features in data [Hal18]. For these models, it is pretty simple to interpret the cause and effect of the final decision. However, these methods fail to deal with linear relationships, producing a drop in the final accuracy. Moreover, these methods tend to be unstable and few changes in the input data may produce completely different models or solutions. Last but not least, interpretability is subject to both the number of rules in rule-based models and the deep of the tree in tree-based models. Recent research studies are proposing alternatives based on a model-agnostic explanation [Bel19], that is, considering a posteriori description of a black-box model through a set of (simpler) rules. A major advantage of this approach lies in its modularity since the machine learning model can be easily replaced.
However, further studies are needed properly select the right set of rules, or even to explain, for example, the cause of deviation in the behavior of some models.
The aim is to extend interpretable models to maximize the interest and interpretability in the following scenarios:
- Extending the recent advances in the generation of interesting and interpretable rules by improving the evolutionary algorithms in terms of appropriate objective function (considering both interest and interpretability of the rules) and the effectiveness of the method to produce diverse sets of solutions.
- Improving the interpretability of black-box models by extracting patterns in the form of rules that describe the cause of changes in the behavior of the predictive model.
The research group has an important background in the development of rule-based predictive models through evolutionary algorithms, and metaheuristics [Can13, Ped14, Pad20]. Recently, the research group has studied the extraction of both interpretable [Gar20] and highly interesting rules [Del20].
[Mil19] T. Miller: Explanation in artificial intelligence: Insights from the social sciences. Artif. Intell. 267: 1-38 (2019).
[Mol20] Christoph Molnar: Interpretable Machine Learning. Lulu.com. ISBN: 978-0244768522
[Bel19] S. Belkoura, M. Zanin, A. LaTorre: Fostering interpretability of data mining models through data perturbation. Expert Syst. Appl. 137: 191-201 (2019)
[Hal18] P. Hall, N. Gill. An introduction to machine learning interpretability. O’Really 2018.
[Can13] A. Cano, A. Zafra, S. Ventura, An interpretable classification rule mining algorithm, Inf. Sci. 240 (2013), 1–20
[Ped14] J.A. Pedraza, C. García-Martínez, A. Cano, S. Ventura, Classification rule mining with iterated greedy, in Proc. of Hybrid Artificial Intelligence Systems workshop, vol. 8480, (2014), pp. 585–596.
[Del20] J.A. Delgado-Osuna, C. García-Martínez, J. Gómez-Barbadillo, S. Ventura. Heuristics for interesting class association rule mining a colorectal cancer database. Information Processing and Management 57:3 (2020) 102207
[Pad20] F. Padillo, J.M. Luna, S. Ventura. LAC: Library for associative classification. Knowledge-Based Systems 193 (2020) pp. 1–3.
[Gar20] C. García-Martínez, S. Ventura. Multi-view genetic programming learning to obtain interpretable rule-based classifiers for semi-supervised contexts. Lessons Learnt. International Journal of Computational Intelligence Systems 13:1 (2020) 576-590.