Symbolic Learning and Rule Extraction with Sole.jl
Date:
Symbolic learning is a branch of machine learning that studies algorithms for building symbolic models, classifiers that can be translated to logical rules. Hence, differently from neural networks and other statistical models, symbolic models are more easily readable and interpretable. Common examples include decision trees and their ensemble counterpart, random forests.
Sole.jl is a Julia package for symbolic learning and rule extraction, aimed at guiding the user throughout the whole process, from the initial data to learning a symbolic model, to the extraction and manipulation of logical rules from such model.
It is also the entry point to the SOLE framework (which stands for SymbOlic LEarning), an open-source project fully developed in Julia, serving as an interface to both its main packages, such as SoleData.jl, SoleModels.jl, SolePostHoc.jl, as well as useful packages developed by the learning community, such as DecisionTree.jl, ModalDecisionTrees.jl, ModalAssociationRules.jl, MLJ.jl, XGBoost.jl, and so on.
For instance, given a dataset, one can use Sole.jl to fit a decision tree model, provided by the DecisionTree.jl package, and then extract and eventually manipulate rules from such model through the SolePostHoc.jl package. Furthermore, if working with more non-tabular data like images or time-series, one can leverage the SoleData.jl package to interpret the dataset as a set of logical interpretations of a more-than propositional logic (e.g., interval or rectangle modal logic) and use a learning package compatible with such logic, such as ModalDecisionTrees.jl or ModalAssociationRules.jl.
In this presentation, we will have a look at how Sole.jl works through a hands-on tutorial, emphatizing on its comprehensiveness and user-friendliness. This will also allow us to introduce two newcomers to the SOLE ecosystem: ModalAssociationRules.jl, a package for mining association rules between instances, and SolePostHoc.jl, a package to extract, interpret and simplify sets of rules starting from a symbolic model.