Conceptual Data Scaling in Machine Learning

Information that is intended for human interpretation is frequently represented in a structured manner. This allows for a navigation between individual pieces to find, connect or combine information to gain new insights. Within a structure, we derive knowledge from inference of hierarchical or logical relations between data objects. For unstructured data there are numerous methods to define a data schema based on user interpretations. Afterward, data objects can be aggregated to derive (hierarchical) structures based on common properties. There are four main challenges with respect to the explainability of the derived structures. First, formal procedures are needed to infer knowledge about the data set, or parts of it, from hierarchical structures. Second, what does knowledge inferred from a structure imply for the data set it was derived from? Third, structures may be incomprehensibly large for human interpretation. Methods are needed to reduce structures to smaller representations in a consistent, comprehensible manner that provides control over possibly introduced error. Forth, the original data set does not need to have interpretable features and thus only allow for the inference of structural properties. In order to extract information based on real world properties, we need methods that are able to add such properties. With the presented work, we address these challenges using and extending the rich tool-set of Formal Concept Analysis. Here, data objects are aggregated to closed sets called formal concepts based on (unary) symbolic attributes that they have in common. The process of deriving symbolic attributes is called conceptual scaling and depends on the interpretation of the data by the analyst. The resulting hierarchical structure of concepts is called concept lattice. To infer knowledge from the concept lattice structures we introduce new methods based on sub-structures that are of standardized shape, called ordinal motifs. This novel method allows us to explain the structure of a concept lattice based on geometric aspects. Throughout our work, we focus on data representations from multiple state-of-the-art machine learning algorithms. In all cases, we elaborate extensively on how to interpret these models through derived concept lattices and develop scaling procedures specific to each algorithm. Some of the considered models are black-box models whose internal data representations are numeric with no clear real world semantics. For these, we present a method to link background knowledge to the concept lattice structure. To reduce the complexity of concept lattices we provide a new theoretical framework that allows us to generate (small) views on a concept lattice. These enable more selective and comprehensibly sized explanations for data parts that are of interest. In addition to that, we introduce methods to combine and subtract views from each other, and to identify missing or incorrect parts.