Distinguishing Dactyls of Crab Species Using Relational Machine Learning

Fossils of decapod crustaceans are more common than published records suggest. Dactyls (movable fingers of claws) are well represented in shell-rich fossil assemblages, but are usually ignored because of the assumption that they can be identified only to high taxonomic levels. Recent studies using outline-based and geometric morphometric methods have demonstrated that closely related species of decapods can be distinguished by their dactyls. Although these techniques allow statistical tests of differences in dactyl morphologies, dactyl shapes must still be described qualitatively.

Our research introduces a new method for distinguishing dactyl shapes by automatically extracting relational features that describe their underlying spatial structure. We first use medial axis techniques, used for shape recognition algorithms in computer vision, to find the shock graph of each dactyl outline. Next, these shock graphs are converted into a first-order logic representation capturing the connections, distances and angles between the nodes in each graph. We then use Aleph, an Inductive Logic Programming algorithm, to find relational classification rules based on the shock graph representations. These relational rules provide a concise and human-understandable way to describe the morphological differences among dactyls of closely related decapods, and can be seen as a first step to creating automatically learned quantitative taxonomic keys.