A novel generic algorithm for cluster split iB-fold cross-validation

This article proposes a point of view (sub-space modelling), a post-processing strategy (relating to sub-optimal model) and a pre-processing strategy involving a novel cross-validation algorithm in machine learning with a focus on small datasets. To overcome the limitations imposed by small datasets, this approach leverages human expertise in being able to visualise or conceptualize clusters at high level using e.g. proximity or similarity, without detailed analysis. Validation in case studies on human motion have shown that it is possible to achieve better modelling results using iB-fold and by leveraging understanding of the data set and the nature of validation.