Cause of anomalous SVM classification performance

Determine the cause of the observed anomalies in balanced accuracy— including results worse than random guessing—when training Support Vector Machine classifiers for detecting violin width reduction using elevation maps and contour-line-derived features under leave-one-out cross-validation, and ascertain whether these anomalies are side-effects of the very small 25-instrument dataset.

Background

The paper compares Support Vector Machines and Decision Trees using two types of geometric representations of violin sound boards: high-dimensional elevation maps and engineered features derived from contour lines. Models are evaluated with leave-one-out cross-validation due to the small dataset of 25 instruments.

In the SVM results, the authors report large variability in balanced accuracy across settings, including values below 50%, which indicate worse-than-chance performance. Despite a robust, non-arbitrary regularization selection procedure, the authors note they cannot fully explain these anomalies and speculate that the very small dataset size may be responsible.

Understanding the origin of these anomalies is important for establishing reliable, reproducible classification procedures for detecting violin width reduction and for guiding methodological choices in small-sample cultural heritage studies.