Applying Path Signatures to Handwritten Digit Classification

Summary

The article demonstrates a machine-learning use of path signatures: classify handwritten digits by treating pen coordinates recorded over time as a path. It describes training a signature-based model on the UCI Pen-Based Recognition dataset, using ordered coordinate points as inputs and digit labels as targets. The model is evaluated on a separate test set, with accuracy compared across signature orders.

The reported best result in the described experiment is 87% accuracy with a ninth-order signature, and performance improves as signature order increases in the plotted comparison. The article presents this as an illustration of signatures' ability to represent sequential shape information, rather than a finished classifier. It suggests rescaling inputs and using cross-validation or neural networks as possible improvements, and cites other signature-based work reporting higher accuracy. The experiment concerns handwriting rather than financial data, so its direct relevance to trading is methodological: signatures can encode sequential observations for machine-learning tasks.

Key ideas

• A handwritten digit can be represented as a time-ordered path of pen coordinates.
• Path signatures transform sequential coordinate data into features for a supervised classifier.
• The experiment trains on one dataset split and evaluates predictions on a separate test split.
• The reported accuracy varies with signature order and reaches its best stated level at ninth order.
• Input rescaling, cross-validation, and neural networks are identified as possible improvements.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.