Our PreDiCT-IDLab researchers recently achieved outstanding success at the HASCA workshop, part of the UbiComp/ISWC Conference, held in Melbourne, Australia. Jonas and Jeroen Van Der Donckt, took first place in both the WEAR and SHL time series machine learning challenges - outperforming the competition by an impressive 10% higher (absolute!) macro F1 score.
Our solution, built on our in-house tsflex toolkit for feature extraction and insights gathered from visualization using our plotly-resampler library, proved highly effective. In particular we relied on a traditional machine learning pipeline consisting of processing, feature extraction, and a gradient boosted trees model.
For the SHL challenge, we proposed a novel feature aggregation technique, with the aim of constructing more “rotation-invariant” features. Whereas, for the WEAR challenge, we investigated a data augmentation scheme by swapping left-right wearables, to train a more robust model.
For those interested in diving deeper, the research papers are available here:
Explore the tools we used:
