My first approach to multivariate PULSAR was to join each recording’s channels end to end. This let me use the original single-channel classifier without first designing a new feature extractor. I ran it on 19 UEA datasets using their original train/test splits and random seed 42.
What the first run showed
The mean test accuracy, weighting each dataset equally, was 77.86%. BasicMotions reached 100%, Handwriting reached 46.35%, and EthanolConcentration reached 55.51%. The average hides those differences, so I kept the individual results as well.
What it didn’t tell me
This was one seed on one split per dataset. It didn’t show whether joining the channels was better than extracting features from them separately. It also didn’t tell me whether relationships between channels would be useful. Those needed their own controls.
Where I went next
I moved on to per-channel feature extraction, then short-window channel relationships and probability blending. The newer experiments use four datasets and five seeds each, so their averages shouldn’t be compared directly with this 19-dataset mean. I’ve kept this note as a record of the starting point; the current project page has the later results and code.