A Multi-Radio Audit of Subject-Dependent Performance in Recent Public WiFi CSI Datasets

Abstract

Random train/test splits of WiFi channel state
information (CSI) sensing data can place the same subjects
and overlapping temporal windows in both partitions, allowing
subject- or session-specific signal structure to inflate estimates
of activity recognition performance. We conduct a multi-radio
evaluation audit of WiStride, ElderAL-CSI, and the activityrecognition
subset of SHARP-80 using random forest and neuralnetwork
probes under random-window and subject-disjoint protocols,
with majority baselines and macro-averaged F1 scores. On
WiStride, random forest accuracy decreases from 0.773 to 0.684,
a gap of 8.9 percentage points (subject-cluster bootstrap 95%
confidence interval: 6.4–12.1 percentage points), while macro-F1
decreases from 0.526 to 0.322. SHARP-80 shows a larger leaveone-
subject-out collapse under campaign-linked person labels,
with RF accuracy falling from 0.954 to 0.181. Its four groups
and ElderAL’s three subjects provide preliminary evidence. Five
repeated source-group seeds show that cross-radio walk/non-walk
transfer does not exceed target majority accuracy. A group-
ID predictability diagnostic on WiStride exceeds its majority
baseline, indicating subject/session-correlated structure without
isolating its causal source. These results motivate subject-disjoint
evaluation, majority baselines, and imbalance-aware metrics
in CSI reporting, and we provide a reporting checklist and
reproducibility artifacts to facilitate their use.