Research & Publications
Applied machine learning research in ambient smart-home activity recognition, self-supervised representation learning, and multimodal fusion at the University of Georgia.
[ Academic Research ]
Thesis
Reading smart-home sensors four ways
Multimodal Self-Supervised Activity Recognition on CASAS Smart-Home Sensor Streams
A smart home can recognise what a resident is doing without cameras, just from ambient motion and door sensors clicking on and off. My thesis investigates how best to read those sparse event logs, finding that the optimal representation depends directly on the home's topology.
Four Synchronized Views of Sensor Streams
Ordered sensor timestamps, states, cyclic time-of-day, and inter-event intervals.
Deterministic natural language narrative summarizing room visits and movement pace.
96x96 spatial activation heatmap projected directly onto the home floor plan.
Spatial-temporal event and sensor topology with local message passing and global attention.
58 pages · 2.3 MB · Advisor: Dr. Fei Dou
4
CASAS Homes
Aruba, Cairo, Kyoto and Milan
4
Synchronized Views
Sequence, text, raster image & graph
30
Events Per Window
Stride of 15 with balanced overlap
3
Random Seeds
Evaluated across seeds 10, 30 and 50
Aruba House
Graph + Image + Sequence
0.850 W-F1
vs SICA (0.847)
Cairo House
Graph + Text + Image
0.842 W-F1
vs SICA (0.755)
Kyoto House
Image + Text + Sequence
0.843 W-F1
vs SICA (0.821)
[ Conference Proceedings ]
Submitted Paper
FACET
FACET: Factorized Ambient-Sensor Experts with Local Competence-Weighted Fusion for Smart-Home Activity Recognition
A machine learning paper on recognizing daily human activities from ambient smart-home sensors using factorized expert models and local competence-weighted fusion.
Joint research in Fei Dou's research lab at the University of Georgia. Submitted to IEEE PerCom 2027 (IEEE International Conference on Pervasive Computing and Communications).