Cross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning
summary
Proposes Adaptive Spatial-Temporal Transfer Learning that selects sources via domain distance computed with MMD, weights marginal vs conditional divergence dynamically, and extends Geodesic Flow Kernel with a temporal weighting inspired by the Markov property (GFK extension). Evaluated on 4 public HAR datasets and a lab-collected client–cloud system (3C2ARS), it yields an average accuracy gain of 8.7%, stabilizes in under 20 iterations, and user testing with 24 participants reports satisfactory performance.
AI-written summary · about
related talks
21mCross-dataset Time Series Anomaly Detection for Cloud Systems
47mA Complex Picture of Multi-task Learning
06mUnsupervised Human Activity Representation Learning with Multi-task Deep Clustering
06mMulti-Head Spatio-Temporal Attention Mechanism For Urban Anomaly Event Prediction
05mLeveraging Activity Recognition to Enable Protective Behavior Detection in Continuous Data
52mLanguage-guided Adaptation
31mIndividual Gaps in Prediction Generalization
44mBuilding Adaptive AI for MRI Diagnostics: A Continual Learning Case Study