Decentralized Attention-based Personalized Human Mobility Prediction
summary
Proposes a decentralized attention-based predictor that personalizes by retrieving similar past trajectories (meta-learning) and trains across phones via federated meta-learning so raw GPS never leaves the device. Implements three pretraining tricks (open datasets, small volunteer set, cross-city) and runs on Android, reporting low-power on-device inference, accuracy comparable to centralized training and better than classic baselines, and no extra per-user or incremental training required.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:54How human mobility is predicted?
- 1:27Location Privacy
- 1:50Two challenges
- 2:12Conventional Predictive Model
- 3:02Personalized Human Mobility Predictor
- 3:52Cross Entropy w.rt time/day
- 4:15Decentralized Human Mobility Prediction
- 5:21Federated Learning Paradigm
- 5:53Pretraining Trick
- 6:25Evaluation
related talks
06mMetaTP: Traffic Prediction with Unevenly-Distributed Road Sensing Data via Fast Adaptation
06mPREFER: Point-of-interest REcommendation with efficiency and privacy-preservation via Federated ...
06mMulti-Head Spatio-Temporal Attention Mechanism For Urban Anomaly Event Prediction
34mDecentralized Machine Learning
44mFrom responsive to anticipatory
07mHERMAS: A Human Mobility Embedding Framework with Large-scale Cellular Signaling Data
28m"Where, When, Why, and How: Directi
57mMachine Learning in the Age of IoT