Multi-Head Spatio-Temporal Attention Mechanism For Urban Anomaly Event Prediction
summary
Proposes M-STAP, a citywide anomaly distribution predictor that combines Multi-Head Self-attention over regions, local convolutional spatial filters, a crowd-flow-to-event spatial attention (CFSAM) and a temporal attention module (TAM). Trains and evaluates on NYC, Chicago and Melbourne crime/incident and crowd-flow traces, reports substantially better heatmap and region-level predictions than standard baselines, and demonstrates that leveraging cross-region crowd-flow impacts improves sparse anomaly forecasting.
AI-written summary · about
related talks
07mDecentralized Attention-based Personalized Human Mobility Prediction
06mCross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning
21mCross-dataset Time Series Anomaly Detection for Cloud Systems
06mGlobalFusion: A Global Attentional Deep Learning Framework for Multisensor Information Fusion
06mMAIL: Multi-Scale Attention-Guided Indoor Localization Using Geomagnetic Sequences
06mAdversarial Multi-view Networks for Activity Recognition
04mLightning Talk: Anomaly Detection - Robust Strategy
34mTalk by Mahdi Soltanolkotabi (University of Southern California)