Uncertainty on Asynchronous Time Event Prediction Marin Bilo * - PowerPoint PPT Presentation

Data Analytics and Machine Learning Group Department of Informatics Technical University of Munich Uncertainty on Asynchronous Time Event Prediction Marin Biloš * • Bertrand Charpentier* • Stephan Günnemann

Setting – Discrete events in asynchronous time • Smart house Lights TV Shower • Social networks • … 𝜐 𝑗−3 𝜐 𝑗−2 𝜐 𝑗−1 𝜐 𝑗 Medical records • Cars What is the next interaction? M. Biloš Uncertainty on Asynchronous Time Event Prediction 2

Setting – Discrete events in asynchronous time • Two main challenges 1. Complex evolution • Smart house 2. Uncertainty in prediction Lights TV Shower • Social networks • … 𝜐 𝑗−3 𝜐 𝑗−2 𝜐 𝑗−1 𝜐 𝑗 Medical records • Cars What is the next interaction? M. Biloš Uncertainty on Asynchronous Time Event Prediction 3

Challenge 1 – Complex evolution of 𝑞 over (continuous) time • Evolution of categorical distribution • Multimodality 100% 𝑞 𝜐, ℋ 𝑗 ) History ℋ 𝑗 50% 𝑞 𝜐, ℋ 𝑗 ) 𝑞 𝜐, ℋ 𝑗 ) 0% … 𝜐 𝑗−3 𝜐 𝑗−2 𝜐 𝑗−1 𝜐 𝑗 M. Biloš Uncertainty on Asynchronous Time Event Prediction 4

Challenge 1 – Complex evolution of 𝑞 over (continuous) time • Evolution of categorical distribution • Multimodality 100% 𝑞 𝜐, ℋ 𝑗 ) History ℋ 𝑗 50% 𝑞 𝜐, ℋ 𝑗 ) 𝑞 𝜐, ℋ 𝑗 ) 0% … 𝜐 𝜐 𝑗−3 𝜐 𝑗−2 𝜐 𝑗−1 𝜐 𝑗 M. Biloš Uncertainty on Asynchronous Time Event Prediction 5

Challenge 2 – Uncertainty in prediction • In classical approaches uncertainty is ignored % % % % M. Biloš Uncertainty on Asynchronous Time Event Prediction 6

Challenge 2 – Uncertainty in prediction Equiprobable Uncertain classes prediction % M. Biloš Uncertainty on Asynchronous Time Event Prediction 7

Challenge 2 – Uncertainty in prediction Equiprobable Uncertain classes prediction % • We distinguish between two scenarios • Instead of outputting one vector → Distribution over the simplex M. Biloš Uncertainty on Asynchronous Time Event Prediction 8

Our approach – Continuously evolving distribution over the simplex ℋ 𝑗 𝒊 𝑗 𝜾 𝜐 𝑗 RNN M. Biloš Uncertainty on Asynchronous Time Event Prediction 9

Our approach – Continuously evolving distribution over the simplex Model 1 – Dirichlet distribution* parameters evolve with basis function decomposition* Model 2 – Logistic-normal* parameters evolve with a weighted Gaussian process* * Technical details during poster session M. Biloš Uncertainty on Asynchronous Time Event Prediction 10

Complex evolution + Uncertainty in prediction • State-of-the-art results • Event prediction Smart house anomaly detection • Anomaly detection 0,6 AUROC 0,5 Poster Wednesday 10:45 – 12:45 0,4 East Exhibition Hall B + C #53 Others Our models Code & Paper www.daml.in.tum.de/uncertainty-event-prediction M. Biloš Uncertainty on Asynchronous Time Event Prediction 11