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

▶
uncertainty on asynchronous time event prediction
SMART_READER_LITE
LIVE PREVIEW

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

Dec 14, 2023 249 likes •373 views

Data Analytics and Machine Learning Group Department of Informatics Technical University of Munich Uncertainty on Asynchronous Time Event Prediction Marin Bilo * Bertrand Charpentier* Stephan Gnnemann Setting Discrete events in

slide-1
SLIDE 1

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

slide-2
SLIDE 2

𝜐𝑗−3 … 𝜐𝑗−2 𝜐𝑗−1 𝜐𝑗

  • Smart house

Lights TV Shower

2

  • M. Biloš Uncertainty on Asynchronous Time Event Prediction
  • Social networks
  • Medical records
  • Cars

What is the next interaction?

Setting – Discrete events in asynchronous time

slide-3
SLIDE 3

…

  • Smart house

3

  • M. Biloš Uncertainty on Asynchronous Time Event Prediction
  • Social networks
  • Medical records
  • Cars

What is the next interaction?

Setting – Discrete events in asynchronous time

  • Two main challenges

1. Complex evolution 2. Uncertainty in prediction 𝜐𝑗−3 𝜐𝑗−2 𝜐𝑗−1 𝜐𝑗

Lights TV Shower

slide-4
SLIDE 4

…

4

  • M. Biloš Uncertainty on Asynchronous Time Event Prediction

100% 0% 50%

History ℋ𝑗

𝑞 𝜐, ℋ𝑗) 𝑞 𝜐, ℋ𝑗) 𝑞 𝜐, ℋ𝑗)

Challenge 1 – Complex evolution of 𝑞 over (continuous) time

  • Evolution of categorical distribution
  • Multimodality

𝜐𝑗−3 𝜐𝑗−2 𝜐𝑗−1 𝜐𝑗

slide-5
SLIDE 5

…

5

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

Challenge 1 – Complex evolution of 𝑞 over (continuous) time

100% 0% 50%

  • Evolution of categorical distribution
  • Multimodality

𝑞 𝜐, ℋ𝑗) 𝑞 𝜐, ℋ𝑗) 𝑞 𝜐, ℋ𝑗)

History ℋ𝑗

𝜐

𝜐𝑗−3 𝜐𝑗−2 𝜐𝑗−1 𝜐𝑗

slide-6
SLIDE 6

6

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

Challenge 2 – Uncertainty in prediction

% % % %

  • In classical approaches uncertainty is ignored
slide-7
SLIDE 7

7

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

Challenge 2 – Uncertainty in prediction

Uncertain prediction Equiprobable classes

%

slide-8
SLIDE 8

8

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

Challenge 2 – Uncertainty in prediction

Uncertain prediction Equiprobable classes

%

  • We distinguish between two scenarios
  • Instead of outputting one vector → Distribution over the simplex
slide-9
SLIDE 9

9

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

Our approach – Continuously evolving distribution over the simplex

𝒊𝑗 𝜾 𝜐 𝑗 ℋ𝑗

RNN

slide-10
SLIDE 10

10

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

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

slide-11
SLIDE 11
  • State-of-the-art results
  • Event prediction
  • Anomaly detection

11

  • M. Biloš

Uncertainty on Asynchronous Time Event Prediction

Complex evolution + Uncertainty in prediction

0,4 0,5 0,6 AUROC

Smart house anomaly detection Our models Others Code & Paper www.daml.in.tum.de/uncertainty-event-prediction Poster Wednesday 10:45 – 12:45 East Exhibition Hall B + C #53