CCVAE: A Variational Autoencoder for Handling Censored Covariates

For time or safety critical scenarios when faulty predictions or decisions can have crucial consequences, such as in certain telecommunications scenarios, reliable prediction models and accurate data are of the essence. When modeling and predicting data in such scenarios, data with censored covariat...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:2022 21ST IEEE INTERNATIONAL CONFERENCE ON MACHINE LEARNING AND APPLICATIONS, ICMLA s. 709 - 714
Hlavní autori: Svahn, Caroline, Sysoev, Oleg
Médium: Konferenčný príspevok..
Jazyk:English
Vydavateľské údaje: IEEE 01.12.2022
Predmet:
Benchmark testing
censored covariates
Data models
deep learning
dynamic range
Estimation
Machine learning
Predictive models
Telecommunications
Uncertainty
variational autoencoder
zero inflated truncated normal
zero inflated truncated normal
deep learning
dynamic range
variational autoencoder
censored covariates
ISBN:1665462833, 9781665462846, 1665462841, 9781665462839
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:For time or safety critical scenarios when faulty predictions or decisions can have crucial consequences, such as in certain telecommunications scenarios, reliable prediction models and accurate data are of the essence. When modeling and predicting data in such scenarios, data with censored covariates remain an issue as ignoring them or imputing them with lack of precision may cause inaccurate or uncertain predictions. In this paper, we provide a fast, reliable Variational Autoencoder framework which can handle covariate censoring in high dimensional data. Our numerical experiments demonstrate that our framework compares favorably to alternative methods in terms of prediction accuracy for both the response and the covariates, while enabling estimation of the prediction uncertainties. We moreover demonstrate that the method is at least 8 times faster than the benchmark models used in this paper, and more robust to initial imputations and noise than existing models. The method can also be used directly for predicting unseen data, which is a challenge for some state-of-the-art methods.
ISBN:1665462833
9781665462846
1665462841
9781665462839
DOI:10.1109/ICMLA55696.2022.00118