Neural Conditional Event Time Models

TL;DR

This paper introduces a neural network-based conditional event time model that separately predicts the probability of event occurrence and the timing of the event, improving accuracy in scenarios with possible non-occurrence.

Contribution

It develops a novel neural network model with a stochastic layer to distinguish event occurrence probability from event timing, trained with maximum likelihood on censored data.

Findings

Superior prediction accuracy on synthetic data

Improved event occurrence and timing predictions on medical data

Effective modeling of social media event timings

Abstract

Event time models predict occurrence times of an event of interest based on known features. Recent work has demonstrated that neural networks achieve state-of-the-art event time predictions in a variety of settings. However, standard event time models suppose that the event occurs, eventually, in all cases. Consequently, no distinction is made between a) the probability of event occurrence, and b) the predicted time of occurrence. This distinction is critical when predicting medical diagnoses, equipment defects, social media posts, and other events that or may not occur, and for which the features affecting a) may be different from those affecting b). In this work, we develop a conditional event time model that distinguishes between these components, implement it as a neural network with a binary stochastic layer representing finite event occurrence, and show how it may be learned from right-censored event times via maximum likelihood estimation. Results demonstrate superior event occurrence and event time predictions on synthetic data, medical events (MIMIC-III), and social media posts (Reddit), comprising 21 total prediction tasks.