NeurIPs Paper Reviews 2023 #6

Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting

Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Schneider, Jasper Zschiegner, Hao Wang, Yuyang Wang

The authors demonstrate how an unconditionally trained generative model can be just as useful as task specific conditional models for various tasks without needing to adapt its training process.

For this purpose, they introduce TSDiff, an unconditionally trained diffusion model for time series and show how it can be used to:

1. Sample from a conditional distribution despite not being trained on it;
2. Enhance predictions from other forecasting models by formulating forecast enhancement as a regularized optimization problem using its learned likelihood function;
3. Provide better synthetic data for downstream forecasting than other time series based generative models.

The authors conduct experiments for each of these properties, comparing TSDiff’s performance against statistical and probabilistic models tailored to specific tasks. The results consistently demonstrate that TSDiff performs at least as well as, if not better than, the selected alternative models for each respective task.

Conformal Prediction for Time Series with Modern Hopfield Networks

Andreas Auer, Martin Gauch, Daniel Klotz, Sepp Hochreiter

Conformal prediction provides distribution- and model-agnostic ways to construct prediction intervals using quantiles of a weighted error distribution from past predictions. However, it assumes the data to be exchangeable (order of observation doesn’t matter), which is violated by time series data with temporal autocorrelation. This paper presents a framework called HopCPT to provide valid prediction intervals with conformal prediction in this scenario.

Given a forecasting model for time series, the idea is that samples with temporal dependencies would have similar prediction errors (regimes). Hence the authors train a modern Hopfield network on the prediction errors to use its associative memory to retrieve past errors from a similar regime and weigh them using trained network weights to construct the prediction interval for a new sample.

The authors demonstrate the efficiency of HopCPT prediction intervals against other CP approaches on different forecasting models and time series.