Unsupervised Latent Debiasing of Time-Series Models

Abstract

Traditional training regimens for time-series models have been shown to encode the biases from their training corpora into the models themselves. We aim to train unbiased time-series models using existing biased datasets. However, most debiasing techniques rely on explicit labels that encapsulate the bias, such as pairs of words along some worrying axis of bias such as race or gender for language models. We propose an unsupervised latent debiasing training regimen based on [2] that simultaneously learns the latent distribution of the dataset and a separate language task; datapoints are selected for training batches by sampling weights inverse to their commonality as determined by their placement in the latent space. We adapt [2] to time-series datasets and show algorithmic improvements to bias identification and bias reduction for models trained on toy and real datasets.