Time Series forecasting has become a pivotal part in many fields to extract information from data collected over time. This task can become challenging when the data exhibits irregularity in its measurements or dimensions. Rather than preprocessing the data and losing information in the process, we propose an approach based on the Kalman Filter to directly model irregularities. We present extensions to the Kalman Filter to establish an expressive model for irregular time series, which, contrary to existing literature, does not have to resort to expensive differential equations solvers to model time in a continuous way. We provide results on both synthetic and medical data, that show superior performance of the presented approach.
The baseline used for this project is the GRU-ODE-Bayes model, which can be found here
datasets/preprocessing
notebooks for preparing the MIMIC4 as well as synthetic 2D Ornstein-Uhlenbeck process data
datasets/utils
dataset class for irregular time series, collate functions for dataloaders, get-data utils
models
implementation for the discrete Kalman Filter, the continuous Kalman Filter with support for varing dimensions,
the Deep Kalman Filter and the Normalizing Kalman Filter
seml
seml files to execute HP tuning for the Kalman Filters
training
Kalman Filter training & evaluation utils
notebooks
includes examples for the discrete kalman filter usage, functionality to discretize the Kalman Filter predict ODEs and a check for normalizing the negative
log-likeihood with varying observation dimensions