UniTS: A Unified Multi-Task Time Series Model

Authors: Shanghua Gao, Teddy Koker, Owen Queen, Tom Hartvigsen, Theodoros Tsiligkaridis, Marinka Zitnik

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Tested on 38 datasets across human activity sensors, healthcare, engineering, and finance, UNITS achieves superior performance compared to 12 forecasting models, 20 classification models, 18 anomaly detection models, and 16 imputation models, including adapted text-based LLMs.
Researcher Affiliation Academia Shanghua Gao Harvard University shanghua_gao@hms.harvard.edu Teddy Koker MIT Lincoln Laboratory tekoker@mit.edu Owen Queen Harvard University owen_queen@hms.harvard.edu Thomas Hartvigsen University of Virginia hartvigsen@virginia.edu Theodoros Tsiligkaridis MIT Lincoln Laboratory ttsili@ll.mit.edu Marinka Zitnik Harvard University marinka@hms.harvard.edu
Pseudocode No The paper describes the model in text and diagrams (e.g., Figure 2c, Figure 4) but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code and datasets are available at https://github.com/mims-harvard/UniTS.
Open Datasets Yes Code and datasets are available at https://github.com/mims-harvard/UniTS. For multi-task learning on forecasting and classification, we compiled 38 datasets from several sources [79, 33, 82].
Dataset Splits No The paper mentions 'splitting the data into training and testing sets' and 'tune the following hyperparameters', which implies the use of a validation set, but it does not explicitly state the validation set split percentage or methodology.
Hardware Specification Yes All experiments are conducted using A100-40G GPUs. Each experiment is conducted with one or two GPUs, and the maximum running time is under 48 hours.
Software Dependencies No The paper does not specify software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions are not listed).
Experiment Setup Yes Supervised training involves 5 epochs using gradient accumulation for an effective batch size of 1024, starting with a learning rate of 3.2e-2 and adjusted with a multi-step decayed schedule.