Towards Transparent Time Series Forecasting

Authors: Krzysztof Kacprzyk, Tennison Liu, Mihaela van der Schaar

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments were conducted on four real-world datasets (Airfoil (Brooks et al., 1989), flchain (Dispenzieri et al., 2012), Stress-Strain (Aakash et al., 2019), and Tacrolimus (Woillard et al., 2011)) and three synthetic ones (Sine, Beta, and Tumor, the latter based on a model from (Wilkerson et al., 2017)). The synthetic datasets are constructed to contain trajectories exhibiting many different trends. Figure 1, Figure 4, Figure 1 show TIMEVIEW fitted to Sine, Beta, and Tumor datasets. As shown in Table 3, TIMEVIEW outperforms the transparent methods and closed-form expression on most datasets and achieves comparable performance to the black boxes.
Researcher Affiliation Academia Krzysztof Kacprzyk University of Cambridge kk751@cam.ac.ukTennison Liu University of Cambridge tl522@cam.ac.ukMihaela van der Schaar University of Cambridge The Alan Turing Institute mv472@cam.ac.uk
Pseudocode Yes Pseudocode. The pseudocode of the model training in TIMEVIEW is shown in Algorithm 1. The pseudocode of composition extraction implemented in TIMEVIEW is shown in Algorithm 2.
Open Source Code Yes The code to reproduce the results and for the visualization tool can be found at https://github.com/krzysztof-kacprzyk/TIMEVIEW and at the wider lab repository https://github.com/vanderschaarlab/TIMEVIEW.
Open Datasets Yes Experiments were conducted on four real-world datasets (Airfoil (Brooks et al., 1989), flchain (Dispenzieri et al., 2012), Stress-Strain (Aakash et al., 2019), and Tacrolimus (Woillard et al., 2011)) and three synthetic ones (Sine, Beta, and Tumor, the latter based on a model from (Wilkerson et al., 2017)).
Dataset Splits Yes All datasets are split into training, validation, and testing sets with ratios (0.7 : 0.15 : 0.15).
Hardware Specification Yes The experiments were performed on 12th Gen Intel(R) Core i7-12700H with 64 GB of RAM and NVIDIA Ge Force RTX 3050 Ti Laptop GPU as well as on the 10th Gen Intel Core i9-10980XE with 60 GB of RAM and NVIDIA RTX A4000.
Software Dependencies No The paper mentions various software packages used (e.g., scipy, scikit-learn, pytorch, py-xgboost, catboost, lightgbm) and their licenses, but does not specify their exact version numbers required for reproduction.
Experiment Setup Yes We perform hyperparameter tuning using Optuna (Akiba et al., 2019) and run it for 100 trials. We describe the hyperparameters we tune and their ranges in Table 6.