Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

TACTiS: Transformer-Attentional Copulas for Time Series

Authors: Alexandre Drouin, ร‰tienne Marcotte, Nicolas Chapados

ICML 2022 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate these properties empirically and show that our model produces state-of-the-art predictions on multiple real-world datasets. 5. Experiments
Researcher Affiliation Industry Alexandre Drouin * 1 Etienne Marcotte * 1 Nicolas Chapados * 1 1Service Now Research. Correspondence to: All authors <EMAIL>.
Pseudocode No The paper describes the model architecture and training procedure in text and with diagrams (e.g., Figure 1, Figure 2) but does not include a formal pseudocode or algorithm block.
Open Source Code Yes Code available at https://github.com/servicenow/tactis.
Open Datasets Yes We consider five real-world datasets from the Monash Time Series Forecasting Repository (Godahewa et al., 2021): electricity, fred-md, kdd-cup, solar-10min, and traffic (see C.1 for details). These were selected for being high-dimensional (107 862 variables), exempt of missing values, and sampled at diverse frequencies (10 min., hourly, monthly). Each dataset was downloaded from the Monash Repository using Gluon TS (Alexandrov et al., 2020) (links to the exact versions are provided in Tab. 3).
Dataset Splits Yes From the remaining time steps, a window at the end of length equals to 7 times the prediction length is reserved to serve as the validation set, which is used to compute the metrics used to select the best hyperparameters. The hyperparameter search described in C.3 is done using only data before ฯ„1, while the i-th model is trained using only the data before ฯ„i.
Hardware Specification Yes Each training is done in a Docker container giving access to an NVIDIA Tesla P100 GPU with 12 GB of GPU RAM, 2 CPU cores, and 32 GB of CPU RAM.
Software Dependencies No The paper mentions software like PyTorch (Paszke et al., 2019), Py Torch TS (Rasul, 2021a), and Gluon TS (Alexandrov et al., 2020) but does not provide specific version numbers for these libraries, only citations to their respective papers.
Experiment Setup Yes In this section, we list the values considered for each hyperparameter of each method. The ranges considered for the baselines are inspired by those published in their respective papers and implementations. Tab. 4 shows the hyperparameters for the TACTi S-TT model.