Loss Shaping Constraints for Long-Term Time Series Forecasting

Authors: Ignacio Hounie, Javier Porras-Valenzuela, Alejandro Ribeiro

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive evaluations comparing constrained and resilient constrained learning against the customary unconstrained training pipeline across eight model architectures and nine popular datasets.
Researcher Affiliation Academia 1University of Pennsylvania. Correspondence to: Ignacio Hounie <ihounie@seas.upenn.edu>.
Pseudocode Yes Algorithm 1 Primal Dual Loss Shaping.
Open Source Code No No explicit statement about the release of source code for the methodology or a link to a code repository was found.
Open Datasets Yes Explicitly, the datasets are: Electricity Consumption Load (ECL), Weather, Exchange Rate (Lai et al., 2018), Traffic, Electricity Transformer Temperature (ETT) (two hourly datasets and two every 15 minutes) (Zhou et al., 2021), and Influenza Like Illness (ILI). ... The data is available at https://archive.ics.uci.edu/ml /datasets/Electricity Load Diagrams20112 014.
Dataset Splits Yes Data is split into train, validation, and test chronologically with a ratio of 7:1:2.
Hardware Specification No No specific hardware details such as GPU/CPU models, processor types, or memory amounts used for experiments were mentioned in the paper.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9) were provided for replication.
Experiment Setup Yes We follow the same setup, including preprocessing, hyperparameters and implementation, as described in (Kitaev et al., 2019; Wu et al., 2021; Liu et al., 2021; Zhou et al., 2021; Liu et al., 2022; Zhou et al., 2022a). ... train for a full 10 epochs... dual learning rate set to 0.01 and duals initialized to 1.0.