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].
iHyperTime: Interpretable Time Series Generation with Implicit Neural Representations
Authors: Elizabeth Fons, Alejandro Sztrajman, Yousef El-Laham, Andrea Coletta, Alexandros Iosifidis, Svitlana Vyetrenko
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through evaluations of fidelity and usefulness metrics, we demonstrate that i Hyper Time outperforms current state-of-the-art methods in challenging scenarios that involve long or irregularly sampled time series, while performing on par on regularly sampled data. Furthermore, we showcase i Hyper Time fast training speed, comparable to the fastest existing methods for short sequences and significantly superior for longer ones. Finally, we empirically validate the quality of the model s unsupervised trend-seasonality decomposition by comparing against the well-established STL method. |
| Researcher Affiliation | Collaboration | Elizabeth Fons EMAIL J.P. Morgan AI Research Alejandro Sztrajman EMAIL University of Cambridge Yousef El-Laham EMAIL J.P. Morgan AI Research Andrea Coletta EMAIL Bank of Italy Alexandros Iosifidis EMAIL Aarhus University Svitlana Vyetrenko EMAIL J.P. Morgan AI Research |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are present in the main text of the paper. The methodology is described in narrative form and with network diagrams. |
| Open Source Code | No | The paper lists publicly available code for baseline methods but does not provide a direct link or explicit statement for the release of the authors' own implementation code for i Hyper Time. |
| Open Datasets | Yes | Datasets We test the performance of i HT using multiple datasets with varying characteristics such as periodicity, level of noise, number of features and length of the series. Stock corresponds to Google stock price data from 2004 to 2019, where each observation has 6 features. Energy is a UCI appliance prediction dataset (Candanedo et al., 2017) with 28 features. Additionally, we also consider Monash dataset (Godahewa et al., 2021), from which we choose FRED-MD, NN5 Daily, Temperature Rain, and Solar Weekly datasets. A complete description of the datasets can be found in Appendix B. ... In Section F we extend the evaluation of our model on two real-world irregularly sampled datasets: Physio Net, which corresponds to health measurements from ICU patients, and USHCN, a climatology network dataset reflecting environmental conditions such as precipitation and temperature. ... The United States Historical Climatology Network (USHCN) (Menne et al., 2016) is a climate dataset ... Physio Net (Silva et al., 2012) is a dataset ... |
| Dataset Splits | Yes | Additionally, we conduct experiments on irregularly sampled time series, achieved by randomly removing fixed percentages of values from each time series. We create the datasets by removing 30, 50 and 70% of each time series. ... In Table 25 we show the evaluation of the approaches while randomly splitting data between training set (i.e., 80% of dataset) and test set (i.e., 20% of dataset). |
| Hardware Specification | Yes | Hardware and Software We implement our method in Python and the experiments are ran using a g4dn.2xlarge AWS instance with a NVIDIA T4 GPU, 8 CPU and 32gb of RAM. |
| Software Dependencies | No | We implement our method in Python and the experiments are ran using a g4dn.2xlarge AWS instance with a NVIDIA T4 GPU, 8 CPU and 32gb of RAM. ... Here, we utilized the PyTorch implementation of the FFT to obtain the DFT for each signal. ... The paper mentions Python and PyTorch but does not provide specific version numbers for these or any other libraries or software dependencies required for replication. |
| Experiment Setup | Yes | Training The training of i HT is performed in three stages to improve stability: 1) we train the Trend Hyper Network, Trend Block for 100 epochs, computing the MSE loss between the ground truth time series y and the output of the block: L1 = P i yi ^ftr(t) 2. This leads to a smooth approximation of the time series, which we use as initial guess for the second stage. 2) We then train the Trend and Seasonality blocks together, computing the MSE reconstruction loss Lrec and the FFT loss LFFT between the ground truth and the added output of both TSnet blocks. 3) Finally, we train the three blocks together. Stage 1 training: Number of epochs: 100 Learning rate: 1e 3 Stage 2 training: Number of epochs: 150 Learning rate: 5e 5 Stage 3 training: Number of epochs: 150 Learning rate: 5e 5 Batch size: 256 lambda1 = 1.0 10 3 lambda3 = 1.0 10 2 We train Energy24, Stock24, Stock72, Stock360 and Solar Weekly datasets for 400 epochs, with Adam optimizer. For the NN5 daily, and Fred MD datasets we trained for 500 epochs, and for the Temperature Rain dataset we train i HT for 1500 epochs. |