Rethinking Time Encoding via Learnable Transformation Functions
Authors: Xi Chen, Yateng Tang, Jiarong Xu, Jiawei Zhang, Siwei Zhang, Sijia Peng, Xuehao Zheng, Yun Xiong
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments across diverse domains including event-based image classification, time series forecasting, dynamic graph representation learning and real-world applications-we demonstrate the effectiveness and versatility of Le TE. |
| Researcher Affiliation | Collaboration | 1Shanghai Key Laboratory of Data Science, College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China 2Tencent Weixin Group, Shenzhen, China 3School of Management, Fudan University, Shanghai, China 4IFM Lab, University of California, Davis, CA, USA. |
| Pseudocode | No | The paper does not contain explicitly labeled pseudocode or algorithm blocks. Appendix D provides implementation details in a mathematical/formulaic format rather than structured, step-by-step pseudocode. |
| Open Source Code | Yes | The codes are available at a Git Hub Repository. |
| Open Datasets | Yes | We conduct long-term forecasting tasks on these baseline models using 4 datasets: ETT, Weather, Exchange (Lai et al., 2018), and Electricity... ETT2 dataset includes time series data... Electricity3 dataset provides hourly electricity consumption data... Exchange (Lai et al., 2018) dataset offers daily panel data... Weather4 dataset contains meteorological time series data... We conduct link prediction experiments on 4 realworld datasets: Wikipedia, Reddit, MOOC, and Last FM (Kumar et al., 2019)... 2https://github.com/zhouhaoyi/ETDataset 3https://archive.ics.uci.edu/ml/datasets/Electricity Load Diagrams20112014 4https://www.bgc-jena.mpg.de/wetter/ |
| Dataset Splits | Yes | The past sequence length is set to 96, while the prediction lengths are {96, 192, 336, 720}. The results are reported in terms of MAE... Table 3. Time Series Dataset Statistics: The dataset size is organized in (Train, Validation, Test). Please refer to (Wu et al., 2023) for the original table. ETTm1, ETTm2 ... (34465, 11521, 11521) ... Electricity ... (18317, 2633, 5261) ... Exchange ... (5120, 665, 1422) ... Weather ... (36792, 5271, 10540) |
| Hardware Specification | No | The paper mentions that "the experiments were conducted locally on Tencent s server" for a real-world application, but this is a general statement and does not specify any particular hardware details such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions using deep learning frameworks and various models like LSTM, Transformer, Pyraformer, etc., but does not provide specific software dependency names with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x, Python 3.x). |
| Experiment Setup | Yes | We then apply an LSTM with a 32-dimensional learnable embedding for the time input... The backbone model we used is a 128-dimensional LSTM, with a batch size of 512, aligned with the settings in (Kazemi et al., 2019). The past sequence length is set to 96, while the prediction lengths are {96, 192, 336, 720}. The hyper-parameters are based on the best configurations reported in the papers, and we keep them unchanged across different experiments for each baseline model to ensure a fair comparison. We rerun the baseline models TGAT with batch size 100 and reuse the baseline results reported in the Dy GFormer paper for other baselines. |