Distributed Hierarchical Decomposition Framework for Multimodal Timeseries Prediction

Authors: Wei Ye, Prashant Khanduri, Jiangweizhi Peng, Feng Tian, Jun Gao, Jie Ding, Zhi-Li Zhang, Mingyi Hong

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate the performance of the proposed algorithm on real-world datasets with a number of time-series forecasting baselines implemented in both centralized and distributed manners. [...] Datasets. Our evaluation is conducted on a number of diverse real-world datasets: [D1] Power demand dataset (POWR) (EIA, 2022) [...] Evaluation metrics. For the multivariate time-series forecasting task, we utilize root mean square error (RMSE), which gauges the disparity within Euclidean space. For the multimodal time-series prediction with the Deep Sense 6G dataset, we use the distance-based accuracy (DBA) score; see the appendix for a detailed illustration.
Researcher Affiliation Collaboration University of Minnesota Twin Cities; Wayne State University; Meta Platforms, Inc.
Pseudocode Yes Algorithm 1 DIVIDE with label sharing [...] Algorithm 2 DIVIDE without label sharing
Open Source Code No Softwares assets. Our codes mainly rely on the following third-party libraries and all of them can be easily installed: Numpy1, scikit-learn2, Pytorch3, Py G4, multi-prophet5, DTAIDistance6, Pandas7, Matplotlib8.
Open Datasets Yes Datasets. Our evaluation is conducted on a number of diverse real-world datasets: [D1] Power demand dataset (POWR) (EIA, 2022), which contains the hourly power demand of thirteen major electricity grid service regions in the United States for 2022; [D2] NYC subway traffic dataset (SUBW) (EDDEN, 2021), which collects the number of people entering and exiting each subway station per hour in NYC, 2017 2021; [D3] Meteorological dataset (WEAT) (BENIAGUEV, 2017), which reports weather data from 36 regions with different indicators from Oct. 2012 to Nov. 2017. [D4] Deep Sense6G (DS6G) (Charan et al., 2022), which is a multimodal dataset containing sensory (e.g., camera, Radar, LiDAR) and radio information. Same with the ITU 2022 challenge (ITU, 2022), we use four scenarios in experiments.
Dataset Splits Yes Hyperparameters Settings. For all the experiments of neural network(NN)-based models, we split the datasets into the training, validation, and test sets with the ratio of 0.5 : 0.2 : 0.3.
Hardware Specification Yes Hardwares. Our workstation runs on the Unbuntu 18 system and is equipped with AMD Ryzen Threadripper PRO 3995WX CPU, 1TB memory, and three Nvidia RTX A6000 GPUs (48GB memory each).
Software Dependencies No Softwares assets. Our codes mainly rely on the following third-party libraries and all of them can be easily installed: Numpy1, scikit-learn2, Pytorch3, Py G4, multi-prophet5, DTAIDistance6, Pandas7, Matplotlib8.
Experiment Setup Yes Hyperparameters Settings. For all the experiments of neural network(NN)-based models, we split the datasets into the training, validation, and test sets with the ratio of 0.5 : 0.2 : 0.3. The min-max scaler is applied for the data normalization. All models are trained using the Adam (Kingma & Ba, 2014) optimizer with a batch size of 128. If there is no additional description, we follow the baseline model s default settings but set 64 hidden cells for all positions. Lastly, we ensure that the training is long enough so that the models converge well; we repeat each of the experiments three times, and report the average results.