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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Enhancing Time Series Forecasting through Selective Representation Spaces: A Patch Perspective
Authors: Xingjian Wu, Xiangfei Qiu, Hanyin Cheng, Zhengyu Li, Jilin Hu, Chenjuan Guo, Bin Yang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To conduct comprehensive and fair comparisons for different models, we conduct experiments on eight well-known forecasting benchmarks as the target datasets, including ETT (4 subsets), Weather, Electricity, Solar, and Traffic, which cover multiple domains see Table 1. ... We compare SRSNet against state-of-the-art models in recent years... We adopt Mean Squared Error (mse) and Mean Absolute Error (mae) as evaluation metrics. ... Comprehensive forecasting results are listed in Table 2... Ablation study of SRS To demonstrate the SRS s capablities of working as a modular plugin to enhance the prediction accuracy of patch-based models, we combine SRS with naive MLP, i.e., SRSNet, classic models Crossformer [Zhang and Yan, 2022] and Patch TST [Nie et al., 2023], and recent state-of-the-art models x Patch [Stitsyuk and Choi, 2025b] and Patch MLP [Tang and Zhang, 2025a]. |
| Researcher Affiliation | Academia | Xingjian Wu, Xiangfei Qiu, Hanyin Cheng, Zhengyu Li, Jilin Hu, Chenjuan Guo, Bin Yang East China Normal University EMAIL, EMAIL |
| Pseudocode | No | The paper describes the Selective Patching and Dynamic Reassembly techniques in Section 3, using textual descriptions and mathematical equations (e.g., Equation 1-14) along with illustrative figures (Figure 2 and Figure 3). However, it does not include a dedicated "Pseudocode" or "Algorithm" block, nor a structured step-by-step procedure formatted like code. |
| Open Source Code | Yes | The resources are available at https://github.com/decisionintelligence/SRSNet. |
| Open Datasets | Yes | Datasets To conduct comprehensive and fair comparisons for different models, we conduct experiments on eight well-known forecasting benchmarks as the target datasets, including ETT (4 subsets), Weather, Electricity, Solar, and Traffic, which cover multiple domains see Table 1. |
| Dataset Splits | Yes | Table 1: Statistics of datasets. ... ETTh1 Electricity 1 hour 14,400 7 6:2:2 Power transformer 1, comprising seven indicators such as oil temperature and useful load ... ETTm2 Electricity 15 mins 57,600 7 6:2:2 Power transformer 2, comprising seven indicators such as oil temperature and useful load ... Weather Environment 10 mins 52,696 21 7:1:2 Recorded every for the whole year 2020, which contains 21 meteorological indicators |
| Hardware Specification | Yes | All experiments of SRSNet are conducted using Py Torch Paszke et al. [2019] in Python 3.8 and executed on an NVIDIA Tesla-A800 GPU. |
| Software Dependencies | Yes | All experiments of SRSNet are conducted using Py Torch Paszke et al. [2019] in Python 3.8 and executed on an NVIDIA Tesla-A800 GPU. |
| Experiment Setup | Yes | Implementation Details To keep consistent with previous works, we adopt Mean Squared Error (mse) and Mean Absolute Error (mae) as evaluation metrics. We consider four forecasting horizon F: {96, 192, 336, 720} for all datasets. Since the size of the look-back window can affect the performance of different models, we choose the look-back window size in {96, 336, 512} for all datasets and report each method s best results for fair comparisons. ... The training process is guided by the MSE loss function and employs the ADAM optimizer. The initial batch size is set to 64, with the flexibility to halve it (down to a minimum of 8) in case of an Out-Of-Memory (OOM) issue. ... The common configurations of Scorers and Scorerr are 2 hidden layers with hidden dimension equals 128. |