Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting
Authors: Yifan Hu, Peiyuan Liu, Peng Zhu, Dawei Cheng, Tao Dai
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments demonstrate our AMD framework not only overcomes the limitations of existing methods but also consistently achieves state-of-the-art performance across various datasets. |
| Researcher Affiliation | Academia | Yifan Hu1,2,*, Peiyuan Liu1,*, Peng Zhu2, Dawei Cheng2, , Tao Dai3, 1Tsinghua Shenzhen International Graduate School 2Tongji University 3Shenzhen University |
| Pseudocode | No | The paper describes the proposed method in prose and mathematical formulas, but it does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Code https://github.com/TROUBADOUR000/AMD |
| Open Datasets | Yes | We conduct experiments on seven real-world datasets, including Weather, ETT (ETTh1, ETTh2, ETTm1, ETTm2), ECL, Exchange, Traffic and Solar Energy for long-term forecasting and PEMS (PEMS03, PEMS04, PEMS07, PEMS08) for short-term forecasting. |
| Dataset Splits | No | The paper specifies input lengths (e.g., "For short-term forecasting, the input length is 96." and "the input sequence length L is searched among {96, 192, 336, 512, 672, 720}") and prediction lengths ("prediction length T is 12") but does not provide explicit training/test/validation dataset splits (e.g., percentages, sample counts, or specific predefined split citations). |
| Hardware Specification | Yes | All experiments are conducted using Py Torch on an NVIDIA V100 32GB GPU and are repeated five times for consistency. |
| Software Dependencies | No | The paper mentions "Py Torch" as a software dependency but does not specify a version number. |
| Experiment Setup | Yes | To ensure fair comparisons, for long-term forecasting, we rerun all baselines with different input lengths L and choose the best results to avoid underestimating the baselines. For short-term forecasting, the input length is 96. We select two common metrics in time series forecasting: Mean Absolute Error (MAE) and Mean Squared Error (MSE). To strike a balance, we set the number of predictors to 8 across all experiments. The total loss function is defined as: L = Lpred + λ1Lselector + λ2 Θ 2 where Θ 2 is the L2-norm, λ1,2 are hyper-parameters. |