Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift

By Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jangho Choi, J. Choo. International Conference on Learning Representations, 2022

1,187 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/0e95008a8c49a4c2538aed62ff61977ff7b47ca5

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

1. Gerald Woo
2. Tian Zhou
3. Zhijian Xu
4. Mingsheng Long
5. Shiyu Wang
6. Jianmin Wang
7. Donghao Luo
8. Chenghao Liu

Citations in the corpus

• Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting (2022, methodology): “RevIN [17] introduces a two-stage instance normalization [33] that transforms model input and output respectively to reduce the discrepancy of each series.”
• TSMixer: An all-MLP Architecture for Time Series Forecasting (2023, methodology): “We apply reversible instance normalization (Kim et al., 2022) to ensure a fair comparison with the state-of-the-art PatchTST (Nie et al., 2023).”
• One Fits All: Power General Time Series Analysis by Pretrained LM (2023, methodology): “In addition to the layer norm utilized in pre-trained LM, we also incorporate a simple data normalization block, reverse instance norm Kim et al. (2022), to further facilitate knowledge transfer.”
• MOMENT: A Family of Open Time-series Foundation Models (2024, unknown): “We train all models for 2 epochs.”
• DCdetector: Dual Attention Contrastive Representation Learning for Time Series Anomaly Detection (2023, methodology): “The input multivariate time series in the Forward Process module is normalized by an instance normalization [38, 69] module.”
• ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis (2024, background): “Stationary Technique RevIN (Kim et al., 2021) is a special instance normalization for time series to mitigate the distribution shift between the training and testing data.”
• Unified Training of Universal Time Series Forecasting Transformers (2024, methodology): “While not visualized, (non-learnable) instance normalization (Kim et al., 2022) is applied to inputs/outputs, aligning with the current standard practice for deep forecasting models.”
• FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting (2022, methodology): “An optional add-on data normalization layer RevIN Kim et al. (2021) is introduced to further enhance the model’s robustness.”
• iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2023, methodology): “In our inverted version, the normalization is applied to the series representation of individual variate as Equation 2, which has been studied and proved effective in tackling non-stationary problems (Kim et al., 2021; Liu et al., 2022b).”
• A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2022, methodology): “This technique has recently been proposed to help mitigating the distribution shift effect between the training and testing data (Ulyanov et al., 2016; Kim et al., 2022).”
• FITS: Modeling Time Series with 10k Parameters (2023, methodology): “Value Zero padding and irFFT then revert this back to the time domain, with iRIN finally reversing the normalization.”
• Foundation Models for Time Series Analysis: A Tutorial and Survey (2024, unknown): “Common tokenization techniques include reversible instance normalization [50] that mitigates distribution shift, patching with channel independence strategy that effectively and efficiently extracts the time series context [70], as well as the joint usage of time series decomposition to explicitly…”
• Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (2023, methodology): “Each input channel X ( i ) is first individually normalized to have zero mean and unit standard deviation via reversible instance normalization (RevIN) in mitigating the time series distribution shift (Kim et al., 2021).”
• TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecasting (2023, background): “The input time series segment goes through reversible instance normalization (RevIN) [17].”
• TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval (2025, unknown): “We apply reversible instance normalization [48] to multivariate time series before tokenizing and embedding.”