Recently, several studies show the powerful capability of neural networks to capture non-linear features from time series which have multiple seasonal patterns. However, existing methods rely on convolution kernels implicitly, hence neglect to capture strong long-term patterns and lack interpretability. In this paper, we propose a memory-augmented neural network named AutoRegressive Memory Network (ARMemNet) for multivariate time series forecasting. ARMemNet utilizes memory components to explicitly encode intense long-term patterns. Furthermore, each encoder is designed to leverage inherently essential autoregressive property to represent short-term patterns. In experiments on real-world dataset, ARMemNet outperforms existing baselines and validates effectiveness of memory components for complex seasonality which is prevalent in time series datasets.
|Title of host publication||Proceedings of the 36th Annual ACM Symposium on Applied Computing, SAC 2021|
|Publisher||Association for Computing Machinery|
|Number of pages||4|
|Publication status||Published - 2021 Mar 22|
|Event||36th Annual ACM Symposium on Applied Computing, SAC 2021 - Virtual, Online, Korea, Republic of|
Duration: 2021 Mar 22 → 2021 Mar 26
|Name||Proceedings of the ACM Symposium on Applied Computing|
|Conference||36th Annual ACM Symposium on Applied Computing, SAC 2021|
|Country||Korea, Republic of|
|Period||21/3/22 → 21/3/26|
Bibliographical noteFunding Information:
is work was supported by Institute of Information ? communications Technology Planning ? Evaluation (IITP) grant funded by the Korea government (MSIT) (IITP-2017–0–00477)
© 2021 Owner/Author.
All Science Journal Classification (ASJC) codes