Concatenation-Informer: Pre-Distilling and Concatenation Improve Efficiency and Accuracy
Jie Yin, Meng Chen, Chongfeng Zhang, Ming Zhang, Tao Xue, Tao Zhang
Abstract
Jie Yin, Meng Chen, Chongfeng Zhang, Ming Zhang, Tao Xue, Tao Zhang
Abstract
Time series widely exist in the real world, and a large part of them are long time series, such as weather information records and industrial production information records. The inherent long-term data dependence of long-time series has extremely high requirements on the feature extraction ability of the model. The sequence length of long time series also directly causes high computational cost, which requires the model to be more efficient. This paper proposes Concatenation-Informer containing a Pre-distilling operation and a Concatenation-Attention operation to predict long time series. The pre-distilling operation reduces the length of the series and effectively extracts context-related features. The Concatenation-Attention operation concatenates the attention mechanism's input and output to improve the efficiency of parameters. The total space complexity of the Concatenation-Informer is less than the complexity and usage of the Informer.
A significance statement is not available in the OpenAlex record.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Time series widely exist in the real world, and a large part of them are long time series, such as weather information records and industrial production information records. The inherent long-term data dependence of long-time series has extremely high requirements on the feature extraction ability of the model. The sequence length of long time series also directly causes high computational cost, which requires the model to be more efficient. This paper proposes Concatenation-Informer containing a Pre-distilling operation and a Concatenation-Attention operation to predict long time series. The pre-distilling operation reduces the length of the series and effectively extracts context-related features. The Concatenation-Attention operation concatenates the attention mechanism's input and output to improve the efficiency of parameters. The total space complexity of the Concatenation-Informer is less than the complexity and usage of the Informer.
Key concepts: Concatenation (mathematics), Computer science, Series (stratigraphy), Context (archaeology), Computational complexity theory, Sequence (biology), Time series, Time complexity