Hierarchy driven time series forecasting

Patent metadata

Field	Value
Publication number	US-2024394522-A1
Application number	US-202318200642-A
Country	US
Kind code	A1
Filing date	May 23, 2023
Priority date	May 23, 2023
Publication date	Nov 28, 2024
Grant date	—

Abstract

A method for lightweight and efficient long sequence time-series forecasting and representation learning includes segmenting a time-series dataset from a plurality of sensors into a plurality of patches. The method further includes applying gated multilayer perceptron (MLP) mixing across different directions of the patched input time-series. The method further includes capturing local and global and interrelated correlations across the plurality of patches and within the plurality of patches. The method further includes applying a patch-time aggregated hierarchy to guide lowest-level predictions based on aggregated hierarchy signals at a patch-level. The method further includes chaining MLP-mixers in a patch length context aware hierarchy fashion to enhance time-series short and long-term correlation capture.

First claim

What is claimed is:

1 . A computer-implemented method comprising:
segmenting a time-series dataset from a plurality of sensors into a plurality of patches; applying gated multilayer perceptron (MLP) mixing across different directions of the patched input time-series; capturing local and global and interrelated correlations across the plurality of patches and within the plurality of patches; and applying a patch-time aggregated hierarchy to guide lowest-level predictions based on aggregated hierarchy signals at a patch-level.

2 . The computer-implemented method of claim 1 , wherein the MLP mixing is channel independent.

3 . The computer-implemented method of claim 1 , wherein the MLP mixing uses layers that are stacked in linear fashion.

4 . The computer-implemented method of claim 1 , wherein the MLP mixing uses layers that are chained in a patch length context aware hierarchy fashion.

5 . The computer-implemented method of claim 1 , wherein the MLP mixing is mixed with respect to patches and features.

6 . The computer-implemented method of claim 1 , further comprising a pretraining task of masking random patches.

7 . The computer-implemented method of claim 4 , further comprising reconstructing the masked random patches.

8 . The computer-implemented method of claim 1 , further comprising a downstream task of forecasting values of the sensors.

9 . The computer-implemented method of claim 1 , further comprising a downstream task of executing regression analysis regarding values of the sensors.

10 . The computer-implemented method of claim 1 , further comprising a downstream task of classifying values of the sensors into one of a variety of predetermined classifications.

11 . A system comprising:
a processor; and a memory in communication with the processor, the memory containing instructions that, when executed by the processor, cause the processor to:
segment a time-series dataset from a plurality of sensors into a plurality of patches;
apply gated multilayer perceptron (MLP) mixing across different directions of the patched input time-series;
capture local and global and interrelated correlations across the plurality of patches and within the plurality of patches; and
apply a patch-time aggregated hierarchy to guide lowest-level predictions based on aggregated hierarchy signals at a patch-level.

12 . The system of claim 11 , wherein MLP the mixing is channel independent.

13 . The system of claim 11 , wherein the MLP mixing is mixed with respect to patches and features.

14 . The system of claim 11 , wherein the MLP mixing uses layers that are either stacked in linear fashion or chained in a patch length context aware hierarchy fashion.

15 . The system of claim 11 , the memory containing additional instructions that, when executed by the processor, cause the processor to execute a pretraining task of masking random patches.

16 . The system of claim 15 , the memory containing additional instructions that, when executed by the processor, cause the processor to reconstruct the masked random patches.

17 . The system of claim 11 , the memory containing additional instructions that, when executed by the processor, cause the processor to execute a downstream task of forecasting values of the sensors.

18 . The system of claim 11 , the memory containing additional instructions that, when executed by the processor, cause the processor to execute a downstream task of executing regression analysis regarding values of the sensors.

19 . The system of claim 11 , the memory containing additional instructions that, when executed by the processor, cause the processor to execute a downstream task of classifying values of the sensors into one of a variety of predetermined classifications

20 . A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to:
segment a time-series dataset from a plurality of sensors into a plurality of patches; apply gated multilayer perceptron (MLP) mixing across different directions of the patched input time-series; capture local and global and interrelated correlations across the plurality of patches and within the plurality of patches; and apply a patch-time aggregated hierarchy to guide lowest-level predictions based on aggregated hierarchy signals at a patch-level.

Assignees

• IBM

Inventors

• Ekambaram Vijay
• Nguyen Nam H
• Jati Arindam
• Sinthong Phanwadee
• Dayama Pankaj Satyanarayan
• Kalagnanam Jayant R

Classifications

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0499 Primary

  Feedforward networks · CPC title