The transformer has been revolutionary in multiple fields. Time series forecasting is one such field and there is a variety of research focused on different aspects of its application. However, to the best of our knowledge, there is no research on how to methodically adapt the transformer architecture to a given time series based task. As such, we create a guide to aid in building task specific adaptations with the goal of increasing a model’s potential. We do this by first describing the elements of the transformer and the intent behind them in an intuitive way. Then we create a generic solution process for relevant tasks and condense it into a short but versatile guide. To show how we intend the guide to be used, we apply it to a simple task. Our guide highlights potential research opportunities for increasing the transformer’s potential in this field.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A Guide for the Systematic Adaptation of the Transformer to Specific Time Series Based Tasks

  • Lukas Helmstaedt,
  • Gernot Frank,
  • Marius Pflüger,
  • Marc Gruber,
  • Marc Roder

摘要

The transformer has been revolutionary in multiple fields. Time series forecasting is one such field and there is a variety of research focused on different aspects of its application. However, to the best of our knowledge, there is no research on how to methodically adapt the transformer architecture to a given time series based task. As such, we create a guide to aid in building task specific adaptations with the goal of increasing a model’s potential. We do this by first describing the elements of the transformer and the intent behind them in an intuitive way. Then we create a generic solution process for relevant tasks and condense it into a short but versatile guide. To show how we intend the guide to be used, we apply it to a simple task. Our guide highlights potential research opportunities for increasing the transformer’s potential in this field.