時間序列資料的大致走向通常稱之為「趨勢線」，然而趨勢線未有精準描述的定義，每個人心中對趨勢線的形狀認知有些許差異，難以用一種趨勢線滿足所有人。另外個別使用者可能也不容易清楚敘述其心中的趨勢線樣貌。 本論文提出一個框架讓個別使用者以「手繪」的方式在十張時間序列資料上標出他認定的趨勢線，讓機器學習模型從中學習該使用者心中的趨勢線樣貌，以應用在其他時間序列資料上。;The tendency of a time series is usually referred to as a “trend line”. However, the precise definition of a trend line is still ambiguous. Given a time series, different users may come up with varying shapes of trend lines – some may prefer smooth lines, while others may hope the trend line responds to local turbulence. Therefore, a single trend line definition is challenging to meet everyone’s needs. Meanwhile, it could be complicated for users to clearly describe the requirements of a trend line in their minds. This thesis proposes a framework to learn the customized trend lines that meet users’ demands. First, the framework asks users to plot the expected trend lines on ten time-series datasets. The framework then learns users’ preferred shapes and automatically draws the customized trend lines for other time-series datasets.
