时间序列的可解释性

时间序列的可解释性

Chen Y, Huang S. TSExplain: Surfacing Evolving Explanations for Time Series[C]//Proceedings of the 2021 International Conference on Management of Data. 2021: 2686-2690.

Abstract

Understanding the underlying explanations for what has happened is more and more crucial in today’s business decision-making processes.

在今天的商业决策过程中，理解对所发生的事情的基本解释越来越重要。

Existing explanation engines focus on explaining the difference between two given sets.

现有的解释引擎着重于解释两个给定集合之间的差异。

However, for time-series, the explanations usually evolve as time advances. Thus, only considering two end timestamps would miss all explanations in between. To mitigate this, we demonstrate TSExplain, a system to help users understand the underlying evolving explanations for any aggregated time-series.

然而，对于时间序列来说，解释通常是随着时间的推移而变化的。因此，只考虑两个结束的时间戳会错过中间的所有解释。为了缓解这个问题，我们展示了TSExplain，一个帮助用户理解任何聚集的时间序列的基本演变解释的系统。

Internally, TSExplain models the explanation problem as a segmentation problem over the time dimension and uses existing works on two-sets diff as building blocks.

在内部，TSExplain将解释问题建模为时间维度上的分割问题，并使用现有的关于两组差异的工作作为构建模块。

In our demonstration, conference attendees will be able to easily and interactively explore the evolving explanations and visualize how these explanations contribute to the overall changes in various datasets: COVID-19, S&P500, Iowa Liquor Sales.

在我们的演示中，会议与会者将能够轻松地、互动地探索不断发展的解释，并直观地看到这些解释对各种数据集的整体变化的贡献。COVID-19, S&P500, Iowa Liquor Sales.

Questions—like “which states make COVID-19 total confirmed case number go up dramatically during the past year?”, “which stocks drive the dramatic crashes of S&P500 in Mar and the quick rebound later?”, and “how does Liquor sales’
trend look like from January 2020 till now and why”—can all get well answered by TSExplain.

像 “在过去一年中，哪些州使COVID-19的总确诊病例数急剧上升？”，“哪些股票推动了3月份S&P500的急剧崩溃和随后的快速反弹？”，以及 "从2020年1月至今，酒类销售的趋势是怎样的，为什么？"这些问题都可以通过TSExplain得到很好的回答。