TIMESIGHT: Discovering Time-driven Insights Automatically and Fairly

EasyChair Preprint 2925

Abstract

Exploratory data analysis (EDA) on time-series data is an indispensable and important process for not only data analysts but also non-expert users. It helps them make data-driven decisions by discovering important patterns of a certain phenomenon. However, it poses 2 challenges for data analysts and decision-makers. First, although a lot of business intelligence tools have been introduced that can help explore the data, they require repeated analytic procedures and most of the procedures rely on users intuition, knowledge and efforts. Second, even though there have been several attempts to quantify insights to automatically detect interesting patterns, they do not consider score fairness among detected patterns. Therefore, they are not suitable when data has the heterogeneity of insight types, attributes scales and time intervals. We attack these challenges by introducing our new proposed system Timesight, which explores data through all possible time units and all attributes automatically. Timesight evaluates various types of time-driven insight, matching the fairness among each type of insight, each attribute, and each time interval. We verify our system using internal application log dataset. Our experiment with data analysts working the same dataset shows that Timesight alleviates the tedious work and is effective in discovering insight.

Keyphrases: Data Mining, Insight discovery, data exploration, time series data.

@booklet{EasyChair:2925,
  author    = {Yohan Bae and Suyeong Lee and Yeonghun Nam},
  title     = {TIMESIGHT: Discovering Time-driven Insights Automatically and Fairly},
  howpublished = {EasyChair Preprint 2925},
  year      = {EasyChair, 2020}}