17 Time-series data

 

This chapter covers

  • Time-series data
  • Performing outlier detection on time series

Though we’ve focused on tabular data so far in the book, another type of data you may often work with, and may need to perform outlier detection with, is time-series data. Time-series data is useful to look at in itself but is also a good example of an important concept in outlier detection: converting data from one format to another. Often, though certainly not always, tabular data can be treated as time series, and time-series data can usually be converted to table format. In general, with outlier detection, any item we examine may be typical in most ways but may, nevertheless, be unusual (and possibly unusual in an interesting way) in one or more other respects. To find these anomalies, we need to look at the data from different perspectives, and an important method to support this is converting data to another format. Although we’ll look specifically at time-series data, this concept can be extended to other types of data as well. First, though, we’ll take a closer look at what time-series data is and how it relates to the tabular data we’ve looked at so far.

17.1 Cross-sectional, panel, and time-series data

17.2 Types of time-series outliers

17.2.1 Extreme values

17.2.2 Contextual point outliers

17.2.3 Outliers based on decomposition

17.2.4 Ruptures

17.2.5 Outliers based on tabular representations

17.2.6 Forecasting-based time-series outlier detection

17.2.7 Unusual shapes of windows in timelines

17.2.8 Tabular outlier detection on windows of data

17.2.9 Mulivariate time-series outliers

17.3 Tools for time-series data

17.3.1 Anomaly Detection Toolkit

17.3.2 DeepOD

17.3.3 Prophet

Summary