5.2 Future work

5.2.1 Process for generating missing data in time

Missing values in multivariate data are typically characterized by the overall, row-wise, and column-wise numbers of missings. However, none of these captures the dynamics in temporal data. A well-defined characteristic is need to characterize temporal missingness, and this could possibly shed light on the processes for generating and imputing missing data in time.

Generating temporal missingness can be decomposed into two steps: (1) injecting missings at time points to reflect the functional form of time, and (2) generating the corresponding run lengths to reflect the temporal dependency. I plan to expand on Chapter 4 to generalize missing data generating processes in temporal contexts. Because of the evolving nature of time, the underpinning mechanisms of missing data may change from one period to another. Applying the new characteristic to the data, on a rolling window basis, could indicate the missing data status and thus lead to appropriate missing data remedies.

5.2.2 Visual methods for temporal data of nesting and crossing interactions

A collection of time series are often structured in a way that allows nesting and crossing interactions (Hyndman and Athanasopoulos 2017). For example, a manufacturing company can add up every store’s sales by region, by state and by country, which gives a strictly hierarchical time series; alternatively, they can gather the sales based on common attributes such as store, brand, price range and so forth, which leads to a crossed configuration. Nesting is a special case of crossing, with parent-children relations involved. Temporal information such as date-times is often also intrinsically hierarchical, seconds nested within minutes, hours, and etc. The new tsibble structure has the neat capability of supporting these structural embeddings.

Numerous nesting and crossing combinations can yield unwieldy plots, in many of which an abundance of information is possibly buried. Focus-plus-context visualization with interactivity comes to the rescue. Dual contexts, structurally informative subjects, and time provide the source and visual clues for elegant navigation. Interactions on contextual plots control what is to be visualized in the main plots. Many kinds of visual displays can be generated to progressively build a richer data picture through guided or self explorations.