Conversations in time:
interactive visualization to explore structured temporal data
Earo Wang
Time series data is rich
1. often intrinsically hierarchical
2. different temporal frequencies
- Nowadays, we often work with a collection of time series instead of univariate series.
- These time series can often be organised in a hier way, e.g. US > states > counties. Here, AU > States > tourism regions
- You collect data at the most disaggregated levels, and sum them up to higher levels.
- A deeper hier, more data available.
- On the other hand, temporal context itself generates loads of information.
- TS data can be break down to different temporal frequencies.
- This chart shows the Australian travelled to Sydney for the Holiday purpose: high in Jan
- But what's happened to Sep-Oct 2000? Yea, 2000 Sydney Olympics
Time series data is rich
1. often intrinsically hierarchical
2. different temporal frequencies
- Now we should look forward to 2032 Brisbane Olympics and work on forecasting the number of visits to Brisbane in 2032 using Sydney's data.
- information are embedded into temporal and structural context.
- Given such rich data, how should we start exploring time series?
Domestic trips in Australia š¦šŗ
library(tsibbletalk)tourism_monthly#> # A tsibble: 80,696 x 5 [1M]#> # Key: State, Region, Purpose [308]#> Month State Region Purpose Trips#> <mth> <chr> <chr> <chr> <dbl>#> 1 1998 Mar ACT Canberra Business 111. #> 2 1998 Apr ACT Canberra Business 93.1#> 3 1998 May ACT Canberra Business 78.1#> 4 1998 Jun ACT Canberra Business 44.3#> 5 1998 Jul ACT Canberra Business 129. #> 6 1998 Aug ACT Canberra Business 71.3#> # ā¦ with 80,690 more rows- first and foremost, we need to arrange ts into a tsibble object.
- here we unlock the domestic trips for lockdown Australia
- What is a tsibble: a tibble represents time series data.
- header
- The idea of key variables is powerful: we know what each time series represents
- The key provides a central hub for identifying and linking each series in different tables.
- You may wonder why tsibble? bc we have a growing ecosystem for the tsibble data structure.
- here tidyverts doesn't mean advocates for the tidyverse; instead it means tidyverse for time series.
Sad charts š
- a tsibble is a data frame behind the scene, we can ggplot it.
- Overplotting, not particular useful
- upward trends for business, and strong seasonal effects for holidays
Viz š š on a feature space
library(feasts)tourism_monthly %>% features(Trips, feat_stl)#> # A tibble: 308 Ć 12#> State Region Purpose trend_strength seasonal_strengā¦#> <chr> <chr> <chr> <dbl> <dbl>#> 1 ACT Canbeā¦ Busineā¦ 0.316 0.332#> 2 ACT Canbeā¦ Holiday 0.220 0.542#> 3 ACT Canbeā¦ Other ā¦ 0.212 0.215#> 4 ACT Canbeā¦ Visitiā¦ 0.262 0.429#> 5 New South Wales Blue ā¦ Busineā¦ 0.185 0.247#> 6 New South Wales Blue ā¦ Holiday 0.426 0.471#> # ā¦ with 302 more rows, and 7 more variables:#> # seasonal_peak_year <dbl>, seasonal_trough_year <dbl>,#> # spikiness <dbl>, linearity <dbl>, curvature <dbl>,#> # stl_e_acf1 <dbl>, stl_e_acf10 <dbl>- The
features()computes seasonal and trend features for all time series at once. - Time series data collapsed to a bunch of descriptive statistics.
- The key variables are there that we can link back to the original tsibble.
Viz š š on a feature space
- each point represents a time series.
Crosstalk between lines and points
- a tisbble -> lines
- the featured table -> scatterplot
- what they share in common are the key identifiers
- for example, we want to highlight SYD and see how they behave on the feature space or vice versa.
- Or find which series corresponds to the highest in trend but lowest in seasonality.
- Static vis stop us doing this, bc writing lots of code, but this is a natural desire for interactive vis.
- When we think about interactive viz in R, oh probably build a Shiny app. but shiny requires for UI and logics in server side.
- What about if without shiny. {crosstalk} for linking between different html widgets.
Let tsibble talk
Syntax sugar: nesting and crossing
tourism_shared <- tourism_monthly %>% as_shared_tsibble(spec = (State / Region) * Purpose)tourism_shared#> # A tsibble: 80,696 x 5 [1M]#> # Key: State, Region, Purpose [308]#> Month State Region Purpose Trips#> <mth> <chr> <chr> <chr> <dbl>#> 1 1998 Mar ACT Canberra Business 111. #> 2 1998 Apr ACT Canberra Business 93.1#> 3 1998 May ACT Canberra Business 78.1#> 4 1998 Jun ACT Canberra Business 44.3#> 5 1998 Jul ACT Canberra Business 129. #> 6 1998 Aug ACT Canberra Business 71.3#> # ā¦ with 80,690 more rows- Turn a normal data frame to a mutable data object for easy interactions.
- spec for specifying structures
hierarchy -> tree
p_l <- plotly_key_tree(tourism_shared, height = 1100, width = 800)p_l
Shared tsibble %>% {ggplot2} & {plotly}
p_tr <- tourism_shared %>% ggplot(aes(x = Month, y = Trips)) + geom_line(aes(group = Region)) + facet_wrap(~ Purpose, scales = "free_y")p_br <- tourism_shared %>% features(Trips, feat_stl) %>% ggplot(aes(x = trend_strength, y = seasonal_strength_year)) + geom_point(aes(group = Region))subplot(p_l, subplot( ggplotly(p_tr, tooltip = "Region", width = 700, height = 800), ggplotly(p_br, tooltip = "Region", width = 700, height = 800), nrows = 2), widths = c(.4, .6)) %>% highlight(dynamic = TRUE)Interactive w/o shiny, enabling quick exploration
Slicing and dicing
Wrapping
- overview
Wrapping
- overview
- time of day
Wrapping
- overview
- time of day
- day of week
A shiny module
library(shiny)p_line <- pedestrian20 %>% ggplot(aes(x = Date_Time, y = Count, colour = Lockdown)) + geom_line(size = .3) + facet_wrap(~ Sensor, scales = "free_y") + labs(x = "Date Time") + scale_colour_brewer(palette = "Dark2") + theme(legend.position = "none")ui <- fluidPage( tsibbleWrapUI("dice"))server <- function(input, output, session) { tsibbleWrapServer("dice", ggplotly(p_line, height = 700), period = "1 day")}shinyApp(ui, server)
Acknowledgements
- {crosstalk}
- {plotly}
- {shiny}