Statcounter is a nice web traffic analysis tool. It collects ISP and geolocation data of visitors of a tracked site. The data is logged on the Statcounter site and can be downloaded by the tracked site’s owner in XLSX or CSV format. In this article I want to show how I managed to visualize geolocation data from the CSV log using R.First of all, I have to admit that I am a newbie in R. I’ve been using Statcounter for years, but have only one month experience in R. All my older scripts were written in awk, perl and bash: they could download and merge Statcounter logs and do some basic data visualization such as plotting bar charts with Gnuplot. I will refer to some of them below in this article, while you can find them in my project statcounter-utils.I discovered R when I decided to put all visits of my blog on a world map. I checked out a plethora of complete GIS solutions, but they all seemed to me unnecessarily heavy-weight and rigid. Then I read somewhere on Stackoverflow about Leaflet and R. This was excellent finding, because this promised programming, and I love to program!Below is the annotated solution of the world map visits with examples (in the statcounter-utils, the R code is located in a file named cities.r).
In lines 1–5 all required libraries are loaded: leaflet for the map, htmltools for function htmlEscape, plyr for function count, dplyr for function case_when, and tidyr for function separate. Function cities (lines 7–55) renders locations (they include not only cities, but also regions and countries, as it will be hinted on the legend of the map) on an interactive world map that shall open in a browser window. This function accepts a list of cities gcities tagged with a count. The gcities can be a file name or a data frame. Geolocation data with locations found in gcities is expected to be passed in parameter geocode: this can only be a CSV file. Other two parameters — len and FUN — define how many top-cities to put on the map and a custom subsetting function.Files for parameters gcities and geocode can be obtained with bash utility group_cities which can extract geolocation data from a Statcounter log and group cities by count. For geocoding, group_cities makes use of the Python Geocoder. Script group_cities can be found in the statcounter-utils.
group_cities -f cities_spells_fix.awk StatCounter-Log.csv > gcities.csv
group_cities -g -f cities_spells_fix.awk StatCounter-Log.csv > geocode.csv
A sample script cities_spells_fix.awk can also be found in the statcounter-utils. This is a manually crafted database of cities and regions synonyms, various transcriptions, misspellings, and apparent errors met in Statcounter logs: the script collapses all variants of a single location to a single value.Files gcities.csv and geocode.csv have schemes with headers Country;Region;City;Count and Country;Region;City;Longitude;Lattitude respectively. In lines 15–16 the data get merged by the first 3 fields (Country, Region, and City) and ordered by field Count from gcities. Then, in line 17, cities with wrong geocode data (more specifically, when field Longitude from geocode is null) get filtered out, and the custom subsetting function FUN is applied. Later, on line 21, top len cities from the survived after all the previous filters set are picked. Basic leaflet construction takes place in line 19. In lines 29–43 cities (as well as regions if there is no city in the record, and countries if there is neither a city nor a region in the record) are marked by circle markers and annotated with popups containing the name and the count. The size of a circle marker is a logarithmic function of the count, whereas its color depends on whether the location is a city or a region or a country. In lines 45–50 a legend with color circles is added to hint a user why circle markers have different colors. Putting circles on a legend is not a trivial task. Function circle_marker_to_legend_color in lines 82–92 accomplishes this using the fact that Leaflet legend’s parameter colors is hackable by supplying a specially crafted HTML code.Selecting cities from a Statcounter log and grouping them by count is a trivial task for R. In other words, there is no need to pass preliminary crafted file gcities.csv, but instead, it makes sense to create a data frame inside R. This makes also possible to apply yet more sophisticated subsettings to the original data because now we are getting access to all the fields in the log directly from R. But remember that we have to apply the cities spells database cities_spells_fix.awk. This seems to be the only complication for function cities_df defined in lines 58–73. This function reads a Statcounter log and returns the desired data frame with grouped cities. Its obscure parameters warn_suspicious and type correspond to whether the awk script should print on the stderr suspicious replacements, and what type of visits to select from the Statcounter log: the default value “page view” is what a user normally expects. A data frame returned from cities_df can be further subset by a custom function as it contains all original data fields. Function gcities (lines 75–80) collapses the data frame fields to the scheme with headers Country;Region;City;Count compatible with input of function cities, and orders the data by count.Let’s run a few examples in an R shell. For all of them, we have to load script cities.r and collect all page view visits from a Statcounter log StatCounter-Log.csv.
pv <- cities_df("StatCounter-Log.csv", "cities_spells_fix.awk")
pvC < gcities(pv)
Now let’s render all collected cities on a world map.
Here is how it looks in my browser.
Seems to be cluttered by myriads of circle markers (function cities printed  1920 cities rendered). No problem! The map is interactive (however, not in this blog) and can be zoomed (look at the buttons at the top-left corner). The legend at the bottom-right corner shows why the markers have different colors.Let’s put on a map cities from the Moscow region.
pvMosk <- pv[grepl("^(Moskva|Moscow)", pv$Region), ]
pvMoskC <- gcities(pvMosk)
In the next examples I won’t show maps any longer to not clutter the article.Render cities from Russian Federation only.
cities(pvC, "geocode.csv", FUN = function(x) grepl("Russia", x$Country))
Render top 10 cities all over the world with total visits from 10 to 20.
cities(pvC, "geocode.csv", 10, FUN = function(x) x$Count %in% 10:20)
Render all the cities with visits in year 2018.
pv2018 <- pv[grepl("^2018", pv$Date.and.Time), ]
Function cities looks good to me, except it seems to make sense to create geocode data frame separately and pass it to the function like gcities data frame. Perhaps I will implement this in the future. Other improvements may also include using library leaflet.extras for searching of marked cities. (Update: both improvements were implemented in the statcounter-utils.)Now let’s turn to the bar charts of cities. I said that I used Gnuplot for that. But R is capable of making them as well! The following solution (which is the rest of cities.r) makes use of ggplot2 and plotly. As such, lines
must be put on the top of the script.
Functions gcities.compound and gcountries (lines 1–14) accept an original data frame read from a Statcounter log and return a data frame with two columns: a location (a compound location comprised from pasted Country, Region and City columns, or from column Country) and the count. They are needed for input to function cities.plot which renders a plot in a browser window. Function cities.plot accepts additionally a title, a tops, and a width. The title and the width correspond to the title and the width of the plot, they can be unset. The tops is a set of counts to form a number of emphasis boxes which help to emphasize top cities on the plot.In the preamble of function cities.plot (lines 17–23) a number of (a bit empirical) factors are configured. The value of w0 corresponds to the standard width for saving a plot as a PNG image and calculating factors for text positions of bar labels, wf is the width factor, the mf is the ultimate width factor which takes into account not only the width of the plot, but also the extents of the Count axis (which depends on the length of the top bar), cw stands for character width, to — text offset — the final y (i.e. Count) position of a text label on the plot, ym is the y extent of the emphasis boxes. All these complicated calculations are needed because in ggplot2 there seems to exist no tool for binding a text label to a bar: there is an appropriate package ggrepel but it is not supported in plotly.The first iteration of the plot creation takes place in lines 25–40. We bind data directly to the plot (cs is an argument of ggplot), flip coordinates (line 29) as soon as bars must lie horizontally, say that the x axis must be discrete (line 26) to enable setting exact positions of the emphasis boxes borders, reverse the x scale (on the same line) to put cities with the same count alphabetically from the top to the bottom, and reorder cs by count (aes settings in line 25). Then put bars (line 30), bar labels (lines 31–32), and the title (line 40). In lines 33–39 not needed elements of x and y axes, as well as the grid, get removed.In lines 42–69 the emphasis boxes get applied on the plot as annotations. Notice that being annotations, they lie on top of the bars and the bar labels like colored glass sheets. This means that the boxes must be very transparent to not obscure the bars and the labels under them, and the colors of the bars must be close to the colors of the boxes. When tops is not specified (lines 42–44), a single fully transparent box is still applied: this is done to ensure that the lengths of the bars will keep the same independently of whether the tops specified or not. Alpha channels of the emphasis boxes decrease steadily from value 0.1 (line 47) by factor 2 (line 64). Each box is additionally annotated by a number at the bottom-right corner which corresponds to the position of the lowest bar it covers (lines 59–62). When the value NA (or any number greater than the number of all the bars) are met in the tops, the current emphasis box gets extended to the bottom of the plot, and the loop over the tops stops (lines 49–51 and 65–67).The height of the plot is limited by the maximum linear size allowed in Cairo (lines 71–73). Finally, in lines 74–78, the plotly toolbar gets configured. Particularly, button toImage (Download plot as a png when hovered) no longer proposes to save the image as newplot.png (lines 75–76), instead, it proposes a name based on the title of the plot. Besides this, the width of the PNG image gets 1200 pixels when saved in case if it has not been specified in the width parameter of function cities.plot (line 78).Now let’s render a bar chart of all visits from the Moscow region with top 10, top 40 and top-to-the-end emphasis boxes.
pvMoskCc <- gcities.compound(pvMosk)
cities.plot(pvMoskCc, paste("Moscow region", format(Sys.time(), "(%F %R)")), c(10, 40, NA), 1200)
I cut out a piece of the image (the white lacuna across its lower part) to conform to the limitations on image sizes in this blog.