Visualisation: Modelling the World

• This talk is an approach to visualisation
• Not many absolutes
• assumptions of vision
• Assumptions of Statistical Graphics
• Understanding data with Visualisation
• Communicating to others with Visualisation

• a tool for understanding the world
• a way to communicate a particular perspective on data
• an adjunct to thought

• The eye is really really good at finding patterns in pictures
• in fact, it's so good that it can find patterns that aren't even

there

m

• You can see one of two things in the previous image
• Which of them can depend on what you expect to see
• It can also depend on what your environment contains

• Europeans and Americans are more susceptible
• Africans are less susceptible
• Possibility that it is due to presence of right angles in urban environments
• appears to be a small difference between urban and rural dwellers
• very very relevant to boxplots (how to lie with boxplots, I guess)

• Shows that how we interpret stimuli is not tabula rasa
• When you gaze into the image, the image also gazes into you…
• We bring our own perception and previous associations into any image ¹

• Property Price Register
  • Kinda a crappy dataset
  • No cleaning or checking done by the authority
  • lots of craziness (1 apartment for 18.6mn)

• We used Google's geocoding service to get more details on each observation
• I updated Shane Lynn's script and ran it on the data up till October 2018
• I also typically break out properties sold for greater than 1e6, as they are often multiple-unit sales (and there's little to no automated way of figuring this out) ²
• Lots of manual fixing required
• the irish text definitely doesn't help

• there are many
• in this section, I'd like to subvert them, in order to make you think

• Normally represent time
• scatterplots don't (always) have the same assumptions
• what is the deepest assumption?

• Here, the violence is that we swap the axes in a fashion only a monster would

• Also encodes a set of base assumptions
• points nearer to each other in space are more related
• more orientation issues

• We have a base level of assumptions that we bring to graphics (especially statistical graphics)
• Most of these appear to have been formed by Descartes
• When these assumptions are subverted, expect problems

• Graphs excel at showing relations between things
• Consider the difference between quantiles of a variable, and a density plot
• For example, the price of houses:

• Useful to get a better sense of the data
• Have a bunch of assumptions (what's the log of -1)
• Can be used to deceive very, very easily
• Really really useful in everyday practice

• Picking the right visualisation for the data is important

• is this a good plot?
• does this depend on the number of points?

• Let's say we remove all properties with prices greater than 2mn

• Better or worse?

• Note the log 10 base
• Some of you may be able to convert from base 2.718, but I missed that class in school
• Still crap though

• The first obvious thing is to split by county, right?

• Oh look, it's lot of little boxes of crap :(

• The obvious answer is summarisation

• But really just washes the whole thing out

• We just group by a higher level variable

• Much clearer :)

• This is one of the major advantages of visualisation:
  • it helps to (dis)confirm your assumptions
  • given that we have too many lines in the various groupings,we know that somethng has gone horribly wrong
  • in this case, it's a mismatch between two different types of data

• This actually works (for me, at least)
• can you explain this to a sales-person?

• This is much, much better
• I definitely don't think I'd try to explain it to a business/sales person

• line plots vs maps
• time versus space
• both provide insight into
• pick one, difficult to do both

• There's a real problem of scale here, in that Dublin City is both responsible for much of the population, but is invisible

• Outliers make the map useless

• A tiny proportion of electoral districts drive the uselessness of the maps

• Just doesn't work
• Even when I account for the outliers, it still doesn't work.

• This shows the trend plus outliers
• Much more useful
• lose the spatial dimension

• Can show both time and space
• for reporting, these are essential
• Much more effort from a software-engineering perspective ³

• Two types of graphs:
  • for yourself
  • for other people (and different audiences need different things)

• These are used to help you understand a problem
• typically created in an iterative fashion
• often move from data transformation to visualisation and back again (like this talk)

• To some extent, your job with presentation visualisations is to tell a story
• hopefully, it will be nuanced, but that isn't a requirement ⁴
• Often good to show smooths as opposed to raw data
• raw data is often ugly
• need for care here, as this should only be done where there is a clear effect

• As few as possible
• One clear message
• Repeat yourself
• Remove nuance

• There should be no extraneous graphs
• Each graph should have a clear purpose
• Smooths are really effective

• You should only be telling one story at a time
• People are easily confused
• Especially in an oral presentation
• Backup docs should contain nuance

• This is the key to helping people retain information
• This is easier once you know the story
• Say what you want to say, say it, then say what you said

• This varies by audience
• Salespeople may just want the results
• colleagues may want to see the code
• most people just want a high level explanation
• Nuance should be present, just not in a presentation

• Everyone bring assumptions to visualisations
• Make sure that you take advantage of this
• Visualisation is primarily a tool for communicating with yourself
• Iterative process, even bad graphs can teach you something
• Secondarily, it's a tool for communicating with others
• When using visualisations with others, keep it simple

• My property article here
• My repository for this talk
• My crazy long notes file with most of my analyses
• the data itself