Following on from my discussions of the design ideas of Edward Tufte and before my discussion on mathematics, maths and graphs I give you my 5 top tips for excellent graphs!
1. Decide your Statement: Absolutely essential! What is it you want to say about the data set you are discussing? Although graphs can show a huge amount of detail and provide fantastic insights into the behaviours being investigated, it is essential when presenting a graph to have one idea of the statement you would like your audience to take away from it. You might find they take away more, but they have to take away at least that argument. Apply this statement to your graph and make sure that this is what it says, remove anything in the graph that distracts from it.
2. Could a table represent the data better?: Why are you including this graph? Just because you have a nice data set and think a bar chart would break things up? Because you want to show off the fact that you have worked out how to do histograms?! NO! STOP! Your job as a data scientist is not just to produce indefinite numbers of graphs so your inner quota is fulfilled. Your job is to analyse the data on hand and allow your audience to fully understand your work. A table may well be better for this than a graph.
3. Remove chart junk: Remove all those grid-lines! why do you have all those colours? Remove background colours, data point colour variations, data point shape variations. Remove data labels! Simplify your legend!
4. 3D Visualisations: 3D visualisations should only be used to represent things that are actually 3D! If you’re showing an MRI of a patient or temperature fluctuations throughout your manifold go ahead. But 3D visualisations should not be used to make things look cool, keep your audience awake or fill more space on your page.
5.Colours: The most effective way to draw the eye to the story graphs tell is to use colour distinctions. Use of strong bold colours against a background of muted colours makes data stand out. Too many colours will distract the reader and the graph will become unreadable. For data sets of trials where one variable is changed, a change in hue of one main colour can aid in describing the relationship between the variable and the output.