Data Visualisation: The Good, The Bad, The Ugly and the Beautiful - EMF 2026

Duration: 19:26 2936 transcript words EN summary

Video summary

Data visualization serves as a powerful tool that transforms complex numerical problems into perceptual ones, leveraging the brain's ability for parallel processing and pattern detection to make information instantly understandable. The core principle discussed is that effective charts must map graphical size directly to data magnitude, ensuring an honest representation where linear changes in data correspond to linear changes in visual length or area. This rule is critical because violating it can severely mislead audiences; for instance, a political poll chart that exaggerates differences between bars creates a false narrative about the closeness of a race. While modern software like Excel or ggplot2 generally handles this mapping correctly automatically, emerging AI image generators often struggle with these basic linear relationships due to hallucination artifacts, highlighting that even simple rules require vigilance in the current technological landscape. Beyond the fundamental rule of linear mapping, the choice of chart type significantly impacts accuracy based on human visual perception capabilities. Humans are highly accurate at comparing positions along a common scale, such as in bar charts or line graphs, but tend to underestimate larger values when comparing areas in treemaps or volumes in 3D charts. Consequently, for tasks requiring precise comparison, simple linear scales are preferred over complex ones like pie charts, which are better suited for showing a part-to-whole relationship without forcing the viewer to mentally stack segments. Furthermore, color perception is not linear; lightness varies non-linearly across the RGB spectrum, meaning yellow and cyan appear much more intense than dark blue or red. To maintain honesty in data representation, designers should use perceptually calibrated color maps that reflect these human limitations rather than assuming a uniform distribution of color intensity. Making visualizations engaging without sacrificing integrity often involves adding contextual benchmarks or embracing the inherent complexity of the data rather than smoothing it over for aesthetic appeal. A compelling example is using a map of the United Kingdom to visualize land use statistics, where overlaying actual geographic areas helps viewers grasp the sheer scale of resources like beef pasture compared to urban areas. Similarly, when dealing with highly complex datasets, such as a national budget split into hundreds of billions, a messy pie chart with many slices can be more effective than a polished but oversimplified graphic because it invites interaction and reveals the true difficulty of the data. In extreme cases, visualizing wealth inequality by scrolling through a bar that stretches far beyond a single screen allows the sheer magnitude of disparity to speak for itself, turning what might seem like an "ugly" or unreadable chart into a profound statement on social reality. Ultimately, data visualization is an art form that balances strict adherence to perceptual rules with creative strategies to engage and educate the audience. The speaker emphasizes that while there are no one-size-fits-all policies, designers should always ask what specific information they want the reader to extract and choose the visual encoding that best supports that goal. Whether utilizing tools like Observable for online exploration or participating in community projects like Tidy Tuesday, the goal remains to create honest representations that respect human cognitive limits. By keeping designs linear where possible, adding relatable benchmarks, and occasionally embracing the complexity of the data, creators can produce visuals that are not only informative but also thought-provoking, ensuring that the audience gains a true understanding of the underlying facts rather than being misled by aesthetic choices.

Read the full video transcript

Uh I'm Charlotte and I'm presenting a talk uh on data visualization, the good, the bad, the ugly, and the beautiful. Um so a little bit of background from me. Um, I'm a PhD student uh at uh there I am at the University of Manchester and I'm studying this basically for a living. Um, the psychology of what makes data visualization good or bad. Um, so I'm residing from home today uh thanks to this little menace. I twisted my ankle in a rabbit hole at a conference a couple weeks ago. Uh, so stay safe out there when you're coming back from null sector. Um yeah they [laughter] uh danger comes in many forms. Um so what is data visualization? Well, we're all geeks here. I think we all know what data viz is. Um it's representation of data in uh graphical form. So whether that's line graphs or bar charts or some of the more uh unusual visualizations here. Um, I what we're going to be talking about today is what makes data visualization effective. What makes a graph good or bad, aesthetically pleasing, or just plain ugly. Um, so data viz is effective because it turns a numerical problem into a perceptual one. It allows us to use utilize different parts of our brain for the same task. So we've got a say we have a very simple representation of data here in an array. Uh these are the count selection results from May. Um if we wanted to get an idea of what this data um of the relationships between these um different parties how well they did relative to another. It's quite hard to do. Uh it takes a lot of serial processing. However, if we encode this data um visually by taking the size of the data point as um the length of a bar in a bar chart, suddenly we can uh obtain a parallel um uh processing capability. Instead, we can instantaneously compare uh the results of the different parties across the visual spect uh the visual space. Um uh data visualization also allows us to use our pattern detection abilities. So we could have a very boring overview of the d a data set like this very um bog standard not much interesting happening here. However, if we instead represented this data set in uh this by variant data set in the xy plane sometimes unusual patterns can emerge. Uh so we can see that data viz is inherently a perceptual process and this is in fact related to why data visualizations can go wrong for the same reasons that they can be effective. Uh there is a trade-off by visualizing data we also render ourselves vulnerable to visual spatial bias. Uh so today I'm going to introduce some rules that you can uh use within your own data visualization um practice to help people um understand the data better and maybe even become more engaged with it as well. However, I'm going to call these rules of thumb rather than strict hard rules because as we'll see um there isn't really a one-sizefits-all policy here. We have to take a huge amount of different considerations into account. Now the first rule of thumb is actually very straightforward. We need to uh simply make sure whenever we create a visualization we map the graphical size to the data size. So going back to the count selection results we need to make sure that there is a linear mapping between the magnitude of the data in this case the number of party seats and the um graphical representation of that data. In this case the length slash area of the bar. Now, uh, this, uh, rule is the most important because it can be used to do the most harm. Um, I'm going to call upon the, uh, king or should I say president of bad visualization design here, uh, Donald Trump. Um, who, uh, in this particular chart has um, uh, visualized a poll lead um, against Hillary Clinton in 2016. Um, so this chart on itself doesn't look too bad. It's not even particularly ugly. However, it is extremely misleading. Uh, the uh difference um in the distance between the bars here does not match up to the data at all. As this post by the Washington Post shows, um, if Trump wanted to visualize this data correctly, it actually looks something more like this. Keep going. Keep going. There we go. Okay, so as you can see uh if you were to visualize this data linearly or honestly uh there really isn't that much of a difference at all. So this this point is the most important for visualizing data honestly um for good charts as it were. Uh but fortunately it's also the easiest rule to follow because with modern visualization software whether using something out of the box like L sorry like Excel or uh something a little more complicated like ggplot or um D3 uh the computer will handle this for you. It's you have to really try and break um the system uh to go against this particular rule of thumb. Now I would normally say uh modern software has made this easy. However, with the a rise of AI, unfortunately, uh even something as simple as this, it gets wrong. Um in the particular case of AI image generators because of uh hallucination artifacts, it's easy to um uh uh it's very easy to mess this simple linear mapping rule up. um even in dedicated uh scientific diagram um image generators um the best of the best can't get this right at the moment. Okay, so we've got a good initial rule. Um but obviously there's so much more to visualization than that such as what type of chart should I use? And this is actually really important because although um mapping data one to one is a good start, it isn't the whole story. Some visualizations are easier for our visual system to compare than others because uh we uh compare some visual variables linearly and other variables less than linearly. For example, we are really good at comparing positions such as line graphs or bar charts. The the rate of error um for these comparisons is pretty low. However, we are less accurate when we compare areas. [clears throat] We tend to underestimate larger values such as in the tree map at the bottom. Um, similarly, we are usually better at processing bar charts than pie charts, and we're really bad at estimating volumes. So, 3D charts aren't the best bet. If you ever been felt like you've been shortch changed at the bar, maybe have a think about that um for next time. Uh, so when accuracy is the most important, it's best to choose a simple comparison over a complex one. However, um, of course, uh, there are times that we might want to perhaps reconsider this. Um, which brings us on, oh, I'm sorry. Um, one final, um, uh, consideration for this rule, um, that can sometimes catch us out is to do with color. Uh, we don't actually perceive, uh, the lightness of color linearly, um, across the RGB space. Yellow and cyan looks a lot more intense than dark blue or red. So if you're using a continuous color map, try to use a perceptually calibrated one instead um like the ones shown here. Although try to also make sure that it's colored line friendly as well. Um so um we know to map the data one to one and we have a good rule of thumb for what type of chart to use. However, some discretion is advised here because it's important when we're visualizing data to ask yourself, what is the um values that you want your readers to compare? What specific um information do you want them to get out of the chart? For example, if you've got lots of different groups and you simply want people to compare them to one another, try a bar chart. Um, however, if you actually want people to compare the size of one group relative to the whole, actually a pie chart is just as good as and sometimes even better for this task. Um, the reason this is the case, um, imagine for the bar chart, you want to compare the size of, um, say the red bar at the left there relative to the size of the whole chart. Um, to do that, you'd have to mentally stack up the size of all the other charts. um and then compare that mental comparison um with the group you you're trying to compare, it's quite difficult. Whereas with a pie chart, the the proportions are instantly available to you and it's much less um cognitively taxing. Um similarly, uh in election maps, this is also very important. This is an honest election map. It shows you where people voted. um pretty pretty interesting, but um it distorts our idea of which party did better [laughter] because um as it turns out, one uh voters of one party tend to live in larger areas. Um but it's not land that votes, [clears throat] it's people. And so, um, if we want to make if what we're actually trying to show is the number of votes, we should make sure that our mapping is actually honest to that, um, uh, task at hand. So, we've got some good rules of thumb for generating honest charts, but it's a bit boring, isn't it? It's not much pizzazz. Uh, how do you make these charts like interesting and engaging, but at the same time not being misleading either? Well, there's a this is where um it becomes a little bit more of an art, but there are still some rules of fun that can get can start to nudge you in the right direction. One of my favorite for these is um adding a benchmark. So, the best way to illustrate this is with a case study um that I found recently um which uh was for uh on the topic of looking at uh land use within the UK. Uh where does our food come from essentially? Um so this um uh paper found that the overall area size of land associated with UK diets is about 24 and 38 million hectares. And one way you could visualize this at least across time here um is with a um stacked chart like this. And it's okay. It's not the prettiest but it is honest. But it doesn't really give us a good idea of the scale of the situation. Um I mean I don't know about you but I don't know how big a hectare is let alone 24 million of them. And so the although I have a good idea of the ratio that grassland is used more than feed crops and food crops. So that's sorry I should say grassland is like beef pasture. I don't really have a good idea of um how important this is. So what the national feud strategy um uh paper did was concretize uh this finding by um supplanting um these results directly on top of the UK and suddenly I have a much better idea of um uh the sheer scale of the situation. I can see that it would take me about eight hours to drive through all the beef and lamb pasture that we use within the UK and comparatively uh builtup areas only uh encompass about Wales roughly. Um this chart actually caused me to um really um reduce my beef consumption as a result here. Uh this is actually how much we use overall. uh quite sobering but very very effective. So if you're unsure of where to start with um making your visualizations more engaging, adding a visual benchmark like this, something that people can relate to, um is a really really good um strategy. And one final rule of thumb here, I wouldn't call it a rule, but just something to to take away, I guess. Uh we've gone on about bad and good and beautiful and ugly charts here. But I think sometimes ugly is kind of beautiful. Um [snorts] as a case in now this might sound a bit strange but hopefully I can try and explain. Um sometimes data visualization in trying to be beautiful can obscure um some of the underlying complexity uh with a data set. Um and uh actually embracing the complexity can produce effective charts in their own right. One example of this is from a really good uh website called where doesit allgo.com which summarizes the UK yearly budget um splitting 1.37 trillion pounds into all these different pots like welfare, health, uh education all in their various billions. And this is a difficult and complex picture. So how do we visualize this? Well, we can make show how difficult and complex it is um with a pie chart like this. Now, on one hand, this chart breaks so many rules visualization. A, it's a pie chart, which some people dislike. I don't really mind them so much. But, you know, um but more importantly, there's so many slices that, you know, it is visually hard to read. But I think that's um what makes it also visually engaging. Um because this plot also has interactivity. Um it allows you to delve into um this all the complexities of this chart and compare different pots with one another in a very effective way. Um, so I I think personally if you try to make this more beautiful, I think it would detract from uh the message or at least the um the object objectivity of the uh overall chart as it were. And uh one final little quirky example here that um I personally uh quite like uh you might have seen Elon Musk is in the news recently. He's recently become a trillionaire. um he now has $833 billion. And what this um uh chart does is visualize the data um visualizes wealth um relative to the size of this tiny little pixel here, which represents $1,000. And from that, it is essentially uh one big bar chart. Um but unlike an average bar chart which usually fits within the size of a page, uh this chart instead um just is stays incredibly objective and just lets the sheer size of his wealth speak for himself. Um so uh you can see here um uh by just simply scrolling along his wealth. I've been talking for about you know 30 seconds here. we're only at 0.40% uh or Opra Win3's net worth and you know you could compress his wealth down into like you know an easily uh readable uh nicely digested uh visualization but I think doing it like this um is so effective because of how ludicrous it is. It's hard to compare all these values definitely and in that sense it's um perhaps not as useful but in but what it gains is um just the sheer size of the d um uh the sheer horror of um the incredible inequality we live in right now. Um so that's uh five little rules of thumb um that I hope you can take away um for your own visualization projects. Um I think the what is important to remember is that data visualization is a perceptual process. And so um when we're visualizing um [clears throat] data, we should try and keep things linear and simple where possible, but also try to keep the visual task in mind. Uh we can try and visualize uh um data with a benchmark where possible. And if um the data is weird enough, sometimes embrace the ugly. Um, I will admit obviously there's uh a lot to um take in and there's still lots of research. That's why I'm doing um my PhD in this area. Um but if you want um some good resources for creating your own visualizations, I have to recommend ggplot 2 if you have uh local data sets. Um, it's definitely, I'd say, the easiest to jump into because it's got a really nice syntax. Um, but also observable and, um, is a good online option as well. Um, and if you want to get your teeth into this kind of thing, uh, Tidy Tuesday is really good. Um, basically it's a project where every week um a shiny interesting new data set is released and you can compare your um results to uh other Tidy Tuesdayers around the world. Um there's some good um uh blogs out there as well which go which drill down into some of the principles I've talked about and um How Charts Lie by Alberta Cairo is a really um good book on the subject as well. Uh finally I'm on LinkedIn if you want to ask any questions. Obviously I can't come to the Q&A tent. Um I know LinkedIn's have that's you know I ideally would like a blue sky. I should probably get on that at some point. Um uh but this is where I'm findable at the moment. And uh yeah, uh thanks for listening and thanks to everyone at EMF for helping me organize this. Um I wasn't sure if I was going to be able to present it, but it was really nice uh to be able to do so. Uh so thanks [applause]