0 0


On Github tlevine / multisensory-data-zipfian

Multisensory data experiences

Thomas Levine (,

CSV Soundsystem (

Information overload

Today, we produce more information than we can handle. To cope with this, we data scientists convert this information to structured data that we can make sense of in a more automated way.

Big data

That helps, but now we have to deal with all of these data. We need new tools to help us with these data.

In previous classes, you've learned about tools for storing insane quantities of data and running calculations on all of the data. We need more tools like that, but we also need new tools for exploring and presenting data.

Data table

Special Operations,2005,2006,2007,2008,2009,Total
Emergency Service,0%,0%,1%,1%,0%,2%
Harbor Unit,0%,0%,0%,0%,0%,0%
Aviation Unit,0%,0%,0%,0%,0%,0%
Taxi Unit,0%,0%,0%,0%,0%,0%
Canine Unit,0%,0%,0%,0%,0%,0%
Mounted Unit,0%,0%,0%,0%,0%,0%
Special Operations Division Total,0%,0%,1%,1%,0%,2%
Percent of All Subject Officers Against Whom Allegations were Substantiated,0%,0%,0.2%,0.3%,0%,0.1%

A standard data visualization tool is the data table. At CSV Soundsystem, we prefer CSV files. As MySQL borat suggests, we could just make bigger tables. Tables are great for finding specific values, but it takes a long time to spot broader trends in tables.

This is a table of complaints against New York City police officers, I think. It's small enough that it fits on the screen and that we can get the general picture that most values are, but this wouldn't work for larger tables.


So we make graphs. If you listen to Tufte, you will use graphs to present data in a multivariate way and to present the relationships among different variables.

Let's talk more about multivariate systems. If you have a lot of variables, it's hard to think about all of them at once. Instead of doing that, we think of all of these variables as one concept. We can use statistical methods to convert many variables into one, or we can use data visualization methods to represent many variables at once.

When we're analyzing a multivariate system, we can first look at the collective trend across all of the variables without worrying about the specific changes.

This graph depicts the turns of New York subway turnstiles. This graph represents about four varibles: Date, year, subway stop, and number of turnstile entries per day. For the past three years, New York city has had one pretty big storm per year, each of them in the fall. The red lines are the dates of these storms, and the x axis is the date relative the storm, going from three weeks before the storm to three weeks after the storm. The y axis is the number of times a person entered a subway station each day, and each line is a different subway station. The first panel is for the 2010 microburst, the second is for the 2011 Huricane Irene, and the third is for the 2012 Hurricane Sandy.

Let's get back to our discussion of multivariate data. We are looking at four variables, but we are thinking about them as one concept.

Look at these bumps that happen about five times per line. When I look at these, I'm not noticing the individual lines or the individual dates; I'm just noticing that there's some pattern. Now that we've found some broader pattern, let's focus on it and start picking apart the individual variables. Looking at the date variable, I see that this bump happens every seven days, so it might be the weekend. I guess people use the subway less on weekends. Looking at the subway station variable (the different lines), I notice that this bump happens across most stations. I also notice that some stations drop really low. That might be an error, or it might be that the stations or closed for maintenance. Or something else.

Now let's look at the elephant in the room that I've been ignoring. We've noticed some pattern in the multivariate system; the lines are mostly straight, with these bumps on the weekends. But this pattern gets inturrupted at the red line. What's going on there? Well that line is the date of the storms. Subway stations got shut during the storms, and that explains the drops in ridership. But when we look more closely, we see what might be errors in the data. These lines going straight across might be missing data for those days that should be represented as zero riders.

You probably didn't need my narration in order to read this graph; I explain it like this to convey this approach of aggregating multivariate data. First, we looked for broader trends without focusing on specific variables. Then, once we identified an interesting trend, we iteratively focused on interesting subsets and interesting variables. This is related to a concept of hierarchy from graphic design; if you make one thing really big, people will know what to look at first, and then they can focus on the details if they're still interested. This is helpful for producing data visualizations, but I also keep this in mind when I'm doing data analyses in general.

I just talked a lot about a how to think of graphs and data analysis in a multivariate way. Following Edward Tufte's advice, I aim to do this in data visualizations, but there are only so many variables I can fit on a graph. The previous visual used four variables, and I think six variables is pretty close to the most that can reliably be fit nicely on one static data visualization. So we need something more.


An obvious first step is video. This video relates a bunch of different variables from some experiment at CERN. I don't know what it means, but you can see how adding a time dimension allows for more variables to be presented at once.

Data sonification

When we add a time dimension, we get more possibilities. Of the five senses, vision is somewhat special in that we can look at many different things at once. It's hard to listen to many different things at once, so it's hard to present multiple variables through a few static notes, but if we add the time dimension, we can represent complex multivariate in sound.

This is CSV Soundsystem's first music video, the FMS Symphony. It's a story about the federal treasury, the federal interest rate, the debt ceiling and the financial crisis. The high note is ..., the chords and flourishes are .... And the video. And the data come from here. And the Chernoff face is from the 15 first principal components of the table. And the thing that happens at the financial crisis.

Data sonification

Ridership Rachenitsa

Here's another CSV soundsystem video. It uses both the New York City subway data we saw before and similar data about Chicago buses. New York subways get ten times the ridership, so these two figures are on different scales, but they have a very consistent pattern of ridership. There's one instrument for each city. High notes indicate more riders, and low notes indicate less riders. There's this dip every measure for the weekend. And once in a while we hear some change that we might want to investigate further. For example, in the winter....

Sound opens up more possibilities, but there are still three more senses we haven't looked at. If we're going to leverage our full sensory capabilities to experience data, we're going to need to use touch, taste and smell too. And I think the best way of doing that is through food.

Data gastronomification

gastronomify R package

To begin my exploration into this concept, I wrote an R package that varies a base recipe based on a dataset. Then I produced two bowls of guacamole to represent New York City math test scores for different years.

Each ingredient corresponds to the average score for a particular school grade that year. There's more guacamole in the 2012 bowl because scores were higher across the board that year. (So grade inflation.) And one of them was spicier than the other because the sixth graders did proportionately better that year.

As interesting as that may be, this is a pretty boring data gastronomification. The texture, visual appearance, smell and sound of the two bowls is quite similar, so we're not really leveraging our sensory capacity.


  • Present the multivariate world; escape Flatland.
  • Explore and present data in a hierarchical way; start with the big picture, then get more detailed.
  • Anything can be turned into data, and data can be mapped to anything.

More resources

One last thing: Data cookies

  • Cookie base: Brownie for garage and chocolate chip cookie for open lot
  • Nuts: Public lots have nuts
  • Length (inches): Log number of parking spaces
  • Toppings correspond with the type of parking lot
    • Oreo: Paid publicly available
    • Reese's Piece: Customer parking only
    • marshmallow: Permit holders only (monthly, employee, student, car share, etc)
    • cherry: Commercial/government only
    • jam: Free parking available
  • Sprinkles
    • Red sprinkles if it has valet
    • Blue sprinkles if it has motorcycle parking

Data guacamole doesn't really use all the senses, and I wanted to make something that did, so we prepared some data cookies for tonight. Each cookie represents a parking lot in San Francisco. [Read the different ingredients]