Skip to content

Latest commit

 

History

History
504 lines (405 loc) · 17.7 KB

readme.md

File metadata and controls

504 lines (405 loc) · 17.7 KB

Data music for big data analysis

Introduction

For handle big data, solution is very simple: buy bigger monitor and use smaller font in the terminal.

Today, we produce more digital content than ever before. Making sense of the data is becoming harder as the data get more complex, so we need to develop new tools. In particular, we need something more powerful than data visualization.

The problem with data visualization

Data visualization cannot support wide datasets, with lots of variables.

Minard's map of Napoleon's march

This plot of Napoleon's march is seen as an exceptional piece of data visualization that conveys the multivariate relationships in the world, and it contains about seven variables: Longitude, latitude, date, temperature, direction branch, army size. (You could arrive at a slightly different number if define the variables differently.)

The best of visualizations only supports a few variables, but today's data might contain thousands of variables. Typically, we deal with this by reducing dimensions before plotting or by making multiple plots, but these approaches lose information. Rather than using these lossy visual approaches, we would like to represent multivariate data in a more raw form. How can we represent more raw variables?

Chernoff faces can get us to about twenty variables.

Chernoff faces

Animations help a bit too.

complicated plot

But we're still only a little bit further.

I think the future is in multisensory data experiences. Food is the prime example of this, as we use all five senses in experiencing it.

Artichoke pizza

Unfortunately, computers don't have particularly good APIs for touch, taste or smell, so it is harder to represent your data for these senses. On the other hand, they do have good sound APIs, so we can start there.

Other reasons for music

As we explained above, music gives lets us use a different sense (sound) that visuals do. If we combine music with visuals, we can consume data through multiple senses and thus experience higher-dimensional data. This ability to represent multivariate data is the main promise I see in data-driven music, but I see some side benefits too.

Accessibility

"Opening Doors to IT" logo

Ambient perception

With traditional data visualizations, we need to focus actively on the visualization. With sound, it is natural to have a data sonification playing in the background all of the time, like music.

Reaching young people

Data is in.

Government representatives

History of ddpy

csv soundsystem makes music from data, and we have developed a few tools to help us do that. We started out by with disgusting hacks in R, which we eventually abstracted into a package called data-driven rhythms (ddr). ddpy is confusingly named after the package ddr, even though there's no pun anymore. The API for ddpy is inspired largely by a prototype we build for making music from Google Spreadsheets.

Today

Today, you'll learn how to transform a dataset into music. We'll use the ddpy package for this tutorial, but the same concepts apply regardless of what tools you use to turn your data into music.

Install

You'll need a means of playing MIDI files. timidity++ is one option. Or just use a website like SolMiRe.

You'll also need ddpy.

pip install ddpy

Imports

The code examples below expect the following to have been run.

import numpy
import pandas
import pandas.io.wb
from ddpy import to_midi

Tables

I see the whole world as collections of things, which I like to represent as data tables. Rows are records, and columns are variables.

I see music the same way. A given sound is a function of the notes that are being played by various instruments, and a song consists of a collection of sounds. Thus, columns are instruments, rows are beats (or some other time-related thing), and cells contain notes.

Here's a passage from Chopin's Étude Op. 10, No. 1 in ordinary sheet music.

Ordinary sheet music

And here it is as comma-separated values. (Well almost; it doesn't include the two chords of dotted half notes.)

left_hand, right_hand
NA,        NA
C2,        E6
G2,        C6
C3,        G5
E3,        C5
C3,        E5
G3,        C5
C4,        G4
E4,        C4
C4,        G4
G4,        C5
C5,        G5
E5,        C6
C5,        G5
G4,        C5
C4,        G5
E4,        C5
C4,        G4
G3,        C4
C3,        E4
E3,        C4
C3,        G3
G2,        E3
C2,        C3

Rather than composing music as traditional sheet music, we can use a table-editing program of our choice to compose this sort of table.

Pandas to MIDI

ddpy provides a to_midi function that converts a pandas data frame to a MIDI file.

to_midi(df, 'output.mid')

It currently only supports integer columns containing integers from 0 to 127.

Text is represented as lyric events, integers are represented as discrete beats, and floats are represented as notes that gradually merge into each other.

The main thing that probably isn't obvious to you is how pitches get created. MIDI files can represent 128 different notes per instrument. If the instrument is a piano with exactly 128 keys (white and black), then zero correspends to the lowest (left-most) key, and 127 corresponds to the right-most key. ddpy just passes these numbers from our data frame into the MIDI file.

Piano with 128 keys, numbered from 0 to 127

Thus, we can compose some simple music by making columns with numbers from 0 to 127. Here's a chromatic scale.

df = pandas.DataFrame({'chromatic':range(50, 63)})
to_midi(df, 'chromatic.mid')

A major (Ionian) scale

df = pandas.DataFrame({'major':[50, 52, 54, 55, 57, 59, 61, 62})
to_midi(df, 'major_scale.mid')

A natural minor (Aeolian) scale

df = pandas.DataFrame({'minor':[50, 52, 53, 55, 57, 58, 60, 62]})
to_midi(df, 'minor_scale.mid')

Some minor chords (multiple instruments)

df = pandas.DataFrame({'low':[50, 57, 64]})
df['middle'] = df['low'] + 3
df['high'] = df['low'] + 7
to_midi(df, 'chords.mid')

Random music of different distributions

df = pandas.DataFrame({'normal':[round(random.normalvariate(55, 7)) for i in range(24)]})
to_midi(df, 'normal.mid')
df = pandas.DataFrame({'gamma':[round(random.gammavariate(2, 3)) for i in range(24)]})
to_midi(df, 'gamma.mid')

You don't always need to play something; here's a Bernoulli rhythm.

df = pandas.DataFrame({'bernoulli':[(52 if random.uniform(0,1) > 0.5 else numpy.nan) for i in range(24)]})
to_midi(df, 'bernoulli.mid')

Exercise

Load a dataset into a pandas data frame, and convert it to MIDI. You can use any dataset you want, but here's an option in case you can't come up with any.

df = pandas.io.wb.download(indicator=['NY.GDP.PCAP.KD','EN.ATM.CO2E.KT'],
                           country=['US', 'CA', 'MX'], start=1900, end=2013)

Don't worry about doing anything that complicated; we'll do that later.

More about MIDI

Let's talk a bit more about MIDI so you get a better feel for what is going on. I think of everything as tables, so I also think of MIDI files as a format for serializating tables, and that's how I'm going to explain it.

A MIDI file contains up to 128 different instruments (columns). Each of these contains up to 16 different tracks. Within each track, we have a bunch of events, like discrete notes and bends in pitch. There are also "meta-events", which include lyrics.

Why do we need this concept of events? We are using a MIDI file, but you can also emit MIDI events directly to other software, live. These live events use the same protocol as the events in our file.

ddpy currently only implements discrete notes. That is, you can say that a note should be played at a constant pitch for a certain period of time.

Preparing our data so the music sounds nice

I've come up with a few elements in the production of interesting data music.

Data must have a noticeable pattern.

Random music gets boring quickly.

df = pandas.DataFrame({'normal':[round(random.gammavariate(2, 3)) for i in range(72)]})
to_midi(df, 'random.mid')

Similarly, empirical data that are effectively random aren't that interesting either.

df = pandas.io.wb.download(indicator='CM.MKT.INDX.ZG',
                           country='US', start=1900, end=2010)
to_midi(df, 'effectively_random.mid')

This gets more important as you add instruments the second instrument normally needs to have some relationship with the first instrument in order for the piece to sound good.

df = pandas.DataFrame({
    'a':[round(random.normalvariate(55, 7)) for i in range(24)],
    'b':[round(random.normalvariate(55, 7)) for i in range(24)],
})
to_midi(df, 'two_random_instruments.mid')
data = pandas.io.wb.download(indicator=['NY.GDP.PCAP.KD','EN.ATM.CO2E.KT'],
                             country='US', start=1900, end=2010)

def scale_for_midi(series, lowest = 0, highest = 127):
    series_int = series.map(float).map(int)
    return lowest + (highest - lowest) * (series_int - series_int.min()) / (series_int.max() - series_int.min())

music = pandas.DataFrame({
    'gdp':scale_for_midi(data['NY.GDP.PCAP.KD'], lowest = 36, highest = 72),
    'co2':scale_for_midi(data['EN.ATM.CO2E.KT'], lowest = 36, highest = 72),
})
to_midi(music, 'two_instruments.mid')

Periodic trends work quite well. Here's transit ridership.

Exercise

Make a simple song from two variables that are somehow related. Again, use any dataset you want.

You're still making music

We started with the example of mapping numbers to keys on a piano. You should treat this as a primitive operation on which more interesting things can be built.

Using Grammar of Graphics terminology, let's say that pitch is one aesthetic that defines our music. We could have other aesthetics, like the key/scale. You could have one column defining the note within a chord, another column defining the base note of the chord, and a third defining whether the chord is major or minor. Then you create one column to convert to MIDI.

gdp_df = pandas.io.wb.download(indicator='NY.GDP.PCAP.KD',country='US', start=1900, end=2012)
gdp = list(reversed(gdp_df['NY.GDP.PCAP.KD']))
df = pandas.DataFrame({
    'gdp':gdp[1:],
    'better.than.last.year': gdp[1:] > gdp[:-1],
})

music = pandas.DataFrame({
    'base.note':scale_for_midi(df['gdp'], lowest = 48, highest = 60),
    'better.than.last.year': df['better.than.last.year'],
    'downbeat':reduce(lambda a,b:a+b,[[12, numpy.nan, numpy.nan, numpy.nan] for i in range(df.shape[0])])[:df.shape[0]],
    'thirdbeat':reduce(lambda a,b:a+b,[[numpy.nan, numpy.nan, 24, numpy.nan] for i in range(df.shape[0])])[:df.shape[0]],
})
music['third'] = music['base.note'] + 4
music[music['better.than.last.year']]['third'] = music[music['better.than.last.year']]['third'] + 1
del music['better.than.last.year']

to_midi(music, 'change_in_gdp.mid')

Also, rows in your dataset could correspond to things other than beats, like a measure, a phrase, or a stanza. This is especially helpful when you're dealing with data of varied resolution (for example, monthly versus daily).

Exercise

Map some data onto musical aesthetics other than pitch. If you know any music theory, do get creative with this.

For something simple, you could try three-note chords (triads). To make a major triad from a base note, play the following notes.

  • the base note
  • the base note plus four
  • the base note plus seven

To make a minor chord, play the following notes.

  • the base note
  • the base note plus three
  • the base note plus seven

Outliers are your solos

If you follow the advice above, you'll have a very coherent piece, where everything within in relates to everything else. This in itself gets boring, but it allows you to create interesting sequences that sharply contrast the rest of the piece. And these interesting sequences naturally arise if you have outliers. For example, check out the financial crisis in FMS Symphony.

This is actually the same for data visuals; people often focus quite strongly on outliers in graphs. Here's a graph of the FMS Symphony data.

A frame from the FMS Symphony video

Data music, just like data visuals, can be set up to emphasize specific parts of a dataset. That is, you could use the same dataset to produce one song or graph that emphasizes an trend and one that completely ignores it.

Anyway, keep in mind that outliers make your music interesting.

Thinking about sound and multivariate analysis

In this tutorial, we only just scratched the surface of how we can represent data as music. I leave you with some thoughts on how to go further.

Given that you're reading this, I suspect that you already know something about how to make meaningful plots. We've been studying data visualization for quite a while, so we've come up with some pretty good theory about how to make good graphics. Our ears work differently from our eyes, so much of this theory won't apply very directly. You'll have to explore different ways of creating sound such that our ears perceive the data properly.

Multivariate analysis

Here's a little tip to get you thinking. The world is multivariate, and we should represent that in our visuals. (As Edward Tufte would say, escape Flatland.) When we are representing dozens of variables at once, we can't expect ourselves to be able to keep track of all of the individual variables; once we get to more than a few variables, we tend to reduce the dimensionality based on some sort of unsupervised learning, like clustering or principal component analysis. We use these multivariate methods to get a bigger picture; once we have the bigger picture, we can choose to delve deeper into specific parts of the dataset and to look at the original variables.

Why vision might not be great for multivariate analysis

When you're producing music, food, or visuals from data, it's good to both present the bigger picture and allow people to delve deeper into specifics. I find that the sense of vision is particularly well suited for delving into specifics. This is because visuals can be static and because we can easily block out certain parts of visuals.

When I say that visuals can be static, I mean that a person can decide with her eyes how long to spend looking at them. Contrast this to sound, where a person has to spend time listening in order to perceive a full song. With a visual, you can easily slow down to focus on just one part.

When I say that we can block out certain parts of visuals, mean that we can cover up parts of the visuals and just focus on the interesting part. For example, we could have a huge scatterplot matrix but choose to focus on only one of the scatterplots. Contrast this to sound and smell; with those two senses, we can focus our perception by walking around or by pointing our heads in different directions, but it's harder for us to focus on a particular range of receptors (a band of frequencies or a set of smells). We can focus our taste by choosing what we eat and to some degree by choosing which part of our tongue we put our food on, but it's still not as much focus as we get with vision. Touch is, perhaps, the closest sense to vision in the ability to focus on particular stimuli.

Review

  • Why music
    • Use other senses
    • High dimensions
    • Accessibility
    • Getting wider audiences interested
  • Seeing music as a table
    • Instruments are columns.
    • Time units are rows.
    • You may have to transform your original dataset to be in this table format.
  • MIDI
  • Tips
    • Data need to have a pattern; random noise is boring.
    • You're still making music, so music theory applies.
    • Gaps along your musical time variable can be annoying.
    • Outliers are your solos.
  • How to think about sound and multivariate analysis

Other resources

Slides

To see just the figures in the present document, run this.

npm install -g reveal-md
make slides