Data Mining: Text Mining, Visualization and Social Media

On Google Image Search, A Paradigm and Metaphor

I've been having fun playing with Google's new image search. Naturally, the first thing I tried was a vanity search. I wondered what it would do with the images in my blogosphere gallery. It was interesting to see how one of the images had propagated around the net, being used by a number of sites to illustrate network effects and to capture the idea of the blogosphere.

However, while it is interesting to see the exact match type of results, what I think is more revealing is the discovery of similar images. Before jumping in to that, here's a quick guess at how Google is implementing this feature.

Firstly, I'm going to guess that it is using an indexing system just like its system for retrieving regular textual documents. This means that images are converted into discrete tokens, the sum of which effectively represent the image. When these tokens are close to identical the images are likely to be identical.

Secondly, the tokens that are used are analytical. That means that they represent qualities of the data, not the thing captured in the image. What I mean by this is that these tokens don't capture features of the object denoted by the image, rather they capture characteristics of the raster data that encodes the image. These features may be aggregates (e.g. the histogram of colours) and also associated with specific subareas of the image.

Finally, the system relies on the user to perceive the similarity. Thus when you put in a picture of The Eye of London, you get back the following images:

and you then remark - wow, it understands what was in the query image - how cool is that!

This is, of course, no mean feet. What Google has excelled at here, as it generally does, is the execution at scale of a reasonably well understood approach to image matching.

However, there is another side to this paradigm. What is this a picture of?

 
A person in a gray shirt?

A person with wavy hair?

A man in a gray shirt?

A man holding a microphone?

A person gesturing?

Sergey Brin?

When we ask Google for similar images, we get the following:

The result set includes, interestingly: transformations of the original image, Bill Gates, Larry Page, people in gray shirts, an agent from The Matrix, both men and women.

So, on the one hand we had the impressive reponse to London landmark query which gives the appearance of intelligence and on the other a rather confusing set of images with a high variance for what precisely is 'similar'. 

Of course, the missing ingreadient here is what is traditionally referred to as semantics. The images are not interpreted in terms of what they denote, they are interpreted in terms of the characteristics of the way in which they are encoded.

Image search as an application of this paradigm makes for a good metahpor for certain approaches to text understanding.

The Data Viz Quiz

As an introduction to a talk I recently gave I threw in some slides to form a data viz quiz. They were approximately as follows.

1. Basic Types : name the type of data presentation

1.1

1.2

1.3

1.4

1.5

2. Semantics - What does this represent?

2.1

2.2

3. Historical Personalities - Who created this?

3.1

3.2

While this was just intended as a warm up before the talk, I believe that anyone working in a data driven organization, or with the title 'data scientist' should be able to nail these questions on the back of their passion and curiosity around data.

Answers

1.1 scatter plot

1.2 bar graph

1.3 tree map

1.4 choropleth

1.5 cartogram

2.1 map of deaths from cholera due to contaminated water supply, by John Snow (1813 - 1858)

2.2 map of Napolean's march on Moscow, by Charles Joseph Minard (1781 - 1870)

3.1 Florence Nightingale (1820 - 1910)

3.2 William Playfair (1759 - 1823)

 

Get Out The Data : US Unemployment Compensation

Read about the Get Out The Data concept here.

Below (from an experimental version of d8taplex) are www.census.gov statistics on unemployment compensation. Recall, red = republican tenure, blue = democrat tenure.

Get Out The Data

It is 18 months until US citizens will have decided to keep their current president or roll in another. While hard many to imagine, this means that chatter is starting now about who will be running for election. This means that we have already started to hear sound bites about why A is better than B and how party X did this and party Y did that. Naturally, bikini statistics will play a major part in the discussion. Or rather, will be used as a tool to bamboozle the electorate.

It doesn't have to be that way.

With services like d8taplex, Timetric, BuzzData, Socrata and other data engines, there is a real opportunity to help people cut through the mumbo-jumbo and go directly to data assets to help make better informed decisions and, perhaps more importantly, to hold the circus accountable for honesty in the use and presentation of data.

A simple idea that I plan to further play with is to create data sets in d8taplex as well as some specialized visualizations to help people understand a number of key points:

• The statistical history of their parties
• Relative measures of different countries (what does a country with good health care look like?)
• Straight forward presentations of scientific data (should we invest in ethanol?)

I rattled out an example of the first area tonight. The graph below shows the spend on national defense in billions of FY 2000 dollars. Overlaid on this data set are coloured areas that represent the party in power at any given time (red = republican, blue = democrat). The data is taken from www.census.gov and is available currently in d8taplex (though not in the form below and not as discoverable as it could be).

I would love to see the other data engines help get out the data!

Commute and Real Estate Maps by Stamen

If you've not yet seen it, take a look at Stamen's latest work on interactive presentations of commuting and real estate.

What I think about when I see this type of application is not actually the fact that I can now make better work / life balance decisions but something bigger. We talk about local search often as if it were something far bigger than looking for restaurants. I believe that it should be. In fact, I believe that something like this application should be part and parcel of what users come to expect from a search experience for a locale.

I'd like to see local search evolve into an engine that can provide information about commuting, house prices, school quality, crime rates, walkability. Oh, and restaurants.

Parasite Spam on Twitter

Whenever I post something on my blog, the post title and URL gets pushed to Facebook and Twitter. The URL is shortened via bit.ly. Almost immediately, I will see a couple of tweets of the post information (not retweets of my sharing) including a shortened link to the post and a piggback link to some other website. This second URL is like a parasite attempting to leverage interest in my post to get links and clicks to the other site.

Due to the sorry behaviour of many Twitter accounts of reciprocal following, these parasite spammers can easily build up modest follower lists.

I really don't like this behaviour.

Visualizing Correlations as Small Multiples - Part 2

An alternate to the first small multiples approach to visualising correlations that I posted about earlier is to group variables vertically. This provides an obvious place for the variable name.

The vertical bar is intended to indicate the groups of correlated terms.

[Thanks to Eytan for discussion on this.]

Visualizing Progress in the Canadian Grand Prix

The official F1 site provides this visualization of the position of the drivers on a lap by lap basis through the race.

With a little more work (or a little eye strain on your part) you should be able to see the progress of Mr. Button.

Planning with Wicked Data

I came across the term wicked problem last night. While the definition is technical and somewhat narrow in scope, a simpler, general paraphrase might be: a problem which can only be understood by solving it (a different syntactic rendering might be - a problem which can only be understood in the solving).

In the world of data driven solutions - that is to say, the creation of systems which derive solutions to problems from large volumes of data - one can never really know what the solution will look like until the characteristics of the data and the aspects of the world that the data represents, are fully understood. Consequently, the idea of planning a system based on inferences over a data set that has not been fully analyzed and grokked is somewhat foolish.

The solution is twofold. Firstly, one has to have a solid platform on which to experiment with the data set. Secondly one has to have an iterative planning process with enough flexibility in it to optimize over the search space represented by the characteristics of the data.

Visualizing Correlations as Small Multiples

Part of the idea behind d8taplex is that by not only discovering but actually interpreting data sets found online the user can start to learn something new from all the data being accessible in one place. Naturally, this will mean some functionality in the future allowing for the discovery of relationships between data sets. But before we get there, I've started coding up some ideas behind visualizing the correlation between variables in a single data set.

The first attempt at this employs small multiples to show sparklines of clustered variables. In the example below, we are looking at the consumption of oil by various nations (according to oilcrisis.com).

Clearly some labeling would be useful here (there are tooltips on the dev version of this), but as a proof of concept I like it. The second line includes South Korea, Singapore, Pakistan and Thailand. The oil consumption of these nations can be compared with the third line which includes Belarus, The Russian Federation and Ukraine.

Here's a couple more interesting correlations (Denmark, Sweden), (Chile, Iceland):

As these data are correlated, it may make sense to display them with a single sparkline per cluster - in some sense the icon of the cluster - then list the names of the variables.