Andrew B. Collier / @datawookie

Creating The Next Rembrandt: using data to touch the human soul. How a team from ING, Microsoft, TU Delft, Mauritshuis and Rembrandthuis used technology to synthesise a painting in the style of the Dutch master, Rembrandt, almost 350 years after his death.

As part of International Open Data Day we spent the morning with a bunch of like minded people poring over some open Census South Africa data. Excellent initiative, @opendatadurban, I’m very excited to see where this is all going and look forward to contributing to the journey!

I used to spend an inordinate amount of time digging through lightning data. These data came from a number of sources, the World Wide Lightning Location Network (WWLLN) and LIS/OTD being the most common. I recently needed to work with some Hierarchical Data Format (HDF) data. HDF is something of a niche format and, since that was the format used for the LIS/OTD data, I went to review those old scripts. It was very pleasant rediscovering work I did some time ago.

Setting up an automated job under Linux is a cinch thanks to cron. Doing the same under Windows is a little more tricky, but still eminently doable.

I previously wrote about some R code for downloading Option Chain data from Google Finance. I finally wrapped it up into a package called flipsideR, which is now available via GitHub. Since I last wrote on this topic I’ve also added support for downloading option data from the Australian Securities Exchange (ASX).

This morning I read Wendy Kan’s interesting post on Creating Santa’s Stolen Sleigh. I hadn’t really thought too much about the process of constructing an optimisation competition, but Wendy gave some interesting insights on the considerations involved in designing a competition which was both fun and challenging but still computationally feasible without military grade hardware.

This seems like an opportune time to jot down some of my personal notes and also take a look at the results. I know that this sort of discussion is normally the prerogative of the winners and I trust that my ramblings won’t be viewed as presumptuous.

I routinely use melt() and cast() from the reshape2 package as part of my data munging workflow. Recently I’ve noticed that the data frames I’ve been casting are often extremely sparse. Stashing these in a dense data structure just feels wasteful. And the dismal drone of page thrashing is unpleasant.

I had a look around for an alternative. As it turns out, it’s remarkably easy to cast a sparse matrix using sparseMatrix() from the Matrix package. Here’s an example.

Walmart Trip Type Classification was my first real foray into the world of Kaggle and I’m hooked. I previously dabbled in What’s Cooking but that was as part of a team and the team didn’t work out particularly well. As a learning experience the competition was second to none. My final entry put me at position 155 out of 1061 entries which, although not a stellar performance by any means, is just inside the top 15% and I’m pretty happy with that. Below are a few notes on the competition.

It’s not my personal choice, but I have to spend a lot of my time working under Windows. Installing MongoDB under Ubuntu is a snap. Getting it going under Windows seems to require jumping through a few more hoops. Here are my notes. I hope that somebody will find them useful.

I was asked to review Learning Shiny (Hernán G. Resnizky, Packt Publishing, 2015). I found the book to be useful, motivating and generally easy to read. I’d already spent some time dabbling with Shiny, but the book helped me graduate from paddling in the shallows to wading out into the Shiny sea.

A question posed by one of my colleagues: can a checksum be used to guess message length? My immediate response was negative and, as it turns out, a simple simulation supported this knee-jerk reaction.

Logarithmic Loss, or simply Log Loss, is a classification loss function often used as an evaluation metric in Kaggle competitions. Since success in these competitions hinges on effectively minimising the Log Loss, it makes sense to have some understanding of how this metric is calculated and how it should be interpreted.

Log Loss quantifies the accuracy of a classifier by penalising false classifications. Minimising the Log Loss is basically equivalent to maximising the accuracy of the classifier, but there is a subtle twist which we’ll get to in a moment.

The recently published 2015 Data Science Salary Survey conducted by O’Reilly takes a look at the salaries received, tools used and other interesting facts about Data Scientists around the World. Download the report as PDF. It’s based on a survey of over 600 respondents from a variety of industries. The entire report is well worth a read, but I’ve picked out some highlights below.

The majority (67%) of the respondents in the survey were from the United States. They also demanded the highest median salaries across the globe. At the other end of the spectrum (and relevant to me personally), only 1% of the respondents were from Africa. These represented only one country: South Africa. The lowest salaries overall were recorded in Africa, while the lowest median salaries were found in Latin America.

Last week I took a high level look at the trends in children’s names over the last century. Today I’ll dig a little deeper and examine the ebb and flow in popularity of some specific names.