I've recently become a lot more interested in what data science involves — to the point of considering careers that use these skills — so when I noticed Coursera's Introduction to Data Science course taught by Prof. Bill Howe at the University of Washington, I decided to take it. I finished the last assignment this weekend, so it's a good time to write up my thoughts while they're still fresh.

The first week gave a quick overview of what data science is and why it's becoming increasingly important. From then on, it covered a lot of topics. Starting with traditional SQL databases, we moved on to covering MapReduce algorithms, NoSQL databases, statistical topics relevant to large data sets, unsupervised and supervised machine learning algorithms, graph analytics, and looked at data visualisation considerations (particularly features that can help make visualisations easy for the viewer to comprehend).

The lecture style was pretty brisk. Lectures were broken up into short 5-15 minute segments, but often information dense. This isn't a bad thing. The fact the lectures are recorded means that you can repeat sections, or go and look something up for more details. Usually, the lectures weren't too difficult to follow. There were a few occasions where I think a little too much was crammed in. One instance: rule learning by sequential covering was discussed in less than five minutes which was far too short to explain clearly. It took me another 20 minutes of reading and working through this explanation to actually grasp the concept properly.

On the other hand, the emphasis on breadth rather than depth in the course fits well for an introductory course. It gives you sufficient grounding and the vocabulary to start understanding concepts you might encounter elsewhere, and points you in the right direction to look further. I was impressed at the considerable amount I learnt about technologies that I knew nothing about previously.

The range of assignments was diverse: analysing tweets, using SQL databases and implementing some very simple MapReduce algorithms. There were also two open ended, assignments: using Tableau Desktop to do some data visualisation, and entering a Kaggle competition. In these, it was up to the student to choose a question and challenge themselves. This approach was a good way of handling the very varied academic backgrounds of students on the course.

Entering a Kaggle competition was interesting. Even with the tutorials Kaggle provides, it's a little intimidating if you haven't tackled those kind of predictive model problems before. Because it was a part of the course, it gave me the motivation to have a go and try using scikit-learn; this is certainly a positive. I'm more aware now of the issues in working with other people's data sets: often a large part of the problem is deciding how to handle missing data.

(There were also a further two optional assignments, which I skipped due to a lack of spare time. One of these was a real world project, so there was no shortage of ways in which you could delve in more deeply should you wish.)

Three of the assignments used Python and required some basic language knowledge to complete them. However, if you weren't familiar with Python prior to the first assignment, being thrown in head first to it was probably not the most fun or easiest way to go about learning it. If you were going to take this course, it's definitely worth working through a few tutorials beforehand to learn the basics in Python (data types, expressions, conditionals, loops, functions); these are good lectures and this is a good free book.

There were some initial grading issues that meant the course didn't get off to the start that was probably hoped. The first assignment was autograded but, with having to deal with Unicode strings, input data sets likely being unique for everyone, and opaque grading feedback, it made for a difficult experience.

Completing the actual assignment itself didn't require any complicated coding. Getting the grader to recognise that you had the right answer was something else entirely. I spent more time on the latter than the former. It didn't help that the grader became increasingly slow to respond as users were subjecting it to repeated attempts, making the whole process a little bit frustrating and to the detriment of the assignment itself (which was actually an interesting one, analysing mood sentiment of tweets).

This was several weeks ago and a fairly distant memory now. In fairness to the staff, they listened to feedback and most of the subsequent assignments were handled more smoothly. Over the past week, I have read some complaints about having the use of Tableau for the final visualisation assignment, but I personally haven't had any problems with it.

There was a post on the course forums suggesting that the course is going to be repeated, tentatively next spring. At the moment, I'm not aware of any free courses that discuss the subject as a whole (there are several that discuss machine learning). With that in mind, one strong aspect was how this course linked ideas together. It never seemed like topics were thrown in there simply for the sake of it and frequently ideas (especially relational algebra) were recurring throughout. Overall, I'm glad I completed the course; I learnt a lot from it. If you want to take a starter course in data science, I wouldn't hesitate to recommend it. Thanks to the staff for providing it!