My first MOOC experience

I was saddened to discover that the course on Quantitative Methods held by my Faculty is no longer run, since I have been starting to realise that I'm sorely in need of some statistics training. After a bit of looking around, I decided that a MOOC was the ideal answer - it kills two birds with one stone. It gives me an opportunity to learn stats, and also gives me first-hand experience in a MOOC, so I can actually claim to have a valid opinion on their usefulness.

I settled on the Johns Hopkins course on Statistical Inference (https://www.coursera.org/course/statinference). It was hard to figure out which course would be right, since I don't know enough about what I don't know to really make a good decision. The description for this course indicated that it covered the things that I needed to know.

My initial experience is that it all seems a bit simple and lightweight compared to previous things I've learned. There just isn't the "hang on, can you explain that bit again" that you have in a face to face course. But things are sinking in, and I'm getting a lot of lightbulb moments as things click and I understand things I've been reading in papers for the last few years. I'm certain I don't know as much as I'd like to from this material, and very uncertain about whether I'd be able to apply it correctly after I finish this course.

In all, I think the MOOC structure isn't a bad way to get some learning done. But there are a lot of gotchas in there. The learning experience isn't great, since if you fall behind on a part there doesn't seem to really be a clear alternative route to understanding (unlike, say, a tutorial where you can ask dumb questions until you figure it out). The teaching is very didactic - it's a series of lectures with a very small quiz and homework attached. Also, the quiz and homework are far to similar. I feel I'd be better served by a lot more example questions, to really stretch my understanding of the material. The week-by-week structure is helpful in driving me to complete the material; without this procrastination would be much easier, so it's good to have that structure. I can understand the huge dropout rate - unless you really need to learn something, it is extremely easy to just give up when it gets too hard.

Learning stats

I've read a lot about the edtech industry implications of MOOCs, and educators' pedagogical concerns regarding their implications, but it's only now (mid-2014!) that I've finally enrolled in one. I've come to the understanding that my exploration of my data in R is hampered not by my knowledge of the tool, but rather by my lack of solid statistical knowledge. It's time to fix that.

I was surprised at the number of stats MOOCs out there - multiple instances of even the same title from multiple universities and multiple MOOC providers. There's a great range of choice. And delightfully, most of them seem to use R as their tool of choice.

I'm going to jump mid-way into a series of courses from Johns Hopkins (https://www.coursera.org/specialization/jhudatascience/1/courses), hoping I'll be fine based on my mathematical background and having had a few months playing with R.

Scale

There has been a lot of interesting discussion online recently about Connectivism and learning theory lately - particularly on Jon Dron's blog and Stephen Downes' responses to them. I'm still very much in two minds about Connectivism as a theory. I'm not sure how it is different to Networked Learning (they seem to be the same practically, if not epistemically), and my inner positivist always screams that "knowledge" is something that resides inside people - that the stuff that resides in these networks is really something very different, even though it is a source for knowledge.

One sentence really stood out for me in Dron's post from last Friday ("What is Connectivism?"):

More is different in a networked system, resulting in a) large scale patterns and emergent behaviours not usually seen in smaller systems, and b) linear benefits of scale - a 1% contribution rate in a network of 10 people will be of a lot less value to all the people in that network than 1% contribution rate in a network of a million people.

It's not something that is a new concept to me - but it made me reflect on my project and how it performs, and has made me think differently other things I've read since then. That 1% contribution rate is not an unreasonable estimate of how likely people are to add ideas into a network if there are no extrinsic motivators; but then the network of ten people most likely has no contributors, and the network of a million has ten thousand. Ten thousand contributors can generate an awful lot of of content, and communities can be formed around that content. My own data bears these numbers out:

I have a group of over a thousand potential contributors - but only 37 (3.3%) have uploaded anything, and only 11 (1%) have contributed five or more resources. There is almost certainly not enough content and activity to generate a significant community around these resources.

I intend to keep improving resource tools in the hope of lowering the barriers to participation to the point where a community can self-start based on these numbers. But this has been a clarion call for me to start thinking about problems of network scale, and what they mean for my project.

Evidence

At last, I've got some evidence that a design change had an actual impact on the usage of my software! This is quite an exciting development. The proof*:












At the end of last year, I changed the way the rating system worked on the student shared resources - rather than a small "RATE" link, which opened up a rating tool, and wasn't very obvious or intuitive, I replaced it with nice big thumbs up/thumbs down rating tool - described here. It turns out, as shown in the graph above, that it has encouraged students to use it. In the 17 months that the first iteration was live, 53 ratings were added. In the last three months, there have been 52 ratings. Given the small numbers, the error bars are going to be large, but this has to be a statistically significant result.


* For probably barely statistically significant values of "proof"

R - loving it

I'm still impressed by R - I'm finding it delightfully powerful. Still trying to figure out what to do with my PhD data, but I'm finding it incredible useful looking at work data to uncover patterns. And the speed with which you can tease pattens out of data means that I can do things like sit down with staff and explore a data set to start to work out why a piece of software is playing up - or at least show patterns in the data which will lead good developers to understand how a piece of code is misbehaving. I'm now annoyed that I hadn't discovered this earlier :)

R

For the last week I've been playing with a new tool for analysing data - R. It's quite an amazing tool. It's a programming language, but one specialized for vector and matrix mathematics - which is ideal for statistical analysis. My initial test data set was ticket tracking data from work - I've been using Excel, and had a very large, clumsy spreadsheet with the same formula repeated thousands of times in order to calculate what tickets were open at what time. After a few hours of reading the R documentation and experimenting, I was able to turn the simple line graph I had in Excel into a really nice stacked bar graph in R - in a fraction of a second it reads my CSV of 128 tickets, calculates from the opening and closing dates how many tickets were open when, and builds a bar chart of how many are open, when, and for which service that we offer (over 60 services are represented).

I was very impressed by R's power and expressiveness. Some things seem a bit harder than they need to be, but that's probably just because I don't know my way around properly yet.

I've now started to work on my log file data. It took quite a while to read in 16 million lines of logs, but now operations across those log file lines are surprisingly quick - in six seconds it'll give me a list of all the IP addresses, and how many accesses can from each (which has lead to one surprising, non-PhD related discovery, which I'll need to follow up at work). Some operations seem to take more memory than I have on my machine (which is now 32GB), but so far I've found a different way to do each calculation, so it's not too bad. And I can do things like a sweet logarithmic scale histogram of lengths of user sessions (spoiler: they're mostly short!), with only a few minutes of playing around to figure out what a histogram is.