Jon Dron

I think the big problem with recommender systems for this kind of thing is that to learn is to change, so past tastes are a very poor predictor of future needs.

RE context switching, absolutely: we are different personas with different needs at different times when we have different intentions and different contexts. To assume that we are a single kind of thing, socially and personally, is just wrong. Quite a bit of the work I have done on recommendation strategies is to try to provide recommendations which do not rely so much on a mined model, but which allow people to create their own disembodied user models. A very simple example would be the use of tags to describe what it is that we value in a resource. Once we have done that, others who value the same things can explicitly choose that facet of their interests to find stuff that helps. Combining this with implicit and explicit ratings in the context of such a tag can lead to recommendations that relate to a particular need in a particular context, rather than an assumption that we are a simple flat set of attributes.

Another approach is to use scrutable learner models, as Judy Kay and colleagues have done. Scrutable systems let you vary the parameters that are used to define you. I quite like that but it runs into difficulties on implementation because it requires people to understand how the user model is being used to provide recommendations, which adds layers of complexity most people are not keen to deal with.

Using clusters such as learning styles might be one way to help simplify that though I am *very* wary of stereotyping of this nature, especially in the light of Coffield et al's wonderful and thorough demolition job on learning style theories a few year's back. Essentially, learning styles are mostly stuff and nonsense and, even when they can reliably identify sets of related learners, offer no reliable means to apply that knowledge to achieve better learning. However, maybe there is scope for emergent, bottom up clusters to be used, identifying similarities through statistical/network analysis techniques like k-means, though it would be hard to give labels that people could reliably identify with using such a technique. Interesting possibility though.

I think this gets to the heart to one of the key problems with using recommender systems to help one to learn. There are two related issues here:
The first is that it is fiendishly hard to match paths of learners because a) there is no guarantee that others took the best path in the first place (because, unlike in matters of taste, they don't know what they don't know) and b) the combination of prior history and future goals is very difficult to model, so it is much harder to match people accurately. Most systems that have been built for this so far, including my own, are more to do with riding a wave than following a path, with sights set on the next thing rather than dealing with the complete learning trajectory.
The second (related) issue is that, unlike for questions of taste, where previous likes are a good predictor of future likes, once we have learned something we no longer have any need to learn the same thing (as a rule).
Just a couple of thoughts. I'm catching up with this conversation so will probably have more to say on the subject!

Depending on the time of week and mood I'm in, I can generally be found in one of Edmonton, Vancouver or Brighton (UK). I work for Athabasca University as an associate professor in the School of Computing and Information Systems as well as being a TEKRI member, and I work for the University of Brighton as a senior lecturer in the Centre for Learning and Teaching. While the course is on I'm mostly going to be in Vancouver where I tend to spend as much time as possible (who wouldn't?).
I'm one of the facilitators of this course and I'm interested in the process, the content and, most of all, the people who are involved and who are likely to be involved with the course. I've been skirting around the area of analytics for a decade or so, mainly finding a home for my interests in the adaptive hypermedia community, which says something about where I look for value: it is not enough for me to know what is happening but also to act on it, and that is where some of the biggest and most intractable issues lie. It is hard enough identifying the facets of behaviour and activity that relate to learning but then how do we turn the knowledge of behaviour gleaned from analytics into something that will affect that behaviour in a way that actually improves learning? And what happens when we iterate that? This leads to a range of questions about the nature of crowds, the nature of learning, the effects of interventions, the nature of creativity and a whole load more.
Like any teaching experience I hope to learn more from others than they learn from me, and that the range and diversity of participants will result in something new, creative and wonderful for all concerned. I don't know quite what that will be yet, but that's why it's interesting!