This is a mixed review of my tour to Buffalo for the Educational Data Mining 2018 (EDM) conference and of my brief but impressive visit to Niagara Falls. I submitted a short paper to EDM and I was invited to give a talk to present the work. I thought that my talk went well and I had some good feedback from some of the participants who attended it. The talk was also good experience for me and I learnt some lessons about what audience I can expect and how I might engage them more actively in the future.
I particularly enjoyed seeing the popular themes at the conference, meeting the other attendees and hearing about some of the innovative ideas that people are pushing forward with. Some of the popular topics were:
- Bayesian Knowledge Tracing (and Deep Knowledge Tracing) - from speaking to an MIT professor called David Pritchard, it seems that this theme is at danger of being overdone at these conferences. Nevertheless, some of the talks about modeling student knowledge using click-stream data were interesting (i.e. tracking what they are clicking on as they navigate a website or a tutorial program).
- Predictive modeling for student dropout - this theme included a range of different prediction algorithms (mostly coming back to logistic regression) and used a range of different student predictors.
- Inference into student activities - most attendees were using LDA (latent Dirichlet allocation) for this task to understand what students are saying in and about MOOCs. I think a really clear and innovative application of the technique was by Ben Motz who used LDA to model popular curriculum choices for undergraduates to see if the degree programs were catering to the desires of the students (see his paper here). He identified a few areas where an administrative update might help students achieve a clearer degree path (e.g. a nice example is that they have no political policy track yet there is a clear theme of students taking history, economics and sociology courses that might form such a track).
- Linguistic modeling - in particular, there was a presentation by Scott Crossley who used linguistic features from a student’s writing to predict math scores. The hypothesis is that the lexical sophistication of a student’s writing might be a proxy to his/her level of cognition about certain topics.
- Deep learning - there were a few talks dealing with some deep learning predictive tasks but honestly, there was nothing noteworthy that I came across.
Ken Koedinger gave an excellent acceptance speech for ongoing service to the EDM community. He discussed a brief history of the community’s research and went on to predict what key investigations the community will work toward. He raised some good points particularly dealing with the transition that the community has made toward purely statistical/probabilistic methods. His prediction is that we may need to return to some of the older work that harnesses logic and, thereafter, combine that with breakthroughs in the statistical methods. He also made a call for the researchers to make their data and implementation repositories more available to the community which I think was notable.
Lastly, I attended a workshop that dealt with ethical issues the community might encounter. Beverly Woolf was especially vocal in this session and the main point was the need for interpretable and understandable models (no surprises here). Beverly was arguing that this means we should keep our focus on simpler (and thus more interpretable) models. There was push-back that the model does not have to be simple to still provide a clear interpretability of its functionality (an example is that for a predictive model, we could study the parameters for each student that might easily change the model’s prediction). There was no particular actionable conclusion that was reached though. Possibly this was not the goal.
Work finished and it was time to play. The University of Buffalo Conference organizers put together a tour to Niagara from the US side. It is about a 30 minute bus-ride from Buffalo to Niagara (that can as easily be done by public transport). The views are spectacular.
The American side of the falls certainly gets a worse rep (when compared to the Canadian side) but they are still very impressive and well worth the trip. I went to the viewing deck that gives you the upper and lower viewing stations. Honestly, this is just a bun-fight with the other tourists and so I’d recommend the more leisurely stroll around the park to absorb the grandeur of the falls. I bumped into Ken and Ben who had been to the “mist of maid” tour. They raved about it but I opted to just enjoy the scenery from the top.
I struggle to not do a little bit of African promotion here but if you think Niagara is great, then you have to head over to Zambia/Zimbabwe. In water volume, the Victoria falls are smaller than Niagara, but in height and dramatic canyon views you’ll find Victoria far surpasses Niagara. That’s just one guy’s opinion though ¯\(ツ)/¯.