The User Modeling, Adaptation, and Personalization (UMAP) conference brings together a wide community of researchers and experts working on systems that adapt to users, model their profiles, and eventually personalize their experiences. The 25th annual conference was held this July in Bratislava, Slovakia.
As Head of Technology here at Cloud Academy, I’m especially interested in systems that can use what they know about individuals and even groups of users to provide a personalized experience. As a first-time attendee, I thought I would share my experience of the conference and 5 takeaways from the event.
Presentations leave time for discussion
As a newbie in the UMAP community, I immediately noticed that the conference format was different compared to similar events. Where typical conferences allow up to 30 minutes for paper presentations, UMAP gives speakers between 10-15 minutes to present their papers. While I’m sure that this is challenging for the presenter, it does force you to go straight to the most important points, ignoring marginal details.
Another advantage of this format is that it allows much more time for discussions, round tables, and interactions between speakers and their audiences. Thus, you have the opportunity to dive into the details of the research activities that interest you most.
User modeling and personalization in education
Many presentations focused on some of the known problems in the education industry. There are aspects of both in-person and online training (MOOCs) that make it difficult to model the user profile, which makes it even more difficult to provide a personalized experience.
One of the main issues in education is the fact that the user profile changes over time. Actually, the time-dependency of the user profile is claimed by many domains. However, this is usually either a slow process, where the user more or less naturally changes her tastes over time, or a temporary status where the user has specific, occasional interests.
In education, changes, among other factors, are driven by the user’s learning activity. While watching a movie does not necessarily impact my skills, watching a video course will definitely impact my knowledge on a given topic. As a consequence, personalization tools have to take into account changes in a user’s profile. This is not only for identifying a user’s interests (e.g., Java, machine learning, etc.), but also to estimate the user’s aptitude for a topic in order to propose learning resources at the right level of difficulty.
Hot topic: IoT
Internet of Things (IoT) devices, where sensors collect the information required to model user behaviors, was a hot topic at UMAP.
I’ll share an example. In personalizing the user experience for a museum visitor, a set of totems is distributed among the rooms in the museum and are used to both gather user data (where the user is, which path she’s following, how long it took to move from one place to another, etc.) and to dispatch information to the user (suggestions about where to go next based on the number of visitors currently in other rooms, etc.).
Explanation and scrutability beyond accuracy
If I had to sum it up, the main message of the conference is: “if you cannot say why, it is meaningless for the user.”
Thus, while advanced algorithms such as deep learning can provide great accuracy, the community’s preference is for solutions that can explain what made the algorithm select certain items.
Personally, I am not drastically against black-box solutions: the key is to make the user trust the system. If I buy a car and the car works as expected, I don’t need to know how the engine works. It’s when the car stops performing as expected that I may start thinking about what may have caused the problem, and what it will take to fix it. A similar concept can be applied to machine learning algorithms.
In attempts to “explain” something that, intrinsically, does not have an explanation, e.g, neural networks, in some cases, we may be trying to rationalize and find the logic behind an unknowable process. The risk is that we force explanations in phenomena that are not under our control. An explanation for the sake of explaining isn’t useful.
The UMAP community
I would say that the different format (e.g., more space for discussion) and the emphasis on key modeling and personalization concepts (e.g., explainability) are two elements that really resound with the community of attendees at this year’s event. (And a love for music didn’t hurt!)
Compared to other events, I noticed that there was a moderate industry presence, both in terms of sponsorship and presenters. The risk of this is that, while many research projects may experiment with appealing solutions, they may do so at the expense of feasibility in the real world. At the same time, this may also result in researchers and projects not having the visibility that they deserve.
Let’s see how the conference will evolve next year!