Bruno Latour just gave the closing keynote for CHI 2013 and he issued four challenges for HCI research. I thought I would get down some thoughts about them before I forget.
Before I talk about the challenges, I should try to describe his central theme. He was arguing against the division of sociology into two scales the unconnected individual and the unindividualised collective. He instead argues that we should think in terms of overlapping and interconnecting “monads” (I won’t try to explain the term). He thinks that digital technology can help to analyse data without going to the two poles of individual qualitative datum or collective, aggregate statistics.
This kind of aligns with my thoughts that interactive machine learning could help to bridge this divide by having human interaction that focuses on the detail of different aspects and items of data within the statistical analysis of machine learning (this is still very vague on my part, but I think there is something there and maybe Bruno Latour does too).
Overall, Latour wants the CHI community to help break down the individual/collective polarity, but in particular he issued four challenges:
Getting rid of data
His first challenge was to help get rid of data from large data sets (presumably so you are only left with “interesting” data in some sense). Given the rest of this talk I interpret this not as wanting to focus on individual items but to pick up connected elements that are important without either aggregating all of the data or removing them from their connections with the rest of the dataset. I can imagine that there could be powerful tool that allows researchers to investigating small snippets of data while a statistical engine runs in the background, clustering or otherwise picking out connections between those snippets and the rest of the dataset.
Capturing the inner narrativity of overlapping monads
I will have to think about this one, but it is something about bridging the gap between human narrative and statistical analysis. He referred to data journalism and how data is used in both interactive and narrative contexts in things like the guardian coverage of the london riots.
Visualizing heritage, process and genealogy
How to visualise these temporal qualities without relying in static structure or loosing connectedness. While answering questions he stressed the importance of not falling back on unchanging structures but acknowledging the changing nature of monads and their connections. This seems to me to relate to another theme that came up quite a lot in CHI (from Bill Buxton to the NIME SIG), the need to have time as a first class concept. This would make it possible to model the evolution of data without relying on static structures (maybe).
Replacing model building and emergent structure by highlighting differently overlapping monads.
I guess that this would require a very dynamic analysis that made it possible to apply many different and changing models to data. I think that interactive machine learning could help a lot here by using the human element to navigate different interpretations, learnt models and views of the data.
Nate Matias has a much more accurate write up of the challenges. I’ll quote them below, but I’ll just warn that I think he’s not quite right about collective phenomena. He says that “Collective phenomena grow out of these collecting sites”, but this isn’t quite what Latour is saying, after all collective phenomena don’t exist so they can’t emerge. Latour is from a Science Studies background so when he talks about collecting sites he is thinking about scientific data collection instruments (or their aggregations) like microscopes, telescopes, mass spectrometers, semi-structured interviews or surveys. While we might think of a survey as collective and an interview as individual they are both just methods of collecting data which both observe and transform (“perform”) the data thus creating a particular view on the world. There is no innate distinction between individual and collective phenomena in the world, just phenomena that are created by methods of data collection (or more precisely by their interaction with the world). This means that there isn’t a division into two (or more scales) collective and individual but a mass of different views of the world, each specific to it’s data collection method.
Anyway, that small criticism aside, well done to Nate, for the otherwise excellent explanation, here is his summary of the challenges:
Visual complexity produces opacity. Massive individualizing data produces beautiful, playful hairballs which show us nothing. How do we get filter and focus data while still appreciating monads?
How can we capture the inner narrativity of overlapping monads? Latour shows us the “512 paths to the White House” visualization by Mike Bostock and Shan Carter. The other, the Guardian’s Rumour tracker, following the 2011 London riots. The idea that quantitative is different from qualitative is an artifact of the history of social science and a fallacy arising from the distinction between the individual and the collective, he tells us.
How can we visualize heritage, process, and genealogies? Latour shows us a paper he worked with on “complex systems science” (I couldn’t find it). To be a monad is to establish connections, but timeseries visualizations can focus on structure rather than connectedness (like the paper on Phylomemetic Patterns in Science Evolution by Chavalarias, Cointet et al)
How can we replace models about emergent structures with models that highlight differentially overlapping monads? He shows us a hairball network diagram and talks about the difficulty of moving beyond the hairball to understand the overlapping monads