May 29, 2009
preventing hospital to hospital infection spread
For the past few months I've been collaborating with Jack Iwashyna, Assistant Professor at UofM's medical school and SI MSI student Umanka Hebbar Karkada. Jack had a fun idea, and a fun data set - hospital to hospital patient transfers, mined from medicare claims. These transfers are a way for highly resistant infections to jump from one critical care unit to another. Mostly hospitals devote resources to preventing infection spread separately from one another.
We posed the question of how resources could be allocated in a coordinated way to maximally stem the spread of infection. Umanka and I tried several stategies - targeting hospitals with the highest degree (number of hospitals they trade patients with), highest betweenness (they are on the "path" between other hospitals), and a greedy allocation based on the number of beds infected at each hospital and downstream from that hospital.
The results are here. Both figures show hospitals as nodes sized by the number of ICU beds they have.
This shows the number of resources allocated by hospital (gray = none, blue = few, red = many).
This shows the relative benefit of a random allocation vs. targeting particular hospitals. (blue = hospital unlikely to become infected, red = likely to be infected)
May 21, 2009
Xiaolin's defense on June 3rd
Xiaolin Shi will be defending her thesis on June 3rd, 2:00 - 4:00 PM, in CSE 3725 (the computer science building). She's the first student of mine to have reached this stage, and so I'm experiencing a bit of anxiety, though she is quite ready. Next up for her is a postdoc with Dan MacFarland at Stanford (she'll be part of an interdisciplinary team of computer scientists, linguists, and education-folk to study how the education environment impacts future scholarly performance).
Info on her thesis:
THE STRUCTURE AND DYNAMICS OF INFORMATION SHARING NETWORKS
Information flows are produced, carried, and directed by information sharing networks. And the evolution of the structure of such networks and the way information diffuses are affected by one another. This thesis studies structural features of information networks and their relationships to information diffusion...
It starts with the
robustness study of topological features when the data sets are sampled from net-
works which are rapidly evolving, as is the case in large-scale online blog networks.
The features of blog networks are found to be stable upon aggregation with compre-
hensive data sets, even as individual network ties are highly intermittent. Another
salient structural feature of such networks is that a small number of vertices play a
disproportionately important role through their position and connectivity. In sev-
eral online information networks studied in this thesis, one can construct subgraphs
of vertices according to different importance measures, while consistently preserv-
ing their attributes such as connectivity and shortest paths in the original networks.
Simple connectivity through arbitrary ties is sometimes insufficient to transmit infor-
mation, because ties may need to be of a given strength in some real-world scenarios.
In this thesis, we further show that strong ties percolate through online social net-
works, and increasing the required threshold strength does not break up the network
into isolated communities, nor lengthen the average shortest paths of the networks.
After examining the structural features which will potentially affect the flow of
information, this thesis further examines the actual relationship between structure
and information flow, and flow and impact. A thorough analysis based on citation
networks indicates that information is less likely to flow across community bound-
aries, and a publication's citing across disciplines is tied to its subsequent impact.
In the case of patents and natural science publications, those that are cited at least
once are cited slightly more when they draw on research outside of their area. In
contrast, in the social sciences, citing within one's own field tends to be positively
correlated with impact. This thesis also studies information diffusion in online com-
munities. The patterns of information diffusion curves reflecting user behavior in
joining groups and the feature factors associated with users or groups that influence
such behavior are studied. Bipartite Markov Random Field (BiMRF) models are
built to help understand the relationships of these features, as well as the differences
in their impact in different types of online forums.
gephi for plotting spiffy-looking network visualizations
* easily subsetting nodes according to attributes
* getting node labels to jiggle around until they no longer overlap
* drawing curved arcs (and controlling the curvature)
May 05, 2009
Social Influence and the Diffusion of User Created Content
Eytan, Brian and I have a paper at EC (to be presented by Eytan @ Stanford in July).
It's on the diffusion of gestures in Second Life (the online massively-multiplayer virtual world). The neat thing about SL users passing assets around is that it leaves digital traces. We were able to surmise that roughly half of the transfers occur between friends (according to the explicit social graph), and that in 38% of the remaining cases a user adopts after their a friend does. Not only that, but as more friends adopt, the hazard of adopting increases. Transfers along the social network are faster, but the overall spread is more limited... Influencers and adopters are distinct groups... What else? Well, you'll have to read the paper :).