SIE Colloquium by Matthew Gerber, Research Assistant Professor in the Systems and Information Engineering Department.
The PTL group has 2 faculty, 10 grad students, and collaborators at the health system.
Predicting crime using twitter:
- Conventional warfare had easily identified forces and open conflict with direct attacks (friends/enemies). The US has no conventional military peers. The US us dealing with asymmetric warfare (asymmetry in size, power, funding, influence). Our enemies have tactical advantages.
- Monitoring via hot-spot maps
- Problems: very specific to the are you're studying and it's retrospective. Can't take yesterday's model and predict on a different place today.
- Overview of the approach
- Gather information on potential crime correlates (Incident Layer, Grid Layer, Demographic Layer, Spatial Layer). Ex: newar military outpost? religious site? Income levels and ethnic tension, and prior history (each on a different layer). Want to take these information and create a statistical model.
- Text provides a problem: unstructured text abounds. These short tweets should be helpful: "The second blast was caused by a motorcycle bomb targeting a minibus in the Domeez area in the south of the city. That needs to be read by a human or automated approach (this talk).
- Automatically integrate unstructured text: add some new layers from the previous model (Twitter Layer, Newswire Layer, ...).
- He's looking at tweets from the Chicago area (collecting in the basement of olsson--time, text, etc). A few topics: 1) flight(0.54), plane(0.2), terminal(0.11),... ; 2) shopping (0.39), buy(--),...
- Mapping these
topics to heat map of Chicago. Can see where certain things are being talked about.
- Unsupervised topic modeling
- Latent Dirichlet allocation (Blei et al 2003)
- A generative story (2 topics). Outside of these documents live topics. We can generate these. Do a similar thing with the documents (grab a dirichlet distribution and produce another--a distribution of topics that the document consists of). Want to pick a topic from that distribution to generate a word. (generate by repeating this process).
- Gather tweets from a neighborhood, tokenize and filter words, identify topic probabilities by LDA, compute probability of crime
. The question what is
?
.
- Find the beta coefficients that give the best function
- Training
- Establish training window (1/1/13-1/31/13)
- Lay down non-crime points
- lay down crime points from training window
- Compute topic neighborhoods
- compile training data (use Kernel Density Estimate (?) that adds historical data to the model)
- Evaluation
- Want to find the smallest place boundaries with the highest crime levels.
- Do people actually talk about crime on twitter? (that's the big question-- but gangs do trash-talk about their crimes, etc)
- Baseline for comparison was the kernel density estimation (based on past, where is crime likely to occur?)
- They do well with twitter data model + KDE over just KDE for certain results: prostitution, battery.
- They are worse with other topics/crime: homicide, liquor law violations.
- AUC improvement for 22 of 25 crime types, with average peak improvement of 11 points
- Mapping these
- Clinical Practice Guidelines
- Want to formalize using natural language processing
- Sentences have a specific order: they're using NLP and parsing English sentences. (concern: context sensitivity of English)
- Want to annotate the text with semantic labels (not XML, though).
- Precisions: temporal identifiers 28% are identified; others average around 50%, with the top around 75-80%
- Warning: need to make sure that fully automated isn't used alone, as there could be things that automated analysis would miss that could be life-threatening.
- The big picture
- Want to get structured information from unstructured text data through Natural Language Processing
Leave a Reply