Structuring a Data Vis Hackathon
Chairman, Create:Fixate Arts Foundation
Tengrade is creating the single central source of popular opinions on everything, utilizing "Startags" to let users submit their ratings on any hashtag directly through Facebook and Twitter. We believe there will be MANY great ways to visualize the data we'll be collecting, and we want to get the Data Vis community excited about our data set. We're looking to throw a hackathon or other contest and want input on how best to structure it.
Bio: Ted Werth is the co-founder and CEO of Tengrade. He previously co-founded TotalNY.com (sold to AOL) and Digital Club Network (merged with eMusic). Ted is also an LED installation artist and Chairman of Create:Fixate, a non-profit that supports emerging artists in the LA area through large-scale exhibitions and education programs.
Visualizing structural change
I will discuss Bayesian dynamic model averaging, the practice of weighting statistical models by their likelihood and prediction accuracy over time, and its role as a basis for visual exploratory analysis. By generating weighted predictions from a collection of models whose weights can change over time, and using a "forgetting" parameter to discount observations progressively as time goes by, this approach to modeling combines aspects of static predictions and time series analyses. Moreover, the strength of a given collection of predictors relative to others is easily grasped by their weight, and once the initial ensemble is trained, online prediction requires minimal additional computation. Last but not least, recent advances in deep neural network training have revealed surprising connections to this approach via so-called "dropout" regularization for faster learning. The key feature of all the above is that they lend themselves to visual representations -- simple inspection of changing model weights over time often leads to much deeper questions and insights about what is being modeled (rather than merely models). I will wrap up with a simple web application showing how active versions of this approach can be used for statistically sound inference about recommendations, predictions, and experimental designs. By embracing change as the most reliable of constants, we free ourselves to explore complicated phenomena from whichever angle is most appropriate to the task at hand.
- 6:00pm networking
- 7:00pm talks start promptly
You will have to arrive between 6:00-6:55pm in order to pass security (see more info below).
Please RSVP as places are limited.
Venue: Cornerstone OnDemand http://www.cornerstoneondemand.com/ will kindly host this meetup. There is no parking provided, but at that time of the day it's not difficult to find parking nearby. Please note that only members with RSVP Yes will be able to clear security.