This paper describes a new way of modeling the changes in a state space over time called Continuous Time Bayesian Networks (CTBNs). It uses a novel graphical model framework which competes with a Dynamic Bayesian Network and is specifically trying to move away from methods which require high frequency discrete updates to "simulate" continuous time. It also attempts to address the correlation problem which Koller describes as the entanglement problem, an effect in DBNs which causes independent random variables to have a dependency due to frequent updates.

At the core of this technique are homogenous Markov processes (HMP) which look like a particular interpretation of Markov Models: The transitions in a Markov process correspond to temporal changes. The transition probabilities do not change over time and are therefore "homogenous". Every state has a self-transition and therefore the distribution over staying in a state is exponential. Instead of a conditional probability table, there is a "conditional intensity matrix" (CIM) which describes the exponential transitions between states.

• Keynote address by Mozelle Thopson, FTC: Spam is bad, Spyware is bad, Porn is bad, I'm busy fixing it all with the help of industry partners.
• Keynote address by James J. Duderstadt, University of Michigan: The Internet is a disruptive technology. It is changing fast. It's affecting Universities. I'm at the forefront of getting universities to "get it."
• The most interesting talk that I've seen this entire conference is a discussion of how memes propogate through blog-space. No big surprises: there are a few very influential individuals and lots of moderately influential people. There were a few off-hand comments made about how to leverage these people for various reasons (e.g., marketing) but I think they were poorly thought out. I think the reason why some people are influential is because they are insightful, they have a monopoly on a information channel, or they provide a service. The second person is the one who can be used to market a product, but the first and last are valuable because they will not market a product that they don't believe in. A better idea is to get the first and last to review your product and listen to what they say before marketing something.
• More on target for this summer's research was an interesting talk on clustering web pages based on tree-edit distances. Since web news sites ( and many others ) are just a template that is filled out with database information it is possible to find the portions of the HTML tree which contain the interesting information. The interesting idea was that there is an algorithm which will figure on the tree-edit distance of two trees. So two trees can be generalized into a single tree with various wildcard nodes. This suggests a method of merging two RFID traces from the anesthesiologists in an interesting way. The presence of a tree might map well to a hierarchical representation of the task. The tree-merge / tree-edit technique might do a good job of merging different expert traces. One of the questions he received was: Why didn't he use XPath to represent the general tree? Good question. Maybe there are XPath algorithms that will generalize two trees.

• Ud Manber Keynote Address about Amazon's A9 search engine
  • Amazon wants to invent new things in search focussed on e-commerce, that are easy to use, relevant and empirically verifiable as better (particularly as quality is concerned). A9 is a wholly owned subsidiary which is set up to do this.
  • "Search inside a book" is one of the first apps. Hundreds of thousands of books scanned then OCR'd in about six months.
  • "Search history" is another feature. This also includes "Click History" and a display of what new sites have been found for a given query since the last time you made that search query. Lots of privacy implications here, so Amazon has an unpersonalized site for the people concerned with privacy that still want to use the search engine.

• Today is "WWW Day" according to Mayor Bloomberg
• Tim Berners-Lee Keynote Address:
  • The new top-level domain names are a bad idea because they aren't a fundamentally new idea. They are supposedly trying to increase the number of short domain names, but in fact someone who owns www.example.com is more or less forced to buy www.example.mobi to maintain trademark control. So at the end of the day it looked like it was just a cash cow for the domain name registers.
  • He used the acronym FOAF: Friend of a Friend. I hadn't heard of that acronym before.
  • Good point: The semantic web will take off as people write thin skins that publish things in RDF format as well as current formats like HTML.
  • A question was posed about how the semantic web could take off because it commoditizes a lot of what web-sites have done to differentiate themselves and it breaks the advertising business model. Tim's response was that the semantic web is a new thing and therefore old business models won't necessarily work. He also made the point that in a lot of cases, communication benefits both parties and in these situations publishing semantic web data will work because both the provider and the receiver benefit. He cautioned against data providers who corrupt the data with advertising such as data integrators that only provide data on restaurants that have paid to be included. This seems like it won't be a problem as long as it's clear that people are paying to be included in the stream. Just like search engines that allow paid placements. Sometimes you want that, sometimes you don't, but you never want to not know which you are using.
• Oren referenced a citation for the Google score that we use in GUIDE to figure out the probability of an object in an activity. He called it point-wise mutual information and said it was introduced by "Turney '01". I'll email him for more details.
• "Haystack" is the name of a semantic web browser that looks pretty cool. It uses semantic information to control the views of information. It looks like it might be something big in an early stage.