paper: SNIF-ACT

I just read this paper by my research group's principal scientist, Peter Pirolli, and former PARC employee Wai-Tat Fu. The paper, entitled "SNIF-ACT: A Model of Information Foraging on the World Wide Web", recently won the Best Theoretical Paper Award at the 9th International Conference on User Modeling.

The paper extends the existing ACT-R cognitive modeling infrastructure to computationally simulate users surfing the web, at a fairly fine-grained psychological level. The system models the user's goals, knowledge, memory, and abilities (in the form of production rules) and combines these with the findings of Information Foraging theory to create a successful model of web surfing and decision making. Information Foraging theory applies the metaphor of animals foraging for food to the task of humans seeking information. In previous work, Pirolli and Card have found that the equations governing the cost structures of the two are the same.

The SNIF-ACT model works by extracting the content and links of a web page and then using a technique known as spreading activation to propagate "activity" through an associative memory network of individual words. Activation proceeds from the modeled user goals through the terms in working memory and out to the currently observed web content. Link weightings between word associations are determined by using word occurence and co-occurrence rates extracted from AltaVista. By finding the highest mutual activity between user goals and available links, the system can compute an estimate of the information scent (much like the scent tracked down by animals in the wild), and use this to construct a probability distribution of the likelihood of following different links. Drop-offs in scent measures are also used to predict when a user will leave the current web site to look elsewhere for a richer information patch (analogous to an animal moving on to greener pastures or hunting an easier prey). The SNIF-ACT model is psychologically richer than previous foraging-influenced systems like Bloodhound, which primarily uses techniques from the information retrieval (IR) field and earlier flawed cognitive approaches.

For more details about information foraging theory and it's applications, check out this essay by pixelcharmer (it even cites my first research paper!), this copy of a talk by Pete Pirolli, and my research group's publication archives.

There are at least two interesting avenues for this work to follow. One is in applications, as successful user models can create better automated usability metrics and could learn from individual behavior to create personalized research and surfing tools. The second is to simultaneously move from the web to other domains, building user models for other content-rich domains (e.g., information visualization). Down the road, I think the integration of content-based and perception-based (e.g. computer vision and audition) analyses will be the next big research leap - creating richer, more realistic, models of user behavior and furthering artificial intelligence research.