09 Apr

SNIF-ACT- A Model of Information Foraging on the World Wide Web

The presentatoin describes a computational model that mimics people interaction with the WWW. This model is important because provides principles for improving usability and helps people to understand the best way to find and use information on the web for significant problems such as health, finance, career, etc.

Authors: Pirolli, P. L. & Fu, W-T.

Year: 2003

Link to the original paper

SNIF-ACT integrates Information Foraging Theory and the ACT-R model. The latter is a computational model developed by Anderson to model human psychology. In order to understand how humans interact with the web, the authors registered all states and events of two users while they interacted with the web. These data were saved in a database and compared via a user-tracing architecture to the data produced by the SNIF-ACT model.

The ACT-R architecture is based on two major components:

Declarative memory includes knowledge represented in chunks that can be either active or not active. Active chunks represent the knowledge required in a particular time. For example, they could represent the current goal of finding an information of the WWW.
Procedural memory includes skills. These skills tell the model how to transform declarative knowledge into behaviour. They are also called conditional production rules. An example of a rule could be “Use-search-engine”: IF the user needs to go on a website and he is in front of a browser, THEN he uses a search engine”. Only one rule at the time can be applied and if there are many matches for one goal, the conflict resolution mechanism decides which rule has the highest utility for the current goal. The utility function is provided by information foraging theory.

The utility function used used by the conflict resolution mechanism is based on information scent. Information scent includes the cues (links, images, and text) that the user process in making judgements. When real-users use the Internet, they mostly focus of the content of the webpages visited. For this reason, information scent is content-based.
The information scent of every page visited activates chunks in declarative knowledge via a spreading activation network. This activation indicates how relevant the page is to the current goal. For example, if the user is looking for a photo of a friend, a webpage with the friend's name would have high information scent, while a page with no content related to the friend would have low information scent. The model predicts that user will choose the rule that leads to the highest information scent.

The authors predicted that novice users would follow links that have high information scent. The second prediction was that users would leave a webpage when the utility of the page diminishes below the utility of moving to a new website.

The model used thes two prediction to match the behaviour of two users performing two different tasks online such as finding some posters to buy or the date of an event. While completing the tasks, user tracing instrumentation was used to register the behaviour of the user interacting with the web. The researhcers recorded eye movement, a file log of the actions on the screen, and a video recording of the user thinking aloud.

After the data was recorded, a user trace comparator controlled the SNIF-ACT simulation model and matched the simulation behaviour with the user data for each step. Each production rule that was selected by SNIF-ACT was compared to the user action. If the two actions matched, the production selected by SNIF-ACT was executed. If they did not match, the production that matched the user action was executed.

The SNIF-ACT model modelled the user behaviour very well. For link-following actions, the model reliably predicted which link the user chose. The distribution of predicted link selection was significantly different from random selection (first graph). For site-leaving actions, the model also correctly matched the user behaviour of leaving a page. The second graph shows the four last pages visited by a user before leaving the webpage. The dotted line represents the mean of the information scent of the page appearing after the user left the website. This was consistent with the prediction: users left the pages when the utility (information scent) of the page was below the utility of moving to a new website.

Pirolli, P. L. & Fu, W-T. (2003)SNIF-ACT: A Model of Information Foraging on the World Wide Web. Ninth International Conference on User Modeling, Johnstown, PA .

Leave a Reply