Discussion Notes

March 23, 2001

(courtesy Marc Vass)

Modeling Users with Spreading Activation from ACT-R Theory

(Anderson - http://act.psy.cmu.edu/)

• Activation Equations

A_i = B_i W_j S_ji

where A_i is the activation of chunk i, B_i is the base-level activation of chunk i (which defaults to 0), and the summation is over the elements j which are the current sources of activation. The W_j are the weightings of these sources, which are referred to as source activations. A unit of source activation is divided up evenly among the connected chunks. The S_ji are the strengths of association from source j to i.

S_ji = ln(m/n) = ln(m) - ln(n)

where m is the number of chunks and n is the number of chunks associated with the source chunk.

B is given by taking the log of the sum of the length of time since the last practice, t_j, for each practice trial to the power of d, which is taken normally to be .5.

• Mapping this to web pages




Ai = Bi Wj Sji

where A_i is the activation of web page i, B_i is the base-level activation of webpage i (which defaults to 0), and the summation is over the webpages j which are being currently viewed. The W_j are the weightings of these pages, which are referred to as source web pages. A unit of source activation is divided up evenly among the connected web pages. The S_ji are the strengths of association from web page j to i.

S_ji = ln(m/n) = ln(m) - ln(n)

where m is the total number of web pages and n is the number of web pages associated with the source web page.

B can be taken to be 0, which means no one has viewed it before, or set to a value representative of its previous usage by the user or users.

• Context is much like adding additional sources of activation to the web page graph.

Context in Web Search

• Context can be modeled using spreading activation by using the context as sources of activation
• Personalization can be viewed as using context to tailor results
• What a user is currently doing in other windows may be related to what they are searching for
• The user can specify what the context of their search is
• Can provide search engines that are built to work within a certain context(s), such as scientific literature, to take advantage of the domain or to provide functionality not given by basic search engines ("hot-rods")

Silk from a Sow’s Ear: Extracting Usable Structures from the Web

• Associative retrieval of web pages
• Automatic Categorization
• Uses ACT theory as basis for spreading activation model of information seeking on the web
• Use approximations such as inter-document text similarity and number of times link has been followed for the matrices in the activation algorithm
• The activation algorithm needs to be iterated a limited number of times before the activation that spreads in negligible
• A chosen activation level determines the value at and above which pages will be returned
• Techniques used in paper assume that user has a single goal in mind. Additional production rules must be added to handle parallel goal seeking