S. E. Preece. A Spreading Activation Network Model for Information Retrieval. PhD thesis, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA, October 1981. [url]

———————–

Despite being an outdated work on information retrieval, this thesis present an innovative associative retrieval method that demonstrate interesting results, and that aims at improving the search process itself.

The basic idea is to use coincidence matrix, whose rows correnspod to the document and whose columns correspond to attributes. On top of this representation of the dataset, the author builds an associative retrieval scheme, which generate a second order level of associations between the terms and that is based on a very simple process of diffusion through the network.

Each node of this connectivity matrix can be marked as active or inactive (or any real value between 0 and 1). Now a simple iterative process of spreading can trigger all the nodes connected to an initial one.

The author propose several types of processing within this framework: Cumulative processing (which deals only with the current level of activation);  Coincidence processing (which deals with activation trails); and Context processing (which uses geometric properties of the pattern of activation).

Chapter three describe the model in details. What follows are some random notes.

Parameters of processing determine how the activation spreads among the nodes. Processing is iterative, consisting of a sequence of spreading pulses. The author propose several kinds of spreafing methods, one is the scource node distribution, which determines the general pattern of spreading; the full strenght spreading (each link attached to an active node receives the full strenght of the node); an inverse source frequency spreading (each link receives an equal share of the node’s strenght). Thresholds may be applied to the links to act as a noise filters. The activation may also assume negative values.

Tags: information retrieval, relevance theory, text data mining