Deriving Recommendations from User Behavior in Digital Libraries

                            Johan Bollen
                   Los Alamos National Laboratory

Abstract:

The proliferation of distributed knowledge systems (e.g. the World Wide Web
(WWW), Digital Libraries and genome databases) has increased the need for
reliable and valid methods of information linking. The latter is required
for efficient human retrieval and the study of the structural properties of
a given domain. Existing Information Retrieval methodologies for
information linking (e.g. vector space models) are largely text-content or
meta-data reliant,  and can therefore only be applied to traditional
document collections. They are furthermore sensitive to synonymy, polysemy
and the use of different languages. Other approaches are limited to the
analysis of the graph-theoretical properties of existing author-defined
document networks (e.g. citation graphs and WWW hyperlink structure).

I present a methodology that derives document relations from an analyses of
user retrieval sequences and therefore operates independent from document
text content or author-based design. Three Hebbian-based learning rules
modify link weights according to sequences of user retrieval requests and
gradually adapt existing hypertext networks or document relations to
changes in user preferences. The generated networks have been shown to
reliably and validly represent the preferences of its community of users.

The presentation consists of three parts.

First, I discuss a prototype of a system that generates hyperlinks for a
collection of reduced hypertext pages based on the retrieval sequences
registered for a large group of users. The results have been evaluated in terms
of network development reliability and validity by a test-retest methodology
using a simulation of user retrieval behavior. I briefly discuss the
computational model of user retrieval behavior used in these simulations. A
recent modification of this model using an associative optimization heuristic
has been shown to yield accurate predictions of user navigation paths in a
large web site.

Second, I present an application of the discussed system to the generation of
information links in Digital Library systems. A prototype has been implemented
which generates large networks of document and journal relations from user
retrieval sequences as they were registered in the Los Alamos National
Laboratory (LANL) Research Library (RL) server logs. One such generated network
was compared to a citation network for the same set of journals. The latter was
derived from the Institute for Scientific Indexing 1998 journal citation data.
Results indicate large discrepancies between the graph-theoretical properties
of the user-determined and the ISI citation network, and raise questions
regarding the validity of ISI Impact Factors.

Third, I provide a demonstration of a prototype of a Spreading Activation
journal recommendation system operating on the generated journal networks. The
system allows users to expand initial keyterm queries by honing lists of
recommended journals and will be connected to the LANL RL MyLibrary
personalization service.

Biography Johan Bollen:

Johan Bollen is a Postdoctoral Research Associate in the Computer and
Computational Science Division at the Los Alamos National Laboratory, where he
works on the Active Recommendation Project. He received his BA and MS in
Experimental Psychology (Specialization Cognitive Science and AI) at the Vrije
Universiteit Brussel with Distinction in 1994. His MS thesis concerned the
implementation of neural behavior systems for autonomous robotics and was
awarded Highest Distinction. He successfully defended his PhD dissertation in
July 2001 and will be awarded a PhD degree in Psychology in October 2001 from
the Vrije Universiteit Brussel. His research focuses on systems for automated
information linking, human factors in web design and navigation, Distributed
Knowledge Systems, Information Retrieval and Digital Library recommendation
systems. He has authored and co-authored numerous articles in peer reviewed
conference proceedings and journals, co-edited an anthology entitled "The
Evolution of Complexity" and co-organized several international symposia and
workshops. His research on distributed knowledge systems for the WWW has
received widespread attention in the scientific and popular press.



BIBTEX:
@ARTICLE{system:bollen1998,
           author = {Johan Bollen and Francis Heylighen},
           year = 1998,
           title = {A system to restructure hypertext networks into valid
user models},
           journal = {The new review of Hypermedia and Multimedia},
           volume = 4,
           pages = {189--213},
}

@INPROCEEDINGS{groupu:bollen2000,
           author = {Johan Bollen},
           title = {Group user models for personalized hyperlink
recommendation},
           year = 2000,
           publisher = {Springer Verlag},
           address = {Trento},
           month = {August},
           booktitle = {LNCS 1892 - International Conference on Adaptive
Hypermedia
and Adaptive Web-based Systems (AH2000)},
           pages = {39--50},
}

@INPROCEEDINGS{adapti:bollen2000,
           author = {Johan Bollen and Luis M. Rocha},
           title = {An adaptive systems approach to the implementation and
evaluation
of digital library recommendation systems},
           year = 2000,
           publisher = {Springer Verlag},
           address = {Lisbon},
           month = {September},
           booktitle = {LNCS - Fourth European Conference on Research and
Advanced
Technology for Digital Libraries (ECDL2000)},
}