Hebbian Algorithms for a Digital Library Recommendation System.

Abstract

This paper proposes a set of algorithms to extract metadata about the documents in a digital library from the way these documents are used. Inspired by the learning of connections in the brain, the system assumes that documents develop stronger associations as they are more frequently co-activated. Co-activation corresponds to consultation by the same user, and decreases exponentially with the time interval between consultations. The strength of activation is proportional to the user's interest for the document, either evaluated explicitly, or inferred implicitly from user actions or the duration of the consultation. Co-activation values are added, producing a matrix of associations. This matrix can be used to recommend the documents that are most strongly related to a given document, most relevant to the user's implicit interest profile, or most interesting to users overall. Moreover, it allows the calculation of document similarity values, which in turn can be used to cluster similar documents. The data needed to feed such a recommendation system are readily extracted from the usage logs of document servers, and can be processed either in a centralized or a distributed manner.

Full text

Page 1

Hebbian Algorithms for a Digital Library Recommendation System
Francis Heylighen
CLEA, Free University of Brussels
http://pcp.vub.ac.be/HEYL.html
Johan Bollen
Computer Science, Old Dominion University
http://www.cs.odu.edu/~jbollen
Abstract
meta-data, so that documents on any specific subject can
be transparently retrieved.
While quality control can in principle still rely on the
traditional methods of peer-refereeing and evaluation by
experts that work relatively well with paper documents,
retrieval on the basis of metadata has some intrinsic
shortcomings, which can only get worse as the number of
documents in the library increases.
The first set of problems derives from the fact that
metadata can never completely capture the subject or
meaning of a document: traditional metadata merely
provide a coarse and rigid categorization, which can never
specify all potentially relevant characteristics [7]. Free-
form keywords provide perhaps the most flexible kind of
traditional metadata. Yet, keywords suffer from the
problems of synonyms (the user may enter a keyword
similar in meaning but different in form to the one by
which documents are classified, and therefore fail to locate
a relevant document), and homonyms (the users may enter
a keyword similar in form but different in meaning, and
therefore receive an inappropriate document). Moreover,
when the subject is new or as yet unclear to the user, the
user may not be able to formulate any relevant keyword.
Combined with huge collections and limited selection this
leads to queries showering the user with material of little
relevance, in which perhaps a few nuggets of true value
are hidden.
A second set of problems derive from the fact that
assigning good (if by definition incomplete) metadata to
documents requires a great effort and a special expertise:
only people who know the domain well, and have studied
the document well can determine the appropriate
keywords or categories. This problem is to some extent
mitigated for electronic texts, since IR algorithms can
extract the most distinctive words from the text to be used
as keywords, as is done by search engines on the web.
However, this method is useless for non-text documents,
such as pictures, movies or sounds.
The present paper proposes a general approach that
seems able to tackle these problems. The idea is to extract
metadata not from the content of the documents but from
the pattern of document usage, assuming that users have
an intuitive grasp of what a document is about and how
valuable it is, and that this intuition guides their actions.
As we will show in the next section, such metadata do not
rely on fixed categories or keywords, but on the variable
This paper proposes a set of algorithms to extract
metadata about the documents in a digital library from the
way these documents are used. Inspired by the learning of
connections in the brain, the system assumes that
documents develop stronger associations as they are more
frequently co-activated. Co-activation corresponds to
consultation by the same user, and decreases
exponentially with the time interval between consultations.
The strength of activation is proportional to the user’s
interest for the document, either evaluated explicitly, or
inferred implicitly from user actions or the duration of the
consultation. Co-activation values are added, producing a
matrix of associations. This matrix can be used to
recommend the documents that are most strongly related
to a given document, most relevant to the user’s implicit
interest profile, or most interesting to users overall.
Moreover, it allows the calculation of document similarity
values, which in turn can be used to cluster similar
documents. The data needed to feed such a
recommendation system are readily extracted from the
usage logs of document servers, and can be processed
either in a centralized or a distributed manner.
1.Introduction
Compared to traditional libraries, the World-Wide Web
has some spectacular advantages: the range of documents
it proposes is much wider, they are easier to consult, they
are available always and everywhere, and their electronic
format makes it easy to search for specific phrases or
keywords. On the other hand, the web’s disadvantages are
obvious too: an almost total lack of organization of the
material, and virtually no selection for quality or
trustworthiness.
Distributed digital libraries hold the promise of
combining the benefits of both the web and traditional
paper libraries. Their electronic documents would be
available to everybody via the Internet, yet a staff of
editors, “cybrarians”, or information scientists would
guarantee that all documents fulfil minimum quality
requirements, and that they are organized according to a
coherent system of categories, keywords or more generally

Page 2

associations that exist between documents in their users’
mind. Since interesting documents will anyway be
consulted by users and this activity is stored in usage logs,
analysing usage patterns allows us to collect metadata
without requiring additional effort from either users or
librarians.
It must be noted that recently a variety of techniques
have been developed to mine knowledge from data about
document usage (see e.g. [15,16]). However, these
techniques constitute a rather ad hoc collection, with little
integration or motivation and no underlying theory to
explain how users navigate through a web of linked
documents [17]. Moreover, most of these techniques are
based on the clustering of navigation paths, user profiles,
or documents into a static set of discrete categories, thus
suffering from all the shortcomings of rigid categorization.
Our approach, on the other hand, which builds on our
experimental results with a website that adapts to or learns
from the way it is used [1, 10, 17, 19], tries to discover the
finely-graded, continuous associations between documents
that trace the users’ constantly changing focus of interest
while browsing a document collection. Clustering in
categories is one possible application of our approach, but
should be seen only as an “afterthought” that is not
fundamental, since different clusterings can be made in
different contexts. Moreover, our approach, whose first
development dates back to 1995 [19], is characterized by
its unique, coherent paradigm, based on the analogy
between the dynamic organization of a document network
and the organization of the brain [18], where concepts or
neurons are connected through variable strength
associations or synapses, whose strength evolves
according to the rule of Hebb [6].
extreme case which is likely to be found only if document
dj is a copy, excerpt or summary of di.
In such an associative network, every node or
document can now be represented by a vector:
di = (ai1, ai2, ..., ain)
The assumption underlying the bootstrapping model is that
this vector captures the essential meaning of the document
relative to other documents. Therefore, the contents of this
vector can be interpreted as representing the associative
metadata about document i. If keywords and categories are
seen as a discrete, symbolic representation of a
document’s meaning, then an association vector provides a
continuous, subsymbolic representation, of the kind used
in distributed or connectionist models of cognition.
2.2.A Hebbian learning rule
Associative networks are inspired by the functioning of
the brain, where on the higher, abstract level concepts are
connected by associations, and on the underlying, physical
level neurons are connected by variable strength synapses.
An association aij represents the degree or probability of
activation of neuron/concept j following the activation of i.
In the brain, associations are learned through the rule of
Hebb [6, 1]: concepts that are activated simultaneously
(co-activation) become more strongly associated. This
strengthening is proportional to the degree of activation
A(i) and A(j) of each concept. Since we can assume that
activation decays with the time that has passed since the
initial stimulus that created the activation, the degree of
co-activation will decrease exponentially with the time
interval (tj - ti) between the activation of i and the
subsequent activation of j.
This brings us to the following formula for the bonus or
association strength added by a particular episode of co-
activation, where d is a constant decay factor:
2. Associative networks that learn
2.1.Association matrices
A(i, ti).A(j, tj).exp (- d.(tj - t i)).
Rather than analysing a document’s content or
components, an alternative way to define meaning is
through bootstrapping: a document’s meaning consists of
the whole of associations it has with other words or
documents [8, 9]. An association between two documents
di and dj can be defined as a measure of the degree of
relatedness between di and dj, or the degree of
“expectancy” of dj, given di.
An associative network is a weighted, directed graph,
whose nodes di represent documents, and whose weights
represent the association between nodes. It can be
represented as a matrix whose components aij ∈ [0, 1]
correspond to the connection weights between nodes d
and dj. [9]
This matrix is generally sparse, since most associations
will have value 0, which means that encountering node i
does not in any way prepare the mind to encounter j. A
maximum weight of 1 means that given di, everything is
already known about dj; dj does not provide any additional
information that isn’t yet contained in di. This is an
The total association strength is then merely the
(possibly normalized) sum over all episodes of co-
activation of the strength bonuses for each co-activation.
Note that a bonus can be negative if we allow for
negative activation values. This means that association
strength can decrease if a positive activation of i is
followed by a negative activation of j, or vice-versa.
Note also that the different co-activations can be
weighted in the calculation of the total association strength
so that more recent co-activations make a larger
contribution. This is useful in circumstances where the
pattern of usage regularly changes depending on new
developments or social or cultural changes in the group of
users. Again, our brain paradigm would suggest an
exponential decay factor d’ that would lessen the impact of
older contributions depending on the time that has passed.
An efficient way to compute the overal value would be
to store only the total association strength S together with
i

Page 3

a time stamp (t1) of when that strength was last updated.
Whenever an updated association strength S(t2) is needed,
this is calculated by multiplying the previously stored
value with the decay factor: S(t2) = S(t1) . exp (-d’(t2–t1)).
A new co-activation at t2 is then simply added, without
decay factor, to this reduced sum of all previous
activations, and this new value is stored together with the
new time stamp t2. Each next time ti that an update is to be
made, the same procedure is applied recursively. In that
way, the value at any moment will reflect the history of
usage, so that older contributions weigh in with a
gradually lower contribution, but without need for the
system to store the full sequence of update episodes.
between duration of consultation and explicit ratings of
value (while—surprisingly—there was no correlation
between duration and size of the document). It must be
noted, though, that the relation between duration and value
will not be strictly proportional or linear: there will be less
difference in value between documents consulted for 90
minutes, respectively 95 minutes, than between documents
consulted for 5, respectively 10 minutes.
A plausible relation might take the form of a sigmoid or
logistic function, which initially increases very slowly to
absorb noise fluctuations due to differences in connection
speed with the document server, then increases almost
linearly, and finally slows down gradually in order to
reach a plateau where further increases in duration
produce virtually no increases in activation. Like in the
case of user actions, the specific shape and parameters of
the function can be derived by determining the best match
with explicit evaluation data.
3. Collecting data from usage
3.1. Document activations
3.3. Co-activation of documents
In the context of libraries or document collections we
can say that a document is activated each time it is being
consulted (opened, downloaded, borrowed or bought) by a
user. In the simplest model, every consultation event
produces a fixed activation unit of, say, 1. In a more
sophisticated model, we can assume that more interesting
documents are used more intensively than others, and
therefore activation values can vary. There are basically
two ways to evaluate activation strength: explicitly or
implicitly.
Explicit evaluation would require the users to indicate
how interesting or relevant the document they are
consulting is. This could be done e.g. with a five point
scale going from “useless” to “just what I needed”. This
can be recomputed to an activation value varying between
0 (or -1) and 1.
The disadvantage of explicit evaluation is that it
demands additional effort from the user, which many users
might not be inclined to perform, especially if they are
browsing through long lists of documents. Explicit
evaluation is likely to be done in practice only for
documents that somehow stand out, because they are
particularly interesting or disappointing.
Now that we know how to get activation values, we
need to determine co-activation. The basic principle is that
documents are co-activated if they are consulted by the
same user, since that user can be assumed to be looking
for mutually relevant documents rather than a random
assortment of unrelated documents. The exponential decay
factor expresses the fact that the more time passes, the
more likely it is that the user has directed his/her attention
elsewhere and has started exploring a different subject.
Still, the fact that people generally have a stable
personality and occupation would imply that two
documents consulted by the same user, even with a ten
year interval, are more likely to be related than two
documents consulted by randomly chosen users.
Therefore, the exponential decay factor might be
complemented by a constant term b so that the co-
activation formula becomes:
A(i, ti).A(j, tj).(a.exp (- d.(tj - ti)) + b)
The situation b = 0 would bring us back to the previous,
purely Hebbian case.
On the other hand, a = 0 would bring us to the case of
collaborative filtering [3,14,10]. This method is used e.g.
by Internet bookshops, such as Amazon.com, which
recommend books on the basis that they have been bought
by the same users, without taking into account the time
interval between the different purchases. Again, the values
of the parameters a, b and d can be determined by
minimizing the difference between the recommendations
derived from the association matrix and the explicit
evaluations by the users.
3.2.Implicit evaluation
Implicit evaluation tries to estimate the degree of
relevance of a document indirectly from the way the user
acts on the document. Different actions such as browsing,
saving, bookmarking, printing, or buying indicate different
degrees of interest [12]. The most straightforward way to
derive activation values from these actions would be to
correlate them with explicit evaluations. E.g., a large
sample of user data might indicate that documents that are
bookmarked get an average evaluation of 3.7 on a 5 point
scale, while documents that are printed get a 4.1
evaluation.
Implicit activation values can be derived even more
simply from the time spent consulting the documents.
Several studies [5,12] have found a strong correlation
4.Applications
Given the co-activation values derived above, we can
compute a matrix of associations between documents by

Page 4

adding together all the collected values for the different
users, documents and moments in time. This matrix can be
used to guide further users in several different ways:
that are associated to other documents that have been co-
consulted.
This implements the general retrieval technique of
recurrent spreading activation [1, 2, 17]: the initial
activation represented by the interest vector is allowed to
spread iteratively through the associative network, so as to
activate all documents that have strong direct or indirect
associations with one or more of the initially selected
documents. Note that the most well-known applications of
spreading activation in information retrieval [13], which
produce rather disappointing results, are not recurrent:
they only allow activation to spread for one or two steps,
in one direction only (i.e. without the possibility of
activation flowing back to previously activated nodes,
which would allow the non-linear accumulation of
activation in the most interesting regions). There exist
many different variations on this spreading activation
algorithm, depending on parameters such as number of
iterations, relative contributions of each iteration phase,
etc. Again, fine-tuning of the result may be obtained by
repeated experiments where recommendations are
compared with explicit evaluations.
The advantage of spreading activation is that the user
may have found only poor examples of relevant
documents, but still receive good recommendations
through indirect association. The only requirement is that
the user be able to distinguish better from worse options.
Thus, with each recommended document that the user
checks, the interest profile and therefore the further
recommendations will be refined, since the system will
now know in how far this additionally consulted document
is really relevant to the task.
4.1.Listing related documents
The most straightforward application is to append to
each document i a list of the documents that are most
strongly associated with it (i.e. that form the largest
components in the document vector di). These are the
documents that not only were frequently consulted by the
same users, but consulted within a relatively short time
interval, and (implicitly or explicitly) evaluated to be most
interesting. In that way, a user who discovers one
document that looks particularly relevant will immediately
get to know all the documents that are most likely to be
relevant as well. To most efficiently guide the user, the
documents can be listed in the order of their degree of
association, the strongest associations first, perhaps with a
graphical indication of that degree.
These links to further documents function like shortcuts
for the otherwise extended exploration sequences that help
users to find other related documents. From these related
documents, users will be offered shortcuts to further
related documents. This may lead the system to create
even shorter shortcuts, from the first document directly to
the third or fourth in the sequence (“transitivity”). Thus,
the use of already learned connections will be assimilated
further into the learning system to create even more direct
connections, creating a positive feedback loop which in
our first experiments was shown to spectacularly enhance
performance [1, 19].
4.3. Determining overall interestingness
4.2. Personalized recommendations
Recurrent spreading activation has more benefits than
fine-tuned, individual recommendations. Associations are
in general asymmetric (aij ≠ a
particular sequence in which a user has moved from one
document to another one. Since users will typically move
from less relevant to more relevant documents, the most
interesting documents will tend to reside at the end of the
association sequence. This means that as activation
spreads further it will encounter documents that are more
and more interesting generally, albeit less directly
associated with the initial preference profile.
If the matrix multiplication is iterated indefinitely, the
output vector will converge to the largest eigenvector of
the matrix. This eigenvector, or “attractor” of the
spreading activation dynamics, represents the equilibrium
distribution of activation. The degree of activation of each
component of that vector can be interpreted as the global
“attractiveness”, “interestingness” or “authority” of that
component, independent of the initial query.
Such “authority” is equivalent to the PageRank
measure that lies behind the surprising effectiveness of the
Google search engine [4], although PageRank starts from a
binary connection matrix (link, no link) rather than a
continuous association matrix, and thus is likely to
ji) since they reflect the
The recommendation of mutually relevant items can be
taken a step further. Users browsing through a library
database generally won’t settle on a single, most
interesting document, but find several documents di that
are relevant in different ways and to varying degrees A(i),
while none of them actually captures the main focus of
interest. This determines an “interest profile” which can be
represented by the activation vector (A(1), A(2), ...) (this
activation vector can also take into account the decrease of
interest of the user with time passing by incorporating an
exponential decay factor, see [9]).
This vector can now be multiplied with the association
matrix to produce a new “recommendation” vector (r1, r2,
...) with:
rj=
aij
i∑
A(i)
This recommendation adds together the contributions
from the previously visited documents in proportion to
their relevancy. This procedure can be repeated,
multiplying the recommendation vector iteratively with
the matrix to get indirectly associated documents, that
perhaps have never been consulted by the same user, but

Page 5

produce less fine-grained results. As demonstrated by
Google, such overall ranking is very useful when ordering
query results before the user has had the time to express
preferences for one document over another.
Generalizing from this observation, we may argue that
the number of iterations is an important parameter that
would allow us to control the generality of the
recommendation: the larger that number, the wider the
public for which the most highly activated documents will
be relevant, but the less direct their relation to the initial
preference profile.
4.5System evaluation and optimization
An essential step in the development of the system that
we envisage is an evaluation of its effectiveness. This can
be easily built into the system itself. If the system allows
for explicit evaluation of recommended documents by
users, then the average score given by users can be
compared with the strength of the recommendation as
calculated by the system. The correlation between the two
scores can be taken as a measure of the system’s
effectiveness. This applies as well to document-centered
recommendations, recommendations based on a user-
profile as to estimates of overall interestingness. (It would
seem that an evaluation of the quality of clusters will have
to be made by domain experts rather than by everyday
users).
To make sure that the recommendations are doing
more than just stating
recommendations can be compared with recommendations
collected from independent sources, such as randomly
selected documents, author-provided references, or simply
lists of the most frequently used documents. In order to
provide an unbiased
recommendations can be randomly interspersed with
recommendations from these other sources, in a way
unknown to the user. The system will have unambiguously
proven its worth if its recommendations get a
systematically higher score than these other possible
recommendations.
Such evaluation can be used to continuously optimize
and fine-tune the system. It suffices to consider the
correlation between system-calculated strength and
average user evaluation as a function to be maximized,
and then vary the different parameters used in the
algorithms (e.g. strength of exponential decay, number of
iterations in spreading activation, ...) so as to achieve the
largest possible value for the correlation. In that way, the
system will not only learn better relevancy judgments
from its users, but moreover learn how to improve its own
learning functions, i.e. it will undergo metalearning
towards ever greater effectiveness.
Moreover, metalearning will allow the system to adapt
to specific contexts: different types of document
collections (e.g. songs vs. lecture notes) will be used in
different ways, and thus require different parameters for
the learning and recommendation algorithms (e.g. the
duration of consultation is likely to be lower for pictures
than for technical documents or movies, and information
search in well-structured databases is likely to be more
focused than in more “associative” collections of artistic
photos, and thus require less iterations during spreading
activation).
4.4.Clustering documents
By multiplying the (asymmetric) matrix with its
transpose we can create a new, symmetric matrix:

sij=
aik
k∑
akj
.
the obvious, system
test, system-generated
sij represents the degree of similarity between the
components i and j. Indeed, sij is the dot product between
the vectors di and dj that represent all the associations that
the documents i and j have with other documents (see 2.1).
The more the association vectors overlap, and thus the
more i and j resemble each other in the way they relate to
other documents, the larger the dot product, and therefore
sij. This similarity measure can now be used as an input to
a variety of clustering algorithms that put documents
together in classes depending on how similar/dissimilar
they are from each other.
One example of such a clustering can be found in the
HITS algorithm developed by Kleinberg [11], that clusters
web pages starting from the product of a connection
matrix with its transpose, by finding the different
orthogonal eigenvectors of that matrix and by considering
components that load strongly on a particular eigenvector
as members of the same cluster. This allowed Kleinberg to
e.g. distinguish “pro life” from “pro choice” pages on
abortion, or pages on the animal “jaguar” from pages on
the car and the sports team with the same name, thus
tackling the problem of homonyms.
More generally, a clustering algorithm should allow us
to automatically create categories of documents, even
when these categories haven’t been formally recognized
yet, thus catching emerging new domains from the very
beginning. The categories can be labelled by extracting the
keywords that appear most frequently in that category
relative to the overall collection. The PageRank or HITS
algorithms can moreover be used to list the documents
most authoritative for each category, which are likely to be
classic papers, general reviews or introductory texts about
the subject.
5. Implementation
The data necessary for the Hebbian algorithm that we
outlined are easy to collect. The document server will

Page 6

normally maintain a log of all consultations, including the
identity of the user, the documents requested, and the
precise date and time at which each document was
requested [2]. This is sufficient to calculate the activation
of each document on the basis of the time spent between
requesting a document and requesting the next document,
which indicates the time spent reading the document and
thus, as we have seen, provides an implicit measure of
interest. The exponential decay factor can be calculated
from the time interval between requests for the two
documents (which may be several steps away from each
other in the request sequence) between which co-
activation needs to be calculated, independently of any
requests in between.
It must be noted that server logs tend to contain a lot of
noise, such as consultations made by webrobots rather
than true users, users whose IP address changes during the
session, different users with the same IP address, sessions
interrupted e.g. because the user went to drink a cup of
coffee, consultations made through backtracking or
bookmarks rather than following sequences of links, etc.
Various techniques have been developed to preprocess
such log data so as to extract only the meaningful
navigation paths (see e.g. [2, 15, 16, 17]). Obviously none
of these will ever be perfect. Yet, we don’t expect the
remaining errors to have a great impact on the results,
because our general algorithms appear quite robust, based
on principles of self-organization that are able to extract
strong patterns from a noisy background [10].
Moreover, the effect of any amount of noise can be
attenuated through the law of large numbers: if a
sufficiently large number of contributions is collected,
summation will drown out any random deviations from the
underlying signal [10]. Because any log file, which for a
typical active webserver contains millions of lines for a
few weeks worth of use, can be used as input, it seems that
in most cases there will be sufficient data to kickstart the
system and quickly produce
recommendations.
If we wish to use other data than duration to estimate
user satisfaction, we will need to establish a protocol that
signals specific user activities, such as printing, saving or
bookmarking, to the server collecting the data. This is
most straightforward for explicit evaluations, where a user
can click on an evaluation bar, and thus pass on the
coordinates of the click to the server. Another approach is
to have a Java applet loaded into the user’s browser when
the server is first contacted, which registers the activity
within the browser and sends this information back to the
server [15].
When there is a single, centralized server for all
documents, this basically completes the information
gathering, since the association matrix can now be directly
extracted from the log of that server, while the related
documents, “authority” measures and clustering can be
computed off-line using the matrix, after which the results
are fed back into the document system, e.g. in the form of
a recommendation list at the bottom of each document
summary, together with a taxonomy of subjects and a list
of the most important pages on the entry page.
Individual recommendations based on spreading
activation are somewhat more involved as they require the
maintenance of a constantly updated interest vector for
each user, which must be multiplied with the matrix to
provide tailored recommendations in real time. One
method is to keep a “cookie” on the user’s browsing
application that keeps track of the user’s sequence of
activities. When desired, this cookie can then be
transferred to a central server to be used as input for a
spreading activation algorithm.
With a truly distributed library system, running on a
variety of independent servers, the main additional step is
the establishment of a protocol for the exchange of data
about user activities between the different machines. Each
time a user moves to a new server, previously consulted
servers should receive a trace of the user’s activities on
that server. Thus, they can update—or create—
associations from the documents kept locally to the
documents kept on other servers.
With many documents spread over many different
servers, the danger is that the information to be kept on
any one server would explode. This can be controlled by
limiting the trace’s extent in time or in number of
consultations, so that e.g. information is no longer sent to
servers whose documents were consulted more than x days
or requests ago. Moreover, each server could locally
decide to maintain not more than y associations for each
document. This can be done by periodically removing
from memory the weakest associations, or—if the system
moreover keeps track of the time each bonus was added—
the associations that received their last bonus the longest
time ago.
The disadvantage of such a distributed implementation
is that there isn’t any single place where the complete
association matrix is stored. In principle, the association
matrix can be reconstructed from the association data that
are kept locally on each server, but this will require
complicated distributed protocols if global computations
must be performed—such as calculating PageRanks or
global clusterings. Local recommendations, such as
proposing documents related to a given document or
spreading activation with few iterations, should be
produced easily.
An alternative for a fully distributed implementation
would be a central server that maintains and manipulates
the overall association matrix. However, since the
complete matrix, while being sparse, will contain a huge
amount of data, an in-depth application will require
extensive computing power together with sophisticated
algorithms for sparse matrix manipulation. Still, similar
matrices, albeit probably
successfully being used by databases such as Google or
the Alexa recommendation service on the web,
demonstrating the feasibility of the project.
a usable list of
less fine-grained, are

Page 7

6.Conclusion
[2] Bollen, Johan, Herbert Van de Sompel, and Luis M. Rocha
(1999) Mining associative relations from website logs and their
application to context-dependent retrieval using spreading
activation, in Workshop on Organizing Web Space (WOWS),
ACM Digital Libraries 99, August 1999, Berkeley, California.
The present paper has sketched a general family of
algorithms to extract meta-data about documents from the
way these documents are
Implementing such a system in a digital library would
automatize much of the hard work that would otherwise
need to be performed by highly trained information
scientists.
However, the results of this system are envisaged to
complement or support traditional methods rather than
fully replace them. The reason is that the proposed system
focuses on otherwise difficult to formalize properties of
documents, namely the subjective associations that exist in
the mind of the users between their different subjects and
contents. The advantage is that these associations allow us
to build a system that emulates human intuition, so that it
can anticipate the desires of its users and provide them
with the information they would find most interesting,
even when these users cannot explicitly formulate what
they are looking for. This is particularly useful for
multimedia documents, which do not contain any
searchable keywords, and for queries that are as yet ill-
defined.
The disadvantage of associative networks is that they
are intrinsically fuzzy, ambiguous, and constantly shifting
[9]. However, the clustering approach that we sketched
might help us to extract discrete categories, which can be
automatically labelled with keywords, although here it is
likely that the system would still need the assistance of a
human operator in order to build a coherent taxonomy.
Another advantage is that the system is designed from
the start to learn, so that its recommendations become
better the more it is used. This applies at the collective
level, where the association matrix becomes more precise
as more usage data are collected, but also at the individual
level, where every (explicit or implicit) evaluation of a
document made by a user helps the system to produce a
more individually tailored recommendation, and even at
the metalevel, where the system adapts its own learning
functions to the circumstances.
The biggest unresolved issue until now is the
implementation of such a system at the level of a
distributed library system. (Smaller scale implementations
have already been made [1, 2, 8, 10, 17], or are under
development.) While a single server centralizing and
processing all incoming and outgoing data seems
straightforward, albeit computationally intensive, the more
interesting challenge will be to distribute both the database
and the processing over a peer-to-peer document server
network.
consulted by users.
[3] Breese J.S., Heckerman D. and Kadie C. (1998), Empirical
Analysis of Predictive Algorithms for Collaborative Filtering,
Proceedings 14th Conference on Uncertainty in Artificial
Intelligence, Madison WI: Morgan Kauffman.
[4] Brin S. and L. Page (1998): The Anatomy of a Large-Scale
Hypertextual Web Search Engine, Proceedings of the 7th
International World Wide Web Conference, April 1998.
[5] Claypool, Mark, Phong Le, Makoto Waseda and David
Brown (2001): Implicit Interest Indicators, In Proc. ACM
Intelligent User Interfaces Conference (IUI), Santa Fe, New
Mexico, USA
[6] Hebb, D. O. 1967 The organisation of behavior: a
neuropsychological theory. Science Editions, New York.
[7] Heylighen F. (1991): Design of a Hypermedia Interface
Translating between Associative and Formal Representations,
International Journal of Man-Machine Studies 35, p. 491-515.
[8] Heylighen F. (2001): Bootstrapping knowledge represen-
tations: from entailment meshes via semantic nets to learning
webs, Kybernetes 30 (5/6), p. 691-722.
[9] Heylighen F. (2001): Mining Associative Meanings from
the Web: from word disambiguation to the global brain, in:
Proceedings of the International Colloquium: Trends in Special
Language and Language Technology, R. Temmerman and M.
Lutjeharms (eds.) (Standaard Editions, Antwerpen), p. 15-44.
[10] Heylighen, Francis (1999) Collective Intelligence and its
Implementation on the Web: algorithms to develop a collective
mental map, Computational and Mathematical Theory of
Organizations 5(3), 253-280.
[11] Kleinberg J. (1998): Authoritative sources in a hyperlinked
environment, Proc. 9th ACM-SIAM Symposium on Discrete
Algorithms.
[12] Nichols, D.M. (1998) Implicit Rating and Filtering, Proc.
Fifth DELOS Workshop on Filtering and Collaborative
Filtering, Budapest, Hungary, 10-12 November 1997, ERCIM,
31-36.
[13] Salton G. and Buckley C. (1988). On the Use of Spreading
Activation Methods in Automatic Information Retrieval, Proc.
11th Ann. Int. ACM SIGIR Conf. on RandD in Information
Retrieval (ACM), 147-160.
7.References
[1] Bollen, Johan and Heylighen, Francis (1998) A system to
restructure hypertext networks into valid user models, New
Review of HyperMedia and Multimedia, 189-213.

Page 8

[14] Shardanand U. and Maes (1995), Social information
filtering: Algorithms for automating ‘word of mouth’,
Proceedings of CHI’95 -- Human Factors in Computing
Systems, 210-217.
[15] Cooley, R. Web Usage Mining: Discovery and Application
of Interesting Patterns from Web Data. Ph.D. Thesis,
University of Minnesota, May 2000.
[16] Shahabi, C., Zarkesh, A.M., Adibi, J., and Shah, V.
Knowledge Discovery from User's Web-page Navigation, in
Proc. 7th IEEE Intl. Conf. On Research Issues in Data
Engineering (1997), 20-29.
[17] Bollen, Johan: A Cognitive Model of Adaptive Web Design
and Navigation: A Shared Knowledge Perspective. PhD Thesis,
Vrije Universiteit Brussel, 2001.
[18] Heylighen F. & Bollen J. (1996) “The World-Wide Web
as a Super-Brain: from metaphor to model”, in: Cybernetics
and Systems '96 R. Trappl (ed.), (Austrian Society for
Cybernetics).p. 917-922.
[19] Bollen J. & Heylighen F. (1996) “Algorithms for the Self-
organisation of Distributed, Multi-user Networks. Possible
application for the future World Wide Web”, in: Cybernetics
and Systems '96 R. Trappl (ed.), (Austrian Society for
Cybernetics), p. 911-916.