=Paper=
{{Paper
|id=Vol-532/paper-8
|storemode=property
|title=Spreading Activation Approach to Tag-aware Recommenders: Modeling Similarity on Multidimensional Networks
|pdfUrl=https://ceur-ws.org/Vol-532/paper8.pdf
|volume=Vol-532
}}
==Spreading Activation Approach to Tag-aware Recommenders: Modeling Similarity on Multidimensional Networks==
https://ceur-ws.org/Vol-532/paper8.pdf
Spreading Activation Approach to
Tag-aware Recommenders:
Modeling Similarity on Multidimensional Networks
Alexander Troussov Denis Parra Peter Brusilovsky
IBM, Ireland University of Pittsburgh and University of Pittsburgh
IBM Software Lab, Bld. 6, CNGL at Trinity College Dublin 135 North Bellefield Ave.,
Mulhuddart, Dublin 15, Ireland 135 North Bellefield Ave., Pittsburgh, Pittsburgh, PA 15260, USA
+353-1-815 1906 PA 15260, USA +1 (412) 624 9404
atrousso@ie.ibm.com +1 (412) 624 9403 peterb@mail.sis.pitt.edu
dap89@pitt.edu
ABSTRACT were created for a bi-modal world of items and users (connected
by rating incidents), social tagging systems present a more
Social tagging systems present a new challenge to the researchers complicated world of users, items, and tags (connected by tagging
working on recommender systems. The presence of tags, which incidents, also known as tagging instances). While some early
uncover the reasons of user interests to tagged items, opens a way works attempted to treat the problem of recommendation in social
to increase the quality of recommendations. Yet, there is no tagging systems in an “old way”, basically ignoring the tags, the
common agreement of how the power of tags can be harnessed for majority of researchers in this new area argued that tags are vital
recommendation. In this paper we argue for the use of spreading for successful recommendation in this new domain and called for
activation approach for building tag-aware recommender systems tag-aware recommenders. They argued that on one hand, tags can
and suggest a specific version of this approach adapted to the compensate the loss of ratings (which are not available in most
multidimensional nature of social tagging networks. We introduce social tagging systems), while on the other hand, tags can make
the asymmetric measure of relevancy (proximity) of two nodes on recommendation more precise because they provide not only the
a multidimensional network as a cumulative strength of information of what items are of interest to a user, but also why
(weighted) multiple connections between two nodes, which they are of interest [8,14,20,26].
includes paths and graph-structures connecting the nodes. This
metric is also applicable to measure relevancy of two sub-graphs. Despite the common agreement that tags should be used as a
Spreading activation methods (SAM), which usually employ successful recommender component of a social tagging system,
breadth first search, are an efficient way to define and compute there is no agreement on how it should be done. As a result, a
such measure taking into account not only links constituent a path, multitude of approaches emerged just over the last three years.
but the properties of nodes in the path such as node’s types and Roughly, these approaches can be classified as an extension of
outdegree. either content-based or collaborative filtering approaches. The
former group emphasizes connections between items and tags
We apply this notion of relevancy to measure similarity of treating tags as an alternative (or additional) way to describe items
collaborative tagging systems users and present the results of and establish a profile of user interests [9, 15]. The latter group
numerical simulation showing that spreading activation methods emphasizes connections between users and tags to establish a
allow us to discriminate between diverse graph-structures better similarity between users in a social tagging system [25, 26].
connecting users via resources and tags. We show that the results
of simulation are stable w.r.t. the variation of parameters of We argue than inherently networked nature of social tagging
spreading activation algorithm used in our experiment. systems calls for some alternative recommender approaches,
which are not just simple extension of either content-based or
Categories and Subject Descriptors collaborative technologies. A successful recommender approach
H.3.4 [Information Storage and Retrieval]: Systems and for this new context should fully employ the complex network
Software – information networks; H.3.5 [Information Storage structure of a typical social tagging system and use all kinds of
and Retrieval]: Online Information Services – data sharing. links: user-tag, item-tag, user-item. We think that the most
promising in this context is the spreading activation approach.
General Terms This approach has been originally developed in the field of
Algorithms, Measurement, Performance, Experimentation. cognitive psychology [3] to model human brain and later explored
in the context of information retrieval [7].
Keywords The power of spreading activation approach was recognized in the
Tagging, relevancy propagation, spreading activation, graph-
area of recommenders and other personalized systems as well;
based mining, structural cohesion, CiteULike.
however, so far these approaches form just a small minority. The
1. INTRODUCTION problem is that the traditional user-item universe does not provide
a sufficiently rich network for spreading activation technology.
Social tagging systems introduced new challenges to the well- Thus most of known recommenders based on spreading activation
established area of recommender systems. While the majority of were built for context where an additional network can be formed
content based, collaborative, and hybrid recommender approaches such as a hypertext network for Web page recommendation [17]
or a network of entities and concepts in semantically enriched result of direct transfer of information retrieval ideas from
recommenders [4, 12, 18]. cognitive sciences to AI. In other domain, [27] created spreading
activation models for trust propagation on the Web.
We believe that social tagging systems will provide a new
promising context as well as new challenges for recommenders In [21] and [23] authors work with the notion of the relevancy of
based on spreading activation. What we consider as the main ontological concepts to a free text. They propagate relevancy of
challenge is the multidimensional nature of a typical social the concepts explicitly mentioned in a document to other
tagging network. Almost all existing applications of spreading ontological concepts using a spreading activation algorithm. Their
activation for personalization and recommendation operated in algorithm works in such a way, that after short number of iteration
relatively homogeneous kinds of networks with 1-2 kinds of the topical foci of a cohesive coherent text become the most
nodes and one kind of links. In contrast, even a simplified social activated concepts (even if they were not explicitly mentioned in
tagging network, where each tagging event is represented by a the text).
group of three links (user-tag, item-tag, and user-item) includes
three types of nodes and three types of asymmetric links. This In [22] authors summarize their experience in creating graph-
organization requires some more sophisticated spreading based related item recommender for activity centric environment
activation approaches. on a Nepomuk Social Semantic Desktop [24]: relevancy of a
“pile” of nodes representing resources and concepts is propagated
Our paper attempts to address this challenge by introducing the to other nodes. Authors in [22] conclude that as a graph-mining
asymmetric measure of relevancy (proximity) of two nodes on a technique, spreading activation combines fuzzy clustering and soft
multidimensional network as a cumulative strength of (weighted) inferencing, and therefore might be suitable for relevancy
multiple connections between two nodes which includes paths and propagation. Propagation should lead to discovery of new nodes
graph-structures connecting the nodes. This metric is also which have short length paths to many (if not all) nodes from the
applicable to measure relevancy of two sub-graphs. Spreading initial set. In other words, newly discovered nodes should
activation methods, as breadth first search, is an efficient way to minimize the “distance” to the initial set of nodes, i.e., nodes
define and compute such measure taking into account not only which might be considered as potential centroids of strong
links constituent a path, but the properties of nodes in the path clusters induced by the initial conditions. Since partitioning of the
such as node’s types and outdegree. nodes according to these clusters is not needed, processing of
polycentric queries [22] for related item recommendation could be
We apply this notion of relevancy to build a tag-aware approach done using soft clustering methods. On the other hand, relevancy
to measure similarity between users in collaborative tagging propagates through links. an alternative view on the related item
systems. The paper presents the results of a numerical simulation recommendation is that newly discovered nodes must be
showing that spreading activation algorithms allow discriminating connected to the initial conditions by particular types of directed
the degree of connectivity of users between certain graph- links. Therefore, propagation of relevancy might be interpreted as
structures connecting users via resources and tags. We fuzzy inference.
demonstrate that the results of the simulation are stable w.r.t. the
variation of parameters of the spreading activation algorithm used In [23], the authors go further in analyzing SAM as a very general
in our experiment. class of iterative algorithms for relevancy propagation, local
search, relationship/association search, and computing of dynamic
The rest of the paper is organized as follows. In section 2 we first local ranking. Authors indicate that the same iterative algorithms
provide a short overview of related work focusing on the use of were used long before in numerical simulation in physics,
spreading activation methods (SAM) to propagating and mechanics, chemistry, and engineering sciences. Hence, the
redistributing relevancy. We also theorize about desired properties algorithm is quite polymorphic: “Using the same iterative
of relevancy propagation on multidimensional network models of algorithm, with one set of parameters one can emulate heat
Web. 2.0 data needed to create efficient and scalable transfer; with another set of parameters the same algorithm will
recommender systems. show us the behavior of oscillating strings”.
In section 3 we render a formal model of folksonomies (tripartite
hypergraph) as a multidimensional network with four types of 2.2 Spreading Activation in Recommender
nodes corresponding to users, resources, tags and instances of Systems
tagging. In section 4 we present the results of numerical Spreading activation approach as a technology for
simulation. Finally, section 5 describes the conclusions and future recommendation in various kinds of networks belongs to a
work broader group, which is typically referred to as graph-based
approaches for recommendation. In addition to several recent
2. RELATED WORK papers mentioned in the introduction, which explicitly use
2.1 Overview of Relevancy Propagation Using spreading activation to build recommender systems, we can a few
Spreading Activation Methods other examples of using various graph-based approaches. In [1],
In neurophysiology interactions between neurons are modeled by the authors presented a theoretic approach where users are
way of activation which propagates from one neuron to another modeled as nodes in a directed graph and the directed links
via connections called synapses to transmit information using represent how representative is a user of another user's behavior.
chemical signals. The first spreading activation models were used In [11], the authors use spreading activation to deal with the
in cognitive psychology to model these processes of memory sparsity problem in collaborative filtering. They try to tackle the
retrieval [5, 3]. This framework was later exploited in Artificial problem finding transitive relationships by comparing three
Intelligence as a processing framework for semantic networks and different methods on a bipartite graph which represented
ontologies, and applied to Information Retrieval [2, 7, 19] as the consumer-product interactions. Other interesting approach was the
one presented in [10], where the authors propose a constrained but is probably much less important and is too coarse-grained
spreading activation algorithm having good results compared with measurement compared to trust propagation.
a traditional memory-based approach over a small subset of the
Movie Lens data set. These approaches show the potential of A final observation on relevancy propagation on multidimensional
spreading activation to be used on recommender systems, but they networks: we don’t assume that all (or many) aspects of such
don't take into account the nature of multidimensional networks, propagation can be properly understood in terms of paths. We
such as folksonomies derived from collaborative tagging systems, assume that there might be structures (like network B on the Fig.
where different types of nodes, links and relationships can have a 1), which might significantly affect the relevancy propagation.
strong influence in the design of the algorithms.
3. THE ALGORITHM
2.3 Propagating Relevancy on The algorithm we used in our experiment in general follows [23]
and employs iterative steps where activation is propagated
Multidimensional Web 2.0 Networks between neighbor nodes. To facilitate comparison of activation
We focus on the applications of SAM to measure similarity distributions on the same or different networks and to account for
between the users of collaborative tagging systems modeled as dissipation of activation caused by list purging step in spreading
multidimensional networks. Indeed, we treat graph-based activation, we introduce the step of normalization (calibration).
“similarity” of users as a particular case of “relevancy” of nodes
on multidimensional networks. In this subsection we provide A multidimensional network can be modeled as a directed graph,
consideration on which properties of a generic class of spreading which is a pair G = (V,E) where
activation algorithms are suitable methods for modeling relevancy V – is the set of vertices vi
propagation.
E – is the set of arcs ej
The general inspiration behind using graph-based methods to init: E → V, is the mapping that provides initial nodes for arcs
model relevancy (energy, trust, risk, etc.) propagation on networks
is probably the same in many domains: the relevancy is treated as term: E → V, is the mapping that provides terminal nodes for arcs
a kind of energy which might be “injected” into some nodes, and imp – is importance value of arcs and nodes.
propagated through links to other nodes: “… the closer node x to w – “weights”
the injection source s, and the more paths leading from s to x, the
higher the amount of energy flowing into x in general” [27].
Therefore, spreading activation methods (SAM), which usually F(E) – is the “activation” real valued function
employ breadth-first search), are an efficient way to propagate The algorithm has the following steps
relevancy. Since according [23] SAM is a broad class of
algorithms, the choice of algorithm’s parameters is crucial and can Initialization
be done taking into account the nature of the target application.
Sets the parameters of the algorithm, network, and initial
First of all, Web 2.0 data could be accurately modeled only by F(E) as a list of non-zero valued nodes V n
multidimensional networks. For instance, formal model of a
folksonomy as tripartite hypergraph [13] converted to network Iterations
representation, has four types of nodes: users, resources, tags, a. List Expansion.
instances of tagging. The shortest possible path between two b. Recomputation: The value at each node in the list is
folksonomy users has the length four (for instance, user1- instance
of tagging1- tag - instance of tagging2 - user2). As compared to recomputed based on the values of the function on
trust propagation in heterogeneous networks, the amount of nodes which have links to the given node and types
relevancy flowing from one node to another should depend not
of connections.
only on types of links, but on properties on nodes in paths.
Connections via resources might be more important than c. List Purging: We exclude the nodes with the values
connections through tags. In our future work we are going to less than a threshold.
exploit what [23] calls “the importance of nodes”, but one
d. Conditions Check To Break Iterations.
property of nodes which should significantly affect the
propagation, can be immediately inferred from the local topology Normalization
of the network, namely from the number of outcoming links from Linear scaling up or down the numerical values of the
a node. Ambiguous and top popular tags might be linked to big
number of tag instances and big number of users. Intuitively, activation level of all nodes in the list of activated nodes to
connections via such tags should provide less (if any) contribution satisfy some conditions of activation conservation
to the similarity of users as compared to the connections through Output
less popular tags.
The list of nodes (value of the function after spread of
In [27], the authors assume that nodes with the higher shortest activation) ranked according F values.
path distance from the injection source should be accorded less
trust in general. This property of trust propagation is probably not Recomputation step is as follows:
applicable to propagating relevancy to measure similarity of
folksonomies users. Moreover, we suggest that for many We have the list of nodes Vn.
applications on multidimensional networks the length of the
shortest path might have positive correlation with the relevancy, Input/Output Through Links Computation.
– For each node v we compute the input signal to each arc
e, such that init(e)=v. This computation can be based on
the value F(v), the outdegree of a node etc. For instance,
if the node v has n outgoing arcs of the same type, each
arc e might get input signal:
I (e) = F(init(e)) ∙ (1 / outdegree(v) ^beta )
where beta might be equal to 1. It could be also less
than one, in which case the node v will propagate more
activation to its neighbors than it has. (This might be
fine for some applications).
– When the signal (“activation”) passes through a link e,
the activation usually experiences decay by a factor
w(e):
O (e) = I(e) ∙ w(e)
Input/Output Of Node Activation
– Before the pulse, the node v has the activation level
F(v).
– Through incoming links v get more activation:
Input(v) = Σ O(e)
for all links e such that init(e) ∈Vn, term(e) = v.
– By dissipating the activation through outgoing
links, the node v might lose activation:
Output(v) = Σ I(e)
for all links e such that init(e) = v, term(e) ∈Vn
Computation Of New Level Of Activation
Fnew(v) = F(v) + Input (v)
To apply spreading activation to measure “similarity” of two Figure 1. Three networks modeling instantiations of
nodes on a network, we put the initial activation 1.0 at the first collaborative tagging systems.
node, and measure the activation at the second node after certain
number of iterations. In [23], the authors view SAM in terms of graph-mining
algorithms as a technique for soft clustering. The major
4. EXPERIMENTS parameters of SAM affecting “the scale” of the phenomena to be
To apply graph-based mining on web 2.0 data we model the data discovered are signal decay and number of iterations (larger
by a multidimensional network (where nodes and links are typed, number of iterations and low decay are needed to discover
and links are “weighted”). “bigger” clusters). Since Web 2.0 applications are at the focus of
In our experiments we use three networks representing this paper, we run the experiments varying these two parameters.
instantiations of collaborative tagging systems. Each of these Our target was to find regions of the parameters which allow us
networks has two actors (A1 and A2), two resources (R1 and R2), consistently to capture structures like that on the Fig.1.
and four instances of tagging (I1, I2, I3 and I4). For instance, the In this paper, we use SAM as a link analysis algorithm for local
network A on the figure 1 has the instance of tagging I1 with ranking, in the same way as PageRank algorithm is used for
links to the actor A1, the resource R1, and the tag T; this sub- global ranking [28]. The major difference between them is that
network shows that the actor A1 used the tag T for the resource PageRank iteratively redistributes the relevancy measure which is
R1. Correspondingly, the links from the instance l2 show that the initially set to each node of the network, while we use SAM to
actor A1 used the tag T for the resource R2. The instances I3 and iteratively redistribute the relevancy measure (the activation) from
I4 show tagging for the user A2. The network A represents the one (or more) nodes sometimes referred to as “seeds”.
situation where both actors used the same tag for both resources.
Diameter of graphs B, and C is 6, with the number of iterations
In the implementation of our algorithm, each of these networks is less than 6 the activation from a node on a network will not
modeled by a directed graph, where for each link we create two necessarily reach all the nodes. The limit distribution (distribution
reciprocal arcs. In each experiment we set initial activation at the of the activation after a number of iterations big enough),
node corresponding to the actor A1 and after several iterations of produced by SAM, in general does not depend on the choice of
the algorithm we compute the “similarity” of actors A1 and A2 the initial seed. This behavior gives us the estimate that local
using the method described in 3. ranking, which is highly sensitive to sub-graphs with the diameter
6, could be achieved when the activation will be redistributed on 5. CONCLUSIONS AND FUTURE WORK
such sub-graphs several times which amounts roughly to 12-48 Our paper argued for the use of spreading activation as a
iterations. recommendation mechanism in multidimensional networks
Our underlying common-sense assumption is that connectivity of produced by collaborative tagging systems. We introduced the
A1 and A2 is bigger in the network A than in B and C; and that new network-based asymmetric measure of relevancy of two
the connectivity of A1 and A2 in the network B is bigger than in nodes on a multidimensional network and applied it to build a tag-
the network C. In other words, if we denote the final activation of aware approach to measure similarity between users in
the node v in the network configuration X as x(v), we would collaborative tagging systems. While it is just one of several
expect that sensible local ranking results should satisfy inequality: possible ways to use spreading activation in collaborative tagging
context, we consider it as the best way to start. As demonstrated
(1) by the stream of recent works, calculating similarity between
The shortest path between the nodes A1 and A2 equals to 4 in the users is a component of the recommendation process where the
network A, and to 6 in networks B and C. So the first part of the use of tags can provide a most valuable impact [25, 26].
inequality is easily achieved with any parameters of the algorithm The results of our experiments show that our metrics can be used
(provided that the number of iterations is not less than 3). To to differentiate activation levels on different network
investigate how the algorithm can discriminate between configurations and they also show a stable behavior when input
configurations B and C we introduce the network discrimination parameters are changed. These results lead us to pass to the next
factor as step on our research on this bottom-up approach, which is to
prove that our results are repeatable in large scale networks. We
(2)
are currently running our experiments on real social network data
that we have collected from the social bookmarking service
We computed the NDF ranging the number of iterations from 1 to
CiteUlike.
50, and the decay factor from 0 to 1. Figure 2 shows the results,
where the X axis represents number of iterations, the Y axis the In this paper we presented applications of spreading activation
decay factor, and the Z axis the network discrimination factor. methods to local ranking on small networks. We didn’t prove yet
that the same “good” properties hold true when the algorithm runs
on massive networks. However, multidimensional networks which
model web 2.0 data and processes usually exhibit small world
phenomena properties, which include small average distance and
clustering effect. According to [23] spreading activation might be
considered as a method for soft clustering. Intuitive justification
of the use of spreading activation for ranking is the same as for
the PageRank algorithm [28]: a node can have a high rank if there
are many nodes that point to it, or if there are some nodes that
point to it and have a high rank. On each iteration strongly
activated nodes continue to support the high level of activation of
nodes to which they have outcoming links, while nodes which
have little connection with strongly activated nodes eventually
lose their activation. Therefore, even if constrained spread of
activation from one node might in several iterations reach
significant portion of the network (small average distance), strong
level of activation will be supported mainly in strong clusters
induced by the node.
Figure 2. Results of the NDF experiment. Axis X shows
iterations, axis Y decay values, and axis Z the NDF. 6. ACKNOWLEDGMENTS
This material is based upon work supported by the National
The results in figure 2 show that we maximize the NDF when Science Foundation under Grant No. 0840597. We also want to
running our spreading activation algorithm with a decay factor thank Dr. Vincent Wade from the CNGL of the Trinity College
between 0.8 and 0.9, and 24 iterations. Additionally, the plot Dublin for his support to work on this collaborative research.
shows stable results for our algorithm, which suggests that
selecting values in close ranges will not return unexpected or Dr. Alexander Troussov's work was done in collaboration with
random activation values. CNGL, which is funded under Science Foundation Ireland's CSET
programme: Grant# 07/CE2/I1142.
We have shown that on small networks SAM might be used to
measure similarity between users. It is part of our future plans to 7. REFERENCES
show that on big multidimensional networks representing Web 2.0
[1] Aggarwal, C. C., Wolf, J. L., Wu, K., and Yu, P. S. 1999.
data activation initiated at one of the nodes could be kept flowing
Horting hatches an egg: a new graph-theoretic approach to
within strong clusters induced by the initial set of activated nodes
collaborative filtering. In Proceedings of the Fifth ACM
(because of high degree of clustering); and therefore the results
SIGKDD international Conference on Knowledge Discovery
could be generalized to real-world data.
and Data Mining (San Diego, California, United States,
August 15 - 18, 1999). KDD '99. ACM, New York, NY, 201-
212. DOI= http://doi.acm.org/10.1145/312129.312230
[2] Aleman-Meza, B., Halaschek, C., Arpinar, I., & Sheth, A. Proceedings of the 13th international conference on WWW,
(2003). Context-Aware Semantic Association Ranking. May 17-20, 2004, New York, NY, USA, 374-383.
Proceedings of SWDB'03, Berlin, Germany, 33-50. [17] Olston, C. and E. H. Chi 2003. "ScentTrails: Integrating
[3] Anderson, J., 1983. A Spreading Activation Theory of browsing and searching on the Web." ACM Transactions on
Memory. Journal of Verbal learning and Verbal Behavior Computer-Human Interaction 10(3): 177-197.
1983, (22), 261-295. [18] Sarini, M. and C. Strapparava 1998. Building a User Model
[4] Bier, E. A., S. K. Card, et al. 2008. Entity-Based for a Museum Exploration and Information-Providing
Collaboration Tools for Intelligence Analysis. IEEE Adaptive System. Second Adaptive Hypertext and
Symposium on Visual Analytics Science and Technology, Hypermedia Workshop at the Ninth ACM International
VAST 2008, Columbus, Ohio, IEEE. Hypertext Conference Hypertext'98, Pittsburgh, PA.
[5] Collins, A.M. & Loftus, E.F. 1975. A spreading-activation [19] Schumacher, K., Sintek, M., Leo Sauermann 2008
theory of semantic processing. Psychological Review. 1975 Combining Fact and Document Retrieval with Spreading
Nov Vol 82(6), 407-428. Activation for Semantic Desktop Search. The Semantic Web:
[6] Contractor, N. 2007. From Disasters to WoW: Using a Multi- Research and Applications, 5th European Semantic Web
theoretical, Multilevel Network Framework to Understand Conference, ESWC 2008, Tenerife, Spain, June 1-5, 2008
and Enable Communities. Retrieved March 8, 2009, from LNCS, Springer Verlag, Volume 5021/2008, 569-583
http://www.friemel.com/asna/keynotes.php [20] Shepitsen, A., J. Gemmell, et al. 2008. Personalized
[7] Crestani, F. 1997. Application of Spreading Activation recommendation in social tagging systems using hierarchical
Techniques in Information Retrieval. Artificial Intelligence clustering. the 2008 ACM conference on Recommender
Review, 11(6), 453-482. systems, RecSys '08 Lausanne, Switzerland, ACM.
[8] Dattolo, A., F. Ferrara, et al. 2009. Supporting Personalized [21] Troussov, A., Judge, J., & Sogrin, M. 2007, December 13).
User Concept Spaces and Recommendations for a IBM LanguageWare Miner for Multidimensional Socio-
Publication Sharing System. 17th International Conference Semantic Networks. Retrieved March 8, 2009, from
on User Modeling, Adaptation, and Personalization (UMAP http://www.alphaworks.ibm.com/tech/galaxy
2009), Trento, Italy, Springer. [22] Troussov, A., Judge, J., Sogrin, M., Bogdan, C., Edlund, H.,
[9] de Gemmis, M., P. Lops, et al. 2008. Integrating tags in a & Sundblad, Y. 2008b. Navigating Networked Data using
semantic content-based recommender. the 2008 ACM Polycentric Fuzzy Queries and the Pile UI Metaphor
conference on Recommender systems, RecSys '08 Lausanne, Navigation. Proceedings of the International SoNet
Switzerland, ACM. Workshop, 5-12.
[10] Griffith, J., O'riordan, C., and Sorensen, H. 2006. A [23] Troussov, A., Levner, E., Bogdan, C., Judge, J., Botvich, D.
constrained spreading activation approach to collaborative "Spread of Activation Methods", in Dynamic and Advanced
filtering. pp. 766-773. Data Mining for Progressing Technological Development, Y.
Xiang and S. Ali (eds) IGI (to appear 2009).
[11] Huang, Z., Chen, H., and Zeng, D. 2004. Applying
associative retrieval techniques to alleviate the sparsity [24] Sauermann, L., Kiesel, M., Schumacher, K., & Bernardi, A.
problem in collaborative filtering. ACM Trans. Inf. Syst. 22, 2009. Semantic Desktop. Social Semantic Web 2009: 337-
1 (Jan. 2004), 116-142. 362
[12] Hussein, T. and J. Ziegler 2008. Adapting web sites by [25] Tso-Sutter, K., L. Marinho, et al. 2008. Tag-aware
spreading activation in ontologies. ReColl '08: Int. Workshop recommender systems by fusion of collaborative filtering
on Recommendation and Collaboration (in conjunction with algorithms. the 2008 ACM symposium on Applied
IUI 2008). computing, SAC '08, Fortaleza, Ceara, Brazil, ACM.
[13] Mika, P 2007. Ontologies are us: A unified model of social [26] Zhao, S., N. Du, et al. 2008). Improved recommendation
networks and semantics. J. Web Sem. 5(1): 5-1 based on collaborative tagging behaviors. the 13th
international conference on Intelligent user interfaces, IUI
[14] Nauerz, A., S. Pietschmann, et al. 2008. Using Collective '08, Gran Canaria, Spain, ACM.
Intelligence for Adaptive Navigation in Web Portals. 3rd
International Workshop on Adaptation and Evolution in Web [27] Ziegler, C.-N. and G. Lausen, 2004. Spreading activation
Systems Engineering at 8th International Conference on Web models for trust propagation. IEEE International Conference
Engineering 2008, Yorktown Heights, New York, USA. on e-Technology, e-Commerce, and e-Service, IEEE
Computer Society Press.
[15] Niwa, S., T. Doi, et al. 2006. Web Page Recommender
System based on Folksonomy Mining for ITNG'06 [28] Brin, S. and Page, L., 1998. The Anatomy of a Large-Scale
Submissions. Third International Conference on Information Hypertextual Web Search Engine. In: Seventh International
Technology: New Generations, ITNG 2006. World-Wide Web Conference (WWW 1998), April 14-18,
1998, Brisbane, Australia.
[16] Rocha, C, Schwabe, D., & Poggi de Aragao, M. 2004. A
Hybrid Approach for Searching in the Semantic Web.