=Paper=
{{Paper
|id=Vol-290/paper-5
|storemode=property
|title=Finding Experts by Link Prediction in Co-authorship Networks
|pdfUrl=https://ceur-ws.org/Vol-290/paper04.pdf
|volume=Vol-290
|dblpUrl=https://dblp.org/rec/conf/fews/PavlovI07
}}
==Finding Experts by Link Prediction in Co-authorship Networks==
https://ceur-ws.org/Vol-290/paper04.pdf
Finding Experts by Link Prediction in
Co-authorship Networks
Milen Pavlov1,2 , Ryutaro Ichise2
1
University of Waterloo, Waterloo ON N2L 3G1, Canada
2
National Institute of Informatics, Tokyo 101-8430, Japan
Abstract. Research collaborations are always encouraged, as they of-
ten yield good results. However, the researcher network contains massive
amounts of experts in various disciplines and it is difficult for the indi-
vidual researcher to decide which experts will match his own expertise
best. As a result, collaboration outcomes are often uncertain and research
teams are poorly organized. We propose a method for building link pre-
dictors in networks, where nodes can represent researchers and links -
collaborations. In this case, predictors might offer good suggestions for
future collaborations. We test our method on a researcher co-authorship
network and obtain link predictors of encouraging accuracy. This leads us
to believe our method could be useful in building and maintaining strong
research teams. It could also help with choosing vocabulary for expert de-
scription, since link predictors contain implicit information about which
structural attributes of the network are important with respect to the
link prediction problem.
1 Introduction
Collaborations between researchers often have a synergistic effect. The
combined expertise of a group of researchers can often yield results far
surpassing the sum of the individual researchers’ capabilities. However,
creating and organizing such research teams is not a straightforward task.
The individual researcher often has limited awareness of the existence of
other researchers with which collaborations might prove fruitful. Further-
more, even in the presence of such awareness, it is difficult to predict in
advance which potential collaborations should be pursued. As a result,
an expert in a particular field might find himself uncertain as to who to
collaborate with.
Experts’ semantic descriptions might be very helpful here. If one knew
to what extent each researcher is an expert in each field, one could po-
tentially use this knowledge to find researchers with compatible expertise
and suggest collaborations. However, semantic descriptions are often un-
available due to lack of supporting vocabulary.
42 2nd International ExpertFinder Workshop (FEWS2007)
Fortunately, semantic descriptions are not the only tool that can re-
alize the goal of suggesting collaborations. We propose a method for link
prediction in networks, where nodes can represent researchers and links
- collaborations. Thus, if we could predict the appearance of new links
with reasonable accuracy, links that have been predicted but do not ap-
pear might be good suggestions for future research collaborations.
Our method extracts structural attributes from the graph of past
collaborations and uses them to train a set of predictors using supervised
learning algorithms. These predictors can then be used to predict future
links between existing nodes in the graph. We test our method on a co-
authorship network and the results confirm that the appearance of new
collaborations is dependent on past network structure and that supervised
learning methods can exploit this dependence to make predictions with
reasonable accuracy. The approach is not specific to the domain of co-
authorship networks and can be easily applied to virtually all types of
networks in which link prediction is desirable.
Section 2 briefly mentions this research and compares it to our method.
Section 3 formally defines how we approach the link prediction problem.
Section 4 describes the experiments we run and discusses their results.
Finally, Section 5 concludes.
2 Related Work
Research in the area of network evolution models [14] is closely related
to the problem of link prediction. Kashima et. al [9] attempt to fit a
parametrized “copy-and-paste” model to a partial snapshot of a network
and use this model to predict the full network structure. The same prob-
lem (though viewed in slightly differing ways) is tackled by [16] and [19]:
they build partial networks from observable data and use them to infer
links that are not directly observable but are likely to exist. These ap-
proaches all differ from our view of the problem, as they only work with a
static snapshot of a network and do not consider network evolution over
time.
A view similar to ours is employed in [12], where a collaboration net-
work is cut into time slices and the network structure of one time slice is
used to predict the structure in the next slice. A link is considered likely
to appear if a structural attribute for the two end-nodes achieves a high
score. The authors test and compare the predictive power of several such
attributes. While generally achieving a low prediction rate, their approach
is promising in the sense that it significantly outperforms a random pre-
2nd International ExpertFinder Workshop (FEWS2007) 43
dictor. This implies there is, in fact, a correlation between the structure of
a collaboration network and the places where collaborations are likely to
appear in the future. Our approach aims to build an improved method for
link prediction by utilizing multiple structural attributes in the prediction
of a single link. The following explains how.
3 Supervised Learning Method for Building Link
Predictors
Consider a network G, defined in the following way. Let G = V, E, W be
a weighted graph with nodes vi ∈ V and edges (vi , vj ) ∈ E, 1 ≤ i, j ≤ |V |,
where wij ∈ W denotes the weight of edge (vi , vj ). For every pair of
nodes in the graph we can compute various graph attributes (such as
shortest path, common neighbours, etc.), which may be correlated with
the probability that a link between the nodes will appear in the future.
Once computed, the set of attribute values for a pair of nodes is said to
form a feature vector. Let fij denote the feature vector for pair (vi , vj ).
Note that vectors fij are defined for all pairs of nodes in V , not only for
the subset E.
There are many structural attributes that can be used to form a fea-
ture vector. [2] and [12] describe a multitude of candidate attributes.
Clearly, the effectiveness of different attributes will depend on the nature
of the network whose links they are utilized to predict. In our method we
choose a variety of attributes that are not specific to any type of network
but we think might be informative for a wide range of different types of
networks. Our selection of attributes is enumerated in Section 3.1.
Let F denote the full set of feature vectors. Formally speaking, F =
{fij |(vi , vj ) ∈ V × V }. The size of F varies, depending on whether the
graph is directed or undirected and whether self-links are allowed. In co-
authorship networks, self-links and directed edges are meaningless and so
|F | = |V |·|V2 −1| ; in web page networks such as the Internet, |F | = |V |2 .
We want to learn how to use the information contained in each feature
vector fij to predict its label yij . The label yij will equal true iff a new link
appears, joining nodes vi and vj . We assume that some labels are already
known and denote them by the set Yr . Our goal is to predict the remaining
labels Ys . We train several supervised learning algorithms on the subset
of feature vectors that correspond to labels in Yr . Formally speaking,
this training set is defined as Fr = {fij |yij ∈ Yr }. Testing is done on
the set of remaining feature vectors Fs = F − Fr , which we will refer to
as the test set. Since there are many learning algorithms available (and
44 2nd International ExpertFinder Workshop (FEWS2007)
Fig. 1. Building link predictors from network structure. Step 1: Extract feature vectors
from network. Step 2: Train predictors on extracted feature vectors using supervised
learning algorithms.
Table 1. Attributes computed for each node pair (vi , vj ).
Attribute name Formula
Shortest path min {s|pathssij > 0}
Common neighbours |Γ (vi ) ∩ Γ (vj )|
|Γ (vi )∩Γ (vj )|
Jaccard’s coefficient |Γ (v )∪Γ (vj )|
i 1
Adamic/Adar vk ∈Γ (vi )∩Γ (vj ) log |Γ (vk )|
Preferential attachment |Γ (vi )| · |Γ (vj )|
∞
Katzβ s=1
β s ·pathssij
Weighted Katzβ Same as above, but paths1ij = wij
PageRankmin
d min {PageRankd (vi ), PageRankd (vj )}
PageRankmax max {PageRankd (vi ), PageRankd (vj )}
d
1 if vi = vj
SimRankγ Σa∈Γ (v ) Σb∈Γ (v ) SimRankγ (a,b)
i j
γ· |Γ (vi )|·|Γ (vj )|
otherwise
Link value wij
none of them are universally effective), we see it reasonable to test and
compare the performance of several popular ones. Section 3.2 discusses
our selection.
For an overview of the method, refer to Fig. 1. The network entity
was discussed here. The feature vectors and predictors are discussed in
the following subsections, respectively.
3.1 Attributes in a feature vector
A feature vector fij ∈ F consists of a number of attributes, computed
for the node pair (vi , vj ). Table 1 shows the attributes we consider in our
method. Γ (vi ) denotes the set of all neighbours of vi and pathssij denotes
the number of paths of length s connecting vi to vj . The following briefly
describes the significance of each attribute.
2nd International ExpertFinder Workshop (FEWS2007) 45
The shortest path between two nodes vi and vj is defined as the mini-
mum number of edges connecting them. If there is no such connecting path
then the value of this attribute is generally assumed to be infinite. For
many types of networks, the shorter the shortest path between two nodes
is, the more likely the nodes are to become linked. Common neighbours
counts the number of neighbours that the two nodes have in common.
E.g., for two researchers in a collaboration network, this is the number
of other researchers, with which both have had some collaboration. Ar-
guably, if two researchers tend to collaborate with the same group of other
researchers then they are likely to collaborate with each other as well. Jac-
card’s coefficient [18] is a normalized measure of common neighbours. It
computes the ratio of common neighbours out of all neighbours (common
or not). This is sometimes a better measure than the (unnormalized)
common neighbours, especially when one end-node has a substantially
larger neighbourhood than the other. Adamic/Adar [1] measures simi-
larity between two nodes by weighing “rarer” common neighbours more
heavily. The rationale is that two nodes that have a common neighbour
that no other node has are often more “similar” than two nodes whose
common neighbours are common for many other nodes. Arguably, the
more similar nodes are, the more likely they are to be linked. Preferential
attachment [13] says that new links will be more likely to connect higher-
degree3 nodes than lower-degree nodes. In a collaboration network, this
means a new collaboration is more likely to occur between authors who
collaborate more often (regardless of who they collaborate with). This
(unnormalized) likelihood is reflected by the simple product of the nodes’
degrees. Katz [10] is a refined measure of shortest path. It considers all
paths between two nodes and weighs shorter ones more heavily. The “non-
attenuation” parameter β ∈ [0, 1] controls the aggressiveness of weighing.
E.g., a very small β yields a measure which is similar to common neigh-
bours, since pathss -values for higher s will not contribute significantly to
the summation. Weighted Katz uses the same core formula as Katz, but
also observes the weight between linked nodes. Thus, two nodes that are
connected by “heavier” paths will achieve a higher score. PageRank is the
same core algorithm used by Google to rank search results [3]. Using our
d (vm )
notation, P ageRankd (vi ) = (1 − d) + dΣvm ∈Γ (vi ) P ageRank
Γ (vm ) . In effect,
the rank of a node in the graph is proportional to the probability that
the node will be reached through a random walk on the graph. d ∈ [0, 1]
is a damping factor which specifies how likely the algorithm is to visit the
3
The degree of a node is equal to the number of edges linking to it. Thus, deg(vi ) =
|Γ (vi )|.
46 2nd International ExpertFinder Workshop (FEWS2007)
node’s neighbours rather than starting over (refer to [3] for more details).
Note that the original algorithm computes ranks over nodes. Since we
need ranks over pairs, we take the minimum and maximum page ranks
for the two nodes in a pair. SimRank [8] is also recursively defined. It
states two nodes are similar to the extent they are connected to similar
neighbours. γ ∈ [0, 1] is a parameter that controls how fast the weight of
connected nodes’ SimRank decreases as they get farther away from the
original nodes. Finally, the link value of a pair of nodes is simply the
weight of the edge between them. If such an edge does not exist then the
value is assumed to be zero.
3.2 Predictors
First, let us formally define what we mean by a predictor 4 . A predictor
p : F → {true, f alse} is a function that maps feature vectors to the
binary space. A good predictor is one for which p(fij ) = yij holds true
for a large proportion of the test feature vectors fij ∈ Fs .
We build predictors by training several learning algorithms on the
training set Fr . Each algorithm can produce different predictors, depend-
ing on the nature of the algorithm and the contents of Fr . The following
briefly outlines the algorithms we employ in our method.
Support vector machines (SVMs) [4] can be used to solve a wide range
of classification problems, including ours. The basic idea is as follows: a
feature vector containing n attributes can be mapped to a point in n-
dimensional space (where every dimension corresponds to an attribute).
Thus, our feature vectors set F is represented by a set of points in this
space. Each point then has its own binary label. The goal is to separate
the points into two groups so that points with the same label are in the
same group. A simple way to do this is by using a linear separator (i.e.,
an n-dimensional hyperplane) and this is the approach we adopt in our
experiments. To minimize generalization error, the hyperplane is usually
chosen in such a way as to maximize the margins on both of its sides. We
use the sequential minimal optimization (SMO) training algorithm, since
it is known to perform well with linear SVMs. [15]
Decision trees are also popular tools for solving classification prob-
lems. A decision tree is a tree whose terminal nodes are classification
outcomes and non-terminal nodes - decisions. Therefore, for each feature
vector, the label is obtained by traversing the tree from the root to a leaf
4
Our definition of a predictor is very similar to what is known in the machine learning
community as a classifier.
2nd International ExpertFinder Workshop (FEWS2007) 47
(terminal) node, following the branches specified by the decision nodes
along the way. An advantage of this approach over SVM is that we can ob-
serve the learned decision tree directly and potentially make meaningful
conclusions about which attributes are truly important for the prediction
problem. To build decision trees we use the J48 (also known as C4.5)
algorithm [17], which first chooses an attribute that best differentiates
the feature vectors and uses it as a decision node in the tree. The set of
feature vectors is then split in two, according to this decision node, and
new decision nodes are chosen in the same way for both partitions. The
process terminates when all nodes in a partition have the same label or
when no attributes can split a partition any further.
A Decision stump is a decision tree with only one decision node. It
is normally not a good classification tool on its own, but in conjunction
with AdaBoost (described below) can sometimes yield good results.
Boosting [5] is a kind of meta-learning: It considers a set of “weak”
predictors and computes a prediction based on their individual opinions.
A popular boosting algorithm is AdaBoost [7]. AdaBoost trains a set of
weak predictors successively, by presenting to each predictor training data
on which the previous predictor performed poorly. In the end, the predic-
tion label is decided by the weighted average of the votes of each predictor.
AdaBoost can be used with any of the previously mentioned predictors,
although it generally achieves a better performance boost when applied
with simpler predictors (e.g., decision stump).
Finally, we consider the random predictor as a baseline for compari-
son. This predictor assigns a random label to each feature vector, regard-
less of attribute values. There is no training involved.
4 Experiments and Results
To evaluate the feasibility and practicality of our method, we conduct a
set of experiments on a Japanese co-authorship network. The data comes
from the Institute of Electronics Information and Communication Engi-
neers (IEICE). The IEICE is considered by many Japanese scholars to
be the Japanese analogue to the Institute of Electrical and Electronics
Engineers [11]. The data is hosted by the National Institute of Informat-
ics (NII), Tokyo. It contains information about 110,210 published articles
by 86,696 authors, collected over the years from 1993 to (and including)
2006. By testing our method on this co-authorship network, we hope to
gain valuable information about which structural attributes are most in-
48 2nd International ExpertFinder Workshop (FEWS2007)
dicative of future collaborations and to help with identifying potentially
fruitful collaborations that have not happened.
Using the notation defined in Secion 3, we construct the network
G = {V, E, W } as follows. V is the set of all authors in the data set. E =
{(vi , vj ) ∈ V × V |vi and vj have co-authored at least one paper}. Finally,
the weight of a collaboration between two authors, wij , is equal to the
number of co-authored papers between them. The full co-authorship net-
work constructed this way is effectively an undirected graph, having 86,696
nodes and 672,332 edges.
However, this network contains 14 years of evolution history clumped
together and it might be difficult to see (or predict) evolution patterns
unless we consider some method of time slicing. We split the data set
into two partitions, each consisting of seven years of evolution data, and
perform the same experiment on both partitions independently. We want
to compare the performances of both sets of predictors thus obtained and
check if there are any differences in the way they use structural attributes
to predict future links. Such differences might suggest that not only the
network structure has evolved but that the evolution dynamics have also
changed.
Before jumping into the experiments, let us define some performance
metrics for predictor evaluation. For each feature vector, a predictor p
can make either a positive (P) or a negative (N) prediction concerning
the corresponding label.5 In the positive case, if p is correct, the predic-
tion is said to be true-positive (TP); otherwise it is false-positive (FP).
Conversely, in the negative case a prediction can be either true-negative
(TN) if correct or false-negative (FN) if wrong. We can now define the
metric recall as the proportion of TP predictions out of all true labels.
Recall will give us an idea of how well p is able to predict collaborations
that will happen in the future. It might also be useful to define the metric
precision as the proportion of TP predictions out of all positive predic-
tions. Precision will be useful in determining how well p fits the whole
data (as opposed to just always predicting true, which guarantees a 100%
recall rate). Using both precision and recall we are able to numerically
evaluate and compare predictors.
|T P | |T P |
precision = , recall =
|T P | + |F P | |T P | + |F N |
5
To disambiguate: A positive prediction means p thinks the label is true.
2nd International ExpertFinder Workshop (FEWS2007) 49
The following two subsections describe the two experiments and their
results individually. Section 4.3 gives some general conclusions regarding
both experiments.
4.1 The data from 1993 to 1999
The co-authorship network for the data set spanning the years 1993 to
1999 is constructed the same way as described above. However, before
extracting feature vectors we draw attention to several things. First, we
note that many collaborations between authors occur only once. This
results in many edges of weight one. These “tenuous” edges do not seem to
carry much information about collaboration tendencies between authors
- in fact, they could often be attributed to chance6 . For this reason, we
find it reasonable to filter out all edges whose weight is equal to one.
From the seven years of evolution data we need to extract feature vec-
tors and corresponding labels. We use the first four years for the former
and the remaining three for the latter. We note, however, that some au-
thors are only active during the first four years but stop publishing after
that. Conversely, others only start publishing in the latter three years and
are inactive in the preceding time period. In the graph, this results in a
variable nodes set V , which leads to a mismatch between feature vectors
and labels. To avoid this, we trim the set V to only contain authors that
are active during both time periods. The resulting network contains 1,380
nodes and 1,620 edges.
We use the following parameter values for feature extraction: Katz’s
β = 0.05, PageRank’s d = 0.85 and SimRank’s γ = 0.8. These seem to
be the popular values in the research community [3, 8, 10]. The count of
thus computed feature vectors is upwards of a million. However, we note
that many feature vectors are duplicates. The most prominent example
is for a pair of nodes that are far apart in the graph, have no common
paths between them, score very low on all attributes and end up never
publishing together. Such pairs usually result in the same feature vectors.
(We consider two feature vectors the same if none of their corresponding
attribute values differ by more than 0.001.) We remove all duplicates and
obtain a set of 29,437 unique feature vectors. This is the set F we will
present to our supervised learning algorithms. Of the labels corresponding
to vectors in this set, we note that only 1.6% are true.
6
E.g., two authors might never personally collaborate, yet still have their names
appear on a paper because of a third party. Unless this happens again, we consider
the link between such authors accidental.
50 2nd International ExpertFinder Workshop (FEWS2007)
For training, we use a subset containing 90% of the vectors in F ,
sampled at random. Testing is performed on the remaining 10%. Imple-
mentations of the learning algorithms described in Section 3.2 are already
available in the Weka software package [20]. We use these implementations
for our experiments.
Let us look at the results shown in Fig. 2. The SMO, J48 and Ad-
aBoost/SMO predictors display outstanding recall rates, with SMO even
going so far as to predict 100% of the true test samples correctly. The
equation used by SMO to accomplish this feat assigns relatively high
weights to the Katz and preferential attachment attributes (respectively,
-3.1 and 1.8) and low weights to the rest of the attributes (between -1
and 1). The shortest path, Jaccard’s coefficient, PageRankmax and Sim-
Rank are almost completely ignored, since each of their weights is less
than |0.06|. The AdaBoost version of SMO uses a weighted average of ten
“weak” SMO predictors. The first one is the same as above, however, the
other nine seem to perform much worse and lower the overall performance
of the compound predictor. The nine predictors seem to favour shortest
path and link value attributes in their weight distributions. J48 performs
almost as well as AdaBoost/SMO. We were hoping that the decision tree
learned by J48 could give us some insight as to which attributes are most
informative to the co-authorship prediction problem, but in actuality the
tree was too large to even visualize. Similar to the case of AdaBoost/SMO,
AdaBoost/J48 achieves a lower performance than J48 alone. Lastly, the
performance of AdaBoost/DecisionStump should be thought of as the
worst of all predictors. Though it achieves a good precision rate, the
reason it does so is simply because it is overly conservative in its true
predictions: This leads to a decrease in FP rates (boosting precision) at
the cost of a considerable increase in FN rates (resulting in the abysmal
recall performance).
4.2 The data from 2000 to 2006
We run this experiment the same way as described in the previous subsec-
tion. During this latter seven-year period however, the number of active
authors, as well as their publishing rates, seem to have increased, result-
ing in a network of 3,603 nodes and 7,512 edges. We extract more than
ten million feature vectors in total and take only the 1.6 million unique
ones. Out of those, we note that only 0.019% have true labels.
We observe some differences when we compare the results from the
previous subsection to those shown in Fig. 3. In the latter, SMO and
J48 both score lower than before, on both measures. This might be an
2nd International ExpertFinder Workshop (FEWS2007) 51
Fig. 2. Classification results for the years 1993 ∼ 1999.
indication that new collaborations are now less dependent on past net-
work structure or that the dependence holds for shorter periods of time.
AdaBoost does not improve the performance of J48, but it achieves a re-
markable 100% recall rate when combined with DecisionStump. It seems
here the case is reversed from before: Now this predictor is being overly
generous in its true predictions, which leads to a decline in precision. All
ten “weak” DecisionStumps base their one-node decisions on either the
shortest path or link value attributes.
Weka failed to provide a result for the AdaBoost/SMO experiment.
We suspect that the size of the training set might have been prohibitively
large for the limited amount of memory (10GB) on the test machine.
4.3 General Discussion
First, we note that virtually all of our predictors outperform the ran-
dom classifier. This is encouraging in itself, considering our predictions
are solely based on structural information from the co-authorship net-
work and do not take into account any node-specific properties (such as
university affiliations, geographical locations, research areas, etc.). It also
confirms the findings of [12], namely, that there is valuable information
to be learned from the structure of a collaboration network, with respect
to its evolution.
52 2nd International ExpertFinder Workshop (FEWS2007)
Fig. 3. Classification results for the years 2000 ∼ 2006.
Some of the predictors’ recall rates might seem overly optimistic. This
might indicate that the evolution dynamics of the tested co-authorship
network are easy to predict or that the proportion of true test samples is
too small to yield reliable metric readings.
Regardless, we observe that our predictors do make meaningful pre-
dictions for the data tested on. For example, the misclassified samples by
J48 in the second experiment include pairs of authors who have a high
number of common neighbours (up to 9), large preferential attachment
value (up to 171) and have co-authored plenty of papers in the past (up to
23), yet end up not collaborating again when it seems likely they would.
Even when making mistakes, the predictor is not entirely unreasonable. In
fact, these “mistakes” might be a fair indication that such collaborations
should be encouraged.
Overall, we see the results of both experiments as evidence that our
method can be very effective in building accurate link predictors in co-
authorship networks. Since the method itself relies solely on structural
attributes of the underlying network and on general supervised learning
algorithms, it should be easily extendible to any kinds of networks in
which link prediction is desirable. Whether similar success can be achieved
in such networks remains to be seen.
2nd International ExpertFinder Workshop (FEWS2007) 53
5 Conclusion
This paper presented a supervised learning method for building link pre-
dictors from structural attributes of the underlying network. In a network
of researchers, where a link represents a collaboration, such predictors
could be useful is suggesting unrealized collaborations and thus help in
building and maintaining strong research teams. In addition, by analyz-
ing the algorithmic structure of predictors build for a specific network, we
could gain valuable information about which attributes are most infor-
mative for the link prediction problem and use this knowledge as a basis
for specifying vocabularies for expert description.
There are many directions future research in this area might take.
We think an important next step would be evaluating the generality of
the method by testing it on various kinds of networks and comparing
the predictive accuracy of the respective link predictors. It would be also
interesting to see if predictors would rely on different structural attributes
in different networks and if there is any evident correlation between the
type of network and informative attributes. We also believe the process
of time slicing in building network data should be investigated further.
For example, in both of our experiments, we use the first four years of
network history to predict the evolution in the following three. However,
many networks do not evolve at a constant rate. It would be interesting to
see if this rate varies significantly and if we can adjust the durations of the
history and evolution time spans to take such variations into account. One
way to do this would be by compiling different pairs of history-evolution
data and comparing predictor performances on them.
References
1. L. Adamic, E. Adar. “Friends and neighbors on the web”. Social Networks, 25(3),
2003.
2. K. Börner, S. Sanyal, A. Vespignani. “Network Science”. In B. Cronin, ed., Annual
Review of Information Science and Technology, Information Today, Inc./American
Society for Information Science and Technology, Medford, NJ, in press.
3. S. Brin, L. Page. “The anatomy of a large-scale hypertextual Web search engine”.
In Proceedings of the 7th international conference on World Wide Web, 107-117.
Brisbane, Australia, 1998.
4. C.J.C. Burges. “A Tutorial on Support Vector Machines for Pattern Recognition”.
Data Mining and Knowledge Discovery, 2(2):1-47. 1998.
5. T.G. Dietterich. “Ensemble Methods in Machine Learning”. In J. Kittler, F. Roli,
eds., Proceedings of the 1st International Workshop on Multiple Classifier Systems,
LNCS, vol. 1857, 1-15. Springer-Verlag, London, 2000.
6. R.O. Duda, P.E. Hart, D.G. Stork. “Pattern Classification” (2nd Edition). Wiley-
Interscience, 2000.
54 2nd International ExpertFinder Workshop (FEWS2007)
7. Y. Freund, R.E. Schapire. “A decision-theoretic generalization of on-line learn-
ing and an application to boosting”. Journal of Computer and System Sciences,
55(1):119-139, 1997.
8. G. Jeh, J. Widom. “SimRank: A measure of structural-context similarity”. In Pro-
ceedings of the 8th ACM SIGKDD International Conference on Knowledge Dis-
covery and Data Mining, Edmonton, Alberta, Canada. 2002.
9. H. Kashima, N. Abe. “A Parametrized Probabilistic Model of Network Evolution
for Supervised Link Prediction”. In Proceedings of the 6th IEEE International
Conference on Data Mining, 2006.
10. L. Katz. “A new status index derived from sociometric analysis”. Psychometrika
18, 39-43, 1953.
11. G. LaRowe, R. Ichise, K. Börner, “Analysis of Japanese Information Systems Co-
authorship Data”. In Proceedings of the 11th International Conference on Infor-
mation Visualization, 459-464, 2007.
12. D. Liben-Nowell, J. Kleinberg. “The Link Prediction Problem for Social Networks”.
In Proceedings of the 12th International Conference on Information and Knowledge
Management, 2003.
13. M.E.J. Newman. “Clustering and preferential attachment in growing networks”.
Phys. Rev. E, 64(2):025102, 2001.
14. M.E.J. Newman. “The structure and function of complex networks”. SIAM Review,
45:167-256, 2003.
15. J.C. Platt. “Fast training of support vector machines using sequential minimal
optimization”. In B. Schölkopf, C. J. Burges, and A. J. Smola, eds., Advances in
Kernel Methods: Support Vector Learning, 185-208. MIT Press, Cambridge, MA,
1999.
16. A. Popescul, L. Ungar. “Statistical relational learning for link prediction”. In
Workshop on Learning Statistical Models From Relational Data, 81-90. ACM Press,
New York, 2003.
17. J.R. Quinlan. “C4.5: Programs for Machine Learning”. Morgan Kaufmann Pub-
lishers, 1993.
18. G. Salton, M.J. McGill. “Introduction to Modern Information Retrieval”. McGraw-
Hill, Inc., New York, NY, 1986.
19. B. Taskar, M.-F. Wong, P. Abbeel, D. Koller. “Link prediction in relational data”.
In Proceedings of the Neural Information Processing Systems, 659-666. MIT Press,
Cambridge, MA, 2003.
20. I.H. Witten, E. Frank, L. Trigg, M. Hall, G. Holmes, S.J. Cunningham. “Weka:
Practical machine learning tools and techniques with Java implementations”.
In H. Kasabov, K. Ko, eds., ICONIP/ANZIIS/ANNES International Workshop,
Dunedin, 1999.
2nd International ExpertFinder Workshop (FEWS2007) 55