USE OF DOMAIN KNOWLEDGE FOR DIMENSION REDUCTION

Application to Mining of Drug Side Effects

Emmanuel Bresso

1,2

, Sidahmed Benabderrahmane

, Malika Smail-Tabbone

, Gino Marchetti

Arnaud Sinan Karaboga

, Michel Souchet

, Amedeo Napoli

and Marie-Dominique Devignes

LORIA UMR 7503, CNRS, Nancy Universit

e, INRIA NGE, 54503 Vandoeuvre-les-Nancy, France

Harmonic Pharma, Espace Transfert INRIA NGE, 615 Rue du Jardin Botanique, 54600 Villers-les-Nancy, France

Keywords:

Dimension reduction, Clustering, Semantic similarity, Drug side effects.

Abstract:

High dimensionality of datasets can impair the execution of most data mining programs and lead to the pro-

duction of numerous and complex patterns, inappropriate for interpretation by the experts. Thus, dimension

reduction of datasets constitutes an important research orientation in which the role of domain knowledge

is essential. We present here a new approach for reducing dimensions in a dataset by exploiting semantic

relationships between terms of an ontology structured as a rooted directed acyclic graph. Term clustering is

performed thanks to the recently described IntelliGO similarity measure and the term clusters are then used

as descriptors for data representation. The strategy reported here is applied to a set of drugs associated with

their side effects collected from the SIDER database. Terms describing side effects belong to the MedDRA

terminology. The hierarchical clustering of about 1,200 MedDRA terms into an optimal collection of 112 term

clusters leads to a reduced data representation. Two data mining experiments are then conducted to illustrate

the advantage of using this reduced representation.

1 INTRODUCTION AND

MOTIVATION

In some domains such as biology, data complexity

is ubiquitous and constitutes a major challenge for

Knowledge Discovery from Databases (KDD) ap-

proaches. One can anticipate that the more complex

the data, the more relevant and interesting the ex-

tracted knowledge is. However, human expertise is

then heavily required to supervise the KDD process

especially for the data preparation and the interpreta-

tion steps. Thus, an interesting research orientation

consists in exploring how domain knowledge can be

used to help the expert and contribute to the KDD pro-

cess in an automated way.

Complex data are usually voluminous and high-

dimensional. In a classical ObjectsXAttributes repre-

sentation, the number of objects reﬂects the volume of

data to handle and the number of attributes provides

the number of dimensions in the dataset. Most data

mining methods become rather inefﬁcient when the

dimensionality is too large. Moreover, the patterns ex-

tracted from the data may also reveal inappropriate for

interpretation by the expert due to plethoric, redun-

dant, or non informative attributes. Data reduction is

therefore a crucial issue for data preparation in com-

plex datasets. Several methods exist and are reviewed

in section 2. An interesting situation is the case where

attributes are terms of a controlled and structured vo-

cabulary. In biology and biomedicine such vocabu-

laries (e.g. the Gene Ontology) have been created to

annotate biological objects in databases (genes, pro-

teins, etc.) in order to facilitate the retrieval of ob-

jects sharing similar annotations. Actually, these vo-

cabularies represent basic domain knowledge in the

form of semantic relationships between terms which

enhance subsequent data analyses such as classiﬁca-

tion or characterization.

The present case-study deals with drugs anno-

tated with respect to their side effects. The objects

are drugs retrieved from the Side Effects Repository

(SIDER) database, which compiles all side effects

described in the drug package inserts (Kuhn et al.,

2010). The annotation terms used in SIDER pertain

from the Medical Dictionary for Regulatory Activities

(MedDRA) terminology. Not less than 1,200 Med-

DRA terms are used in SIDER to annotate the drugs.

Dealing with annotation terms taken from a hierar-

chical vocabulary allows to use existing strategies for

attribute reduction such as generalization (Han and

271

Bresso E., Benabderrahmane S., Smail-Tabbone M., Marchetti G., Sinan Karaboga A., Souchet M., Napoli A. and Devignes M..

USE OF DOMAIN KNOWLEDGE FOR DIMENSION REDUCTION - Application to Mining of Drug Side Effects.

DOI: 10.5220/0003662602630268

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2011), pages 263-268

ISBN: 978-989-8425-79-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Kamber, 2001). However, available domain vocabu-

laries ore often organized as a rooted Directed Acyclic

Graph (DAG) rather than a tree structure in which

generalization is intractable. In the present study,

we propose alternatively to cluster annotation terms

based on their similarity in the rooted DAG using

the IntelliGO similarity measure that was initially de-

ﬁned on the Gene Ontology (Benabderrahmane et al.,

2010).

The application of the IntelliGO measure to the

MedDRA vocabulary resulted in the clustering of the

1,200 MedDRA terms into an optimal collection of

112 term clusters. These term clusters led to a reduced

data representation in which drugs are annotated with

term clusters instead of MedDRA terms (Section 3).

Several data mining experiments are then conducted

to show the advantage of using the reduced represen-

tation of the data (Section 4).

2 DATA REDUCTION FOR KDD:

A BRIEF STATE OF THE ART

Methods for data reduction divide into two types: fea-

ture selection and dimension reduction. Feature se-

lection (or attribute selection) is a means of data re-

duction without altering the original data representa-

tion (Guyon and Elisseeff, 2003). Feature selection

methods fall into two categories: ﬁlters and wrap-

pers. Filter methods perform feature selection as a

pre-processing step based on a search in the feature

space. Feature subset evaluation relies on measures

that use the training data properties and is carried out

independently of any classiﬁer (John et al., 1994). A

typical example is the correlation-based feature selec-

tion method which eliminates redundant and irrele-

vant attributes by selecting those that individually cor-

relate with the class but have little inter-correlation

(Witten et al., 2011). Concerning the wrapper meth-

ods, they evaluate candidate feature subsets on the ba-

sis of their predictive accuracy (classiﬁcation perfor-

mance) using a learning algorithm (Kohavi and John,

1997). Most feature selection methods work with su-

pervised classiﬁcation and use the class information

of the training examples to select the relevant features.

However, Wrapper methods were recently proposed

for feature selection upstream unsupervised learning,

namely clustering (Kim et al., 2000; Dy and Brodley,

2004). Such studies focus on how to evaluate the re-

sults of clustering for each candidate feature subset.

Also,a recent study suggests a ﬁlter method indepen-

dent of both the learning algorithm and any predeﬁned

classes by guiding the attribute selection with formal-

ized domain knowledge (Coulet et al., 2008).

Alternative data reduction methods alter the data

representation by encoding the data into a smaller rep-

resentation space. Such dimension reduction is also

called feature compression. The principal compo-

nent analysis is a popular example of such methods

which deals with numeric and continuous data. Clus-

tering was proposed for grouping the attributes in or-

der to improve the classiﬁcation or the clustering of

textual documents (Kyriakopoulou, 2008). Here the

attributes are binary and correspond to words annotat-

ing textual documents. Word clustering then relies on

comparing their joint distributions in the documents

over the classes (Koller and Sahami, 1996; Slonim

and Tishby, 2000). Thus, the similarity measures used

for clustering the word attributes are corpus-based.

In this paper, we propose that when the attributes

belong to a domain-speciﬁc structured vocabulary, a

better clustering of these attributes could be achieved

by using a suitable semantic similarity measure.

3 SEMANTIC CLUSTERING OF

ATTRIBUTES

3.1 The MedDRA Terminology

The MedDRA medical terminology is used to classify

adverse event information associated with the use of

drugs and vaccines. MedDRA is a part of the Uni-

ﬁed Medical Languages System (UMLS) and is of-

ten presented as a hierarchy consisting of ﬁve lev-

els: System Organ Class, High Level Group Term,

High Level Term, Preferred Term, Lowest Level Term

(MedDRA, 2007). Lowest level terms correspond

to different terms for the same preferred term. In

the MedDRA terminology, each term has an identi-

ﬁer and all the paths to the root can be download.

For example, for the term C0000733, we have two

paths: C1140263.C0017178.C0947761.C0947846

and C1140263.C0947733.C0021502.C0851837. A

careful review of the parent-child relationships shows

that the MedDRA is actually not a hierarchy: about

37% of the MedDRA terms have more than one di-

rect parent. This together with the natural oriented

of term-term relationship and the absence of cycle,

confers to the MedDRA terminology the status of a

rooted DAG.

3.2 Term-term Semantic Similarity

Measure

Two approaches exist for term-term semantic sim-

ilarity measures: structure-based approaches which

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

272

exploit the structure of the vocabulary (depth, path

length) and corpus-based approaches which exploit

the term distribution in a corpus (annotation fre-

quency, information content). The IntelliGO measure

is a structure-based approach in which the generalized

cosine similarity measure proposed by Ganesan for

hierarchical vocabularies has been adapted to rooted

DAGs (Benabderrahmane et al., 2010). The IntelliGO

measure was initially deﬁned for quantifying simi-

larity between Gene Ontology (GO) terms. For two

terms t

, t

, it takes into account the maximal depth of

their common ancestors (CA) and the minimal path-

length (SPL) between them:

Sim

IntelliGO

) =

2MaxDepth(CA)

MinSPL(t

) + 2MaxDepth(CA)

(1)

To calculate the semantic similarity between two

MedDRA terms, the algorithm starts by retrieving for

each term all their paths to the root node. Then, the

set of common ancestors is deﬁned as the intersec-

tion between the two sets containing the ancestors of

the two terms. In the next step, the algorithm identi-

ﬁes the common ancestors having the maximal depth

from the root node (MaxDepth(CA)). Note that the

Depth of a MedDRA term can be calculated as the

maximal length of a path from this term to the root

node. After that, the algorithm calculates the shortest

path length (MinSPL) separating the two terms. Fi-

nally, the semantic similarity between the two terms

is computed using the equation (1).

As the values of Sim

IntelliGO

range from 0 to 1, we

also deﬁne the distance Dist

IntelliGO

as its complement

to 1.

3.3 Clustering MedDRA Terms

A total of 1,288 terms from the 20,037 MedDRA

terms are used in the SIDER database for annotating

drug side effects. Pairwise distances were calculated

for these 1,288 terms. We then used the Ward’s hier-

archical agglomeration algorithm (Ward, 1963) with

an optimization step necessary to select the best level

where to cut the dendrogram in order to obtain a set of

clusters (Kelley et al., 1996). This method deﬁnes a

penalty value which is function of the cluster number

and the intra-cluster distance. When this value is min-

imal, the resulting clusters are as highly populated as

possible while simultaneously maintaining the small-

est average intra-cluster distance. In our case the min-

imal penalty value was obtained with 112 clusters

which were subsequently inspected and validated by

two experts. In the rest of this paper, these clusters

will be deﬁned as the term clusters (TC) which will

be used as attributes for data mining.

In order to label each TC with its most represen-

tative term, we introduce a function AvgDist

asso-

ciating to each TC term its average distance to all the

TC terms:

AvgDist

) =

|TC|

∑

j=1

dist(t

) (2)

Then the representative element R of a TC is the

term which minimizes AvgDist

. For example in

Table 1, Erythema is the representative element of

the TC. The label of a given TC is the concatenation

of the TC number and the representative term (e.g.,

54 Erythema for the TC described in Table 1). Once

TC are built, they can be used for dimension reduc-

tion.

Table 1: Example of TC with the average distance function

calculated for each term t.

Term Cluster Element t AvgDist

(t)

Decubitus ulcer 0.35

Rash 0.35

Lichen planus 0.32

Parapsoriasis 0.32

Pruritus 0.35

Psoriasis 0.37

Sunburn 0.35

Erythema 0.31

Pityriasis alba 0.32

Photosensitivity reaction 0.37

Rash papular 0.32

Dandruff 0.37

Lupus miliaris disseminatus faciei 0.35

Vulvovaginal pruritus 0.35

4 EVALUATION OF THE IMPACT

OF FEATURE CLUSTERING ON

DATA MINING

4.1 Experimental Design

In order to evaluate the impact of our dimension re-

duction strategy, two data mining experiments were

conducted. The ﬁrst experiment aims at retrieving fre-

quent associations of side effects shared by drugs in a

given drug category. The second experiment aims at

discriminating drugs belonging to two categories in

terms of side effects. Datasets consist of binary (Ob-

jects X Attributes) relations between drugs (Objects)

and their side effects (Attributes). The side effects are

represented either as individual MedDRA terms or as

TC leading to two data representations.

USE OF DOMAIN KNOWLEDGE FOR DIMENSION REDUCTION - Application to Mining of Drug Side Effects

273

In the ﬁrst experiment we search for Frequent

Closed Itemsets (FCIs) in order to compare the two

data representations with respect to computation time,

number and relevance of the extracted FCIs. In our

context, a FCI of length n and support s corresponds

to an association of n terms, respectively term clus-

ters, shared by the maximal group of drugs corre-

sponding to a percentage value s of the whole dataset.

The Zart program was used for FCI extraction on the

Coron platform (Szathmary et al., 2007). The experi-

ment was ran on a 2.6GHz processor with 1GB mem-

ory.

As for the second experiment we use the CN2-SD

subgroup discovery algorithm (Lavrac et al., 2004)

with the two data representations in order to check

the impact of term clustering on the computation time

and the produced subgroups. Given a population of

objects and a property of those objects that we are

interested in, subgroup discovery allows to ﬁnd sub-

groups of objects that are statistically most interest-

ing, i.e., as large as possible and having the most un-

usual distributional characteristics with respect to the

property of interest. In our case, two categories of

drugs are investigated with this method for identify-

ing subgroups of drugs sharing discriminative side-

effects in one category versus the other. The CN2-SD

algorithm implementation used is the one of the Keel

software (Alcala-Fdez et al., 2009) and was executed

on a 8-core 1.86GHz processor with 8GB memory.

4.2 Datasets Description

The category of a drug refers to its therapeutic uses.

The categories of the drugs present in the SIDER

DB are available in the DrugBank DB (Knox et al.,

2011). We chose to perform the data mining experi-

ments on the drugs corresponding to the two largest

categories: the Cardiovascular Agents (CA) and the

Anti-Infective Agents (AIA) containing respectively

94 and 76 drugs.

For each category, we built two datasets : the All

dataset has for attributes the 1,288 MedDRA terms

annotating the drugs in SIDER and the TC dataset has

for attributes the 112 TC (as described in section 3)

note that a TC is assigned to a drug if at least one

member of the TC is reported as annotating the drug

in SIDER. This gives four datasets CA

All

, AIA

All

, and AIA

4.3 Frequent closed Itemset Extraction

with and without Term Clustering

The Zart program was executed on each dataset with

a minimal support varying from 50 to 100%. Table

Table 2: Number of FCIs for each dataset when varying the

minimal support value.

Minimal

support

50% 60% 70% 80% 90% 100%

AIA

all

178 41 9 2 0 0

AIA

654 154 30 3 0 0

all

386 94 41 11 1 0

5,564 1,379 256 62 6 0

2 summarizes the number of FCIs produced in each

case.

The ﬁrst observation is the increase in the num-

ber of FCIs for a given minimal support when term

clusters are used. For the AIA

All

and AIA

datasets,

this increase varies from more than 3-fold for mini-

mal supports between 50 and 70% to about 1.5-fold

for higher minimal supports. For CA

All

and CA

datasets this increase goes from about 14-fold at min-

imal support 50 and 60% to 6-fold for higher min-

imal supports. This increase clearly reﬂects the ex-

pected role of clustering in feature reduction, namely

increasing the density of the binary (Objects X At-

tributes) relation by aggregating object properties.

Accordingly, the computation time of the program

was two-fold longer with the TC than with the All rep-

resentation.

Content analysis of the FCIs was done after rank-

ing FCIs according to their support. The ﬁve top-

ranked FCIs (plus ex-aequo) are listed for each dataset

in Figure 1. With the All representation, FCIs can be

very redundant. For example in the AIA

All

dataset

(left panel), three from the ﬁve displayed FCIs con-

tain very similar term: nausea, vomiting, nausea and

vomiting symptoms. On the contrary with the TC rep-

resentation (right panel), a unique FCI contains the at-

tribute 64 nausea and vomiting symptoms which rep-

resents the cluster of terms containing all three at-

tributes cited before. Thus, data reduction by term

clustering allows less redundant FCI extraction and

therefore leads to the presentation of more potentially

interesting itemsets to the expert.

Figure 1 also shows that the FCI supports are gen-

erally higher with the TC than with the All data rep-

resentation. A correspondence can be established be-

tween individual terms in the All representation and

the term clusters in the TC representation as illus-

trated in Figure 1 (remember that a term cluster is

labeled with its number and its most representative

term). For example, the {54 Erythema} FCI found

in the AIA

dataset with a support of 88% contains

the term cluster labeled 54

Erythema that includes

the Pruritus individual term and has therefore been

matched with the {Pruritus} FCI found in the AIA

All

dataset with a support of 79%.

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

274

Figure 1: The ﬁve most frequent FCIs are extracted from

the total list of FCIs obtained for each dataset. In case of

ex-aequo, all itemsets with the same support are listed.

Furthermore, combined with the lack of redun-

dancy, this increase of FCI support leads to the dis-

covery in the TC datasets of frequent term clusters for

which all individual terms are less frequent and there-

fore not considered in the All datasets.

4.4 Subgroup Discovery with and

without Term Clustering

The second experiment aims at discriminating the

drugs belonging to two categories in terms of pres-

ence or absence of side effects. Indeed, the absence of

side effects may also be important for drug characteri-

zation. This is possible with the CN2-SD algorithm as

the input data is in attribute-value format. We ran the

CN2-SD program on the following unions of datasets

in which an additional attribute was added for the cat-

egory information: CA

All

versus AIA

All

, and CA

versus AIA

The ﬁrst observation concerns the computation

time. When term clusters are used the execution time

is less than ten minutes whereas it does not resume

within six days when all side effects are used. Thus,

data reduction is necessary for successful execution

of the CN2-SD algorithm.

In a second stage, the rules extracted from CA

versus AIA

dataset were analyzed. The left part of a

rule is veriﬁed for a number of drugs (support) among

which a certain fraction (coverage) are of the category

indicated in the right part of the rule. The resulting

subgroup therefore identiﬁes a subset of drugs from

this category sharing a speciﬁc proﬁle of side effects

with regard to the other category. The best rules in

terms of coverage are shown in Table 3.

To sum up, these results show that the data reduc-

Table 3: Best rules (with coverage / support) extracted by

the CN2-SD program for CA

versus AIA

50 Angina pectoris = T AND 93 Bacteraemia = F

AND 52 Ichthyosis = F AND 54 Erythema = T AND

49 Folate deﬁciency = F => CA (0.96/56)

31 Splenic infarction = T AND 41 Neutropenia = T

AND 42 Penile discharge = F AND 77 Facial pain =

T AND 79 Cachexia = T => AIA (0.88/26)

tion used allows subgroup discovery which was im-

possible with the extended data representation. Fur-

ther investigation by domain experts is ongoing.

5 DISCUSSION AND

PERSPECTIVES

In this paper we have reported a method for dimension

reduction guided by domain knowledge. The method

is based on attribute clustering using a semantic sim-

ilarity measure. We took advantage of our recently

deﬁned IntelliGO similarity measure which applies to

the rooted DAG structure encountered in many vocab-

ularies.We believe that our strategy can be applied in

various other biomedical context (Leva et al., 2005;

Pakhomov et al., 2007). We tested our method on

a dataset of 170 drugs annotated with 1,288 terms

taken from the MedDRA terminology and represent-

ing the drugs possible side effects. Using IntelliGO-

based term-term distances and hierarchical clustering,

we reduced data representation from 1,288 individ-

ual terms down to 112 term clusters. In this work we

adopted a binary representation for the reduced data

representation, i.e., a TC is assigned to a drug if at

least one of its elements has been associated with the

drug in the SIDER database. This representation ig-

nores the impact of multiple associations between a

drug and TC elements. A many-valued relation could

be produced to take into account such situations. Re-

cently described extension of formal concept analysis

may help us handling such data representation (Mes-

sai et al., 2008; Kaytoue-Uberall et al., 2009).

The dimension reduction method we have devel-

oped was tested with two data mining algorithms: FCI

extraction and subgroup discovery. The results show

that FCIs extracted from the TC data representation

are less redundant and display higher supports than

from the All representation. Another consequence of

the data reduction is that the expert’s task is facili-

tated because more relevant and explicit side effects

are found among FCIs displaying high support. As

for subgroup discovery, dimension reduction revealed

to play a crucial role. Indeed, the program was unable

to resume with the All data representation whereas it

USE OF DOMAIN KNOWLEDGE FOR DIMENSION REDUCTION - Application to Mining of Drug Side Effects

275

provided, with the reduced TC representation, quite

interesting rules characterizing subgroups of one drug

category versus another one. Complementary experi-

ments can now be carried out to identify rules speciﬁc

of a given category versus all other categories.

REFERENCES

Alcala-Fdez, J., Snchez, L., Garca, S., del Jesus, M., Ven-

tura, S., Garrell, J., Otero, J., Romero, C., Bacardit,

J., Rivas, V., Fernndez, J., and Herrera, F. (2009).

KEEL: a software tool to assess evolutionary algo-

rithms for data mining problems. Soft Computing -

A Fusion of Foundations, Methodologies and Appli-

cations, 13(3):307–318–318.

Benabderrahmane, S., Smail-Tabbone, M., Poch, O.,

Napoli, A., and Devignes, M. (2010). IntelliGO: a new

vector-based semantic similarity measure including

annotation origin. BMC Bioinformatics, 11(1):588.

Coulet, A., Smail-Tabbone, M., Benlian, P., Napoli, A., and

Devignes, M. (2008). Ontology-guided data prepa-

ration for discovering genotype-phenotype relation-

ships. BMC Bioinformatics, 9(Suppl 4):S3.

Dy, J. G. and Brodley, C. E. (2004). Feature selection for

unsupervised learning. J. Mach. Learn. Res., 5:845–

889.

Guyon, I. and Elisseeff, A. (2003). An introduction to

variable and feature selection. J. Mach. Learn. Res.,

3:1157–1182.

Han, J. and Kamber, M. (2001). Data Mining: Concepts

and Techniques. Morgan Kaufmann, San Francisco, 1

edition.

John, G. H., Kohavi, R., and Pﬂeger, K. (1994). Irrelevant

Features and the Subset Selection Problem. In In-

ternational Conference on Machine Learning, pages

121–129.

Kaytoue-Uberall, M., Duplessis, S., Kuznetsov, S. O., and

Napoli, A. (2009). Two fca-based methods for mining

gene expression data. In ICFCA, pages 251–266.

Kelley, L. A., Gardner, S. P., and Sutcliffe, M. J. (1996).

An automated approach for clustering an ensemble

of NMR-derived protein structures into conforma-

tionally related subfamilies. Protein Engineering,

9(11):1063 –1065.

Kim, Y., Street, W. N., and Menczer, F. (2000). Feature

selection in unsupervised learning via evolutionary

search. In Proceedings of the sixth ACM SIGKDD in-

ternational conference on Knowledge discovery and

data mining, pages 365–369, Boston, Massachusetts,

United States. ACM.

Knox, C., Law, V., Jewison, T., Liu, P., Ly, S., Frolkis, A.,

Pon, A., Banco, K., Mak, C., Neveu, V., Djoumbou,

Y., Eisner, R., Guo, A. C., and Wishart, D. S. (2011).

DrugBank 3.0: a comprehensive resource for Omics

research on drugs. Nucleic Acids Research, 39(suppl

1):D1035 –D1041.

Kohavi, R. and John, G. H. (1997). Wrappers for feature

subset selection. Artiﬁcial Intelligence, 97(1-2):273–

324.

Koller, D. and Sahami, M. (1996). Toward optimal feature

selection. In Saitta, L., editor, Proceedings of the Thir-

teenth International Conference on Machine Learning

(ICML), pages 284–292. Morgan Kaufmann Publish-

ers.

Kuhn, M., Campillos, M., Letunic, I., Jensen, L. J., and

Bork, P. (2010). A side effect resource to capture phe-

notypic effects of drugs. Mol Syst Biol, 6.

Kyriakopoulou, A. (2008). Text classiﬁcation aided by clus-

tering: a literature review. In Tools in Artiﬁcial Intel-

ligence, chapter 14. Paula Fritzsche, intech edition.

Lavrac, N., Kavsek, B., Flach, P., and Todorovski, L.

(2004). Subgroup discovery with CN2-SD. J. Mach.

Learn. Res., 5:153–188.

Leva, A. D., Berchi, R., Pescarmona, G., and Sonnessa,

M. (2005). Analysis and prototyping of biological

systems: the abstract biological process model. In-

ternational Journal of Information and Technology,

3(4):216–224.

MedDRA (2007). Meddra maintenance and support ser-

vices organization. introductory guide, meddra ver-

sion 10.1.

Messai, N., Devignes, M.-D., Napoli, A., and Sma

ıl-

Tabbone, M. (2008). Many-valued concept lattices

for conceptual clustering and information retrieval. In

ECAI, pages 127–131.

Pakhomov, S. S., Hemingway, H., Weston, S. A., Jacob-

sen, S. J., Rodeheffer, R., and Roger, V. L. (2007).

Epidemiology of angina pectoris: Role of natural lan-

guage processing of the medical record. American

Heart Journal, 153(4):666–673.

Slonim, N. and Tishby, N. (2000). Document clustering

using word clusters via the information bottleneck

method. In Proceedings of the 23rd annual interna-

tional ACM SIGIR conference on Research and de-

velopment in information retrieval, pages 208–215,

Athens, Greece. ACM.

Szathmary, L., Napoli, A., and Kuznetsov, S. O. (2007).

Zart: A multifunctional itemset mining algorithm. In

CLA, pages 26–37.

Ward, J. H. (1963). Hierarchical grouping to optimize

an objective function. Journal of the American Sta-

tistical Association, 58(301):236–244. ArticleType:

research-article / Full publication date: Mar., 1963 /

Witten, I. H., Frank, E., and Hall, M. A. (2011). Data

Mining: Practical Machine Learning Tools and Tech-

niques. Morgan Kaufmann, Burlington, MA, 3 edi-

tion.

KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval

276