AM-2004 - Osaka University

2 downloads 0 Views 4MB Size Report
Jun 1, 2004 - 13:00-13:30 A Novel Hybrid Approach for Interestingness Analysis of ..... P. Berka: ECML/PKDD 2002 Discovery Challenge, Download Data about Hepati- ... http://www.slab.dnj.ynu.ac.jp/erratumicml2003.pdf) (2003). 6. ...... Case3 = {event 1, event 5, …, event m3}. ...... http://ist.ksc.kwansei.ac.jp/~kitamura/.
Proceedings of the Third International Workshop on Active Mining (AM-2004) in Conjunction with the 18th Annual Conference of the Japanese Society for Artificial Intelligence, 2004 Organizers Hiroshi Motoda Takahira Yamaguchi Shusaku Tsumoto Masayuki Numao

Ishikawa Kousei Nenkin Kaikan, Kanazawa, JAPAN June 1, 2004

ORGANIZERS Hiroshi Motoda

Osaka University, Japan

Takahira Yamaguchi

Keio University, Japan

Shusaku Tsumoto

Shimane University, Japan

Masayuki Numao

Osaka University, Japan

PROGRAM COMMITTEE Hiroki Arimura

Hokkaido University, Japan

Stephen D. Bay

Stanford University, U.S.A.

Hendrik Blockeel

Katholieke Universiteit Leuven, Belgium

Robert H.P. Engels

CognIT, Norway

Shoji Hirano

Shimane University, Japan

Tu Bao Ho

JAIST, Japan

Akihiro Inokuchi

IBM Japan, Japan

Hiroyuki Kawano

Kyoto University, Japan

Boonserm Kijsirikul

Chulalongkorn University, Thailand

Ross D. King

The University of Wales, Aberystwyth

Yasuhiko Kitamura

Kwansei Gakuin University, Japan

Ravi Kothari

IBM - India Research Lab, India

Marzena Kryszkiewicz

Warsaw University of Technology, Poland

Tsau Y. Lin

San Jose State University, U.S.A.

Huan Liu

Arizona State University, U.S.A.

Masayuki Numao

Osaka University, Japan

Miho Ohsaki

Doshisha University, Japan

Takashi Okada

Kwansei Gakuin University, Japan

Takashi Onoda

CRIEPI, Japan

Yukio Ohsawa

University of Tsukuba, Japan

Luc de Raedt

University of Freiburg, Germany

Henryk Rybinski

Warsaw University of Technology, Poland

Masashi Shimbo

NAIST, Japan

Einoshin Suzuki

Yokohama National University, Japan

Yoshimasa Takahashi

Toyohashi University of Technology, Japan

Masahiro Terabe

MRI, Japan

Ljupico Todorovski

Jozef Stefan Institute, Slovenia

Shusaku Tsumoto

Shimane University, Japan

Stefan Wrobel

Fraunhofer AIS and University Bonn, Germany

Seiji Yamada

NII, Japan

Takahira Yamaguchi

Keio University, Japan

Yiyu Yao

University of Regina, Canada

Kenichi Yoshida

University of Tsukuba, Japan

Tetsuya Yoshida

Hokkaido University, Japan

SPONSORS Japanese Society for Artificial Intelligence Active Mining Project (Grant-in-Aid for Scientific Research on Priority Areas, No.759)

Program and Table of Contents

June 1st (Tuesday) 9:30-10:00

Registration

10:00-10:10

Opening Session 1: Mining Medical Data

10:10-10:40

Spiral Discovery of a Separate Prediction Model from Chronic Hepatitis Data Masatoshi Jumi, Einoshin Suzuki, Muneaki Ohshima, Ning Zhong, Hideto Yokoi, Katsuhiko Takabayashi ................................................... 1

10:40-11:10

Evaluation of Rule Interestingness Measures with a Clinical Dataset on Hepatitis Miho Ohsaki, Shinya Kitaguchi, Kazuya Okamoto, Hideto Yokoi, Takahira Yamaguchi..............................................................................11

11:10-11:40

Process to Discovering Iron Decrease as Chance to Use Interferon to Hepatitis B Yukio Ohsawa, Hajime Fujie, Akio Saiura, Naoaki Okazaki, Naohiro Matsumura ............................................................................. 21

11:40-12:10

Preliminary Analysis of Interferon Therapy by Graph-Based Induction Tetsuya Yoshida, Warodom Geamsakul, Akira Mogi, Kouzou Ohara, Hiroshi Motoda,Takashi Washio, Hideto Yokoi, Katsuhiko Takabayashi ............................................................................................................... 31

12:10-13:00

Lunch Session 2: Mining and Information Gathering

13:00-13:30

A Novel Hybrid Approach for Interestingness Analysis of Classification Rules Tolga Aydin and Halil Altay Guvenir ................................................... 41

13:30-14:00

Preliminary Evaluations of Discovered Rule Filtering Methods Yasuhiko Kitamura, Akira Iida, and Keunsik Park............................ 53

14:00-14:30

Proposal of Relevance Feedback based on Interactive Keyword Map Yasufumi Takama and Tomoki Kajinami ............................................ 63

14:30-14:50

Coffee break Session 3: Mining Chemical Compound Structure

14:50-15:20

A Correlation-Based Approach to Attribute Selection in Chemical Graph Mining Takashi Okada ...................................................................................... 73

15:20-15:50

Combining

Partial

Rules

and

Winnow

Algorithm:

Results

on

Classification of Dopamine Antagonist Molecules Sukree Sinthupinyo, Cholwich Nattee, Masayuki Numao, Takashi Okada, Boonserm Kijsirikul................................................... 83 15:50-16:20

Identification of Activity Classes of Drugs under Existing Noise Compounds by ANN and SVM Yoshimasa Takahashi, Satoshi Fujishima, Katsumi Nishikoori, Hiroaki Kato, Takashi Okada .............................................................. 93

16:20-16:50

Mining of Three-Dimensional Structural Fragments in Drug Molecules Hiroaki Kato, Takashi Koshika, Yoshimasa Takahashi, Hidetsugu Abe..................................................................................... 102

16:50-17:00

Closing

Spiral Discovery of a Separate Prediction Model from Chronic Hepatitis Data Masatoshi Jumi1 , Einoshin Suzuki1 , Muneaki Ohshima2 , Ning Zhong2 , Hideto Yokoi3 , and Katsuhiko Takabayashi3 1. Electrical and Computer Engineering, Yokohama National University, Japan 2. Faculty of Engineering, Maebashi Institute of Technology, Japan 3. Division of Medical Informatics, Chiba University Hospital, Japan [email protected], [email protected], [email protected], [email protected], [email protected], [email protected]

Abstract. In this paper, we summarize our endeavor for spiral discovery of a separate prediction model from chronic hepatitis data. We have initially proposed various learning/discovery methods including time-series decision tree, PrototypeLines, and peculiarity-oriented mining method for mining the data. This experience has led us to believe that physicians tend to consider typical cases with the specific disease and rule out from the clearly exceptional cases. We have developed a spiral discovery system which learns a prediction model for each type of cases, and obtained promising results from experiments.

1

Introduction

Medical data are challenging to data miners since they show problems related to data quantity, data quality, and data form [5]. Chronic hepatitis data [1], which show high diversities in various aspects of cases, satisfy this nature and necessitate us to develop novel data mining methods. For the chronic hepatitis data, providers presented several objectives including progress of chronic hepatitis (difference between type B and type C) and effect of interferon therapy. Among them, LC (liver cirrhosis) prediction from blood test data can be regarded as one of the most important objectives since it can potentially substitute routine tests for a biopsy, which represents a highly invasive test. We have been working on knowledge discovery from the chronic hepatitis data including the LC prediction. Our methods include a decision tree learner for time-series classification problem [5], a visualization method for medical test data [4], and a peculiarity-oriented mining method [6]. The methods have proved to be effective for various purposes including detection of exceptional cases. From the endeavor, we have come to believe that physicians tend to start their differential diagnoses by distinguishing typical cases from apparently exceptional cases. In this paper, we propose a discovery method which distinguishes the two types of cases in its separate prediction model based on the abovementioned data mining methods.

1

2

Knowledge Discovery from the Chronic Hepatitis Data

2.1

Liver Cirrhosis Prediction

Chronic hepatitis represents a disease in which liver cells become inflamed and harmed by virus infection. In case the inflammation lasts a long period, the disease comes to an end which is called a liver cirrhosis (LC). During the process to an LC, the degree of fibrosis, which consists of five stages ranging from F0 (no fibrosis) to F4 (LC), represents an index of the progress. The degree of fibrosis can be inspected by biopsy which picks liver tissue by inserting an instrument directly into liver. A biopsy, however, cannot be frequently performed since it requires a short-term admission to a hospital and involves danger such as hemorrhage. Therefore, if we can predict the degree of fibrosis with a conventional medical test such as a blood test, it would be highly beneficial in medicine. A time sequence A represents a list of values α1 , α2 , · · · , αI sorted in chronological order. A data set D consists of n examples e1 , e2 , · · · , en , and each example ei is described by m attributes a1 , a2 , · · · , am and a class attribute c. We assume that an attribute aj represents a time-series attribute which takes a time sequence as its value1 . The class attribute c represents a nominal attribute and its value is called a class. We show an example of a data set which consists of time-series attributes in Figure 1. The data set consists of examples 84, 85, 930, each of which is described with time-series attributes GPT, ALB, PLT, and a class. time-series attributes

GPT

ALB

200

examples 84

PLT

6

300

5 100

200 4

0 -500 -250

0

250 500

200

0

250

500

6

250 500

200

0

250 500

non-LC

100

3 -500 -250

0

250

500

6

0 -500 -250

0

250 500

400 300

5 100

200 4

0 -500 -250

0

300

4

200

930

0 -500 -250 400

100

0 -500 -250

non-LC

100

3 -500 -250

5

85

class

400

0

250 500

LC

100

3 -500 -250

0

250

500

0 -500 -250

0

250 500

Fig. 1. Data set which consists of time-series attributes

In classification from time-series data, the objective represents induction of a classifier, which predicts the class of an example e, given a training data set D. 1

Though it is straightforward to include other kinds of attributes in the definition, we exclude them for clarity.

2

In this paper, we assume liver cirrhosis (LC) and non-liver cirrhosis (non-LC) as classes. 2.2

Our Endeavor for the Chronic Hepatitis Data

Our time-series decision tree was proposed since physicians requested to use time sequences which exist in data in a classifier [5]. We applied the method to the LC prediction problem and the results attracted their interests. The physicians commented that the obtained decision tree is highly valid although we used medical knowledge only for selecting medical tests in the data [5]. Moreover, most of the cases who are mispredicted by the decision tree were recognized as exceptions by the physicians. Our PrototypeLines represents a visualization method which is considered to enable discovery of interesting knowledge without extensive training [4]. Application of PrototypeLines to the chronic hepatitis data revealed interesting characteristics. A student in computer science discovered at least two exceptions which were both recognized as missing diseases in the original data by a domain expert. Moreover, a physician who was introduced PrototypeLines for the first time discovered an exception condition of a case after a 5-minute explanation. Our peculiarity-oriented mining method obtains tendencies among examples with peculiar data [6]. The peculiar data are detected based on a distance measure, and the number of the peculiar data for an example is called the number of peculiar attributes. The method has successfully discovered exceptional cases in terms of interferon therapy.

3 3.1

Discovery of a Separate Prediction Model Overall Architecture

Based on the motivation presented in Section 1, we propose a discovery method which learns a separate prediction model. Our method employs a naive Bayes learner and a 1-NN (1-nearest neighbor) classification method for typical cases and exceptional cases respectively, and refines the separate prediction model by a spiral interaction with a medical expert. The term “spiral” is employed since it represents iterative refinement of discovered knowledge. In the interaction, exceptional cases are identified with the data mining methods mentioned in Section 2.2 and a likelihood-based method to be proposed in the next Section. Figure 2 shows the architecture of our method. In our method, detection of exceptional cases with our previous methods as well as update of a naive Bayes learner for typical cases are iterated in a spiral manner with interaction with an expert. A naive Bayes classifier predicts a class cˆN Bayes,i of an example ei assuming that each attribute aj is independent. Here vij represents the value for attribute aj of p. cˆN Bayes,i = argmaxc Pr(c)

m  j=1

3

Pr(aj = vij |c)

(1)

As the result, we obtain a list of typical cases, conditional probabilities for typical cases, and a list of exceptional cases. Classification of a new case begins by detection of exceptional cases based on a 1-NN method, which will be presented in Section 3.3. If a case who is close to the new case is found, our method outputs the class of the former case. Otherwise, the class of the new case is predicted with a naive Bayes classifier.

Prototypelines

peculiarity oriented mining method

interaction

expert

time-series decision tree

detection of exceptional cases

spiral iteration

odds-based judgement

NB learner learning list of typical cases

conditional probabilities NB classifier

prediction 2

list of exceptional cases

prediction 1 1-NN for time-series classification

list of novel cases

Fig. 2. Architecture of the proposed method

3.2

Likelihood-based Judgment for Exception Degrees

In a classification problem, an example can be intuitively regarded as “typical” or “atypical”. The typicalness Φ(p), which is based on the right-hand side of Eq. (1), of a case p represents a degree to which p belongs to its class c compared with the other class c. Here vij represents the value for attribute aj of p.  Pr(aj = vij | c) Pr(c) m j=1 (2) Φ(p) = m Pr(c) j=1 Pr(aj = vij | c) The larger Φ(p) is, the more certain that p belongs to its class c. If a naive Bayes classifier is relatively accurate for the class c, the typicalness Φ(p) tends to be large and vice versa. Thus the typicalness Φ(p) is relative since it depends on the preciseness of a naive Bayes classifier in terms of the class c. The degree how a naive Bayes classifier is precise for c can be measured by the degree of correct classification and the degree of incorrect classification. For c and c, we represent the number of correctly-predicted examples by the naive Bayes method L and n respectively. Likewise, for c and c, we represent the number of incorrectly-predicted examples by the naive Bayes method Le and

4

ne respectively. The preciseness Ψ (ˆ c) of a naive Bayes classifier of the estimated class cˆ represents the ratio of the precision for c and the precision for c. Ψ (ˆ c) =

(L + 1)(n + ne + 2) (L + Le + 2)(ne + 1)

(3)

For our LC prediction problem, we define a degree of exception E(p, cˆ) for a case p and his/her estimated class cˆ as follows in order to discriminate exceptional cases from typical cases. E(p, cˆ) = − logΨ (ˆc) Φ(p)

(4)

Intuitively, E(p, cˆ) represents an evaluation index which is equal to the number of upvaluated digits below the decimal point when we measure the typicalness Φ(p) of a case p in terms of the preciseness Ψ (ˆ c) of a naive Bayes classifier of the estimated class cˆ. When prediction of the naive Bayes method is accurate, Ψ (ˆ c) tends to be large, and the absolute value of E(p, cˆ) is relatively small even if the absolute value of Φ(p) is large. This fits our intuition that certain information rarely leads to an extreme degree of exception. In each loop, cases whose degrees of exception are no less than a user-specified threshold are detected as exceptional cases. The loop continues until prediction accuracy estimated with 10-fold cross validation either decreases or reaches 100 %. 3.3

Similarity between a Pair of Cases Based on Dynamic Time Warping

Dynamic time warping (DTW) represents an index of dissimilarity between a pair of time sequences [3]. Unlike Euclidean distance, DTW can allow distortion along the time axis since a point in a time sequence can correspond to multiple points in the other sequence. A 1-NN method based on DTW constantly showed high accuracy in our experiments for time-series classification including the LC prediction problem [5]. Due to this good result, we have chosen this method for detecting exceptional cases in test data. Following [5], the window width [3] of DTW was settled to 10 % of the length of the time sequence, and a dissimilarity measure H(ei , ej ) between a pair of examples ei , ej has been employed. This measure normalizes the results G(ei (ak ), ej (ak )) of DTW for ei and ej in terms of a time-series attribute ak with the maximum value q(ai ), where q(ai ) ≥ ∀j∀kG(e(aj ), e(ak )). H(ei , ej ) =

m  G(ei (ak ), ej (ak ))

q(aj )

k=1

4 4.1

(5)

Experiments Application of Conventional Methods

In the experiments, we used data from 180 days before the first biopsy to the day of the first biopsy following advice of medical experts. As the result, the num-

5

bers of LC and non-LC cases are 159 and 112 respectively. The blood tests that we use are GOT (glutamic oxaloacetic transaminase = AST (asparate aminotransferase)), GPT (glutamic pyruvic transaminase = ALT (alanine aminotransferase)), TTT (thymol turbidity test), ZTT (Zinc sulfate turbidity test), D-BIL (direct bilirubin), I-BIL (indirect bilirubin), T-BIL (total bilirubin), ALB (albumin), CHE (cholineesterase), TP (total protein), T-CHO (total cholesterol), WBC (white blood cell), PLT (platelet), and HGB (hemoglobin). In our experiment, we used 46 LC cases and 55 non-LC cases each of whom has test values for all of these blood tests. For the classifier, we first averaged each time sequence then discretized each value as we show in Table 1, where we use for each attribute value U: extremely high, V: very high, H: high, N: normal, L: low, v: very low, and u: extremely low. Each conditional probability is estimated using Laplace correction in order to cope with the 0-occurrence problem [2]. A missing value is ignored both in estimating probabilities and in classifying an example.

Table 1. Blood tests in the chronic hepatitis data

attribute GOT GPT TTT ZTT D-BIL I-BIL T-BIL ALB CHE TP T-CHO WBC PLT HGB

intuitive explanation amount of broken liver cells degree of immune activity disorder of bile excretion decrease of protein generation

good order of a liver hemoglobin

discretization N≤ 40< H≤ 100< V≤ 200< U N≤ 40< H≤ 100< V≤ 200< U N≤ 5 < H≤ 10< V≤ 15< U N≤ 12< H≤ 24< V≤ 36< U N≤ 0.3< H≤ 0.6< V≤ 0.9< U N≤ 0.9< H≤ 1.8< V≤ 2.7< U N≤ 1.2< H≤ 2.4< V≤ 3.6< U v≤ 3.0< L≤ 3.9< N≤ 5.1< H≤ 6.0 1     $            8$ ,   FE G EIHJE KL>@MON P7Q1R@?7G SUT VXWZY[\]=^ _Y`^UabIcdqar_sFq[queFt^ Y`F]Ilf!\`F]=g\ih `Fv!YeY]w]=lh `Ygp ij7h k lY`^Um;`h nY]=o@h ^ pi xXy(z({}|~(y€~I‚„ƒ…d†(‡;ˆ‰Š!‹ZŒL7†(|(Žƒ…}zuŒƒIIŒz}‡ Z‘›= !’1©œ“@”=¨•° –›—£@”=£@˜šŸ §%™œŸ ¨›=›=@@›dŸ ª¬ž œŸ­= !³!Ÿ  ° ¥¡¢£;Ÿ›=£(­¥d ¤!¯!¥¦¥=­=¥=¤!§w§%›´Ÿ¯!¨£(Ÿ ¥= ¤!;¥7©žŸ £@ª¨ª¬£ ;«(¥=Ÿ ž¦­=®D¯!ª¬ª¬­=§„@¯!›= ›=( @ ¥=®µ­= !¯!£;¥¦ŸªL O§„°¯!›=­w›´¡( @¥¦¥=­=© !›=£;@›=Ÿ  O±!›©£@›=¥@£@› ² ž¡(¥=Ÿ ! !¥=Ÿ  ­›¡ ° ®J›=¯«(¯!¥=­´° Ÿ®J¨›=±!@ŸŸ ¡Jª@ ! ¤!£@¶¦›= ·œ£@ª¬¥¸ ° ¥¨=¥=@«(Ÿ ¥=ª¬­ ¢²ª¬ §C¤!¥OŸ  @ ¥=³­¥ž¦£ @±!Ÿ  ¥=­¹¡Dª¬ª¬ !§º¥£ £ ³!›´¨=ž¤»ª¬ ¨¡D° ›£@›£@° IŸ §%©Ÿ ¨Ÿ ›=£¬@¨Ÿ ª¬ª¬«( ¥=­­w¥³!©;° ¥­=³!£'° ¥Ÿ ££ ±!ª¬§ ¥¨›(ª¬­=ž³!°¥¥;£u›=›=­ ¼¥;ŸŸ  ž¦«(¯!¥ª¬£ ­=@@Ÿ ¡(›= ›=@}¥@©›£ ›=½( !¶F©¾= £@ª¬ž¤(Ÿ¥(£} !¯!¥=›=¸¿¯!¥=§%­›²F¨=§%@›ª¬­¨=@£dª¬›=­­£}¥(­¯¥­° ›=ª¬@¯!¥ª©u£@@¥ªu©F¶L Àd¤!ª¥° Ÿ  Fª@@¸ ¥=­Ÿ  ¥¡¦£ @Ÿ  ¤!¡FŸ  !£@²#¥›=£@  £ Ÿ›= ³@@¥=ª¬­ž›¨=›=@@Ÿ «(Ÿ ¨¥+›° ­=®É³!° ¥° ›=±!Ÿ  ¥@°¥=­¥ ¤!£ @¥ÊŸ  ¡F !¨° ¥›£@£@£@£@Á=Ÿ °§%¥Ÿ ›=¨­=›= !@Ÿ Ÿ  ª¬ Ë¡Â›­=³!° ¡(° ª¬¥­£ÌŸ  ¤¥žÄŸ  ¤!¥=à ¾=­ÊÅ7¾=Æ(›£ÎÇ͚Ÿ £Ÿ  È©¥=¥=­«(¥£¥@° Ÿª¬ ¯!¡F¥ÏЩÈª¬ª­ ÍÀ7³¥ !›= Ÿ ³ F@­¥=¥Ö­¥Õ£ @­%Ÿ ª ×F¡F¥Ï¿¨=@Ÿ¸ ª@ !Ÿ  £¤ÑDzذ Ÿ ›ÖžŸ§%@¥¥›=©Ò ³³!­¥Ö£@¥=­Ó¯­¯!ª ×F›=­=¥@¨=Ÿ @¨Ÿ Ÿ ª@¯! Ù›=@Ÿ ±!ª@ !›£@¶ ¥¾w©Ù ÑŸ  !ª@³¨=­­¿¥=ž£  ³(¥= ©¬@®I›² °OÔC¨Àš° ›Õ £@£@Ÿ §%à ԜŸ ¨ª¬›=@@ŸŸ  ª¬¡  °©¥¥›=£@­=¨= !­Ÿ Ÿ ¯¡¿@Ÿ ª@› Ö° ¡(°ª¬¥­›=Ÿ ­= ¤ !ž¥©u²„±Ÿ £®Û› ¤!° £@¥ ªØÔC©Àš¥=«(Õ+¥° ›ª¬° ¯!¡(¥ª¬­©ØŸ  ¤Ÿ  Øž» ¤!¨¥Zª¬ !§w­£ ›´@žÓŸ  ³¥´@¸ ¥£uª¬­=› ½¿ (ª¬ª¬§;«(¾=¥Å7°U¾´¤Æ;®F±¶š­Ú(Ÿ ¤!©¥œ›w¯¨¯ª¬­% !ª¨›¥=¨=¯¤  §%ª¬­ Ÿ  @¥=­¥£ @Ÿ  ¡F !¥£@£„›= !›° ®I£@Ÿ £„ªL§¨° ›£@£@Ÿ §%Ÿ ¨›=@Ÿ ª¬ ¦­=³(° ¥£@¶ ÜÞÝ#ßuàáIâ7ãuäuåIàæÈâ7ß çDó7è@è@éLéLè@èû7êèFîwð(ë ì1îýë ì7ü íÛþ7ÿ ëLî;Zé ï1è@ðÓé é ðFðFöÈñÈì7ñÈë î¢òFë ëðFìÑì1é é óšï1ëLð îwòFó7ô1è@õ1éLðFè¿öÈ÷Òò(ô1è@ìÑñuø7û1è@ð é é ðFöÈöÈðFì7øšî%öÈù7ð(îwèFðFî;ì1ô1é ðFø1óÑø1ô7ë îìÑðFóÒêÖé ôÒè@ì1ó7÷Ñè@éLó1èë ñÈë éLñÈðFîwðFöÈúðFñµì1ù1éDë ìÒñ ô1úLöÈè@êÖöµíîLð ù ë îwì1ðFòFöÈëú ð(1î%ó1ùë òFì7ô1íýì1é òFë úLì7èFíî%ðFîwì1ë ñÈòFë ò(÷Ûè@éLé ëè@ô1û1ì ú ð(î%ö ù11è@ú ð(ìšóœî%ùô1èFé ï1î%îwðFô1öµòFî;ëLü è@7é ë ÿ ô1ìJô ö J1ú ðFð(õ1î%ðFùuöµòFùúé 7ï1ð¦îwé ðFì öµ1î%êùû1îwððFöd1ô1ðFñ„ì1ó1é ëLëLîwè@òFúÓô1ø7õ1ðFè@é öÈé ðFðFóœöÈì7ø7î%è@ùué é é ðFë êöÈì7ð î ëLî î1û77îwî ðF7éCè@ô1ú ú ñJ÷ýé ï1õ1ðFðFêöÈ÷ýû17ëLï1íýðFöÈèÈðFì1ñÈóýôšöÈéð(ï1ù;ð îwðµ7ú îwðFðFòFöÒé ë ô1è@ì7ì¿è@ô1ú ÷ ñJFëë ì1ì7é í¢ðFöÈé ð(ï1îwðÒé ë ì7ø7íè@é ø7é ðFè@öÈé ì7é ðFîöÈì7ë¬îØîCíëLîðFì1è@ðFì¿öµè@ë ú êú ÷ýø1ô1ë ì1öÈéLé ðFè@ì1öÈð(éCîwöÈé ðFð(óýîwð(ëè@ì öÈòµï è é ô1ø1ë Èò ìé ï1ëLî!ø7è@ø1ðFöµù JðòFô1ì1òFðFì1é öµè@é ðô1ìé ï1ðø7è@é é ðFöÈì7î!öÈðFø1öÈð(îwðFì1é ðFóûš÷Ûé ï1ðòFúLèFî%îwë ñÈë ò(è¬é ë ô1ìö 1ú ð(î è@è ì11óÐé ô1êÖó1è@ðFé õ1ë ò(ðFè@ú úô1ú øÞ è@úLìÐèFî%îwë ë ñÈì1÷¹é ðµé öµï1è@ð(òFéîwëð¿õ1ð ö 1ú öð(1î¢ú ð èFî ë ë ì1ì1é é ðFðFöÈöÈð(ð(îwîwé é ë ë ì7ì7ííÒì1ð(ô1î%ö î 1ú ð(ì1è@ë ì1öÈì1é ðFë ìdö ð(íÐîwé ë ì7è@úLí1íù ô1JöÈë é ë ï1é ï¹ê ú ë êë é ðFó 7îwðFé ô ö ¹ ÷ F ò ø7û7è@èFöÈîwéðFë óÌòFë ø7ë è@ì1é òFë ô1öÈðFì êðFÈìºì1éLôè@úº1ö¼òFúLîwèFé 1î%ó1îwë ÷7ñÈë ù ò(è@é ë ô1ì ú ð(Jè@ô1öÈì1é ë ë ì7ìdí íÌ!è@ð(úLè@íé ô11öÈöÈë ð é ï1!ê¼öÈôù LðFDòFéèFë îÑô1ì7è@î úLµîwùôÌèÓó1ñÈðFð(õ1è@éðF1ú ô1öÈðÖø1ðFø1óÌöÈô Lë ðFìÌòFé ëé ô1ï1ìð ñÈñµöµè@è@òFêé ô1ð öµJîuô1èFî„ö ÛñÈð(ô1è@ñ é 1öÈð(î CðFë ì7íîwø1ðFòFë ñÈë òé ôô 1ö7òµô1ì1òFðFö ì7î%ù DéLè 1ð(î!é ï1ðö 1ú ðë ì1é ðFöÈð(îwé ë ì7íì1ð(î%î 







 

 

































"







"#%$&' #

$





,



 *-

 



(



*)



!



*)

+



#%$&

,



./ 021 354 3607982: 6.;/27 < 6=;4 : 6>1 /796;?63 791 / @*/"6?33BA8*/ A6C7./ 07982A: .331 D E

7 < 6F/ 6G%: EH: 6.;/ 60IA: .331 D1 A.791 8*/I;4 : 63JLKNM"OQPR*S"POTUWVOXPY?Z TU9Z R*S\[ O]YS"OV_^a`(U M"OFb>c&d

][ efR*Y?Z U MghPRS"X UZ U iU9OX(]S"R*j"Ok[lM`f^Y?Z Vm]TTY?R]PM'noR*YpZ SfU9OY?OkX U9Z S e*S"OX X(]qS"][ `"XZ X(RnFU M"O P[ ]XXZ noZ P]U9Z RSFYi"[ OkXr s&tuv9w x*y{z|t}u~?w  t}v € t_w yv9t~?t?} v9w y *y t}}‚w }} ƒ t+x„… †v v9t~y }‡ˆs&tuv9w x*y{‰w }‚|t?Šx*v9t|vqx v € tF‹y xŒ% t|tF~?t…~?t}tyv9†v9w x*y(ƒ }t|Iw yHx*ƒ~Ž} v ƒ |* ‡Ls&tuv9w x*yI\†y |I‘I†~?tF~?t †v9t|_v9x_v €"t

41

’ “?”• – • – —˜”– ™˜š› ”œœ•  ž • – —!Ÿ  ”?œ¡œ£¢’   ¡\¤ˆ¥§¦¨”› —¢*“?• ’  ©ª2« ¬‚« ­®• œ£¡¯*Ÿ › ”• – ¡™˜• –˜’   ¡ ¢› › ¢°%• – —‚œ¡š’9• ¢*– ª±%• ²• – —’   ¡l¡¯*Ÿ ¡“?• ©¡–’9”›?“?¡œ ³ › ’9œ=• –´&¡š’9• ¢ –µ¶*Ÿ ”Ÿ ¡“l• œ=š¢*–"š› ³ ™¡™fª

·¹¸»ºB¼ ½N¾N½N¿»¼ ÀºBÁ ºB½N¿»¿W¸»¿»¿»ÂB½ŽÃ ÄÅ>Æ ¼ ¼ ½¾NºB¿

ÇaÈ ÉWÊ ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎ Ê ÎÎ Ð ÉBÈ ÑÎÒ ÉÉË£ÑË£Ê ÓWÔ Õ*ÍÌ9ÑËÌ"ÔÍ?Õ Ò Ö ÉÓ×É?ØÉ͂ÎÊ Ë ÙÉBÌ È ÉBÒ É?ÏÊ ËË Ê Ë"ώÕÚ Û ÑÌ9ÑÓÊ Ë Ê Ë ÏBÍ?ÉÎÉÑÍ?ÙÈÝÜÞ*ßàaÇÈ ÉÍÉ ÑÍ?ÉÓÑËáÝÚÑÙÌ9Õ Í?ÎLÙÕ*ËÌ Í?Ê ÒÐÌ9Ê Ë"ÏBÌ9ÕBÌ È É Ê ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎLÕÚ Û Ê ÎÙÕØÉÍ?É Û Ñ

Ô ÑÌ Ì9ÉÍËIÜ?Þ*â&ãâ&äfßàå&Õ*ÓÉFÕÚÌ È ÉÓæÑÍ?ÉFÙÕØÉÍ?ÑÏÉâÙÕ*Ë ÚÊ Û ÉË ÙÉ?â&ÙÕ*ÓÔ Ö ÉÌ9ÉË ÉÎÎâ

ÑÙÌ9Ê Õ*Ë£ÑÒ Ê Ö Ê Ì á£ÑË Û

ÐË Éç Ô ÉÙÌ9É Û Ë ÉÎÎà5ÇaÈ ÉWÚÊ Í?Î Ì Ì ÈÍ?ÉÉWÚÑÙÌ9Õ*Í?Î ÑÍÉWÕ*Òè9ÉÙÌ9Ê ØÉâ5ÑÙÌ9Ê Õ*Ë£ÑÒ Ê Ö Ê Ì á

Ê ÎÎ ÐÒè9ÉÙÌ9Ê ØÉQÑË Û

ÐfË Éç*Ô"ÉÙÌ9É Û Ë ÉÎÎÊ ÎÎÕ*ÓÉÌ9Ê ÓÉÎlÍ?ÉÏÑÍ Û É Û

Ñ?ÎÎ ÐÒè9ÉÙÌ9Ê ØÉQÜéâ=êâ=ëß>ÑË Û

ÎÕ*ÓÉÌ9Ê ÓÉÎÑÎÕ*Òè9ÉÙÌ9Ê ØÉQÜ?Þ*ìâ%ÞÞ*ßàLíˆÒè9ÉÙÌ9Ê ØÉŽÊ ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎÚÑÙÌ9Õ*Í?ÎÙÑË(Ò ÉFÓÉÑÎ ÐÍ?É Û Ê Ë Û ÉÔ ÉË Û ÉËÌ9Ö áîÕfÚïÌ È ÉhÐ ÎÉÍðÑË ÛñÛ Õ*ÓÑÊ ËóòË Õô%Ö É Û ÏÉ?àæõ+Õô%ÉØÉÍ?âöÎ ÐÒè9ÉÙÌ9Ê ØÉ Ê ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎ÷ÚÑÙÌ9Õ*Í?Î÷ÑÍ?ÉøË Õ*Ì{Ð ÎÉÍùÑË Û×Û Õ*ÓÑÊ Ë¨òË Õfô%Ö É Û ÏÉúÊ Ë Û ÉÔ ÉË Û ÉËÌ9àpÇaÈ"É ÓÉÑÎ ÐÍ?ÉÓÉËÌLÕÚÑ%Î ÐÒè9ÉÙÌ9Ê ØÉ%Ê ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎÚÑÙÌ9Õ*ÍÓÑáØÑÍoáÑÓÕ*Ë ÏBÐ ÎÉÍ?ÎÑË ÑÖ á ûÊ Ë ÏWÑ Ô ÑÍÌ9Ê ÙÐ Ö ÑÍ

Û Õ*ÓÑÊ Ë â>ÓÑáüØÑÍoáæÑÓÕ*Ë Ï

ÑË ÑÖ á ûÊ Ë ÏøÑË Û

Û Ê ÚÚÉÍ?ÉËÌ

Û Õ*ÓÑÊ Ë ÎÌ È ÑÌ!Ñ£Ô ÑÍÌ9Ê ÙÐ Ö ÑÍýÐ ÎÉÍmÊ Î

ÓÑáùØÑÍáùÉØfÉËùÚÕ*ÍÝÌ È É{ÎÑÓÉÐ ÎÉÍþÑË ÑÖ á"ûÊ Ë ÏpÌ È É{ÎÑÓÉ

Û Õ*ÓÑÊ ËùÑÌ

Û Ê ÚÚÉÍ?ÉËÌNÌ9Ê ÓÉÎà ÿ Ë(Õ*Òè9ÉÙÌ9Ê ØÉ>Ê ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎ5ÓÉÑÎ ÐÍ?É>Ê ÎBÙÕ*Ë Î Ì ÍÐ ÙÌ9É Û

Òá(ÙÕ ÓÒ Ê Ë Ê Ë Ï2Ñ=ÔfÍ?Õ*Ô ÉÍ%Î ÐfÒ ÎÉÌ

  

          !" #$ %&' %!  #(%! )   *+   #$ ,  -!  .&' %!#0/21343!51768 9:6 9;3!5 18 ?=@8 91= A$=B,18 9 C!9 =BBED:=6B,FA$=@3G1 H =2G'3!AIDKJLNM'O P QSR T4UV T WYX WT T R Z V [\QWY] T [^U _W!Z`[a QR b[^R _Q[ c$[T,QR _ d!_ [T T4e'Ua QW!c@fhgi j!k l^gm4g kjonplq rs tl s krl u$lm,rs k v!k lm mxw:lgm,yu$l{z'l| v+| }.~!!€‚ ƒ „  …„†'‡‰ˆ‹Š Œ Ž ‘$$’4“ ” • •7 ‘'– Œx— ”Š –"˜;™.š!›!œ‚ ž Ÿ ›  ŸY¡ ¢ £¤ ¥ ¦§£¨© ª¬« ­$®2­$®¯« ­$°®°;«± ² ³´® ­$®±,´² ³ ¯ µ0¶@· ´ ¸ ¹!º ¯!¸;´ ¸ ®2º ± ® ­Y°® ´® ­'»:² ³ ®±x´ ¸ ®2´ ¸­$®±,¸ ¹· °¼ ´ ¸ ² ±^² ±½­$®¯« ­$°®°¾«±^±,»:«· ·¿º ± ®7­:² ³´® ­$À® ³´² ¹!³« ³"°Á´ ¸ ®Â² ³´® ­$®$±,´² ³ ¯!³ ®± ±\»:®«±,º­$®Â² ±Ã±,´²,· · «± ±,º»:®°.´¹.Ä ®Å« ³Ã¹!ÄÆ®Ç ´² À®È¹!³"®µ É ¸ ®{®Ê² ±,´² ³ ¯;±,ºÄÆ®Ç ´² À®{² ³´® ­$®±,´² ³ ¯!³ ®± ±x»:®«±,º­$®±S² ³;´ ¸ ®{· ² ´®7­$« ´ º­$®{« ­$®{ǹ!³ ±,´ ­'º Ç ´®° ºË ¹!³½º³ ®ÊoË ®Ç ´®°!³ ®± ±« ³ °S«Ç ´² ¹!³Á« Ä ² · ² ´ ÌÁÍ'«Ç ´¹!­$± µ¶@± ±,º»:² ³ ¯.´ ¸ ®Å°² ± ǹÀ® ­$®°.Ë « ´ ´® ­I³S´¹.Ä"® «\± ® ´.¹ÍN­'º · ®±Î² ³ °!º Ç®°YÍ ­$¹!»Ï«@°¹!»:«² ³ ¼´ ¸ ®º ± ® ­@¯² À®±Ð¸ ® ­¿Ñ³"¹Ò@· ®°¯®\« Ä ¹oº´E´ ¸ ®\°¹!»:«² ³ ² ³´® ­'»:±.¹Í0Í º ÓÓ Ì­'º · ®±.Ô'ÕÖ'¼N¯® ³ ® ­$«·×² »Ë­$®± ± ² ¹!³ ±.Ô'ØÖ0¹!­0­'º · ®¿´® »Ë · « ´®±.Ô'ÙÖ'µ É ¸ ®@² ³ °!º Ç®$° ­'º · ®±Ú« ­$®Û´ ¸ ® ³Üǹ!»Ë « ­$®$°ÝÒ@² ´ ¸Ýº ± ® ­ ±Ú®Ê² ±,´² ³ ¯Ý°¹!»:«² ³Ýѳ ¹Ò@· ®°¯®Û´¹Ý°® ´®'­'»:² ³"® ±,ºÄÆ®Ç ´² À®· ÌSº³ ®ÊoË ®Ç ´®°S«'³ °!Þ¹o­È«Ç ´² ¹!³ « Ä · ® ­Iº · ®± µ ß ¹!´ ¸à´ ÌË ®±¹;² ³´® ­$®±,´² ³ ¯!³ ®± ±á»:®«±,º­$®±á¸ «À®K± ¹!»:®K°!­$«âÒ¿Ä «Ç Ñ ± µ;¶ãË « ­'´² Ç º · « ­ ¹!ÄÆ®Ç ´² À®Á² ³´® ­$®±,´² ³ ¯!³ ®± ±Å»:®«±,º­$®Á² ±Å³ ¹!´±,º Í'Í'² Dz ® ³´ÈÄÌÚ² ´± ®·,ͽÔ'ÕÖ'µ É ¸ ®'ÌÚ« ­$®Á¯® ³ ® ­$«· · Ì º ± ®°h«±{«;Í'² · ´® ­$² ³ ¯^»:®Ç ¸ « ³ ² ±,»äÄ ®Í'¹!­$®;« Ë+Ë · Ì ² ³ ¯h«;±,ºÄÆ®Ç ´² À®4»:®«±,º­$®µÎ忳h´ ¸ ®;¹!´ ¸ ® ­ ¸ « ³ °¼.±,ºÄÆ®Ç ´² À®4»:®«±,º­$®±{«'­æ®;± ¹!»:® ´² »:®±2º ± ®°hÒ@² ´ ¸ ¹!º´ÈË­æ² ¹!­Sº ± «¯®Á¹ÍS«'³Ú¹!ÄÆ®Ç ´² À® ¹!³ ®µ\ç ³´ ¸ ®hÇ«± ®h¹+͋±,ºÄÆ®Ç ´² À®h² ³´® ­$®±,´² ³ ¯!³ ®± ±4»:®«±,º­$®± ¼¿º ± ® ­‹»:« ÌÛ³ ¹!´SÄ ®èÒ@®· ·S² ³ ®Ê!Ë­$®± ± ² ³ ¯.¸ ® ­È°+¹!»:«² ³Sѳ"¹Ò@· ®°¯®Å« ´N´ ¸ ® Ä ®¯² ³+³ ² ³ ¯S¹Í+´ ¸ ®Å² ³´® ­$®±,´² ³ ¯!³ ®± ±« ³ «· Ì ± ² ± µç ´ ° Ä ®\Ä ® ´ ´® ­2´¹Â«'º´¹!»:« ´² Ç«· · Ì· ®«'­'³Â´ ¸ ² ±ÎÑ+³ ¹Ò@· ®°¯®@Ä «± ®°:¹!³;¸ ® ­‹Ç· «± ± ² Í'² Ç« ´² ¹!³¹Í2± ¹!»:® Ë­$®± ® ³´®°\­Iº · ®±2«± ² ³´® ­$®±,´² ³ ¯ ¹!­ º³ ² ³´®'­$®±,´² ³ ¯ µÐ¶¿³"¹!´ ¸ ® ­½°!­$«Ò¿Ä «Ç Ñè¹Íx«4±,ºÄÆ®Ç ´² À+® »:®«±,º­$®h² ±4´ ¸ « ´S´ ¸ ®h² ³ °!º Ç®$°:­'º · ®±;« ­$®hǹo»Ë « ­$®°ÛÒ@² ´ ¸´ ¸ ®è°¹!»:«² ³Ûѳ"¹Ò@· ®°¯®^´ ¸ « ´ «°°!­$®± ± ®±0´ ¸ ®Îº³ ®Ê!Ë ®Ç ´®°!³ ®± ±2« ³ °!Þ¹!­S«Ç ´² ¹!³h« Ä ² · ² ´ Ìh² ± ±,º ®± µÐç ³´® ­$®±,´² ³ ¯!³ ®± ±2² ±2«â± ±,º»:®° ´¹¾°® Ë ® ³ °¾¹!³¾´ ¸ ®± ®Y´Ò@¹¾² ± ±,º0®± µ É ¸ « ´‹² ± ¼2² Í«Y­'º · ®Â²,±^Í'¹!º³ °Á´¹ÁÄ ®Yº³ ®Ê!Ë ®Ç ´®°¼Å² ´‹² ± « º´¹!»:« ´² Ç«· · Ì­$®¯« ­æ°®°«±:« ³é² ³´® ­$®±,´² ³ ¯á­'º · ®µxê{¹Ò@®À® ­$¼½² ´;Ò@¹!º · °áÄ ®^Ä ® ´ ´® ­h² Í^Ò2® · ®« ­'³ ®°Á«Yǹo³ Ç® Ë´@°®± Ç ­$² Ë´² ¹!³Ã´ ¸ « ´@°®«· ´@Ò@² ´ ¸Á´ ¸ ®Y² ³´® ­$®±,´² ³"¯!³ ®± ±\² ± ±,º ®Y°² ­$®Ç ´· ̾« ³ °Á² Í ÕÚ%Ì È É£Õ Òè9ÉÙÌ9Ê ØÉ£Ê ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë"ÉÎÎÝÚÑÙÌ9Õ*Í?ÎÝÊ Ë÷Ñ£Î Ð Ê Ì9ÑÒ Ö É£ô%Ñá à LÕ*ÍýÉçÑÓÔ Ö Éâ>Õ Òè9ÉÙÌ9Ê ØÉ Ê ËÌ9ÉÍ?ÉÎ Ì9Ê Ë Ï*Ë ÉÎÎøÚÑÙÌ9Õ*Í

42

ë@ìSí ì î ìï'ð ñìòèï ó$ôoõöî ì$÷!ø ìù ñì$ò!î ìú úYû î òèûù ñð ô!îÚû í ð ü ð ñ ýÚûúÅñë@ôèôïxñ þ ìÁï'ûù ñô!ó$úÅö ú ìò‹ñô ì÷!øó$ìú úñ þ ìYùô!î ùì øñ@òìú ù ó$ð øñð ô!î ÿ ‰þ û ñ@ð ú ð îñì ó$ìú,ñð î!î ìú ú\ôïÈûÅø û ñ ñì ó'îÁõ:û ý¾òì ø ì î"òÁô!î ï'ûù ñô!ó$úô!ñ þ ì ó"ñ þ û îSö+î ì÷!ø ìù ñì$òoî ìú úû î òSûù ñð ô!îÁû í ð ü ð ñ ýÁð ú ú,ö ìú ÿ ‰þ ìÁð òìûÁôï½ûÁùô!î ùì'øñòìú ù ó$ð øñð ô!îèñ þ û ñð úYû öñô!õ:û ñð ùûü ü ýÚòì ñì ó'õ:ð î ìòèû î òèòð ó$ìù ñü ý ó$ìü û ñìò¬ë@ð ñ þ¬ñ þ ìáð îñì ó$ìú,ñð î !î ì$ú úÃð úoú,ö ì{õ:ô!ñð û ñìò;ö ú½ñô¬òìú ð ! î    Ïûü ô!ó$ð ñ þ+õ: ÿ  

        ! " $#&%'( *),+.-/! (01  243156   758!96!    :      (   ;  !