Professional Documents
Culture Documents
Keynotes:
Reflex Delay and Its Estimation by Analyzing Brain Waves
Goutam Chakraborty
Iwate Prefectural University, Japan
New Paradigm of Mixed Signal LSI based on Time-domain Circuits for Extremely-scaled Technologies
Kunihiro Asada
University of Tokyo, Japan
Research papers:
8
Study on Associative Classification Model
Dong Gyu Lee, In Bae Oh, Keun Ho Ryu
Derivation of Classification Rules from Incremental Decision Tree Committee without Feature
Redundancy
16
YongjunPiao, MinghaoPiao, Ho Sun Shon, Keun Ho Ryu
The energy expenditure monitoring system of trail hiking using smart phones
Mungunbagana Batzorig, Yung-Fu Chen, Hsuan-Hung Lin, John Y. Chiang, Tan-Hsu Tan, Erdenetsogt
56
Davaasambuu, Sukhbaatar Bilgee,
Risk Factors According to gender and age in Patients with Non ST Elevation Myocardial Infarction
from KAMIR
76
Ho Sun Shon, Jong Yun Lee, Soo Ho Park, Keun Ho Ryu
A Study on Utilization of Feature-based Seamless Data Model in Geospatial Information Service Aspect
Hyeongsoo Kim, Hyun Woo Park, Keun Ho Ryu 82
Assessment of the Prediction Capability of Healthcare System Using Carotid Image and HRV analysis
HeonGyu Lee, Jin Hyoung Park, Keun Ho Ryu 87
A Gearbox Fault Detection Method under Varying Rotate Speed Conditions Based on Auxiliary
Particle Filter
195
Pan Hong-xia, Yuan Ku-mei
Analysis and the current situation of Information Communication Technology (ICT) in Mongolia
Naranchimeg Bold, Batnyam Battulga 208
Geospatial Data Indexing for University Management System using Rraster Storage
Bulganchimeg Enkhtaivan, Oyun-Erdene Namsrai 211
The 5th International Conference FITAT 2012
16
The 5th International Conference FITAT 2012
17
The 5th International Conference FITAT 2012
where is the correlation between feature and the redundancy analysis (line 9) and repeat the procedure
class. is the correlation between feature and until there are no features in Drem (line 4-15). Drem is
feature. the subset of features which are removed in each
redundancy analysis step (line 13). At first iteration, as
3. Methodology there are no features in the Drem, so we choose the most
relevant feature as the starting point (line 6-7). The
pseudo code of our algorithm is shown in Figure 2. As
Feature ranking a result, we can obtain a number of feature subsets,
then in a cascading manner, each top ranked feature
(line 17) in each subset uses as the root node of the ith
tree (Tree committees1 committees2, committeesk).
Subset1 Subset2 … Subsetk Tree induction is done by the skeleton of ITI
incremental learning algorithm [16]. The tree
committees are constructed at the point when there are
test examples or unknown instances that need to assign
Top K ranked Top K ranked Top K ranked
feature feature feature a class label. Because we found that using incremental
manner to construct the tree committees was just
wasting time and storage. Subsequently, from the tree
committees, we share the rules and the aggregate score
Tree Tree Tree
[9] is used to perform the prediction task (line 20-21).
committees1 committees2 committeesk
The classification score for a specific class, say class C,
is calculated as:
(4)
18
The 5th International Conference FITAT 2012
4. Experimental results
19
The 5th International Conference FITAT 2012
From Table 2, one can easily observe that in most cases [4] Pagallo,G. and Haussler,D., ―Boolean feature discovery
the prediction performance of the proposed algorithm in empirical learning‖, Machine Learning, 1993, pp. 71-99.
is better than other classifiers. However, the accuracy [5] Rakkrit Duangsoithong and Terry Windeatt, ―Hybrid
Correlation and Causal Feature Selection for Ensemble
of our algorithm on Heart Disease dataset is a little
Classifiers‖, Studies in Computational Intelligence, 2011, pp.
lower than other classifiers. The reason is that after 97-115.
redundancy analysis, the greater part of the features is
removed and only two features are used to construct the [6] Breiman, L., ―Bagging predictors‖, Machine Learning 24,
tree committees. Consequently thus, our algorithm 1996, pp. 123-140.
cannot greatly improve the accuracy of the classifiers
for this kind of dataset. The experiment results show [7] Freund, Y. and Schapire, R.E., ―Experiments with a New
that removing redundant features is a good way to Boosting Algorithm‖, The Thirteenth International
improve the accuracy of classifiers and it is worthwhile Conference on Machine Learing, 1996, pp. 148-156.
to use different predominant features as a starting point
[8] Li, J. Y., Liu, H. A., See-Kiong Ng and Limsoon Wong,
to remove other features. ―Discovery of significant rules for classifying cancer
diagnosis data‖, Bioinformatics, 2003, pp. 93-102.
5. Conclusions
[9] Minghao Piao, Jong Bum Lee, Khalid E.K. Saeed and
In this paper, an efficient incremental decision tree Keun Ho Ryu, ―Discovery of Significant Classification Rules
ensemble method is presented to deal with high from Incrementally Inducted Decision Tree Ensemble for
Diagnosis of Disease‖, ADMA, 2009, pp. 93-102.
dimensional features and feature redundancy. During
the experiments, we used four datasets which are taken [10] Lei Yu and Huan Liu ―A case study of incremental
from the UCI machine learning repository and we have concept induction‖, SIGKDD, Washington DC, USA, 2003.
met our objective which is to evaluate and investigate
the performance of the classifiers by using different [11] Duangsoithong, R. and Windeatt, T., ―Relevant and
size of training and test dataset. The results show that Redundant Feature Analysis with Ensemble Classification‖,
dividing the redundant features into several parts for Advances in Pattern Recognition, 2009, pp. 247-250.
ensemble construction can obtain better performance
for classification and our proposed algorithm has [12] Schlimmer, J. C. and Fisher, D., ―Induction of Decision
Trees‖, Proceedings of the Fifth National Conference on
higher prediction accuracy than other classification
Artificial Intelligence, 1986, pp. 496-501.
algorithms in most cases. Furthermore, among the
mutually redundant features, it is not a best choice to [13] Utgoff, P. E., ―Incremental Induction of decision trees‖,
select the most relevant feature as an input feature of Machine Learning, pp. 161-186, 1989.
the classifier. Some future works are planned to extend
the above algorithm by applying the datasets which [14] Tan, P. N., Steinbach, M., Kumar, V., Introduction to
have high dimensionality. data mining, 2006, pp. 276-277.
20