Tratamiento del desbalance en problemas con múltiples clases con ECOC

Handling the Multi-Class Imbalance Problem using ECOC

Rosa María Valdovinos Rosas¹, Rosalinda Abad Sánchez, Roberto Alejo Eleuterio², Edgar Herrera Arteaga ^1,3, Adrián Trueba Espinosa⁴

¹ Universidad Autónoma del Estado de México, Facultad de Ingeniería, Ciudad Universitaria, Toluca, México. li_rmvr@hotmail.com

³ Instituto Nacional de Investigación Nuclear ININ, La Marquesa, Ocoyoacac, México. edgar.herrera@inin.gob.mx

⁴ Centro Universitario UAEM Texoco, México.

Article received on 05/11/2012
Accepted on 21/06/2013

Resumen

El problema del desbalance de clases puede producir un deterioro importante en la efectividad del clasificador, en particular con los patrones de las clases menos representadas. El desbalance en el conjunto de entrenamiento (CE) significa que una clase es representada por una gran cantidad de patrones mientras que otra es representada por muy pocos. Los estudios existentes se encuentran orientados principalmente a tratar problemas de dos clases, no obstante, un importante número de problemas reales se encuentran representados por múltiples clases, donde resulta más difícil su discriminación para el clasificador. El éxito de la Mezcla de Expertos (ME) se basa en el criterio de "divide y vencerás". En su funcionamiento general, el problema es dividido en fragmentos más pequeños que serán estudiados por separado. De este modo, el modelo general es poco influenciado por las dificultades individuales de sus componentes. La idea principal del estudio aquí mostrado, es construir una Mezcla de expertos cuyos miembros serán entrenados en una parte del problema general y de este modo, mejorar el rendimiento del clasificador en el contexto de múltiples clases. Para este fin, se hace uso de los métodos conocidos como Error-correcting output codes (ECOC), que permiten realizar una codificación en parejas de clases el problema de estudio. Resultados experimentales sobre conjuntos de datos reales, muestran la viabilidad de la estrategia aquí propuesta.

Palabras clave: Desbalance de clases, mezcla de expertos, fusión, múltiples clases, error correcting output codes (ECOC).

Abstract

Imbalanced training sample means that one class is represented by a large number of examples while the other is represented by only a few. This problem may produce an important deterioration of the classifier performance, in particular with patterns belonging to the less represented classes. The majority of the studies in this area are oriented, mainly, to resolve problems with two classes. However, many real problems are represented by multiple classes, where it is more difficult to discriminate between them. The success of the Mixture of Experts (ME) strategy is based on the criterion of "divide and win". The general process divides the global problem into smaller fragments which will be studied separately. In this way, the general model has few influences of the individual difficulties (of their members). In this paper we propose a strategy for handling the class imbalance problem for data sets with multiple classes. For that, we integrate a mixture of experts whose members will be trained as a part of the general problem and, in this way, will improve the behavior of the whole system. For dividing the problem we employ the called Error-correcting output codes (ECOC) methods, when the classes are codified in pairs, which are considered for training the mixture of experts. Experiments with real datasets demonstrate the viability of the proposed strategy.

Keywords: Class imbalance, fusion, mixture of experts, error correcting output codes (ECOC).

DESCARGAR ARTÍCULO EN FORMATO PDF

Agradecimientos

Este trabajo fue realizado gracias al apoyo recibido de los proyectos: 3072/2011 de la UAEM, PROMEP/103.5/12/4783 de las SEP, SDMAIA-010 del TESjo, UR-001 del ININ.

1. Alejo, R., Sotoca, J.M., & Casañ, G.A. (2008). An empirical study for the multi-class imbalance problem with neural networks. Pattern Recognition, Image Analysis and Applications. Lecture Notes in Computer Science, 5197, 479-486.         [ Links ]

2. Barandela, R., Sánchez, J.S., García, V., & Rangel, E. (2001). Fusion of techniques for handling the imbalanced training sample problem. 6^th Ibero-American Symposium on Pattern Recognition, Florianópolis, Brazil, 34-40.         [ Links ]

3. Batista, G., Patri, R.C., & Monard, M.C. (2004). A study of the Behavior of several methods for Balancing Machine Learning Training data. ACM SIGKDD Explorations Newsletter, 6(1), 20-29.         [ Links ]

4. Breiman, L. (1998). Arcing classifiers. The Annals of Statistics. 26(3), 801 -849.         [ Links ]

5. Chawla, N.V., Bowyer, K.W., Hall, L.O., & Kegelmeyer. W.P. (2000). SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16(1), 321-357.         [ Links ]

6. Dietterich, T.G. & Bakiri, G. (1994). Solving multiclass learning problems via error-correcting output codes. Journal of Artificial Intelligence Research, 2(1), 263-286.         [ Links ]

7. Dietterich, T.G. (1997). Machine learning research: four current directions. AI Magazine, 18(4), 97-136.         [ Links ]

8. Kong, E.B. & Dietterich, T.G. (1995). Error-Correcting Output Coding Corrects Bias and Variance. 12^th International Conference on Machine Learning. California, USA, 313-321.         [ Links ]

9. Eavis, T. & Japkowicz, N. (2000). A recognition-based alternative to discrimination-base multilayer perceptrons. Advances in Artificial Intelligence, Lecture Notes in Computer Science, 1822, 280-292.         [ Links ]

10. Kuncheva, L.I. (2000). Clustering-and-selection model for classifier combination. 4^th International Conference on Knowledge-Based Intelligent Engineering Systems and Allied Technologies (KES'2000). Brighton, UK, 1, 185-188.         [ Links ]

11. Kuncheva, L.I. (2001). Using measures of similarity and inclusion of multiple classifier fusion by decision templates. Fuzzy Sets and Systems, 122(3), 401 -407.         [ Links ]

12. Kuncheva, L.I. & Whitaker, C.J. (2003). Measures of diversity in classifier ensembles and their relationship with the ensemble accuracy. Machine Learning, 51(2), 181 -207.         [ Links ]

13. Kuncheva, L.I. (2005). Using diversity measures for generating error-correcting output codes in classifier ensemble. Pattern Recognition Letters, 26, 83-90.         [ Links ]

14. Congalton, R.G. & Green, K. (1999). Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. Boca Raton: Lewis Publications.         [ Links ]

15. Daqi, G., Wei, W., & Jianliang, G. (2007). Class-modular multi-layer perceptions, task decomposition and virtually balanced training subsets. International Joint Conference on Neural Networks, Orlando, Florida, USA, 2153-2158.         [ Links ]

16. Soda, P. & lannello, G. (2010). Decomposition Methods and Learning approaches for Imbalanced Dataset: An Experimental Integration. 20^th International Conference on Pattern Recognition, Istanbul, Turkey, 3117-3120.         [ Links ]

17. Demsar, J. (2006). Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7(2006), 1-30.         [ Links ]

18. Garcia, S. & Herrera, F. (2008). An extension on statistical comparisons of classifiers over multiple data sets for all pairwise comparisons. Journal of Machine Learning Research, 9(12), 2677-2694.         [ Links ]

19. Dasarathy, B.V. (1991). Nearest Neighbor (NN) Norms: nn Pattern Clasification Techniques. Los Alamitos, CA: IEEE Computer Society Press.         [ Links ]