Solving the slate tile classification problem using a DAGSVM multiclassification algorithm based on SVM binary classifiers with a one-versus-all approach

https://doi.org/10.1016/j.amc.2013.12.087Get rights and content

Highlights

  • It combines DAG with SVM binary classifiers with one-versus-all approach.

  • To validate UCI Machine Learning Repository and slate tiles data have been used.

  • Error rates of different models have been analysed.

  • Results show a good performance of proposed DAG-one-versus-all model.

Abstract

We describe a new classification methodology based on binary classifiers constructed using support vector machines and applying a one-versus-all approach supported by the use of the directed acyclic graphs. The new methodology, which is computationally less costly because a smaller number of binary classification problems have to be resolved, was validated using UCI Machine Learning Repository data sets. Results point to the improved performance of the proposed model compared to approaches based on the one-versus-one and directed acyclic graph techniques. This new multiclassification strategy successfully applied to a slate tile classification problem produced favourable results compared to other validated techniques.

Introduction

Cortes and Vapnik [1] developed the support vector machine (SVM) model in 1995 to solve a binary classification problem. Since then, this model has become a benchmark technique for pattern recognition problems [2] because of its flexibility, prediction capacity, parsimony and the global optimum character.

Different strategies for solving the multiclassification problem have been developed based on constructing SVM binary classifiers [3], [4] according to either the one-versus-one approach, which consists of confronting each class with each other class [5], or the one-versus-all approach, which consists of confronting each class with all the other classes together [6]. Another classification strategy is the directed acyclic graph (DAG) [7] with a perceptron in each node, along with the adaptive version [8] and the integration of neural networks to form a DAG-multilayer perceptron structure [9], [10]. Other studies from around the same period made adaptations to the SVM algorithm; these include the approach by Cramer and Singer [11], [12], [13] based on the design of output codes and the approach by Manikandan [14], which explored the problem of isolated digit recognition.

We describe a new SVM binary classifier strategy for solving the multiclassification problem based on combining the one-versus-all and DAG. The proposed methodology was validated using data sets from the UCI Machine Learning Repository [15], frequently used to evaluate new algorithms [16], [17], [18], [19], [20].

The new classification strategy described in this study was also validated for the problem of classifying slate tiles. In the slate production process, the final stage is the manual classification of tiles in commercial categories. Martinez et al. [21] describe the application of machine learning techniques to the problem of classifying slate tiles on the basis of objective information extracted by a 2D–3D laser scanner [22]. This research work demonstrated the possibility of characterizing the slate tiles according to their defects described by the Spanish UNE-EN 12326–1 standard using a hybrid system that acquires images of the tiles and applies a series of mathematical algorithms to detect the defects.

The article is laid out as follows: first we outline the mathematical background, next we describe the validation of the methodology, then we discuss the results and finally we outline the main lessons learned from our research.

Section snippets

The classification problem

The classification problem can be interpreted as the partitioning of a population of individuals in mutually exclusive subpopulations Cj,j=1,,c. If the individuals are characterized by a vector of features x=(x1,x2,,xd) with values in a set X, and the classes are coded, for instance, by a variable YY={1,2,,c}, the classification problem consists of determining classification rule g:XY :g(x)=1ifxX1cifxXcaccording to some optimal criterion. That is, each example x is assigned to the

Model validation

The model proposed in this study, based on binary SVM classifiers and a DAG with one-versus-all approach was validated using different UCI Machine Learning Repository data sets [15]. The error criterion used for evaluation was the error rate (ER),ER=1ni=1n1yiyi^where 1yiyi^ represents the indicator function for the set yiyi^ which takes the value 1 when the real value yi and the estimated value yi^ do not coincide and where n represents the number of points in the test set.

To evaluate the

Conclusions

We have validated a novel classification strategy for solving a multiclassification problem based on one-versus-all SVM binary classifiers and the construction of a DAG.

The model was validated using different UCI Machine Learning Repository data sets covering different classification problem scenarios. For these four data sets four multiclassification models were implemented based on the construction of binary SVM classifiers using the one-versus-one, one-versus-all and DAG one-versus-one

Acknowledgements

J.M. Matias’ research and J. Taboada’s research are both funded by the Spanish Ministry of Education and Science under Grant MTM2008-03010 and Grant No. BIA2007-66218, respectively, and J. Martinez’ research is funded by the Spanish Ministry of Science and Technology through research project ECO2011-22650. C. Iglesias is acknowledged to Spanish Ministry of Education for the FPU 12-02283 grant.

References (45)

  • A. Lorena et al.

    A review on the combination of binary classifiers in multiclass problems

    Artif. Intell. Rev.

    (2008)
  • U. Krebel, Pairwise classification and support vector machines, in: B. Scholdopf, C. Burges, A.J. Smola (Eds.),...
  • J. Weston, C. Watkins, Multi-class support vector machines, in: M. Verleysen (Ed.), Proceedings of ESANN99, D. Facto...
  • J.C. Platt, N. Cristianini, J. Shawe-Taylor, Large margin DAGs for multiclass classification, in: S.A. Solla, T.K....
  • B. Kijsirikul, N. Ussivakul, Multiclass support vector machines using adaptative directed acyclic graph, in:...
  • A. Niyom, S. Chiewchanwattana, K. Sunat, C. Lursinsap, The DAGs-MLP structure to the efficiency of neural network...
  • K. Cramer et al.

    On the learnability and design of output codes for multiclass problems

    Mach. Learn.

    (2002)
  • T.G. Dietterich et al.

    Solving multi class learning problems via error-correcting output codes

    J. Artif. Intell. Res.

    (1995)
  • R.E. Schapire, Using output codes to boost multi class learning, in: Mach. Learn.: Proceedings of the Fourteenth...
  • A. Frank, A. Asuncion, UCI Machine Learning Repository [http://archive.ics.uci.edu/ml], University of California, Sch....
  • S. Suresh et al.

    A sequential learning algorithm for complex-valued self-regulating resource allocation network-CSRAN

    IEEE Trans. Neural Networks

    (2011)
  • P. Kraipeerapun et al.

    Solving regression problem with complementary neural networks and an adjusted averaging technique

    Memet. Comput.

    (2010)
  • Cited by (18)

    • An Advanced Genetic Algorithm with Improved Support Vector Machine for Multi-Class Classification of Real Power Quality Events

      2021, Electric Power Systems Research
      Citation Excerpt :

      It can evaluate the results with only (c-1) decision sets. In Comparison With other decision based classifiers, the execution time is fast and it has a special feature of fault tolerance due to its unique decision based structure [31]. The proposed multiple power quality events classification can set the automatic feature selection with random determination of topology shape and attend excellent fault tolerance mechanisms by reducing error accumulation of the higher nodes.

    • Automated vision system for quality inspection of slate slabs

      2018, Computers in Industry
      Citation Excerpt :

      The initial system developed was a hybrid 2D-3D laser scanner system, which, with two cameras, captured images of slabs for subsequent analysis and grading. Slate slabs were automatically graded using supervised and unsupervised learning techniques [47] and a novel classification algorithm based on SVMs was developed [48] (a description and results for the full 2D-3D hybrid system can be found in [49]). The hybrid 2D-3D system developed in 2010, in which two cameras captured 3D and monochrome 2D data from slate slabs [4,46,49] was the starting point for this research, aimed at developing a system to detect five traits in slate slabs: material defects, false squaring, warping, kink-bands (fissuring) and white (flowerlike) staining.

    • A membrane-inspired bat algorithm to recognize faces in unconstrained scenarios

      2017, Engineering Applications of Artificial Intelligence
    View all citing articles on Scopus
    View full text