On the design of error adaptive ECOC algorithms Elizabeth Tapia 1 , José Carlos González 1 , Javier García-Villalba 2 1 Department of Telematics Engineering - Technical University of Madrid, Spain {etapia, jcg}@gsi.dit.upm.es 2 Department of Computer Systems - Complutense University of Madrid (U.C.M.), Spain javiergv@sip.ucm.es Abstract. A new type of adaptive boosting algorithms suitable for multiclassific a- tion problems is proposed. By means of recursive coding models and the ECOC framework, learning in multivalued output domains is formulated in terms of binary learning algorithms. In such a system, learning algorithms play a decoding func- tion. The use of recursive coding allows the design of error adaptive ECOC learning algorithms. This approach generalizes the boosting concept in arbitrary multivalued output domains in a transparent way and gives rise to the so-called RECOC learning models. A RECOC instance based on Turbo Codes is pre- sented. 1 Introduction Redundancy is a well-known concept in coding theory. Transmission of information through noisy channels can be carried out at arbitrarily small error rates by suitable channel encoding. Redundancy is also present in the learning field. It is observable in the repetition of concepts implemented by boosting [1] algorithms and it is ex- plicitly used in Dietterich’s Error Correcting Output Codes (ECOC) approach for classification problems in multivalued output domains [2]. In his work, Dietterich recognized that ECOC schemes should resemble random coding in order to achieve successful learning. Let us denote by M the number of classes involved. Good ECOC schemes for low to medium M values have been found by exhaustive search. However, the design of suitable ECOC schemes in arbitrary output domains has remained as an open problem. Despite of their common motivation, namely the design of low complexity learning algorithms, both ECOC and boosting appear to be conceptually different for Machine Learning literature. Boosting algorithms are assumed to exhibit an error adaptation capacity, which appears to be absent in ECOC algorithms. Because of this apparent lack of error adaptation capacity, ECOC algo- rithms have been used in the core of multiclass boosting algorithms. On the opposite side, binary boosting has not been used in the core of ECOC algorithms. This is a striking question considering that the ECOC strategy reduces M classification problems to a related set of binary ones.