PRlab GRFIA

The main worklines of our group are:

Fast nearest neighbor search algorithms for metric spaces
Given a training set of objects classified into classes, the nearest neighbor (NN) classifier assigns an unknown sample to the class of its nearest neighbor in the set. The nearest neighbor may be found using an exhaustive search (brute force approach), or using one of the many fast NN search algorithms. Our group has developed several algorithms for finding the nearest neighbor that do not require a vector representation of objects. These algorithms are specially suitable for tasks where the distance between two objects is computationally demanding, e.g. the Levenshtein distance (edit distance).

Classification rules based on neighborhood
The nearest neighbor (NN) rule and the k-NN rule define a neighborhood around the sample. Although these are simple and theoretically have a good behaviour, in practice its results may be improved using another neighborhood definition. In our group we have developed another neighborhood definition, the k-NSN rule, that uses the candidates to NN in a fast NN search algorithm in order to approximate to k-NN rule error rates with the cost of a 1-NN search. Also, we are currently working on reducing the k-NN error rates using alternative neighborhood definitions.

Combination of classifiers
The combination of a group of classifiers often obtain better results than a single classifier. There are many ways of combining classifiers such as fusion of outputs, using different training sets or different feature sets. Recently, our group has developed several ways of training weights for classifier output combination.