Francesc J. Ferri

Prototype selection for the nearest neighbour rule through proximity graphs

Abstract In this paper, the Gabriel and Relative Neighbourhood graphs are used to select a suitable subset of prototypes for the Nearest Neighbour rule. Experiments and results are reported showing the effectiveness of the method and comparing its performance to those obtained by classical techniques.

The Imbalanced Training Sample Problem: Under or over Sampling?

The problem of imbalanced training sets in supervised pattern recognition methods is receiving growing attention. Imbalanced training sample means that one class is represented by a large number of examples while the other is represented by only a few. It has been observed that this situation, which arises in several practical domains, may produce an important deterioration of the classification accuracy, in particular with patterns belonging to the less represented classes. In this paper we present a study concerning the relative merits of several re-sizing techniques for handling the imbalance issue. We assess also the convenience of combining some of these techniques.

On the use of neighbourhood-based non-parametric classifiers

Alternative non-parametric classification schemes, which come from the use of different definitions of neighbourhood, are introduced. In particular, the Nearest Centroid Neighbourhood along with the neighbourhood relation derived from the Gabriel Graph and the Relative Neighbourhood Graph are used to define the corresponding (k-)Nearest Neighbour-like classifiers. Experimental results are reported to compare the performance of the approaches proposed here to the one obtained with the k-Nearest Neighbours rule.

Considerations about sample-size sensitivity of a family of edited nearest-neighbor rules

The edited nearest neighbor classification rules constitute a valid alternative to k-NN rules and other nonparametric classifiers. Experimental results with synthetic and real data from various domains and from different researchers and practitioners suggest that some editing algorithms (especially, the optimal ones) are very sensitive to the total number of prototypes considered. This paper investigates the possibility of modifying optimal editing to cope with a broader range of practical situations. Most previously introduced editing algorithms are presented in a unified form and their different properties (acid not just their asymptotic behavior) are intuitively analyzed. The results show the relative limits in the applicability of different editing algorithms.

An efficient prototype merging strategy for the condensed 1-NN rule through class-conditional hierarchical clustering ☆

Abstract A generalized prototype-based classification scheme founded on hierarchical clustering is proposed. The basic idea is to obtain a condensed 1-NN classification rule by merging the two same-class nearest clusters, provided that the set of cluster representatives correctly classifies all the original points. Apart from the quality of the obtained sets and its flexibility which comes from the fact that different intercluster measures and criteria can be used, the proposed scheme includes a very efficient four-stage procedure which conveniently exploits geometric cluster properties to decide about each possible merge. Empirical results demonstrate the merits of the proposed algorithm taking into account the size of the condensed sets of prototypes, the accuracy of the corresponding condensed 1-NN classification rule and the computing time.

Another move toward the minimum consistent subset: a tabu search approach to the condensed nearest neighbor rule

This paper presents a new approach to the selection of prototypes for the nearest neighbor rule which aims at obtaining an optimal or close-to-optimal solution. The problem is stated as a constrained optimization problem using the concept of consistency. In this context, the proposed method uses tabu search in the space of all possible subsets. Comparative experiments have been carried out using both synthetic and real data in which the algorithm has demonstrated its superiority over alternative approaches. The results obtained suggest that the tabu search condensing algorithm offers a very good tradeoff between computational burden and the optimality of the prototypes selected.

Combining similarity measures in content-based image retrieval

The purpose of content based image retrieval (CBIR) systems is to allow users to retrieve pictures from large image repositories. In a CBIR system, an image is usually represented as a set of low level descriptors from which a series of underlying similarity or distance functions are used to conveniently drive the different types of queries. Recent work deals with combination of distances or scores from different and usually independent representations in an attempt to induce high level semantics from the low level descriptors of the images. Choosing the best method to combine these results requires a careful analysis and, in most cases, the use of ad-hoc strategies. Combination based on or derived from product and sum rules are common approaches. In this paper we propose a method to combine a given set of dissimilarity functions. For each similarity function, a probability distribution is built. Assuming statistical independence, these are used to design a new similarity measure which combines the results obtained with each independent function.

DECISION BOUNDARY PRESERVING PROTOTYPE SELECTION FOR NEAREST NEIGHBOR CLASSIFICATION

The excessive computational resources required by the Nearest Neighbor rule are a major concern for a number of specialists and practitioners in the Pattern Recognition community. Many proposals for decreasing this computational burden, through reduction of the training sample size, have been published. This paper introduces an algorithm to reduce the training sample size while preserving the original decision boundaries as much as possible. Consequently, the algorithm tends to obtain classification accuracy close to that of the whole training sample. Several experimental results demonstrate the effectiveness of this method when compared to other reduction algorithms based on similar ideas.

Perceptual feedback in multigrid motion estimation using an improved DCT quantization

In this paper, a multigrid motion compensation video coder based on the current human visual system (HVS) contrast discrimination models is proposed. A novel procedure for the encoding of the prediction errors has been used. This procedure restricts the maximum perceptual distortion in each transform coefficient. This subjective redundancy removal procedure includes the amplitude nonlinearities and some temporal features of human perception. A perceptually weighted control of the adaptive motion estimation algorithm has also been derived from this model. Perceptual feedback in motion estimation ensures a perceptual balance between the motion estimation effort and the redundancy removal process. The results show that this feedback induces a scale-dependent refinement strategy that gives rise to more robust and meaningful motion estimation, which may facilitate higher level sequence interpretation. Perceptually meaningful distortion measures and the reconstructed frames show the subjective improvements of the proposed scheme versus an H.263 scheme with unweighted motion estimation and MPEG-like quantization.

A naive relevance feedback model for content-based image retrieval using multiple similarity measures

This paper presents a novel probabilistic framework to process multiple sample queries in content based image retrieval (CBIR). This framework is independent from the underlying distance or (dis)similarity measures which support the retrieval system, and only assumes mutual independence among their outcomes. The proposed framework gives rise to a relevance feedback mechanism in which positive and negative data are combined in order to optimally retrieve images according to the available information. A particular setting in which users interactively supply feedback and iteratively retrieve images is set both to model the system and to perform some objective performance measures. Several repositories using different image descriptors and corresponding similarity measures have been considered for benchmarking purposes. The results have been compared to those obtained with other representative strategies, suggesting that a significant improvement in performance can be obtained.