Search

In the process of designing pattern recognition systems one may choose a representation based on pairwise dissimilarities between objects. This is especially appealing when a set of discriminative features is difficult to find. Various classification systems have been studied for such a dissimilarity representation: the direct use of the nearest neighbor rule, the postulation of a dissimilarity space and an embedding to a virtual, underlying feature vector space. It appears in several applications that the dissimilarity measures constructed by experts tend to have a non-Euclidean behavior. In this paper we first analyze the causes of such choices and then experimentally verify that the non-Euclidean property of the measure can be informative. [ABSTRACT FROM AUTHOR]

Generative kernels represent theoretically grounded tools able to increase the capabilities of generative classification through a discriminative setting. Fisher Kernel is the first and mostly-used representative, which lies on a widely investigated mathematical background. The manufacture of a generative kernel flows down through a two-step serial pipeline. In the first, ˵generative″ step, a generative model is trained, considering one model for class or a whole model for all the data; then, features or scores are extracted, which encode the contribution of each data point in the generative process. In the second, ˵discriminative″ part, the scores are evaluated by a discriminative machine via a kernel, exploiting the data separability. In this paper we contribute to the first aspect, proposing a novel way to fit the class-data with the generative models, in specific, focusing on Hidden Markov Models (HMM). The idea is to perform model clustering on the unlabeled data in order to discover at best the structure of the entire sample set. Then, the label information is retrieved and generative scores are computed. Experimental, comparative test provides a preliminary idea on the goodness of the novel approach, pushing forward for further developments. [ABSTRACT FROM AUTHOR]

The strength of classifier combination lies either in a suitable averaging over multiple experts/sources or in a beneficial integration of complementary approaches. In this paper we focus on the latter and propose the use of group-induced vector spaces (GIVSs) as a way to combine unsupervised learning with classification. In such an integrated approach, the data is first modelled by a number of groups, found by a clustering procedure. Then, a proximity function is used to measure the (dis)similarity of an object to each group. A GIVS is defined by mapping an object to a vector of proximity scores, computed with respect to the given groups.In this study, we focus on a particular aspect of using GIVSs in a mode of building a trained combiner, namely the integration of generative and discriminative methods. First, in the generative step, we model the groups by simple generative models, building the GIVS space. The classification problem is then mapped in the resulting vector space, where a discriminative classifier is trained. Our experiments show that the integrated approach leads to comparable or better results than the generative methods in the original feature spaces. [ABSTRACT FROM AUTHOR]

Sometimes novel or outlier data has to be detected. The outliers may indicate some interesting rare event, or they should be disregarded because they cannot be reliably processed further. In the ideal case that the objects are represented by very good features, the genuine data forms a compact cluster and a good outlier measure is the distance to the cluster center. This paper proposes three new formulations to find a good cluster center together with an optimized ℓp-distance measure. Experiments show that for some real world datasets very good classification results are obtained and that, more specifically, the ℓ1-distance is particularly suited for datasets containing discrete feature values. Keywords: one-class classification, outlier detection, robustness, ℓp-ball. [ABSTRACT FROM AUTHOR]

Statistical learning algorithms often rely on the Euclidean distance. In practice, non-Euclidean or non-metric dissimilarity measures may arise when contours, spectra or shapes are compared by edit distances or as a consequence of robust object matching [1,2]. It is an open issue whether such measures are advantageous for statistical learning or whether they should be constrained to obey the metric axioms. The k-nearest neighbor (NN) rule is widely applied to general dissimilarity data as the most natural approach. Alternative methods exist that embed such data into suitable representation spaces in which statistical classifiers are constructed [3]. In this paper, we investigate the relation between non-Euclidean aspects of dissimilarity data and the classification performance of the direct NN rule and some classifiers trained in representation spaces. This is evaluated on a parameterized family of edit distances, in which parameter values control the strength of non-Euclidean behavior. Our finding is that the discriminative power of this measure increases with increasing non-Euclidean and non-metric aspects until a certain optimum is reached. The conclusion is that statistical classifiers perform well and the optimal values of the parameters characterize a non-Euclidean and somewhat non-metric measure. [ABSTRACT FROM AUTHOR]

Pairwise proximities describe the properties of objects in terms of their similarities. By using different distance-based functions one may encode different characteristics of a given problem. However, to use the framework of statistical pattern recognition some vector representation should be constructed. One of the simplest ways to do that is to define an isometric embedding to some vector space. In this work, we will focus on a linear embedding into a (pseudo-)Euclidean space. This is usually well defined for training data. Some inadequacy, however, appears when projecting new or test objects due to the resulting projection errors. In this paper we propose an augmented embedding algorithm that enlarges the dimensionality of the space such that the resulting projection error vanishes. Our preliminary results show that it may lead to a better classification accuracy, especially for data with high intrinsic dimensionality. [ABSTRACT FROM AUTHOR]

A common way of expressing string similarity in structural pattern recognition is the edit distance. It allows one to apply the kNN rule in order to classify a set of strings. However, compared to the wide range of elaborated classifiers known from statistical pattern recognition, this is only a very basic method. In the present paper we propose a method for transforming strings into n-dimensional real vector spaces based on prototype selection. This allows us to subsequently classify the transformed strings with more sophisticated classifiers, such as support vector machine and other kernel based methods. In a number of experiments, we show that the recognition rate can be significantly improved by means of this procedure. [ABSTRACT FROM AUTHOR]

Statistical inference of sensor-based measurements is intensively studied in pattern recognition. It is usually based on feature representations of the objects to be recognized. Such representations, however, neglect the object structure. Structural pattern recognition, on the contrary, focusses on encoding the object structure. As general procedures are still weakly developed, such object descriptions are often application dependent. This hampers the usage of a general learning approach. This paper aims to summarize the problems and possibilities of general structural inference approaches for the family of sensor-based measurements: images, spectra and time signals, assuming a continuity between measurement samples. In particular it will be discussed when probabilistic assumptions are needed, leading to a statistically-based inference of the structure, and when a pure, non-probabilistic structural inference scheme may be possible. [ABSTRACT FROM AUTHOR]

For learning purposes, representations of real world objects can be built by using the concept of dissimilarity (distance). In such a case, an object is characterized in a relative way, i.e. by its dissimilarities to a set of the selected prototypes. Such dissimilarity representations are found to be more practical for some pattern recognition problems. When experts cannot decide for a single dissimilarity measure, a number of them may be studied in parallel. We investigate two possibilities of combining either dissimilarity representations themselves or classifiers built on each of them separately. Our experiments conducted on a handwritten digit set demonstrate that when the dissimilarity representations are of different nature, a much better performance can be obtained by their combination than on individual representations. [ABSTRACT FROM AUTHOR]