1,721,052 research outputs found
Weighted K-Nearest Neighbor revisited
No full textIn this paper we show that weighted K-Nearest Neighbor, a variation of the classic K-Nearest Neighbor, can be reinterpreted from a classifier combining perspective, specifically as a fixed combiner rule, the sum rule. Subsequently, we experimentally demonstrate that it can be rather beneficial to consider other combining schemes as well. In particular, we focus on trained combiners and illustrate the positive effect these can have on classification performance
Dissimilarity-Based Multiple Instance Learning
No full textMultiple instance learning (MIL) is an extension of supervised learning where the objects are represented by sets (bags) of feature vectors (instances) rather than individual feature vectors. For example, an image can be represented by a bag of instances, where each instance is a patch in that image. Only bag labels are given, however, the standard assumption is that that a bag is positive if and only if it contains a positive, or concept instance. In other words, only concept instances are informative for the bag label. The goal is to learn a bag classifier, although an instance classifier may also be desired. This scenario is suitable for applications where objects are heterogeneous and representing them as a single feature vector may lose important information, and/or in cases where only weakly labeled data is available. Several approaches to MIL exist. Instance-based approaches rely on stronger assumptions about the relationship of the instance labels and the bag labels, and define a bag classifier through an instance classifier. Bag-based approaches learn a bag classifier directly, often by converting the problem into a supervised problem. These methods often disregard the standard assumption, and instead use the collective assumption, where all instances are informative. One way to convert the problem into a supervised one, is to describe each bag by a vector of its distances to a set of reference prototypes. In this so-called dissimilarity representation, supervised classifiers can be used. The goal of this thesis is to study the dissimilarity representation as a method for dealing with multiple instance learning problems. We address the questions of defining a dissimilarity function and choosing a reference set of prototypes, while considering the assumptions that these choices implicitly make about the problem.Intelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
Moment Constrained Semi-Supervised LDA (extended abstract)
No full textThis BNAIC compressed contribution provides a summary of the work originally presented at the First IAPR Workshop on Partially Supervised Learning and published in [5]. It outlines the idea behind supervised and semi-supervised learning and highlights the major shortcoming of many current methods. Having identified the principal reason for their limitations, it briefly sketches a conceptually different take on the matter for linear discriminant analysis (LDA). Finally, the contribution hints at some of the results obtained. For any details, the reader is of course referred to [5].Intelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
Sample reusability in importance-weighted active learning
No full textRecent advances in importance-weighted active learning solve many of the problems of traditional active learning strategies. But does importance-weighted active learning also produce a reusable sample selection? This thesis explains why reusability can be a problem, how importance-weighted active learning removes some of the barriers to reusability and which obstacles still remain. With theoretical arguments and practical demonstrations, this thesis argues that universal reusability is impossible: because every active learning strategy must undersample some areas of the sample space, classifiers that depend on the samples in those areas will learn more from a random sample selection. This thesis describes several reusability experiments with importance-weighted active learning that show the impact of the reusability problem in practice. The experiments confirm that universal reusability does not exist, although in some cases – on some datasets and with some pairs of classifiers – there is sample reusability. This thesis explores the conditions that could guarantee the reusability between two classifiers.Pattern Recognition LabIntelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
Combining and evaluating regression methods for galaxyredshift estimates
No full textIn this thesis I develop a Machine Learning method to combine galactic redshift estimates of previous authors into an aggregate estimate. I investigate weather combining earlier results into a single estimate is a worthwhile effort. Disagreement between the earlier results is used as a metric the quality of estimation. This disagreement is exploited using kernel density estimation to iteratively selecting a subset of the problem that is smaller, but harder to solve. The iteration is repeated until either the problem is solved, or the remaining subset is too difficult to solve reliably.Electrical Engineering, Mathematics and Computer ScienceIntelligent System
Comments: On Distributional Assumptions and Whitened Cosine Similarities
No full textMediamaticsElectrical Engineering, Mathematics and Computer Scienc
Constrained parameter estimation for semi-supervised learning: The case of the nearest mean classifier
No full textA rather simple semi-supervised version of the equally simple nearest mean classifier is presented. However simple, the proposed approach is of practical interest as the nearest mean classifier remains a relevant tool in biomedical applications or other areas dealing with relatively high-dimensional feature spaces or small sample sizes. More importantly, the performance of our semi-supervised nearest mean classifier is typically expected to improve over that of its standard supervised counterpart and typically does not deteriorate with increasing numbers of unlabeled data. This behavior is achieved by constraining the parameters that are estimated to comply with relevant information in the unlabeled data, which leads, in expectation, to a more rapid convergence to the large-sample solution because the variance of the estimate is reduced. In a sense, our proposal demonstrates that it may be possible to properly train a known classification scheme such that it can benefit from unlabeled data, while avoiding the additional assumptions typically made in semi-supervised learning.MediamaticsElectrical Engineering, Mathematics and Computer Scienc
Approximate pairwise accuracy criteria for multiclass linear dimension reduction: Generalisations of the Fisher Criterion
No full textElectrical Engineering, Mathematics and Computer Scienc
Improving Cross-Validation Classifier Selection Accuracy through Meta-learning
No full textIn order to choose from the large number of classification methods available for use, cross-validation error estimates are often employed. We present this cross-validation selection strategy in the framework of meta-learning and show that conceptually, meta-learning techniques could provide better classifier selections than traditional cross-validation selection. Using various simulation studies we illustrate and discuss this possibility. Through a collection of datasets resembling real-world data, we investigate whether these improvements could possibly exist in the real-world as well. Although the approach presented here currently requires significant investment when applied to practical applications, the concept of being able to outperform cross-validation selection opens the door to new classifier selection strategies.Pattern Recognition LabIntelligent SystemsElectrical Engineering, Mathematics and Computer Scienc
- …
