Search

This paper examines the prediction of disturbances based on their past measurements using k -nearest neighbours. The aim is to provide a prediction of a measured disturbance to a controller, in order to improve the feed-forward action. This prediction method works in an unsupervised way, it is robust against changes of the characteristics of the disturbance, and its functioning is simple and transparent. The method is tested on data from industrial process plants and compared with predictions from an autoregressive model. A qualitative as well as a quantitative method for analysing the predictability of the time series is provided. As an example, the method is implemented in an MPC framework to control a simple benchmark model. [ABSTRACT FROM AUTHOR]

The k-nearest neighbour ( kNN) problem appears in many different fields of computer science, such as computer animation and robotics. In crowd simulation, kNN queries are typically used by a collision-avoidance method to prevent unnecessary computations. Many different methods for finding these neighbours exist, but it is unclear which will work best in crowd simulations, an application which is characterised by low dimensionality and frequent change of the data points. We therefore compare several data structures for performing kNN queries. We find that the nanoflann implementation of a k-d tree offers the best performance by far on many different scenarios, processing 100,000 agents in about 35 ms on a fast consumer PC. [ABSTRACT FROM AUTHOR]

A popular and easy to implement classifier is the k-Nearest Neighbour (k-NN). However, sequentially searching for nearest neighbours in large datasets leads to inefficient classification because of the high computational cost involved. This article presents an adaptive hybrid and cluster-based method for speeding up the k-NN classifier. The proposed method reduces the computational cost as much as possible while maintaining classification accuracy at high levels. The method is based on the well-known k-means clustering algorithm and consists of two main parts: (i) a pre-processing algorithm that builds a two-level, cluster-based data structure, and (ii) a hybrid classifier that classifies new items by accessing either the first or the second level of the data structure. The proposed approach was tested on seven real life datasets and the experiential measurements were statistically validated by the Wilcoxon signed ranks test. The results show that the proposed classification method can be used either to achieve high accuracy with slightly higher cost or to reduce the cost at a minimum level with slightly lower accuracy. [ABSTRACT FROM AUTHOR]

Abstract: Nearest neighbour queries (or kNN) have been used in many disciplines, including spatial databases. kNN queries have gone beyond simply finding nearest objects. It is, therefore, important to understand the full spectrum of kNN queries, even before starting to work on how to process and optimise such queries. The aim of this paper is to give a complete picture of what kNN queries are capable of. In this study, we present a survey of kNN queries, in which we propose a taxonomy of kNN queries, comprising four perspectives: (i) a Space perspective, (ii) a Result perspective, (iii) a Query-Point perspective, and (iv) a Relationship perspective. These give a comprehensive overview of kNN queries. As kNN is a large area of research, in this paper, we confine the discussion to kNN queries on stationary objects. [Copyright &y& Elsevier]

In this paper, we consider first an estimator of the residual variance treated by Evans [(2005), ‘Estimating the Variance of Multiplicative Noise’, in 18th International Conference on Noise and Fluctuations, ICNF, in AIP Conference Proceedings, 780, pp. 99–102], Evans and Jones [(2008), ‘Non-Parametric Estimation of Residual Moments and Covariance’, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 464, 2831–2846] and by Liitiäinen, Corona, and Lendasse [(2008), ‘On Nonparametric Residual Variance Estimation’, Neural Processing Letters, 28, 155–167; (2010), ‘Residual Variance Estimation Using a Nearest Neighbour Statistic’, Journal of Multivariate Analysis, 101, 811–823], based on first and second nearest neighbours given an independent and identically distributed sample. Its strong consistency and almost sure convergence of the arithmetic means sequence are shown under mere boundedness and square integrability, respectively, of the response variable Y. Moreover, in view of the local variance, a correspondingly modified estimator of local averaging (partitioning) type is proposed, and strong L 2-consistency for bounded Y, weak L 2-consistency and optimal rate of convergence (for bounded X under suitable Hölder continuity conditions on regression and local variance functions) under moment conditions on Y are established. Simulation studies illustrate the behaviour of the local variance estimates. [ABSTRACT FROM AUTHOR]

K-nearest neighbours algorithms are among the most popular existing classification methods, due to their simplicity and good performances. Over the years, several extensions of the initial method have been proposed. In this paper, we propose a K-nearest neighbours approach that uses the theory of imprecise probabilities, and more specifically lower previsions. We show that the proposed approach has several assets: it can handle uncertain data in a very generic way, and decision rules developed within this theory allow us to deal with conflicting information between neighbours or with the absence of close neighbour to the instance to classify. We show that results of the basic k-NN and weighted k-NN methods can be retrieved by the proposed approach. We end with some experiments on the classical data sets. [ABSTRACT FROM AUTHOR]

Query expansion is the process of reformulating an original query to improve retrieval performance in information retrieval systems. Relevance feedback is one of the most useful query modification techniques in information retrieval systems. In this paper, we introduce query expansion into ligand-based virtual screening (LBVS) using the relevance feedback technique. In this approach, a few high-ranking molecules of unknown activity are filtered from the outputs of a Bayesian inference network based on a single ligand molecule to form a set of ligand molecules. This set of ligand molecules is used to form a new ligand molecule. Simulated virtual screening experiments with the MDL Drug Data Report and maximum unbiased validation data sets show that the use of ligand expansion provides a very simple way of improving the LBVS, especially when the active molecules being sought have a high degree of structural heterogeneity. However, the effectiveness of the ligand expansion is slightly less when structurally-homogeneous sets of actives are being sought. [ABSTRACT FROM AUTHOR]

Abstract: An analytical result should be expressed as x ± U, where x is the experimental result obtained for a given variable and U is its uncertainty. This uncertainty is rarely taken into account in supervised classification. In this paper, we propose to include the information about the uncertainty of the experimental results to compute the reliability of classification. The method combines k-nearest neighbours (kNN) with a nested bootstrap scheme, in which a new bootstrap training set is generated using the classical bootstrap in the first level (B times) and a new bootstrap method, called U-bootstrap, in the second level (D times). Two bootstraps are used to reduce the effect of sampling in the first level and the effect of the uncertainty in the second one. These B × D new training bootstrap sets are used to compute the reliability of classification for an unknown object using kNN. The object is classified into the class with the highest reliability. In this method, unlike the classical kNN and Probabilistic Bagged k-nearest neighbours (PBkNN), the reliability of classification changes (increases or decreases) when the uncertainty is increased. These changes depend on the position of the unknown object with respect to the training objects. For the benchmark Wine dataset, we found similar values of classification error rate (CER) than for kNN (5.57%), but lower than Probabilistic Bagged k-nearest neighbours using Hamamoto''s bootstrap (7.96%) or Efron''s bootstrap (8.97%). [Copyright &y& Elsevier]

Abstract: This paper presents a new method for computing the probability of correct classification for the k-Nearest Neighbours (kNN) method. The method uses bootstrap to provide the posterior probability which a new object is classified with. This is a measure of the reliability of the classification; it increases as the test object is closer to the training objects of a given class and is more sensitive to the position of the test object in the calibration space than the classical measure of posterior probability in kNN. This reliability of the classification is also used to derive a new rule for classification. [Copyright &y& Elsevier]

This paper explores the ability of factor models to predict the dynamics of US and UK interest rate swap spreads within a linear and a non-linear framework. We reject linearity for the US and UK swap spreads in favour of a regime-switching smooth transition vector autoregressive (STVAR) model, where the switching between regimes is controlled by the slope of the US term structure of interest rates. We compare the ability of the STVAR model to predict swap spreads with that of a non-linear nearest-neighbours model as well as that of linear AR and VAR models. We find some evidence that the non-linear models predict better than the linear ones. At short horizons, the nearest-neighbours (NN) model predicts better than the STVAR model US swap spreads in periods of increasing risk conditions and UK swap spreads in periods of decreasing risk conditions. At long horizons, the STVAR model increases its forecasting ability over the linear models, whereas the NN model does not outperform the rest of the models. Copyright © 2007 John Wiley & Sons, Ltd. [ABSTRACT FROM AUTHOR]

• A new algorithm for performing one-class classification. • First large-scale evaluation of one-class classification without tuning. • Identification of recommended hyperparameter values for range of algorithms. One-class classification is a challenging subfield of machine learning in which so-called data descriptors are used to predict membership of a class based solely on positive examples of that class, and no counter-examples. A number of data descriptors that have been shown to perform well in previous studies of one-class classification, like the Support Vector Machine (SVM), require setting one or more hyperparameters. There has been no systematic attempt to date to determine optimal default values for these hyperparameters, which limits their ease of use, especially in comparison with hyperparameter-free proposals like the Isolation Forest (IF). We address this issue by determining optimal default hyperparameter values across a collection of 246 one-class classification problems derived from 50 different real-world datasets. In addition, we propose a new data descriptor, Average Localised Proximity (ALP) to address certain issues with existing approaches based on nearest neighbour distances. Finally, we evaluate classification performance using a leave-one-dataset-out procedure, and find strong evidence that ALP outperforms IF and a number of other data descriptors, as well as weak evidence that it outperforms SVM, making ALP a good default choice. [ABSTRACT FROM AUTHOR]

This paper proposes a nonlinear term structure model that nests the discrete and Continuous time models as special Cases. I estimate the model non-parametrically using nearest-neighbours regression. In sample, the non-linear model matches the standard theories, but out of sample, it offers substantial improvement. Linear models fail to track future interest rates: a random walk dominates the forward rate as a predictor for 3-month Treasury bills. A non-linear forecast based on the spread is shown statistically to be the best forecast. [ABSTRACT FROM AUTHOR]