Indra Adrianto

Kernel Logistic Regression Using Truncated Newton Method

Kernel logistic regression (KLR) is a powerful nonlinear classifier. The combination of KLR and the truncated-regularized iteratively re-weighted least-squares (TR-IRLS) algorithm, has led to a powerful classification method using small-to-medium size data sets. This method (algorithm), is called truncated-regularized kernel logistic regression (TR-KLR). Compared to support vector machines (SVM) and TR-IRLS on twelve benchmark publicly available data sets, the proposed TR-KLR algorithm is as accurate as, and much faster than, SVM and more accurate than TR-IRLS. The TR-KLR algorithm also has the advantage of providing direct prediction probabilities.

Active Learning with Support Vector Machines for Tornado Prediction

In this paper, active learning with support vector machines (SVMs) is applied to the problem of tornado prediction. This method is used to predict which storm-scale circulations yield tornadoes based on the radar derived Mesocyclone Detection Algorithm (MDA) and near-storm environment (NSE) attributes. The main goal of active learning is to choose the instances or data points that are important or have influence to our model to be labeled and included in the training set. We compare this method to passive learning with SVMs where the next instances to be included to the training set are randomly selected. The preliminary results show that active learning can achieve high performance and significantly reduce the size of training set.

Missing Data Imputation Through Machine Learning Algorithms

How to address missing data is an issue most researchers face. Computerized algorithms have been developed to ingest rectangular data sets, where the rows represent observations and the columns represent variables. These data matrices contain elements whose values are real numbers. In many data sets, some of the elements of the matrix are not observed. Quite often, missing observations arise from instrument failures, values that have not passed quality control criteria, etc. That leads to a quandary for the analyst using techniques that require a full data matrix. The first decision an analyst must make is whether the actual underlying values would have been observed if there was not an instrument failure, an extreme value, or some unknown reason. Since many programs expect complete data and the most economical way to achieve this is by deleting the observations with missing data, most often the analysis is performed on a subset of available data. This situation can become extreme in cases where a substantial portion of the data are missing or, worse, in cases where many variables exist with a seemingly small percentage of missing data. In such cases, large amounts of available data are discarded by deleting observations with one or more pieces of missing data. The importance of this problem arises

A spatiotemporal approach to tornado prediction

Automated tornado detection or prediction techniques in the literature have all been based on analyzing signatures of tornadoes that appear in Doppler radar velocity data. Attributes of these signatures are derived from radar data, as well as the near-storm environment, and associated with observed tornadoes. This associated database has been used to train neural networks and support vector machines to automatically classify radar signatures. In this paper, we formulate the tornado prediction problem differently. Instead of devising a machine intelligence approach to classify detections, we formulate the problem as a spatiotemporal one: of estimating the probability of a tornado event at a particular spatial location within a given time window. In this paper, we also describe our initial approach to addressing this differently formulated problem. We use a least-squares methodology to estimate shear, morphological image processing to estimate gradients, fuzzy logic to generate compact measures of tornado possibility and a classification neural network to generate the final spatio-temporal probability field.

Support vector machines for spatiotemporal tornado prediction

The use of support vector machines (SVMs) for predicting the location and time of tornadoes is presented. In this paper, we extend the work by Lakshmanan et al. (Proceedings of 2005 IEEE international joint conference on neural networks (Montreal, Canada), 3, 2005a, 1642–1647) to use a set of 33 storm days and introduce some variations that improve the results. The goal is to estimate the probability of a tornado event at a particular spatial location within a given time window. We utilize a least-squares methodology to estimate shear, quality control of radar reflectivity, morphological image processing to estimate gradients, fuzzy logic to generate compact measures of tornado possibility and SVM classification to generate the final spatiotemporal probability field. On the independent test set, this method achieves a Heidkes skill score of 0.60 and a critical success index of 0.45.

The p-Centre machine for regression analysis

Support vector machines (SVMs) have become one of the most powerful methods in machine learning for solving classification and regression problems. Finding the SVM solution can be regarded as estimating the centre of the largest hypersphere that can be inscribed in the set of consistent hypotheses called the version space. However, this solution can be inaccurate if the version space is asymmetric or elongated. Several approaches have been proposed to utilize other possible centres of the version space that can improve the generalization performance. Morreti in 2003 proposed an algorithm for finding the centre of a general polytope, the so called p-Centre, using weighted projections. By applying this method, Brückner in 2001 introduced a formulation for solving binary classification problems based on an approximation of the p-Centre of the version space, the so called p-Centre machine. In this paper, we extend the work by Brückner and propose a kernel-based algorithm for regression analysis using the p-Centre method. The concept of the p-Centre of a polytope and version space is also explained. Furthermore, the applications of the proposed method are presented. The preliminary results indicate that the p-Centre-based kernel machine for regression has promising performance compared with the SVM for regression.