Search

This work investigates into cost behaviors of binary classification measures in a background of class-imbalanced problems. Twelve performance measures are studied, such as F measure, G-means in terms of accuracy rates, and of recall and precision, balance error rate (BER), Matthews correlation coefficient (MCC), Kappa coefficient, etc. A new perspective is presented for those measures by revealing their cost functions with respect to the class imbalance ratio. Basically, they are described by four types of cost functions. The functions provides a theoretical understanding why some measures are suitable for dealing with class-imbalanced problems. Based on their cost functions, we are able to conclude that G-means of accuracy rates and BER are suitable measures because they show "proper" cost behaviors in terms of "a misclassification from a small class will cause a greater cost than that from a large class". On the contrary, F1 measure, G-means of recall and precision, MCC and Kappa coefficient measures do not produce such behaviors so that they are unsuitable to serve our goal in dealing with the problems properly., Comment: 7 pages, 5 figures, 3 Tables

Unsupervised ranking faces one critical challenge in evaluation applications, that is, no ground truth is available. When PageRank and its variants show a good solution in related subjects, they are applicable only for ranking from link-structure data. In this work, we focus on unsupervised ranking from multi-attribute data which is also common in evaluation tasks. To overcome the challenge, we propose five essential meta-rules for the design and assessment of unsupervised ranking approaches: scale and translation invariance, strict monotonicity, linear/nonlinear capacities, smoothness, and explicitness of parameter size. These meta-rules are regarded as high level knowledge for unsupervised ranking tasks. Inspired by the works in [8] and [14], we propose a ranking principal curve (RPC) model, which learns a one-dimensional manifold function to perform unsupervised ranking tasks on multi-attribute observations. Furthermore, the RPC is modeled to be a cubic B\'ezier curve with control points restricted in the interior of a hypercube, thereby complying with all the five meta-rules to infer a reasonable ranking list. With control points as the model parameters, one is able to understand the learned manifold and to interpret the ranking list semantically. Numerical experiments of the presented RPC model are conducted on two open datasets of different ranking applications. In comparison with the state-of-the-art approaches, the new model is able to show more reasonable ranking lists., Comment: This paper has 14 pages and 9 figures. The paper has submitted to IEEE Transactions on Knowledge and Data Engineering (TKDE)

Let $g$ be a Riemannian metric for $\mathbf{R}^d$ ($d\geq 3$) which differs from the Euclidean metric only in a smooth and strictly convex bounded domain $M$. The lens rigidity problem is concerned with recovering the metric $g$ inside $M$ from the corresponding lens relation on the boundary $\partial M$. In this paper, the stability of the lens rigidity problem is investigated for metrics which are a priori close to a given non-trapping metric satisfying "strong fold-regular" condition. A metric $g$ is called strong fold-regular if for each point $x\in M$, there exists a set of geodesics passing through $x$ whose conormal bundle covers $T^*_{x}M$. Moreover, these geodesics contain either no conjugate points or only fold conjugate points with a non-degeneracy condition. Examples of strong fold-regular metrics are constructed and are expected to be generic. Our main result gives the first stability result for the lens rigidity problem in the case of anisotropic metrics which allow conjugate points. The approach is based on the study of the linearized inverse problem of recovering a metric from its induced geodesic flow, which is a weighted geodesic X-ray transform problem for symmetric 2-tensor fields. A key ingredient is to show that the kernel of the X-ray transform on symmetric solenoidal 2-tensor fields is of finite dimension. It remains open whether the kernel space is trivial or not., Comment: A minor modification compared to the previous version to take into account the update of the references

In this work, we define cost-free learning (CFL) formally in comparison with cost-sensitive learning (CSL). The main difference between them is that a CFL approach seeks optimal classification results without requiring any cost information, even in the class imbalance problem. In fact, several CFL approaches exist in the related studies, such as sampling and some criteria-based pproaches. However, to our best knowledge, none of the existing CFL and CSL approaches are able to process the abstaining classifications properly when no information is given about errors and rejects. Based on information theory, we propose a novel CFL which seeks to maximize normalized mutual information of the targets and the decision outputs of classifiers. Using the strategy, we can deal with binary/multi-class classifications with/without abstaining. Significant features are observed from the new strategy. While the degree of class imbalance is changing, the proposed strategy is able to balance the errors and rejects accordingly and automatically. Another advantage of the strategy is its ability of deriving optimal rejection thresholds for abstaining classifications and the "equivalent" costs in binary classifications. The connection between rejection thresholds and ROC curve is explored. Empirical investigation is made on several benchmark data sets in comparison with other existing approaches. The classification results demonstrate a promising perspective of the strategy in machine learning., Comment: Classification, class imbalance, cost-free learning, cost-sensitive learning, abstaining, mutual information, ROC

The regularized near-cloak via the transformation optics approach in the time-harmonic electromagnetic scattering is considered. This work extends the existing studies mainly in two aspects. First, it presents a near-cloak construction by incorporating a much more general conducting layer between the cloaked and cloaking regions. This might be of significant practical importance when production fluctuations occur. Second, it allows the cloaked contents to be both passive and active with an applied current inside. In assessing the near-cloaking performance, comprehensive and sharp estimates are derived for the scattering amplitude in terms of the asymptotic regularization parameter and the material tensors of the conducting layer. The scattering estimates are independent of the passive/active contents being cloaked, which implies that one could nearly cloak arbitrary contents by using the proposed near-cloak construction., Comment: Any comments are welcome

The existing upper and lower bounds between entropy and error are mostly derived through an inequality means without linking to joint distributions. In fact, from either theoretical or application viewpoint, there exists a need to achieve a complete set of interpretations to the bounds in relation to joint distributions. For this reason, in this work we propose a new approach of deriving the bounds between entropy and error from a joint distribution. The specific case study is given on binary classifications, which can justify the need of the proposed approach. Two basic types of classification errors are investigated, namely, the Bayesian and non-Bayesian errors. For both errors, we derive the closed-form expressions of upper bound and lower bound in relation to joint distributions. The solutions show that Fano's lower bound is an exact bound for any type of errors in a relation diagram of "Error Probability vs. Conditional Entropy". A new upper bound for the Bayesian error is derived with respect to the minimum prior probability, which is generally tighter than Kovalevskij's upper bound., Comment: 11 pages, 5 figures, 2 Appendixes including Maple code. A refined version based on the previous version entitled "Analytical bounds between entropy and error probability in binary classifications", appeared as arXiv:1205.6602v1[cs.IT] in May 30, 2012

The paper investigates the sensitivity of the inverse problem of recovering the velocity field in a bounded domain from the boundary dynamic Dirichlet-to-Neumann map (DDtN) for the wave equation. Three main results are obtained: (1) assuming that two velocity fields are non-trapping and are equal to a constant near the boundary, it is shown that the two induced scattering relations must be identical if their corresponding DDtN maps are sufficiently close; (2) a geodesic X-ray transform operator with matrix-valued weight is introduced by linearizing the operator which associates each velocity field with its induced Hamiltonian flow. A selected set of geodesics whose conormal bundle can cover the cotangent space at an interior point is used to recover the singularity of the X-ray transformed function at the point; a local stability estimate is established for this case. Although fold caustics are allowed along these geodesics, it is required that these caustics contribute to a smoother term in the transform than the point itself. The existence of such a set of geodesics is guaranteed under some natural assumptions in dimension greater than or equal to three by the classification result on caustics and regularity theory of Fourier Integral Operators. The interior point with the above required set of geodesics is called "fold-regular"; (3) assuming that a background velocity field with every interior point fold-regular is fixed and another velocity field is sufficiently close to it and satisfies a certain orthogonality condition, it is shown that if the two corresponding DDtN maps are sufficiently close then they must be equal., Comment: 32 pages

The approximate cloaking is investigated for time-harmonic Maxwell's equations via the approach of transformation optics. The problem is reduced to certain boundary effect estimates due to an inhomogeneous electromagnetic inclusion with an asymptotically small support but an arbitrary content enclosed by a thin high-conducting layer. Sharp estimates are established in terms of the asymptotic parameter, which are independent of the material tensors of the small electromagnetic inclusion. The result implies that the `blow-up-a-small-region' construction via the transformation optics approach yields a near-cloak for the electromagnetic waves. A novelty lies in the fact that the geometry of the cloaking construction of this work can be very general. Moreover, by incorporating the conducting layer developed in the present paper right between the cloaked region and the cloaking region, arbitrary electromagnetic contents can be nearly cloaked. Our mathematical technique extends the general one developed in [30] for nearly cloaking scalar optics. In order to investigate the approximate electromagnetic cloaking for general geometries with arbitrary cloaked contents, new techniques and analysis tools must be developed for this more challenging vector optics case.

The existing upper and lower bounds between entropy and error probability are mostly derived from the inequality of the entropy relations, which could introduce approximations into the analysis. We derive analytical bounds based on the closed-form solutions of conditional entropy without involving any approximation. Two basic types of classification errors are investigated in the context of binary classification problems, namely, Bayesian and non-Bayesian errors. We theoretically confirm that Fano's lower bound is an exact lower bound for any types of classifier in a relation diagram of "error probability vs. conditional entropy". The analytical upper bounds are achieved with respect to the minimum prior probability, which are tighter than Kovalevskij's upper bound., Comment: 7 Pages, 2 Figures, Maple code

This work presents a systematic study of objective evaluations of abstaining classifications using Information-Theoretic Measures (ITMs). First, we define objective measures for which they do not depend on any free parameter. This definition provides technical simplicity for examining "objectivity" or "subjectivity" directly to classification evaluations. Second, we propose twenty four normalized ITMs, derived from either mutual information, divergence, or cross-entropy, for investigation. Contrary to conventional performance measures that apply empirical formulas based on users' intuitions or preferences, the ITMs are theoretically more sound for realizing objective evaluations of classifications. We apply them to distinguish "error types" and "reject types" in binary classifications without the need for input data of cost terms. Third, to better understand and select the ITMs, we suggest three desirable features for classification assessment measures, which appear more crucial and appealing from the viewpoint of classification applications. Using these features as "meta-measures", we can reveal the advantages and limitations of ITMs from a higher level of evaluation knowledge. Numerical examples are given to corroborate our claims and compare the differences among the proposed measures. The best measure is selected in terms of the meta-measures, and its specific properties regarding error types and reject types are analytically derived., Comment: 25 Pages, 1 Figure, 10 Tables

In this study, both Bayesian classifiers and mutual information classifiers are examined for binary classifications with or without a reject option. The general decision rules in terms of distinctions on error types and reject types are derived for Bayesian classifiers. A formal analysis is conducted to reveal the parameter redundancy of cost terms when abstaining classifications are enforced. The redundancy implies an intrinsic problem of "non-consistency" for interpreting cost terms. If no data is given to the cost terms, we demonstrate the weakness of Bayesian classifiers in class-imbalanced classifications. On the contrary, mutual-information classifiers are able to provide an objective solution from the given data, which shows a reasonable balance among error types and reject types. Numerical examples of using two types of classifiers are given for confirming the theoretical differences, including the extremely-class-imbalanced cases. Finally, we briefly summarize the Bayesian classifiers and mutual-information classifiers in terms of their application advantages, respectively., Comment: (2nd version: 19 pages, 5 figures, 7 tables. Theorems on Bayesian classifiers are extended to multiple variables. Appendix B for "Tighter bounds between the conditional entropy and Bayesian error in binary classifications" are added, in which Fano's bound is shown numerically to be very tight)

This correspondence studies the basic problem of classifications - how to evaluate different classifiers. Although the conventional performance indexes, such as accuracy, are commonly used in classifier selection or evaluation, information-based criteria, such as mutual information, are becoming popular in feature/model selections. In this work, we propose to assess classifiers in terms of normalized mutual information (NI), which is novel and well defined in a compact range for classifier evaluation. We derive close-form relations of normalized mutual information with respect to accuracy, precision, and recall in binary classifications. By exploring the relations among them, we reveal that NI is actually a set of nonlinear functions, with a concordant power-exponent form, to each performance index. The relations can also be expressed with respect to precision and recall, or to false alarm and hitting rate (recall)., Comment: 8 pages, 8 figures, and 2 tables

The electrostatic charging of the test mass in ASTROD I (Astrodynamical Space Test of Relativity using Optical Devices I) mission can affect the quality of the science data as a result of spurious Coulomb and Lorentz forces. To estimate the size of the resultant disturbances, credible predictions of charging rates and the charging noise are required. Using the GEANT4 software toolkit, we present a detailed Monte Carlo simulation of the ASTROD I test mass charging due to exposure of the spacecraft to galactic cosmic-ray (GCR) protons and alpha particles (3He, 4He) in the space environment. A positive charging rate of 33.3 e+/s at solar minimum is obtained. This figure reduces by 50% at solar maximum. Based on this charging rate and factoring in the contribution of minor cosmic-ray components, we calculate the acceleration noise and stiffness associated with charging. We conclude that the acceleration noise arising from Coulomb and Lorentz effects are well below the ASTROD I acceleration noise limit at 0.1 mHz both at solar minimum and maximum. The coherent Fourier components due to charging are investigated, it needs to be studied carefully in order to ensure that these do not compromise the quality of science data in the ASTROD I mission., Comment: 20 pages, 14 figures, submitted to International Journal of Modern Physics D

As ASTROD I travels through space, its test mass will accrue charge due to galactic cosmic-rays and solar energetic particles incident on the spacecraft. This test mass charge will result in Coulomb forces between the test mass and the surrounding electrodes. In earlier work using the GEANT4 toolkit, we predicted a net charging rate of nearly 9.0 +e/s from cosmic-ray protons between 0.1 and 1000 GeV at solar maximum, and rising to 26.5 +e/s at solar minimum. Here we use GEANT4 to simulate the charging process due to solar energetic particle events and to estimate the magnitude of acceleration noise due to this charging. The predicted charging rates range from 2840 to 64300 +e/s, at peak intensity, for the 4 largest SEP events in September and October 1989. For the 2 larger events, the acceleration disturbances due to charging exceeds the ASTROD I acceleration noise budget requirement. Continuous discharge should be considered for suppressing this charging noise. The acceleration noise during the 2 small events is well below the design target, although during these events, the net charging rate will be dominated by these solar fluxes., Comment: 10 pages,8 figures, COSPAR2006 H0.1-1, submitted

ASTROD I is planned as a single spacecraft mission. It will use interferometric and pulse ranging techniques between the spacecraft and ground stations, to make high precision measurements of the parameters that describe the solar system, and to test relativistic gravity with improved accuracy. At the heart of the spacecraft is a test mass, which the spacecraft will follow using a drag-free control system. The mission critically depends on maintaining the geodesic motion of the test mass. Charging of the test mass due to cosmic rays and solar particles will disturb its geodesic motion. We have modelled the charging process using the GEANT4 toolkit and a simplified, geometrical model and estimate that the ASTROD I test mass will charge positively, at a rate of 24 +/-7 e+/s, due to cosmic ray protons and alpha particles (3He and 4He) at solar minimum. We have used the results of this simulation to estimate the magnitude of disturbances associated with test mass charging, for the worst-case scenario, taking into account uncertainties in the model and potential charging contributions from minor cosmic-ray components., Comment: 17 pages, 11 figures, submitted to General Relativity and Gravitation