Our aim in this paper is to detect bank clients involved in suspicious activities related to money laundering, using the graph of transactions of the bank. Although we have a labeled real dataset, our target is not only to obtain relevant results on it, but also on random graphs in which typical anomaly patterns have been injected. So, we want simultaneously adequacy to the real data and robustness. Our method is based on designing new features; the most important are those resulting from the reduced egonet, which is the subgraph that remains from an egonet after eliminating the nodes connected with a single edge to the center; another feature is built by appealing to random walks and serves as indicator of circular flows. Our features are added to usual egonet features and a general anomaly detection algorithm, in our case Isolation Forest, serves to detect the anomalies. Experiments on the real data and a comprehensive set of synthetic data show that our approach is adequate, robust and better than some previous methods.

An algorithm for moving average (MA) parameter estimation was proposed by Stoica et al. (see ibid. vol.48, p.1999-2012, 2000). Its key step (covariance fitting) is a semidefinite programming (SDP) problem with two convex constraints: one reflecting the real positiveness of the desired covariance sequence and the other having a second-order cone form. We analyze two parameterizations of a positive real sequence and show that there is a one-to-one correspondence between them. We also show that the dual of the covariance fitting problem has a significantly smaller number of variables and, thus, a much reduced computational complexity. We discuss in detail the formulations that are best suited for the currently available semidefinite quadratic programming packages. Experimental results show that the execution times of the newly proposed algorithms scale well with the MA order, which are therefore convenient for large-order MA signals.

In this paper, we consider infinite impulse response (IIR) filter design where both magnitude and phase are optimized using a weighted and sampled least-squares criterion. We propose a new convex stability domain defined by positive realness for ensuring the stability of the filter and adapt the Steiglitz-McBride (SM), Gauss-Newton (GN), and classical descent methods to the new stability domain. We show how to describe the stability domain such that the description is suited to semidefinite programming and is implementable exactly; in addition, we prove that this domain contains the domain given by Rouche/spl acute/'s theorem. Finally, we give experimental evidence that the best designs are usually obtained with a multistage algorithm, where the three above methods are used in succession, each one being initialized with the result of the previous and where the positive realness stability domain is used instead of that defined by Rouche/spl acute/'s theorem.

The problem of training a dictionary for sparse representations from a given dataset is receiving a lot of attention mainly due to its applications in the fields of coding, classification and pattern recognition. One of the open questions is how to choose the number of atoms in the dictionary: if the dictionary is too small then the representation errors are big and if the dictionary is too big then using it becomes computationally expensive. In this letter, we solve the problem of computing efficient dictionaries of reduced size by a new design method, called Stagewise K-SVD, which is an adaptation of the popular K-SVD algorithm. Since K-SVD performs very well in practice, we use K-SVD steps to gradually build dictionaries that fulfill an imposed error constraint. The conceptual simplicity of the method makes it easy to apply, while the numerical experiments highlight its efficiency for different overcomplete dictionaries.

We propose a characterization of multivariate trigonometric polynomials that are positive on a given frequency domain. The positive polynomials are parameterized as a linear function of sum-of-squares polynomials and so semidefinite programming (SDP) is applicable. The frequency domain is expressed via the positivity of some trigonometric polynomials. We also give a bounded real lemma (BRL) in which a bounding condition on the magnitude of the frequency response of a multidimensional finite-impulse-response (FIR) filter is expressed as a linear matrix inequality (LMI). This BRL avoids the problem of a lack of spectral factorization in the multidimensional case. All the proposed theoretical contributions can be implemented only as sufficient conditions, due to degree limitations on the sum-of-square polynomials. However, the two-dimensional (2-D) FIR filter designs we study numerically suggest that these limitations have negligible impact on the optimality