Greedy feature selection: Classifier-dependent feature selection via greedy methods
Authors:
Fabiana Camattari,
Sabrina Guastavino,
Francesco Marchetti,
Michele Piana,
Emma Perracchione
Abstract:
The purpose of this study is to introduce a new approach to feature ranking for classification tasks, called in what follows greedy feature selection. In statistical learning, feature selection is usually realized by means of methods that are independent of the classifier applied to perform the prediction using that reduced number of features. Instead, greedy feature selection identifies the most…
▽ More
The purpose of this study is to introduce a new approach to feature ranking for classification tasks, called in what follows greedy feature selection. In statistical learning, feature selection is usually realized by means of methods that are independent of the classifier applied to perform the prediction using that reduced number of features. Instead, greedy feature selection identifies the most important feature at each step and according to the selected classifier. In the paper, the benefits of such scheme are investigated theoretically in terms of model capacity indicators, such as the Vapnik-Chervonenkis (VC) dimension or the kernel alignment, and tested numerically by considering its application to the problem of predicting geo-effective manifestations of the active Sun.
△ Less
Submitted 8 March, 2024;
originally announced March 2024.
Unbiased CLEAN for STIX in Solar Orbiter
Authors:
Emma Perracchione,
Fabiana Camattari,
Anna Volpara,
Paolo Massa,
Anna Maria Massone,
Michele Piana
Abstract:
Aims: To formulate, implement, and validate a user-independent release of CLEAN for Fourier-based image reconstruction of hard X-rays flaring sources. Methods: CLEAN is an iterative deconvolution method for radio and hard X-ray solar imaging. In a specific step of its pipeline, CLEAN requires the convolution between an idealized version of the instrumental Point Spread Function (PSF), and a map co…
▽ More
Aims: To formulate, implement, and validate a user-independent release of CLEAN for Fourier-based image reconstruction of hard X-rays flaring sources. Methods: CLEAN is an iterative deconvolution method for radio and hard X-ray solar imaging. In a specific step of its pipeline, CLEAN requires the convolution between an idealized version of the instrumental Point Spread Function (PSF), and a map collecting point sources located at positions on the solar disk from where most of the flaring radiation is emitted. This convolution step has highly heuristic motivations and the shape of the idealized PSF, which depends on the user's choice, impacts the shape of the overall reconstruction. Here we propose the use of an interpolation/extrapolation process to avoid this user-dependent step, and to realize a completely unbiased version of CLEAN. Results: Applications to observations recorded by the Spectrometer/Telescope for Imaging X-rays (STIX) on-board Solar Orbiter show that this unbiased release of CLEAN outperforms the standard version of the algorithm in terms of both automation and reconstruction reliability, with reconstructions whose accuracy is in line with the one offered by other imaging methods developed in the STIX framework. Conclusions: This unbiased version of CLEAN proposes a feasible solution to a well-known open issue concerning CLEAN, i.e., its low degree of automation. Further, this study provided the first application of an interpolation/extrapolation approach to image reconstruction from STIX experimental visibilities.
△ Less
Submitted 18 July, 2023;
originally announced July 2023.
