arXiv subjects

Vlasios Vasileiou

Publications and source records attributed to Vlasios Vasileiou.

7 recordsLinked to original sources

arXiv2020

Metrics and methods for a systematic comparison of fairness-aware machine learning algorithms

Understanding and removing bias from the decisions made by machine learning models is essential to avoid discrimination against unprivileged groups. Despite recent progress in algorithmic fairness, there is still no clear answer as to which bias-mitigation approaches are most effective. Evaluation strategies are typically use-case specific, rely on data with unclear bias, and employ a fixed policy to convert model outputs to decision outcomes. To address these problems, we performed a systematic comparison of a number of popular fairness algorithms applicable to supervised classification. Our study is the most comprehensive of its kind. It utilizes three real and four synthetic datasets, and two different ways of converting model outputs to decisions. It considers fairness, predictive-performance, calibration quality, and speed of 28 different modelling pipelines, corresponding to both fairness-unaware and fairness-aware algorithms. We found that fairness-unaware algorithms typically fail to produce adequately fair models and that the simplest algorithms are not necessarily the fairest ones. We also found that fairness-aware algorithms can induce fairness without material drops in predictive power. Finally, we found that dataset idiosyncracies (e.g., degree of intrinsic unfairness, nature of correlations) do affect the performance of fairness-aware approaches. Our results allow the practitioner to narrow down the approach(es) they would like to adopt without having to know in advance their fairness requirements.

cs.LG↗

arXiv2020

Fairness by Explicability and Adversarial SHAP Learning

The ability to understand and trust the fairness of model predictions, particularly when considering the outcomes of unprivileged groups, is critical to the deployment and adoption of machine learning systems. SHAP values provide a unified framework for interpreting model predictions and feature attribution but do not address the problem of fairness directly. In this work, we propose a new definition of fairness that emphasises the role of an external auditor and model explicability. To satisfy this definition, we develop a framework for mitigating model bias using regularizations constructed from the SHAP values of an adversarial surrogate model. We focus on the binary classification task with a single unprivileged group and link our fairness explicability constraints to classical statistical fairness metrics. We demonstrate our approaches using gradient and adaptive boosting on: a synthetic dataset, the UCI Adult (Census) dataset and a real-world credit scoring dataset. The models produced were fairer and performant.

cs.LG↗

arXiv2020

Counterfactual fairness: removing direct effects through regularization

Building machine learning models that are fair with respect to an unprivileged group is a topical problem. Modern fairness-aware algorithms often ignore causal effects and enforce fairness through modifications applicable to only a subset of machine learning models. In this work, we propose a new definition of fairness that incorporates causality through the Controlled Direct Effect (CDE). We develop regularizations to tackle classical fairness measures and present a causal regularization that satisfies our new fairness definition by removing the impact of unprivileged group variables on the model outcomes as measured by the CDE. These regularizations are applicable to any model trained using by iteratively minimizing a loss through differentiation. We demonstrate our approaches using both gradient boosting and logistic regression on: a synthetic dataset, the UCI Adult (Census) Dataset, and a real-world credit-risk dataset. Our results were found to mitigate unfairness from the predictions with small reductions in model performance.

cs.AI↗

arXiv2013

Constraints on Lorentz Invariance Violation with Fermi-LAT Observations of Gamma-Ray Bursts

Some Quantum Gravity (QG) theories allow for a violation of Lorentz invariance (LIV), manifesting as a dependence of the velocity of light in vacuum on its energy. If such a dependence exists, then photons of different energies emitted together by a distant source will arrive at the Earth at different times. High-energy (GeV) transient emissions from distant astrophysical sources such as Gamma-ray Bursts (GRBs) and Active Galaxy Nuclei can be used to search for and constrain LIV. The Fermi collaboration has previously analyzed two GRBs in order to put constraints on the dispersion parameter in vacuum, and on the energy scale at which QG effects causing LIV may arise. We used three different methods on four bright GRBs observed by the Fermi-LAT to get more stringent and robust constraints. No delays have been detected and strong limits on the QG energy scale are derived: for linear dispersion we set tight constraints placing the QG energy scale above the Planck mass; a quadratic leading LIV effect is also constrained.

astro-ph.HE↗

arXiv2010

Constraining Lorentz Violation with Fermi

One of the cornerstones of special relativity is the postulate that all observers measure exactly the same photon speeds independently on the photon energies. However, a hypothesized structure of spacetime may alter this conclusion at ultra-small length scales. Even a tiny energy-dependent variation in the speed of light may be revealed, when accumulated over cosmological light-travel times, by high temporal-resolution measurements of sharp features in Gamma-Ray Burst lightcurves. We report the results of a study of the emission from GRB 090510 as detected by Fermi's LAT and GBM instruments that set unprecedented limits on the dependence of the speed of light on its energy.

hep-ph↗

arXiv2010

Constraining Lorentz Invariance Violation with Fermi

A cornerstone of special relativity is Lorentz Invariance, the postulate that all observers measure exactly the same photon speeds independently on the photon energies. However, a hypothesized structure of spacetime may alter this conclusion at ultra-small length scales, a possibility allowed in many of the Quantum-Gravity (QG) formalisms currently investigated. A generalized uncertainty principle suggests that such effects might occur for photon energies approaching the Planck energy, $E_{Planck}=M_{Planck} c^2 \simeq 1.22\times10^{19} GeV$. Even though all photons yet detected have energies $E_{ph}<<E_{Planck}$, even a tiny variation in the speed of light, when accumulated over cosmological light-travel times, may be revealed by high temporal-resolution measurements of sharp features in Gamma-Ray Burst (GRB) lightcurves. Here we report the results of a study using the emission from GRB 090510 as detected by \textit{Fermi}'s LAT and GBM instruments, that set unprecedented limits on the dependence of the speed of photons on their energy. We find that the mass/energy scale for a linear in energy dispersion must be well above the Planck scale, something that renders any affected QG models highly implausible.

astro-ph.HE↗

arXiv2008

A Blind Search for Bursts of Very High Enery Gamma Rays with Milagro

Milagro is a water-Cherenkov detector that observes the extended air showers produced by cosmic gamma rays of energies E>100GeV. The effective area of Milagro peaks at energies E~10TeV, however it is still large even down to a few hundred GeV (~10m^2 at 100GeV). The wide field of view (~2sr) and high duty cycle (>90%) of Milagro make it ideal for continuously monitoring the overhead sky for transient Very High Energy (VHE) emissions. This study searched the Milagro data for such emissions. Even though the search was optimized primarily for detecting the emission from Gamma-Ray Bursts (GRBs), it was still sensitive to the emission from the last stages of the evaporation of Primordial Black Holes or to any other kind of phenomena that produce bursts of VHE gamma rays. Measurements of the GRB spectra by satellites up to few tens of GeV showed no signs of a cutoff. Even though multiple instruments sensitive to $GeV/TeV$ gamma rays have performed observations of GRBs, there has not yet been a definitive detection of such an emission yet. One of the reasons for that is that gamma rays with energies E>~100GeV are attenuated by interactions with the extragalactic background light or are absorbed internally at the site of the burst. There are many models that predict VHE gamma-ray emission from GRBs. A detection or a constraint of such an emission can provide useful information on the mechanism and environment of GRBs. This study performed a blind search of the Milagro data of the last five years for bursts of VHE gamma rays with durations ranging from 100musec to 316sec. No GRB localization was provided by an external instrument. Instead, the whole dataset was thoroughly searched in time, space, and duration. No significant events were detected. Upper limits were placed on the VHE emission from GRBs.

astro-ph↗