-
Metrics and methods for a systematic comparison of fairness-aware machine learning algorithms
Authors:
Gareth P. Jones,
James M. Hickey,
Pietro G. Di Stefano,
Charanpal Dhanjal,
Laura C. Stoddart,
Vlasios Vasileiou
Abstract:
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…
▽ More
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.
△ Less
Submitted 8 October, 2020;
originally announced October 2020.
-
Fairness by Explicability and Adversarial SHAP Learning
Authors:
James M. Hickey,
Pietro G. Di Stefano,
Vlasios Vasileiou
Abstract:
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 definit…
▽ More
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.
△ Less
Submitted 26 June, 2020; v1 submitted 11 March, 2020;
originally announced March 2020.
-
Counterfactual fairness: removing direct effects through regularization
Authors:
Pietro G. Di Stefano,
James M. Hickey,
Vlasios Vasileiou
Abstract:
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 deve…
▽ More
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.
△ Less
Submitted 26 February, 2020; v1 submitted 25 February, 2020;
originally announced February 2020.
-
Intermittency and dynamical Lee-Yang zeros of open quantum systems
Authors:
James M. Hickey,
Christian Flindt,
Juan P. Garrahan
Abstract:
We use high-order cumulants to investigate the Lee-Yang zeros of generating functions of dynamical observables in open quantum systems. At long times the generating functions take on a large deviation form with singularities of the associated cumulant generating functions - or dynamical free energies - signifying phase transitions in the ensemble of dynamical trajectories. We consider a driven thr…
▽ More
We use high-order cumulants to investigate the Lee-Yang zeros of generating functions of dynamical observables in open quantum systems. At long times the generating functions take on a large deviation form with singularities of the associated cumulant generating functions - or dynamical free energies - signifying phase transitions in the ensemble of dynamical trajectories. We consider a driven three-level system as well as the dissipative Ising model. Both systems exhibit dynamical intermittency in the statistics of quantum jumps. From the short-time behavior of the dynamical Lee-Yang zeros we identify critical values of the counting field which we attribute to the observed intermittency and dynamical phase co-existence. Furthermore, for the dissipative Ising model we construct a trajectory phase diagram and estimate the value of the transverse field where the stationary state changes from being ferromagnetic (inactive) to being paramagnetic (active).
△ Less
Submitted 7 January, 2015; v1 submitted 15 July, 2014;
originally announced July 2014.
-
Signatures of many-body localisation in a system without disorder and the relation to a glass transition
Authors:
James M. Hickey,
Sam Genway,
Juan P. Garrahan
Abstract:
We study a quantum spin system with local bilinear interactions and without quenched disorder which seems to display characteristic signatures of a many-body localisation (MBL) transition. From direct diagonalisation of small systems, we find a change in certain dynamical and spectral properties at a critical value of a coupling, from those characteristic of a thermalising phase to those character…
▽ More
We study a quantum spin system with local bilinear interactions and without quenched disorder which seems to display characteristic signatures of a many-body localisation (MBL) transition. From direct diagonalisation of small systems, we find a change in certain dynamical and spectral properties at a critical value of a coupling, from those characteristic of a thermalising phase to those characteristic of a MBL phase. The system we consider is known to have a quantum phase transition in its ground-state in the limit of large size, related to a first-order active-to-inactive phase transition in the stochastic trajectories of an associated classical model of glasses. Our results here suggest that this transition is present throughout the spectrum of the system in the large size limit. These findings may help understand the connection between MBL and structural glass transitions.
△ Less
Submitted 31 May, 2016; v1 submitted 22 May, 2014;
originally announced May 2014.
-
Timescales, dynamical phase transitions and 3rd order phase transitions in the 1d anisotropic XY model
Authors:
James M. Hickey
Abstract:
Recent research into time-integrated observables has revealed a special class of states which cap- ture the singular features of the generating functions of those observables, as estimated by full counting statistics (FCS). In this work we extend the results of [Phys. Rev. B 87 184303 (2013)] to the 1d anisotropic XY -model and find a set of FCS critical points associated with the time-integrated…
▽ More
Recent research into time-integrated observables has revealed a special class of states which cap- ture the singular features of the generating functions of those observables, as estimated by full counting statistics (FCS). In this work we extend the results of [Phys. Rev. B 87 184303 (2013)] to the 1d anisotropic XY -model and find a set of FCS critical points associated with the time-integrated transverse magnetization and anisotropy. We show dynamical phase transitions (DPTs) as defined in [Phys. Rev. Lett. 110 135704 (2013)] do not emerge on quenching the states associated with the time-integrated anisotropy across FCS critical points. We also study the timescales required to prepare the associated special states of the transverse magnetization using appropriate Markovian baths and find they are independent of the number of spins in the chain. However the probability to evolve to such a state decreases drastically with increasing system size due to dissipation. Thus when preparing such states it is preferable to use few body systems and to reach the thermodynamic limit it is necessary to use an ancillary system in conjunction with the Markovian baths.
△ Less
Submitted 21 March, 2014;
originally announced March 2014.
-
Cumulants of time-integrated observables of closed quantum systems and PT-symmetry, with an application to the quantum Ising chain
Authors:
James M. Hickey,
Emanuele Levi,
Juan P. Garrahan
Abstract:
We study the connection between the cumulants of a time-integrated observable of a quantum system and the PT-symmetry properties of the non-Hermitian deformation of the Hamiltonian from which the generating function of these cumulants is obtained. This non-Hermitian Hamiltonian can display regimes of broken and of unbroken PT-symmetry, depending on the parameters of the problem and on the counting…
▽ More
We study the connection between the cumulants of a time-integrated observable of a quantum system and the PT-symmetry properties of the non-Hermitian deformation of the Hamiltonian from which the generating function of these cumulants is obtained. This non-Hermitian Hamiltonian can display regimes of broken and of unbroken PT-symmetry, depending on the parameters of the problem and on the counting field that sets the strength of the non-Hermitian perturbation. This in turn determines the analytic structure of the long-time cumulant generating function (CGF) for the time-integrated observable. We consider in particular the case of the time-integrated (longitudinal) magnetisation in the one-dimensional Ising model in a transverse field. We show that its long-time CGF is singular on a curve in the magnetic field/counting field plane that delimits a regime where PT-symmetry is spontaneously broken (which includes the static ferromagnetic phase), from one where it is preserved (which includes the static paramagnetic phase). In the paramagnetic phase, conservation of PT -symmetry implies that all cumulants are sub-linear in time, a behaviour usually associated to the absence of decorrelation.
△ Less
Submitted 18 March, 2014;
originally announced March 2014.
-
Fluctuation Theorems and the Generalised Gibbs Ensemble in Integrable Systems
Authors:
James M. Hickey,
Sam Genway
Abstract:
We derive fluctuation relations for a many-body quantum system prepared in a Generalised Gibbs Ensemble subject to a general nonequilibrium protocol. By considering isolated integrable systems, we find generalisations to the Tasaki-Crooks and Jarzynski relations. Our approach is illustrated by studying the one-dimensional quantum Ising model subject to a sudden change in the transverse field, wher…
▽ More
We derive fluctuation relations for a many-body quantum system prepared in a Generalised Gibbs Ensemble subject to a general nonequilibrium protocol. By considering isolated integrable systems, we find generalisations to the Tasaki-Crooks and Jarzynski relations. Our approach is illustrated by studying the one-dimensional quantum Ising model subject to a sudden change in the transverse field, where we find that the statistics of the work done and irreversible entropy show signatures of quantum criticality. We discuss these fluctuation relations in the context of thermalisation.
△ Less
Submitted 14 March, 2014;
originally announced March 2014.
-
Dynamical phase transitions, time-integrated observables and geometry of states
Authors:
James M. Hickey,
Sam Genway,
Juan P. Garrahan
Abstract:
We show that there exist dynamical phase transitions (DPTs), as defined in [Phys. Rev. Lett. 110 135704 (2013)], in the transverse-field Ising model (TFIM) away from the static quantum critical points. We study a class of special states associated with singularities in the generating functions of time-integrated observables found in [Phys. Rev. B 88 184303 (2013)]. Studying the dynamics of these s…
▽ More
We show that there exist dynamical phase transitions (DPTs), as defined in [Phys. Rev. Lett. 110 135704 (2013)], in the transverse-field Ising model (TFIM) away from the static quantum critical points. We study a class of special states associated with singularities in the generating functions of time-integrated observables found in [Phys. Rev. B 88 184303 (2013)]. Studying the dynamics of these special states under the evolution of the TFIM Hamiltonian, we find temporal non-analtyicities in the initial-state return probability associated with dynamical phase transitions. By calculating the Berry phase and Chern number we show the set of special states have interesting geometric features similar to those associated with static quantum critical points.
△ Less
Submitted 6 September, 2013;
originally announced September 2013.
-
Trajectory phase transitions and dynamical Lee-Yang zeros of the Glauber-Ising chain
Authors:
James M. Hickey,
Christian Flindt,
Juan P. Garrahan
Abstract:
We examine the generating function of the time-integrated energy for the one-dimensional Glauber-Ising model. At long times, the generating function takes on a large-deviation form and the associated cumulant generating function has singularities corresponding to continuous trajec- tory (or "space-time") phase transitions between paramagnetic trajectories and ferromagnetically or anti-ferromagneti…
▽ More
We examine the generating function of the time-integrated energy for the one-dimensional Glauber-Ising model. At long times, the generating function takes on a large-deviation form and the associated cumulant generating function has singularities corresponding to continuous trajec- tory (or "space-time") phase transitions between paramagnetic trajectories and ferromagnetically or anti-ferromagnetically ordered trajectories. In the thermodynamic limit, the singularities make up a whole curve of critical points in the complex plane of the counting field. We evaluate analyti- cally the generating function by map** the generator of the biased dynamics to a non-Hermitian Hamiltonian of an associated quantum spin chain. We relate the trajectory phase transitions to the high-order cumulants of the time-integrated energy which we use to extract the dynamical Lee-Yang zeros of the generating function. This approach offers the possibility to detect continuous trajectory phase transitions from the finite-time behaviour of measurable quantities.
△ Less
Submitted 13 May, 2013;
originally announced May 2013.
-
Dynamical Phases in the Full Counting Statistics of the Resonant-Level Model
Authors:
Sam Genway,
James M. Hickey,
Juan P. Garrahan,
Andrew D. Armour
Abstract:
We present a thermodynamic formalism to study the full counting statistics (FCS) of charge transport through a quantum dot coupled to two leads in the resonant-level model. We show that a close analogue of equilibrium phase transitions exists for the statistics of transferred charge; by tuning an appropriate `counting field', crossovers to different dynamical phases are possible. Our description r…
▽ More
We present a thermodynamic formalism to study the full counting statistics (FCS) of charge transport through a quantum dot coupled to two leads in the resonant-level model. We show that a close analogue of equilibrium phase transitions exists for the statistics of transferred charge; by tuning an appropriate `counting field', crossovers to different dynamical phases are possible. Our description reveals a map** between the FCS of a given device and current measurements over a range of devices with different dot-lead coupling strengths. Further insight into features in the FCS is found by studying the occupation of the dot conditioned on the transported charge between the leads.
△ Less
Submitted 20 December, 2012;
originally announced December 2012.
-
Time-integrated observables as order parameters for dynamical phase transitions in closed quantum systems
Authors:
James M. Hickey,
Sam Genway,
Igor Lesanovsky,
Juan P. Garrahan
Abstract:
The dynamical behaviour of many-body systems is often richer than what can be anticipated from their static properties. Here we show that in closed quantum systems this becomes evident by considering time-integrated observables as order parameters. In particular, the analytic properties of their generating functions, as estimated by full-counting statistics, allow to identify dynamical phases, i.e…
▽ More
The dynamical behaviour of many-body systems is often richer than what can be anticipated from their static properties. Here we show that in closed quantum systems this becomes evident by considering time-integrated observables as order parameters. In particular, the analytic properties of their generating functions, as estimated by full-counting statistics, allow to identify dynamical phases, i.e. phases with specific fluctuation properties of time-integrated observables, and to locate the transitions between these phases. We discuss in detail the case of the quantum Ising chain in a transverse field. We show that this model displays a continuum of quantum dynamical transitions, of which the static transition is just an end point. These singularities are not a consequence of particular choices of initial conditions or other external non-equilibrium protocols such as quenches in coupling constants. They can be probed generically through quantum jump statistics of an associated open problem, and for the case of the quantum Ising chain we outline a possible experimental realisation of this scheme by digital quantum simulation with cold ions.
△ Less
Submitted 20 November, 2012;
originally announced November 2012.
-
Thermodynamics of Quadrature Trajectories in Open Quantum Systems
Authors:
James M. Hickey,
Sam Genway,
Igor Lesanovsky,
Juan P. Garrahan
Abstract:
We apply a large-deviation method to study the diffusive trajectories of the quadrature operators of light within a reservoir connected to dissipative quantum systems. We formulate the study of quadrature trajectories in terms of characteristic operators and show that in the long time limit the statistics of such trajectories obey a large-deviation principle. We take our motivation from homodyne d…
▽ More
We apply a large-deviation method to study the diffusive trajectories of the quadrature operators of light within a reservoir connected to dissipative quantum systems. We formulate the study of quadrature trajectories in terms of characteristic operators and show that in the long time limit the statistics of such trajectories obey a large-deviation principle. We take our motivation from homodyne detection schemes which allow the statistics of quadrature operator of the light field to be measured. We illustrate our approach with four examples of increasing complexity: a driven two-level system, a `blinking' three-level system, a pair of weakly-coupled two-level driven systems, and the micromaser. We discuss how quadrature operators can serve as alternative order parameters for the classification of dynamical phases, which is particularly useful in cases where the statistics of quantum jumps cannot distinguish between such phases. The formalism we introduce also allows us to analyse the properties of the light emitted by quantum jump trajectories which fluctuate far from the typical dynamics.
△ Less
Submitted 25 June, 2012;
originally announced June 2012.
