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.

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.

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.

Primordial Black Holes (PBHs) are gravitationally collapsed objects that may have been created by density fluctuations in the early universe and could have arbitrarily small masses down to the Planck scale. Hawking showed that due to quantum effects, a black hole has a temperature inversely proportional to its mass and will emit all species of fundamental particles thermally. PBHs with initial masses of similar to 5.0 x 10(14) g should be expiring in the present epoch with bursts of high-energy particles, including gamma radiation in the GeV-TeV energy range. The Milagro high energy observatory, which operated from 2000 to 2008, is sensitive to the high end of the PBH evaporation gamma-ray spectrum. Due to its large field-of-view, more than 90% duty cycle and sensitivity up to 100 TeV gamma rays, the Milagro observatory is well suited to perform a search for PBH bursts. Based on a search on the Milagro data, we report new PBH burst rate density upper limits over a range of PBH observation times. In addition, we report the sensitivity of the Milagro successor, the High Altitude Water Cherenkov (HAWC) observatory, to PBH evaporation events. Published by Elsevier B.V.

The Large Area Telescope (LAT) on board the Fermi Gamma-ray Space Telescope routinely detects the MeV-peaked flat-spectrum radio quasar PKS 1830−211 (z = 2.507). Its apparent isotropic γ-ray luminosity (E > 100 MeV), averaged over ∼3 years of observations and peaking on 2010 October 14/15 at 2.9 × 1050 erg s−1, makes it among the brightest high-redshift Fermi blazars. No published model with a single lens can account for all of the observed characteristics of this complex system. Based on radio observations, one expects time-delayed variability to follow about 25 days after a primary flare, with flux about a factor of 1.5 less. Two large γ-ray flares of PKS 1830−211 have been detected by the LAT in the considered period, and no substantial evidence for such a delayed activity was found. This allows us to place a lower limit of about 6 on the γ-ray flux ratio between the two lensed images. Swift XRT observations from a dedicated Target of Opportunity program indicate a hard spectrum with no significant correlation of X-ray flux with the γ-ray variability. The spectral energy distribution can be modeled with inverse Compton scattering of thermal photons from the dusty torus. The implications of the LAT data in terms of variability, the lack of evident delayed flare events, and different radio and γ-ray flux ratios are discussed. Microlensing effects, absorption, size and location of the emitting regions, the complex mass distribution of the system, an energy-dependent inner structure of the source, and flux suppression by the lens galaxy for one image path may be considered as hypotheses for understanding our results.

Gamma-ray bursts can be used to test for the presence of spacetime foam—postulated in theories of quantum gravity. Quantum fluctuations would cause the photon speeds to vary, leading to 'fuzziness' and, consequently, Lorentz invariance violation. Wheeler's 'spacetime-foam'1 picture of quantum gravity (QG) suggests spacetime fuzziness (fluctuations leading to non-deterministic effects) at distances comparable to the Planck length, LPl ≈ 1.62 × 10−33 cm, the inverse (in natural units) of the Planck energy, EPl ≈ 1.22 × 1019 GeV. The resulting non-deterministic motion of photons on the Planck scale is expected to produce energy-dependent stochastic fluctuations in their speed. Such a stochastic deviation from the well-measured speed of light at low photon energies, c, should be contrasted with the possibility of an energy-dependent systematic, deterministic deviation. Such a systematic deviation, on which observations by the Fermi satellite set Planck-scale limits for linear energy dependence2, is more easily searched for than stochastic deviations. Here, for the first time, we place Planck-scale limits on the more generic spacetime-foam prediction of energy-dependent fuzziness in the speed of photons. Using high-energy observations from the Fermi Large Area Telescope (LAT) of gamma-ray burst GRB090510, we test a model in which photon speeds are distributed normally around c with a standard deviation proportional to the photon energy. We constrain the model's characteristic energy scale beyond the Planck scale at >2.8EPl(>1.6EPl), at 95% (99%) confidence. Our results set a benchmark constraint to be reckoned with by any QG model that features spacetime quantization.

Simultaneous Swift and Fermi observations of gamma-ray bursts (GRBs) offer a unique broadband view of their afterglow emission, spanning more than 10 decades in energy. We present the sample of X-ray flares observed by both Swift and Fermi during the first three years of Fermi operations. While bright in the X-ray band, X-ray flares are often undetected at lower (optical), and higher (MeV to GeV) energies. We show that this disfavors synchrotron self-Compton processes as the origin of the observed X-ray emission. We compare the broadband properties of X-ray flares with the standard late internal shock model, and find that in this scenario, X-ray flares can be produced by a late-time relativistic (Gamma > 50) outflow at radii R similar to 10(13)-10(14) cm. This conclusion holds only if the variability timescale is significantly shorter than the observed flare duration, and implies that X-ray flares can directly probe the activity of the GRB central engine.

We present the detections of 18 solar flares detected in high-energy gamma-rays (above 100 MeV) with the Fermi Large Area Telescope (LAT) during its first 4 yr of operation. This work suggests that particle acceleration up to very high energies in solar flares is more common than previously thought, occurring even in modest flares, and for longer durations. Interestingly, all these flares are associated with fairly fast coronal mass ejections (CMEs). We then describe the detailed temporal, spatial, and spectral characteristics of the first two long-lasting events: the 2011 March 7 flare, a moderate (M3.7) impulsive flare followed by slowly varying gamma-ray emission over 13 hr, and the 2011 June 7 M2.5 flare, which was followed by gamma-ray emission lasting for 2 hr. We compare the Fermi LAT data with X-ray and proton data measurements from GOES and RHESSI. We argue that the gamma-rays are more likely produced through pion decay than electron bremsstrahlung, and we find that the energy spectrum of the proton distribution softens during the extended emission of the 2011 March 7 flare. This would disfavor a trapping scenario for particles accelerated during the impulsive phase of the flare and point to a continuous acceleration process at play for the duration of the flares. CME shocks are known for accelerating the solar energetic particles (SEPs) observed in situ on similar timescales, but it might be challenging to explain the production of gamma-rays at the surface of the Sun while the CME is halfway to the Earth. A stochastic turbulence acceleration process occurring in the solar corona is another likely scenario. Detailed comparison of characteristics of SEPs and gamma-ray-emitting particles for several flares will be helpful to distinguish between these two possibilities.

Because they are bright and distant, Gamma-ray Bursts (GRBs) have been used for more than a decade to test propagation of photons and to constrain relevant Quantum Gravity (QG) models in which the velocity of photons in vacuum can depend on their energy. With its unprecedented sensitivity and energy coverage, the Fermi satellite has provided the most constraining results on the QG energy scale so far. In this talk, the latest results obtained from the analysis of four bright GRBs observed by the Large Area Telescope will be reviewed. These robust results, cross-checked using three different analysis techniques set the limit on QG energy scale at EQG,1>7.6 times the Planck energy for linear dispersion and EQG,2>1.3×1011GeV for quadratic dispersion (95% CL). After describing the data and the analysis techniques in use, results will be discussed and confronted to latest constraints obtained with Active Galactic Nuclei.

The Fermi Large Area Telescope (LAT) observed two bright X-class solar flares on 2012 March 7, and detected gamma-rays up to 4 GeV. We detected gamma-rays both during the impulsive and temporally-extended emission phases, with emission above 100 MeV lasting for approximately 20 hours. Accurate localization of the gamma-ray production site(s) coincide with the solar active region from which X-ray emissions associated with these flares originated. Our analysis of the >100 MeV gamma-ray emission shows a relatively rapid monotonic decrease in flux during the first hour of the impulsive phase, and a much slower, almost monotonic decrease in flux for the next 20 hours. The spectra can be adequately described by a power law with a high energy exponential cutoff, or as resulting from the decay of neutral pions produced by accelerated protons and ions with an isotropic power-law energy distribution. The required proton spectrum has a number index 3, with minor variations during the impulsive phase, while during the temporally extended phase the spectrum softens monotonically, starting with index 4. The >30 MeV proton flux and spectra observed near the Earth by the GOES satellites also show a monotonic flux decrease and spectral softening during the extended phase, but with a harder spectrum, with index 3. Based on the Fermi-LAT and GOES observations of the flux and spectral evolution of these bright flares, we explore the relative merits of prompt and continuous acceleration scenarios, hadronic and leptonic emission processes, and acceleration at the solar corona by the fast Coronal Mass Ejections (CME) as explanations for the observations. We conclude that the most likely scenario is continuous acceleration of protons in the solar corona which penetrate the lower solar atmosphere and produce pions that decay into gamma-rays.

The discovery of rapidly variable Very High Energy (VHE; E > 100 GeV) gamma-ray emission from 4C +21.35 (PKS 1222+216) by MAGIC on 2010 June 17, triggered by the high activity detected by the Fermi Large Area Telescope (LAT) in high energy (HE; E > 100 MeV) gamma-rays, poses intriguing questions on the location of the gamma-ray emitting region in this flat spectrum radio quasar. We present multifrequency data of 4C +21.35 collected from centimeter to VHE during 2010 to investigate the properties of this source and discuss a possible emission model. The first hint of detection at VHE was observed by MAGIC on 2010 May 3, soon after a gamma-ray flare detected by Fermi-LAT that peaked on April 29. The same emission mechanism may therefore be responsible for both the HE and VHE emission during the 2010 flaring episodes. Two optical peaks were detected on 2010 April 20 and June 30, close in time but not simultaneous with the two gamma-ray peaks, while no clear connection was observed between the X-ray an gamma-ray emission. An increasing flux density was observed in radio and mm bands from the beginning of 2009, in accordance with the increasing gamma-ray activity observed by Fermi-LAT, and peaking on 2011 January 27 in the mm regime (230 GHz). We model the spectral energy distributions (SEDs) of 4C +21.35 for the two periods of the VHE detection and a quiescent state, using a one-zone model with the emission coming from a very compact region outside the broad line region. The three SEDs can be fit with a combination of synchrotron self-Compton and external Compton emission of seed photons from a dust torus, changing only the electron distribution parameters between the epochs. The fit of the optical/UV part of the spectrum for 2010 April 29 seems to favor an inner disk radius of <6 gravitational radii, as one would expect from a prograde-rotating Kerr black hole.

This paper reports the results from three targeted searches of Milagro TeV sky maps: two extragalactic point source lists and one pulsar source list. The first extragalactic candidate list consists of 709 candidates selected from the Fermi-LAT 2FGL catalog. The second extragalactic candidate list contains 31 candidates selected from the TeVCat source catalog that have been detected by imaging atmospheric Cherenkov telescopes (IACTs). In both extragalactic candidate lists Mkn 421 was the only source detected by Milagro. This paper presents the Milagro TeV flux for Mkn 421 and flux limits for the brighter Fermi-LAT extragalactic sources and for all TeVCat candidates. The pulsar list extends a previously published Milagro targeted search for Galactic sources. With the 32 new gamma-ray pulsars identified in 2FGL, the number of pulsars that are studied by both Fermi-LAT and Milagro is increased to 52. In this sample, we find that the probability of Milagro detecting a TeV emission coincident with a pulsar increases with the GeV flux observed by the Fermi-LAT in the energy range from 0.1GeV to 100GeV.

M. Ackermann, M. Ajello, A. Albert, A. Allafort, L. Baldini, G. Barbiellini, 6 D. Bastieri, 8 K. Bechtol, R. Bellazzini, R. D. Blandford, E. D. Bloom, E. Bonamente, 10 E. Bottacini, A. Bouvier, T. J. Brandt, M. Brigida, 14 P. Bruel, R. Buehler, S. Buson, 8 G. A. Caliandro, 16 R. A. Cameron, P. A. Caraveo, C. Cecchi, 10 E. Charles, R.C.G. Chaves, A. Chekhtman, J. Chiang, G. Chiaro, S. Ciprini, 21 R. Claus, J. Cohen-Tanugi, J. Conrad, 24, 25, 26 S. Cutini, 21 M. Dalton, F. D’Ammando, A. de Angelis, F. de Palma, 14 C. D. Dermer, S. W. Digel, L. Di Venere, E. do Couto e Silva, P. S. Drell, A. Drlica-Wagner, C. Favuzzi, 14 S. J. Fegan, E. C. Ferrara, W. B. Focke, A. Franckowiak, Y. Fukazawa, S. Funk, ∗ P. Fusco, 14 F. Gargano, D. Gasparrini, 21 S. Germani, 10 N. Giglietto, 14 F. Giordano, 14 M. Giroletti, T. Glanzman, G. Godfrey, G. A. Gomez-Vargas, 34, 35 I. A. Grenier, J. E. Grove, S. Guiriec, 36 M. Gustafsson, D. Hadasch, Y. Hanabata, A. K. Harding, M. Hayashida, K. Hayashi, J.W. Hewitt, D. Horan, X. Hou, R. E. Hughes, Y. Inoue, M. S. Jackson, 24 T. Jogler, G. Jóhannesson, A. S. Johnson, T. Kamae, T. Kawano, J. Knödlseder, 45 M. Kuss, J. Lande, S. Larsson, 24, 46 L. Latronico, F. Longo, 6 F. Loparco, 14 M. N. Lovellette, P. Lubrano, 10 M. Mayer, M. N. Mazziotta, J. E. McEnery, 48 J. Mehault, P. F. Michelson, W. Mitthumsiri, 49, † T. Mizuno, A. A. Moiseev, 48 C. Monte, 14 M. E. Monzani, A. Morselli, I. V. Moskalenko, ‡ S. Murgia, R. Nemmen, 51, 53 E. Nuss, T. Ohsugi, A. Okumura, 54 M. Orienti, E. Orlando, J. F. Ormes, D. Paneque, 3 J. H. Panetta, J. S. Perkins, M. Pesce-Rollins, F. Piron, G. Pivato, T. A. Porter, S. Rainò, 14 R. Rando, 8 M. Razzano, 57 S. Razzaque, A. Reimer, 3 O. Reimer, 3 S. Ritz, M. Roth, M. Schaal, A. Schulz, C. Sgrò, E. J. Siskind, G. Spandre, P. Spinelli, 14 A. W. Strong, H. Takahashi, Y. Takeuchi, J. G. Thayer, J. B. Thayer, D. J. Thompson, L. Tibaldo, M. Tinivella, D. F. Torres, 65 G. Tosti, 10 E. Troja, 48 V. Tronconi, T. L. Usher, J. Vandenbroucke, V. Vasileiou, G. Vianello, V. Vitale, 66 M. Werner, B. L. Winer, K. S. Wood, M. Wood, and Z. Yang 24

Recent accurate measurements of cosmic-ray (CR) species by ATIC-2, CREAM, and PAMELA reveal an unexpected hardening in the proton and He spectra above a few hundred GeV, a gradual softening of the spectra just below a few hundred GeV, and a harder spectrum of He compared to that of protons. These newly discovered features may offer a clue to the origin of high-energy CRs. We use the Fermi Large Area Telescope observations of the γ-ray emission from Earth's limb for an indirect measurement of the local spectrum of CR protons in the energy range ∼90  GeV-6  TeV (derived from a photon energy range 15 GeV-1 TeV). Our analysis shows that single power law and broken power law spectra fit the data equally well and yield a proton spectrum with index 2.68±0.04 and 2.61±0.08 above ∼200  GeV, respectively.

TeV-flaring activity with timescales as short as tens of minutes and an orphan TeV flare have been observed from the blazar Markarian 421 (Mrk 421). The TeV emission from Mrk 421 is believed to be produced by leptonic synchrotron self-Compton (SSC) emission. In this scenario, correlations between the X-ray and the TeV fluxes are expected, TeV orphan flares are hardly explained, and the activity (measured as duty cycle) of the source at TeV energies is expected to be equal to or less than that observed in X-rays if only SSC is considered. To estimate the TeV duty cycle of Mrk 421 and to establish limits on its variability at different timescales, we continuously observed Mrk 421 with the Milagro observatory. Mrk 421 was detected by Milagro with a statistical significance of 7.1 standard deviations between 2005 September 21 and 2008 March 15. The observed spectrum is consistent with previous observations by VERITAS. We estimate the duty cycle of Mrk 421 for energies above 1 TeV for different hypotheses of the baseline flux and for different flare selections and we compared our results with the X-ray duty cycle estimated by Resconi et al. The robustness of the results is discussed.

Observations of occultations of bright γ-ray sources by the Sun may reveal predicted pair halos around blazars and/or new physics, such as, e.g., hypothetical light dark matter particles-axions. We use Fermi Gamma-Ray Space Telescope (Fermi) data to analyze four occultations of blazar 3C 279 by the Sun on October 8 each year from 2008 to 2011. A combined analysis of the observations of these occultations allows a point-like source at the position of 3C 279 to be detected with significance of ≈3σ, but does not reveal any significant excess over the flux expected from the quiescent Sun. The likelihood ratio test rules out complete transparency of the Sun to the blazar γ-ray emission at a 3σ confidence level.