|
|
Bertossi, L., & Geerts, F. (2020). Data Quality and Explainable AI. ACM J. Data Inf. Qual., 12(2), 11.
Abstract: In this work, we provide some insights and develop some ideas, with few technical details, about the role of explanations in Data Quality in the context of data-based machine learning models (ML). In this direction, there are, as expected, roles for causality, and explainable artificial intelligence. The latter area not only sheds light on the models, but also on the data that support model construction. There is also room for defining, identifying, and explaining errors in data, in particular, in ML, and also for suggesting repair actions. More generally, explanations can be used as a basis for defining dirty data in the context of ML, and measuring or quantifying them. We think dirtiness as relative to the ML task at hand, e.g., classification.
|
|
|
|
Carrasco, J. A., Carrasco, M., & Yanez, R. (2022). An inexpert expert. Appl. Econ. Lett., Early Access.
Abstract: We explore strategic information transmission when there is noise at the observation stage, when an expert observes signals, before he advises a policymaker. That is, the expert might be inexpert. We account for the fact that his signals might be totally uninformative, which is commonly known by players. We find that this inexpertise translates into a greater preference misalignment between players and that this yields a less informative equilibrium. We show that our results follow from the fact that the strategic effect of noise – the welfare change exclusive due to changes in the equilibrium partition – is always negative. Numerical simulations show that noise might be beneficial if the policymaker openly disagrees about noise chances. This makes the point that whether noise is beneficial or not crucially depends on how early in the game it arises, and also whether noise chances are commonly known by players or not.
|
|
|
|
Chadwick, C., Gironas, J., Gonzalez-Leiva, F., & Aedo, S. (2023). Bias adjustment to preserve changes in variability: the unbiased mapping of GCM changes. Hydrol. Sci., Early Access.
Abstract: Standard quantile mapping (QM) performs well, as a bias adjustment method, in removing historical climate biases, but it can significantly alter a global climate model (GCM) signal. Methods that do incorporate GCM changes commonly consider mean changes only. Quantile delta mapping (QDM) is an exception, as it explicitly preserves relative changes in the quantiles, but it might present biases in preserving GCMs changes in standard deviation. In this work we propose the unbiased quantile mapping (UQM) method, which by construction preserves GCM changes of the mean and the standard deviation. Synthetic experiments and four Chilean locations are used to compare the performance of UQM against QDM, QM, detrended quantile mapping, and scale distribution mapping. All the methods outperform QM, but a tradeoff exists between preserving the GCM relative changes in the quantiles (QDM is recommended in this case), or changes in the GCM moments (UQM is recommended in this case).
|
|
|
|
Fustos-Toribio, I., Manque-Roa, N., Vasquez Antipan, D., Hermosilla Sotomayor, M., & Gonzalez, V. L. (2022). Rainfall-induced landslide early warning system based on corrected mesoscale numerical models: an application for the southern Andes. Nat. Hazards Earth Syst. Sci., 22(6), 2169–2183.
Abstract: Rainfall-induced landslides (RILs) are an issue in the southern Andes nowadays. RILs cause loss of life and damage to critical infrastructure. Rainfall-induced landslide early warning systems (RILEWSs) can reduce and mitigate economic and social damages related to RIL events. The southern Andes do not have an operational-scale RILEWS yet. In this contribution, we present a pre-operational RILEWS based on the Weather and Research Forecast (WRF) model and geomorphological features coupled to logistic models in the southern Andes. The models have been forced using precipitation simulations. We correct the precipitation derived from WRF using 12 weather stations through a bias correction approach. The models were trained using 57 well-characterized RILs and validated by ROC analysis. We show that WRF has strong limitations in representing the spatial variability in the precipitation. Therefore, accurate precipitation needs a bias correction in the study zone. We used accurate precipitation simulation and slope, demonstrating a high predicting capacity (area under the curve, AUC, of 0.80). We conclude that our proposal could be suitable at an operational level under determined conditions. A reliable RIL database and operational weather networks that allow real-time correction of the mesoscale model in the implemented zone are needed. The RILEWSs could become a support to decision-makers during extreme-precipitation events related to climate change in the south of the Andes.
|
|
