Linear Dynamical Systems

ldsr is an R package implementing two algorithms (Expectation-Maximization and Global Search) for learning a linear dynamical system. The algorithms are applied to the problem of reconstructing streamflow and catchment dynamics based on climate proxies (e.g., tree rings).

Publications

• Nguyen, H.T.T., Galelli, S. (2018). A linear dynamical systems approach to streamflow reconstruction reveals history of regime shifts in northern Thailand. Water Resources Research, 54 (3), 2057-2077.

AutoEncoders for Event Detection

AEED is a Python implementation of an event detection scheme based on AutoEncoders (a deep learning architecture). The scheme is conceived to detect anomalies in data simulated/observed in water distribution systems, but it can be readily applied to other domains.

Publications

• Taormina, R., Galelli, S. (2018) A deep learning approach to the detection and localization of cyber-physical attacks on water distribution systems. Journal of Water Resources Planning and Management, 144(10): 04018065. 2019 Best Research-Oriented Paper Award

Iterative Input variable Selection

IIS is a variable (feature) selection algorithm developed by Stefano Galelli and Andrea Castelletti. It is based on a regression/classification method–Extremely Randomised Trees–that ensures computational efficiency and scalability to high dimensional problems.

Publications

• Galelli, S., Castelletti, A. (2013) Tree-based Iterative Input variable Selection for hydrological modelling. Water Resources Research, 49(7), 4295-4310.
• Galelli, S., Castelletti, A. (2013) Assessing the predictive capability of randomized tree-based ensembles in streamflow modelling. Hydrology and Earth System Sciences, 17, 2669-2684.

(Quasi) Equally Informative Subsets Selection

This library implements the Wrapper for Quasi Equally Informative Subset Selection (W-QEISS) algorithm–developed by Gulsah Karakaya, Stefano Galelli, Selin Ahipasaoglu and Riccardo Taormina. The algorithm solves variable selection problems (for both classification and regression) and returns multiple subsets having similar predictive performance.

Publications

• Taormina, R., Galelli, S., Karakaya, G., Ahipasaoglu, S.D. (2016) An information theoretic approach to select alternate subsets of predictors for data-driven hydrological models. Journal of Hydrology, 542, 18-34.
• Karakaya, G., Galelli, S., Ahipasaoglu, S.D., Taormina, R. (2016) Identifying (quasi) equally informative subsets in feature selection problems for classication: a max-relevance min-redundancy approach. Cybernetics, IEEE Transactions on, 46(6), 1424-1437.