Moving Targets is novel hybrid technique which can be used to inject external constraints and domain knowledge into any machine learning model. The algorithm is based on an iterative schemes that alternates between two steps: a learner step, which is in charge of training the machine learning model, and a master step, which refines the model predictions by projecting them into the feasible space, so that the model can be learnt again on these adjusted targets. Since the model is general-purpose, it can address any type of learner and constraints, given that these constraints lead to a solvable constrained optimization problem (see the original paper for details).
You can install the Moving Targets' library via pip by pasting this command in your terminal:
pip install moving-targets
The library contains various classes and utility functions, but the core of the algorithm can be found in the MACS
class, which needs both a Learner and a Master instance in order to be instantiated — additional parameters
such as the initial step or a list of Metric objects may be optionally passed as well. Once the MACS object
has been built, it can be used as a classical machine learning model along with its methods fit(),
predict(), and evaluate().
A predefined set of learners, metrics, and callbacks can be found in the respective modules. Instead, as regards the master, you will need to extend one of the template ones in order to formulate your constrained optimization problem, even though the master formulation is facilitated by a set of predefined loss functions and backend objects that wrap some of the most famous constraint optimization solvers — e.g., Cplex, Gurobi, CvxPy, etc.
More specific tutorials to learn how to use Moving Targets can be found in the examples module.
If you use Moving Targets in your works, please cite the following article:
@article{Detassis2020,
title={Teaching the Old Dog New Tricks: Supervised Learning with Constraints},
author={Detassis, Fabrizio and Lombardi, Michele and Milano, Michela},
journal={arXiv preprint arXiv:2002.10766},
year={2020}
}
- Maintainer: luca.giuliani13@unibo.it