This repository provide a realization of iP_DCA on Bilevel-SVM problem in Python with a simple test demo and a code for reproduce the numerical result of experiments.
The algorithm and the model are presented in the paper Difference of convex algorithms for bilevel programs with applications in hyperparameter selection
Beside "usual" packages (numpy), iP_DCA is built upon cvxpy.
In consideration of large-scale problem, scipy is recommanded for the usage of sparse matrix.
For ease of data reading, sklearn is also recommanded.
When sklearn.datasets.fetch_openml is used as in demo_realdata.py, pandas is also required due to the package dependency.
cvxpy contains several open-source solver including ECOS and SCS we used in our code, but it also provides easy-use interface for other solvers;
Here we recommand researchers to try the commercial solvers for moderate-size problem to attain a better efficiency;
However, for large-scale problem, SCS based on OpenMP may perform better with the default setting.
To give a try of our algorithm without any local data, one could call
python demo_realdata.py
To reproduce the numerical results in paper, one can run the following commands respectively
python experiments_realdata.py --solvers GUROBI --repeat_time 20 --data_scale moderate
python experiments_realdata.py --solvers open_source --repeat_time 20 --data_scale moderate
python experiments_realdata.py --solvers open_source --repeat_time 20 --data_scale large