Randomized Information Coefficient

The Randomized Information Coefficient: Assessing Dependencies In Noisy Data

Please refer to this website:

https://sites.google.com/site/randinfocoeff/

Each experiments is available in its own folder: e.g. experiments about noisy relationships between variables are available in the folder ./sec51.

RIC source code as well as other statistic code can also be found under the folder ./Statistics. With regards to RIC:

• ./Statistics/RIC contains RIC between variables;
• ./Statistics/RIC/RICmultivariable/RICrndFerns contains RIC between sets of variables with random fern discretization;
• ./Statistics/RIC/RICmultivariable/RICrndSeeds contains RIC between sets of variables with random seeds discretization.

The code is supposed to run in Matlab and it was tested on a 64bit Linux system. Some statistics are implemented either in C or C++, thus their code might be re-compiled if another system is used. We provide the script compile_X_matlab.m to compile measures implemented in C or C++. Those are:

• The Randomized Information Coefficient (RIC) - compile_RIC_matlab.m;
• RIC with random fern discretization - compile_RICrndFerns_matlab.m;
• RIC with random seeds discretization requires to compile the procedure computeNMI - compile_computeNMI_matlab.m;
• The Maximal Information Coefficient (MIC) - compile_MINE_matlab.m;
• The Maximal Information Coefficient (e) (MICe) - compile_MINE_e_matlab.m;
• The Generalized Mean Information Coefficient (GMIC) - compile_MINE_matlab.m;
• The Total Information Coefficient (e) (TICe) - compile_MINE_e_matlab.m;
• The Mutual Information Dimension (MID) - compile_MID_matlab.m;
• The kNN Kraskov's Mutual Information Estimator - compile_MI_Kraskov_matlab.m.

To plot the results of any experiment use plotX.m, where X is a particular experiment present in a specific folder. You might want to run experiments on your own, some might be time demanding: e.g. the power experiment in Section 5.1 takes around 30 minutes on a server Xeon E5-2666 v3 with 36 cores . If you want to speed up results try to: decrease the number of simulations, bootstrap repetitions, folds for cross-validation or number of records n in the simulations.