Random Forest probabilities to uncertainty measure

Description

This repository contains a piece of Python code used to compute "uncertainty level" for a Random Forest (RF) probabilistic output.

From a input .csv similar to the following :

cat	ACS	BARE	PLAN	UNPLAN	VEG
1	0.00100000000	0.26700000000	0.00100000000	0.00300000000	0.73000000000
2	0.00000000000	0.53100000000	0.00050000000	0.00300000000	0.46550000000
3	0.00000000000	0.13100000000	0.00550000000	0.01400000000	0.84950000000
4	0.10900000000	0.32000000000	0.11200000000	0.16650000000	0.29250000000
5	0.08000000000	0.00500000000	0.52000000000	0.39100000000	0.00400000000

It return a new .csv file :

cat	ACS	BARE	PLAN	UNPLAN	VEG	first_label	second_label	first_prop	second_prop	uncert_level
1	0.001	0.267	0.001	0.003	0.73	VEG	BARE	0.73	0.267	0.463
2	0.0	0.531	0.0005	0.003	0.4655	BARE	VEG	0.531	0.4655	0.0655
3	0.0	0.131	0.0055	0.014	0.8495	VEG	BARE	0.8495	0.131	0.7185
4	0.109	0.32	0.112	0.1665	0.2925	BARE	VEG	0.32	0.2925	0.0275
5	0.08	0.005	0.52	0.391	0.004	PLAN	UNPLAN	0.52	0.391	0.129

Here after are the meaning of the different new columns:

• first_label : Column with the highest probability
• second_label : Column with the second highest probability
• first_prop : Value of the highest probability
• second_prop : Value of the second highest probability
• uncert_level : "Uncertainty level" which is simply the difference between first_prop and second_prop.

A low value of the "uncert_level" indicates a high uncertainty of classification from RF.

Versions

V1.0 -> Python code used in the publication GRIPPA & al. 2018

Todos

• Add a version of the code that enables for the same table manipulation (cross-tab) with Postgresql (for very large dataset management).