Entropy and Gini Trees

Basic decision tree model on gas sensor array data. Two trees were constructed based on Entropy and Gini split criteria in R. Results are presented below.

Examples of how to read feature names on the tree

S2DR - Direct Resistance of sensor 2

S11NDR - Normalized Direct Resistance of sensor 11

S9I_1 - Exponential Moving Average with a smoothing parameter of 0.1 when current is increasing for Sensor 9

S9D_01 - Exponential Moving Average with a smoothing parameter of 0.01 when current is decreasing for Sensor 9

CONC - Concentration (ppmv)

Entropy Split Tree

Optimum Number of Splits - 14

Misclassification Rate - 16.28% (on test data)

Entropy Confusion Matrix

Classes - 1: Ethanol; 2: Ethylene; 3: Ammonia; 4: Acetaldehyde; 5: Acetone; 6: Toluene

---

Gini Split Tree

Optimum Number of Splits - 15

Misclassification Rate - 15.49% (on test data)

Gini Confusion Matrix

Classes - 1: Ethanol; 2: Ethylene; 3: Ammonia; 4: Acetaldehyde; 5: Acetone; 6: Toluene

ROC Curves

Area Under Curve

Entropy Tree - 0.9548684

Gini Tree - 0.9592732