In this paper we propose a kernel based COBRA which is a direct approximation of the original COBRA. We propose a novel tuning procedure for original COBRA parameters based on this kernel approximation. We show that our proposed algorithm provides much better accuracy than other COBRAs and faster than usual Gridsearch COBRA. We use two datasets to illustrate our proposed methodology over existing COBRAs.
Split of the dataset
Dataflow
Pretrain Dataset
Train Set (or Proximity Set)
Original COBRA
$D_{k}$
$D_{l}$
Split Full Prox
$D_{k}$
$D_{k} \cup D_{l}$
No Split
$D_{k} \cup D_{l}$
$D_{k} \cup D_{l}$
Search Methods
Search Method
Description
Grid Search
Grid Search over the 100 grid points between the minimum and maximum prediction of Proximity Set Predictions
Random Search
Random Search over the 100 random points between the minimum and maximum prediction of Proximity Set Predictions
Gradient Descent
Gradient Descent over the 100 iterations both in proposed and Gradient COBRA
Datasets
The dataset taken from UCI regression dataset.
Dataset
Observations
Dimensions
airfoil
1503
5
autompg
398
7
breastcancer
198
33
california housing dataset
20640
8
concreteslump
103
7
energy
768
8
forest
517
12
servo
167
4
skillcraft
3395
19
sml
4137
26
yacht
308
6
File Description
File
Description
final.py
Run the base models, store the results, and distances for all dataset except California Housing and Boston Housing
final_other.py
Run the base models, store the results, and distances for California Housing and Boston Housing
final2.py
Run the proposed models, and othee COBRA search methods, store the results, and distances for all dataset except California Housing and Boston Housing
final2_other.py
Run the proposed models, and othee COBRA search methods, store the results, and distances for California Housing and Boston Housing
final2_randomized.py
Run the randomized search methods, store the results, and distances for all dataset except California Housing and Boston Housing
final2_other_randomized.py
Run the randomized search methods, store the results, and distances for California Housing and Boston Housing
