kernels-2021

Kaggle-Ranking of Kernels (Public Leaderboard):

Method	Kaggle Accuracy
Linear-Regression	0.60466
SVM	0.55866
Kernel-Ridge-Regression	0.61933
Spectrum (k=6)	0.62600
Gappy (q=8, l=6)	0.62866
mismatch (m=1, k=8)	0.65400
mismatch (m=1, k=9)	0.67333

Final Ranking
Private Leaderboard: 0.67733
(mismatch (m=1, k=9))

Details about this code-base

Dependencies: cvxopt, scipy, numpy

You can run main.py as: python main.py or with: python main.py --dry for validation. In the latter case it needs sciki-learn for train_test_split and accuracy_report. All the submitted kaggle files are in folder predictions and all the logs from validation are in the folder logs. You can find the extensive evaluation logs with the most promising of the tested kernel parameters and C from 2^{-2}, ..., 2^{2} in

The only kernel that is not in validation and in predictions and which has been implemented is the substring kernel, which can be found at substring.py.

Files corresponding to kernels:

• spectrum.py
• svm.py and gaussian.py
• gappy.py
• mismatch.py
• substring.py

Files corresponding to classifiers:

• kernel_svm.py
• ridge_regression.py
• kernel_ridge_regression.py

Other files:

• training.py corresponds to all the train and predict pipeline, for all methods.
• classes.py used to define base classes for kernels and classifiers.
• argparser.py used to define the simple command line argument of dry.
• utils.py used to define a thresholding function.