Kernel Method Project

Author: Mathis Petrovich, Evrard Garcelon

Project from the MVA course Kernel methods for machine learning.

Introduction

This Kaggle competition is focused on the prediction of binding sites to a certain transcription factor for DNA sequences using kernel methods. To answer this question, we implemented several kernel classifiers and several kernels designed for protein sequences.

Leaderboard

The model we choose for the final submission was the sum of spectrum kernel for k = 7 to 20. This final submission scored 0.71333 on the public leaderboard (ranked at the 21th place) and scored 0.71 on the private leaderboard (ranked 3rd place).

Recreate this submission

Just run this command in a terminal:

./start

python start.py

Test other kernels/methods

String kernels available:

mismatch
spectral
spectralconcat (for adding spectral together)
wd
la
wildcard
etc

Methods availables:

kknn
klr
ksvm

Exemple

python -m src.tools.test spectral ksvm allseq --kparam "--k 7" --mparam "--C 1 --tol 0.01" --submit True --cvsname "Yte.csv" --show True --pcadim 3

General usage

usage: test.py [-h] [--kparams KPARAMS] [--mparams MPARAMS]
               [--dparams DPARAMS] [--pcadim {2,3}] [--show SHOW]
               [--csvname CSVNAME] [--submit SUBMIT]
               {mismatch,spectral,wd,la,wildcard,gappy,spectralconcat,exp,gaussian,laplacian,linear,poly,quad,sigmoid,multikernel}
               {kknn,klr,ksvm,simple_mkl} {allseq,seq,synth}

positional arguments:
  {mismatch,spectral,wd,la,wildcard,gappy,spectralconcat,exp,gaussian,laplacian,linear,poly,quad,sigmoid,multikernel}
                        kernel name
  {kknn,klr,ksvm,simple_mkl}
                        method name
  {allseq,seq,synth}    data name

optional arguments:
  -h, --help            show this help message and exit
  --kparams KPARAMS     kernel parameters
  --mparams MPARAMS     method parameters
  --dparams DPARAMS     data parameters
  --pcadim {2,3}        pca dimention for visualization
  --show SHOW           True to draw the pca
  --csvname CSVNAME     Path to the csv file to write
  --submit SUBMIT       Submit a csv or not

Mathux / Kernel-Method-From-Scratch