slides.md

title	author	theme	date
Group 117 Machine Learning Project	Tuomo Lehtonen, Qianqian Xia	Antibes	December 14, 2022

Overview

Model selection

• Separately for binary and multi-label
  • Train multi-label classifier using only samples with associated event
• Considered models
  • Logistic regression (softmax function for multi-label)
  • Random forests
  • Support vector machine
• Feature selection
  • Separately for each model
  • Recursive feature elimination with 5-fold CV
    • a variant of backwards feature elimination

Chosen models and results

• Binary:

  • Random Forest classifier
  • 15 selected features
  • Increased number of trees to 1500
  • Decent accuracy
    • 0.86 CV accuracy, 0.89 validation set accuracy
    • Actual test accuracy: 0.862

• Multi-label

  • Random Forest classifier
  • Trained on samples with an event class
  • 1500 trees
  • Quite poor multi-label accuracy
    • Actual test accuracy: 0.666
    • Probably bad strategy chosen