Melvin's digital garden

Team YNC Hacks (Yale-NUS) at Data Science Game 2016

[2016-11-14 Mon 19:43:02] event: DataScienceSG

Approach for the image classification task

  • preprocessing
    • uniform image size
    • rotate images to get more training data
    • L2 regularization
  • did not access to much hardware resources
  • models
    • Alexnet
    • Inception V3
    • Residual Network
  • implementation
    • Lasagne
    • Tensorflow
    • Theano
  • ensemble
    • majority vote
    • weighted vote based on accuracy on test data

Approach for the quote conversion challenge

  • important to understand the data attributes
  • customer looking for quotes for multiple cars likely not going to convert
  • create features to identify users who are not serious about buying
  • ensembling of several XGBoost models
  • difference from the best teams
    • they did not work overnight
    • were mostly monitoring the running processes
    • most of the work was automated
    • went through the AXA quote process

Links to this note