Team Data Ninja (NUS) at Data Science Game 2016

[2016-11-14 Mon 19:00:00] event: DataScienceSG

Qualifier is an image classification challenge held on Kaggle

• orientiation of roof affects solar production
  • north-south
  • east-west
  • flat
  • others
• determine orientation of roof from satellite images
• opensolarmap.org
• evaluation metric: accuracy

Approaching the qualifier problem

• augment the training dataset
  • rotate north-south into east-west, vice-versa
  • slight rotations, shearing, zoom
• XGBoost got 56% accuracy
• various convolution neural net
  • ResNet produced the highest accuracy
• ensemble by allowing other weaker models to overturn the result from ResNet
• if the majority class is different from ResNet

Top 20 teams will qualify for the final round in Paris.

Finals is a quote conversion challenge from AXA

• user gets quote for insurance from various brokers
• evaluation metric: log-loss
• feature engineering
  • user more likely to convert on anniversay
  • topic modelling with LDA
• decided on XGBoost