bolaik / RenthopApartmentInterestLevel

Solution posts to Kaggle competition Two sigma connect Renthop for rental listing inquries

RenthopApartmentInterestLevel

Solution posts to Kaggle competition Two sigma connect Renthop for rental listing inquries

features selected

The basic features include features from Branden Murray's post and some further revisions.

advanced features

• encoded features: This is based on the fact that some categorical turns out to have pretty high impact on prediction.
  • encode some numerical features, such as bedrooms, distance_city on manager_id, by taking the mean of features grouped by the categorical feature.
  • inspired by this, one can also encode important numerical features, such as price, on other features.
  • encode prediction (i.e. interest_level) on manager_id, requires cross validation.
  • encode (group mean, other statistics could also be defined) other numerical features (e.g. price) on categorical features (e.g. manager_id) conditioned on interest_level, also request cross validation.
• geological features: local price fluctuation, from plantsgo
• images features: also see magic feature

classifiers

• Trained different machine learning models, including xgboost, lightgbm, nn, adaboost, gb, rf, et, lsvc, lr, knn
• check out https://github.com/bolaik/RenthopApartmentInterestLevel/tree/master/my_final_version/classifiers_ipynb

Ensemble

• Train level-2 models, including xgb, nn, knn, lr, lightgbm, where xgb prediction is submitted because of best cv score.

many helpful links for reference

• Check out top-1 solution and github repo for many insightful ideas about feature engineering
• List of solution posts
• My best solution only using manager_id
• Unsupervised and supervised neighborhood encoding
• Price compared to neighborhood median
• A proxy for $/sqft and the interest on 1/2-baths