/Pump-it-Up-Data-Mining-the-Water-Table-Competition

🏆 Pump it Up: Data Mining the Water Table - Competition

Primary LanguageJupyter Notebook

Pump it Up Data Mining the Water Table - Competition

Implementation Link : https://github.com/HimashiRathnayake/Pump-it-Up-Data-Mining-the-Water-Table-Competition

Experiments

Preprocessing Techniques :

  • Fix typo errors in the Dataset - installer
  • Drop duplicated features - source_class, source_type, management_group, scheme_management, quantity_group, quality_group, payment_type, extraction_type_class, extraction_type, waterpoint_type_group, region_code
  • Identify and fix the feature types.
  • Handling missing Values
    • Add a new Category for missing values or add missing values to an existing category like Unknown. - For categorical features.
    • Replace 0 which seems to be unknown by mean. - longitude, population.
    • Replace NaN values by most common value / median value. - public_meeting, permit
  • Remove features with less details.
    • wpt_name, scheme_name - Lots of None values, & most of the labels in this feature have very low value count.
    • amount_tsh - Most of the data (70%) is 0.0.
    • date_recorded - Most of the dates are beetween 2011-2013 & seems to be non-informative.
    • num_private - Most of the data (99%) is 0.
    • sub_village - Very low value count.
    • recorded_by - One unique value for all.
  • Reduce feature's category counts - Only keep few categories & group other categories to Other, for the features having large number of categories - installer, funder
  • Converting target values to ternary values. - status_group

Feature Engineering Techniques :

  • Check Mutual Information
  • Binning to groups - construction_year
  • Clustering - longitude, latitude
  • Principal Component Analysis - gps_height, population
  • Target Encoding - ward

Experimented Models :

Experiments Results :

Modal Train Accuracy Test Accuracy
Random Forest 0.9089 0.7913
XGBoost 0.9060 0.8111

Submission

result2 result3