sbavon / kaggle-nyc-taxi-fare-prediction

My solution for Kaggle NYC Taxi Fare Prediction ( ranked 21st/1463)

Kaggle NYC Taxi Fare Prediction Kaggle Solution (Top 2% Ranked 21st/1478)

This repository is the solution that obtains the top 2% ranking of NYC Taxi Fare Prediction competition in Kaggle.

Data cleaning ( refer to data_cleaning.py)

• remove null records
• remove records whose locations are not within range provided in test data
• remove data points in sea
• eliminate outlier according to fare distribution

Data preprocessing ( refer to Data_preprocessing.py)

• the new feature cluster is added.
  • during data exploration, I found that the fare/distance ratio is varying according to the location. So, I add the new categorical feature to specify the area of the dropoff location and pickup location
  • I used HDBScan to get the clustering model. Then, I use this model to predict the area of each record.
• the new feature distance is added
• the new feature distance to airport is added
• categorical data are changed to float32 to prevent memory surge due to Lightgbm python package. (The library will convert all data to float. So if the data is integer, new data will be created)

train and predict ( refer to train_predict.py)

• lightgbm is used, and it was trained in the Amazon EC2 instance
• With this model, the test score is 2.85311