170676P-DrivenData-CS4622

Highest score = 0.8182 Rank = 1399

Github Code Link = https://github.com/GihanAyesh/170676P-DrivenData-CS4622.git

Exploratory Data Analysis

The dataset contained 40 features.

Main findings from EDA

• The majority of the variables are categorical and a small number of numerical, boolean and temporal features exist.
• num_private has zeros for 99% of the rows
• recorded_by has only one value.
• scheme_name has around 50% of null values

The following features have null values

• funder
• installer
• subvillage
• public_meeting
• scheme_management
• scheme_name
• permit

The following features have high correlation

• Region and region_code
• extraction_type, extraction_type_group and extraction_type_class
• management and management_group
• payment and payment_type
• water_quality and quality_group
• quantity and quantity_group
• source source_type and source_class
• waterpoint_type and waterpoint_type_group

Univariate distribution of population distribution

Scatterplot of longitude and latitude

Preprocessing

Following columns were dropped

• Id - Every row has a unique value
• subvillage, lga, ward, wpt_name - Too many unique values
• extraction_type_group and extraction_type_class - Highly correlated with extraction_type
• Num_private - 99% of the rows are zeros
• region- Highly correlated with region_code
• Recorded_by - Has only one value for every row
• scheme_name - Around 50% of the rows has null
• Payment - Highly correlated with payment_type
• quality_group- Highly correlated with water_quality
• quantity_group- Highly correlated with quantity
• source_type source_class- Highly correlated with source
• waterpoint_type_group- Highly correlated with waterpoint_type

Null values were filled on following columns

• funder, installer, permit - Using value ‘other’
• public_meeting - Using value True

Encodings

• gps_height - Binning
• basin, funder, scheme_management, installer, extraction_type, payment_type, management, management_group, water_quality, quantity, source, waterpoint_type - One hot encoding
• Status_group - Label encoding

Other preprocessing steps

• Min max scaler is used before feeding the data to neural network
• Standard scaler is used before feeding to XGBoost
• Longitude and latitude are feature crossed to get a new feature for neural network
• funder, scheme_management, management, installer - labels with low counts were changed to other category
• date_recorded is turned to date-time dtype and the year is filtered out
• Amount_tsh, population - Normalized using standard scaler

Models used with parameters

Random Forest

n_estimators=200, max_depth=None, min_samples_split=6, min_samples_leaf=1, max_features='auto', bootstrap=True, warm_start=True

XGBoost

base_score=0.5, booster='gbtree', colsample_bytree=0.4, gamma=0.0, importance_type='gain', learning_rate=0.05,max_depth=3, min_child_weight=7, n_estimators=100,n_jobs=1,num_class=3, objective='multi:softmax', random_state=0, reg_lambda=1, scale_pos_weight=1,subsample=1, verbosity=1

Ada Boost

n_estimators=50, learning_rate=1.0, random_state=None

Logistic Regression

multi_class='multinomial'

Neural network

test_size=0.2, learning_rate = 0.001, epochs = 150, batch_size = 4096, validation_split = 0.1

Hyper parameter tuning is done to Random Forest, XGBoost and Neural Network

Feature importance (XGBoost)