Pump It Up - ML Project

Github Repo Link: https://github.com/minullamahewage/pumpitup_ML_project

• Driven Data Username: moracse_170342n
• Name: M.H. Lamahewage
• Index No.: 170342N

EDA

• Went through the train dataset and observed the features using various tools.
• The dataset was found to be imbalanced. Weighted F1-score was used for evaluation.

functional	non-functional	functional needs repair
32259	22824	4317

• Features with missing values, null values, zeros and other inconsistencies were identified.
• Features with high cardinality were identified.
• Features with high correlation/similar features were identified.

Preprocessing

Handling Missing Values

• Replaced null values with the mode for categorical features.
• Replace other missing values such as unknown, None and 0 with mean for numerical features and mode for categorical features.

Features with high correlation

These features were found to have high correlation or were similar upon manual inspection.

• extraction_type, extraction_type_group, extraction_type_class
• payment, payment_type
• water_quality, quality_group
• quantity, quantity_group
• source, source_type, source_class
• region, region_code
• waterpoint_type, waterpoint_type_group One of each were chosen based on several factors and the remaining features were dropped.

Ordinal Encoding

• Categorical features which are ordered or had too many categories were encoded using the OrdinalEncoder.
• Features: "funder","installer", "subvillage", "lga", "ward", "public_meeting", "permit", "management", "payment_type", "water_quality", "date_recorded".

OneHot Encoding

• Categorical features which aren't ordered and had a smaller number of categories were encoded using the OneHotEncoder
• Features: "basin", "scheme_management", "extraction_type_class", "management_group", "quantity", "source_type", "waterpoint_type".

Label Encoding

• The labels were encoded using the LabelEncoder

Feature Selection

• Corelation of different features were inspected and leaving only one each of similar features, the rest were dropped.
• Two features were removed as all the values were the same in every row.
• Features with too many missing values were also dropped.

Feature Engineering

• Added new feature active_time which represents the total time the pump has been active by the recorded date.

Model Selection

• RandomForestClassifier
• XGBoost and RandomForest were considered and RandomForest was used due to better results.
• parameters: n_estimators=600, max_depth=40, min_samples_split=10

Submissions

Score	Current Rank	# competitiors	Submission Date
0.8170	1624	12453	Sept. 16, 2021