Goals

• Exploratory data analysis (EDA) of the Portugese Bank Marketing dataset (UCI, kaggle).
• Classification goal of predicting whether or not a customer will subscribe a term deposit.

Info

• This was a group project part of the course "DSA5101: Introduction to Big Data for Industry". I did the EDA and pre-processing bits.
• I've also included some models to demo performance. A more thorough study of different models and parameters was performed by the other members.
• The presentation also included discussions of the business implication, insights and findings, and future ideas. I'll not be updating the slides here.
• Please download the dataset from the links provided. I'll not be uploading the csv files here.

Summary

EDA

• Get metadata about the data frame (like number of entries) and its columns (like names and type).
• Distinguish between numerical and categorical columns and treat them separately.
• Get the distribution of the output. Here, it's imbalanced classification (a.k.a outlier/novelty detection).
• Get the distributions of the individual columns and any relevant conditional distributions (e.g.: column | output = 1).
• Impute unknown/null values.
• Check for (and remove) outliers.
• Drop columns that seem logically irrelevant (carefully).
• Create new features.
• Find correlations using mutual information and correlation matrices.

Preprocessing

• Use pipelines to streamline processing.
• Transform the data:
  • Incorporate adding/dropping columns from EDA into pipelines and leave the original data untouched.
  • Use column transformers to scale numerical columns and encode + scale categorical columns.
  • If your model requires it, consider normalizing.
• Select features:
  • Remove noisy features to avoid overfitting.
  • Remove interdependent features.
  • Use dimensionality reduction, univariate feature selection or select from model.
• Consider oversampling to tackle the imbalanced dataset.

The pipeline looks like this:

model = IMBPipeline([
    ('cleaner', data_cleaner),  # to drop columns and imputation
    ('feature_engineer', feature_engineer),  # to add new columns
    ('column_transformer', column_transformer),
    ('feature_selector', 'passthrough'),
    ('oversampler', SMOTE(random_state=0)),
    ('predictor', RandomForestClassifier(random_state=0)),
])

Models

• Trees:
  • Boosting
  • Bagging
  • Balanced
  • Ensemble
• Linear and NNs

Performance Analysis

• Scoring
  • Choose scoring metric (more than one preferred).
  • Cross validate or simple train test split.
  • Generate report (e.g., confusion_matrix).
• Model Selection:
  • Use grid search for the best hyper parameters.
  • Use tuner if you are using deep learning.