Kaggle challenge : Adult Census Income

Installation

• Setup environment

virtualenv venv
source venv/bin/activate
pip install -r requirements

• Jupyter python 2 kernel (if necessary)

python2 -m ipykernel install --user

• Download the data

mkdir data

Download the zip files into this folder.

• Run the jupyter notebook

jupyter notebook

Main goal

The main goal is to predict if someone earns more or less than 50k per year based on static features of the person : the age, the occupation, the native country etc.

The data

There are multiple kinds of features :

• The static description
  • native country (+ continent)
  • race
  • sex
• The conjugal situation
  • marital status
  • relationship
  • single
• The educational background (education / education.num)
• The current job informations
  • hours worked per week
  • capital (can be positive or negative)
  • workclass
  • occupation

We first transformed and clean the data by :

• Removing duplicate informations : eduction / education.num
• Refine features : continent, is_single
• Merge the capital data into one and use a logarithmic transformation
• Convert sex and income into float data
• Replace ? by None

The exploratoring part can be seen using exploratory.ipynb

This visualization part taught us that some features seem to be useful to split the data easily but it would be difficult to be perfect.

The prediction

The algorithm

We used regressors instead of classifiers to be able to plot the ROC curves and thus be more flexible on the threshold choice.

We chose to benchmark 3 algorithms: - The Linear Regression for its simplicity, quickness and performances - The Random Forest for its global performances and the computation of feature importance - The Gradient Boosting to compare bagging with boosting (and the same reasons as the RF)

We also add an aggregated regressor that uses the previous predictions and average them.

#### Performance metrics

We describe our results in prediction.ipynb

Our predictions are based on a 30 KFold cross validation. We used :

• ROC Curve
• Confusion matrix
• Precision / Recall

to compare ourto have an overhaul quite good predictor, with good f-score.

The problem is to be accurate in both class (>50k, <=50k) that is why we didn't emphasize on the precision metric.

#### Future work

• Feature engineering
  • Try to transform categorical data into numerical data
  • Find new features based on the actual one
  • Interpret FNLWGT influence
• Be smarter with the prediction aggregation (learning overlay)
• Use the OOB score as performance metric
• The data is a bit imbalanced : oversampling / undersampling
• Add unit tests

Appendix

• FNLWGT

FNLWGT = SSCA * BW * DCF * NAF :

The \# of people the census takers believe that observation represents

SSCA (the second stage calibration adjustment) :

[...] equal to the ratio of the pre-second-stage weight and the final weight after the calibration process is completed.

BW (Base Weight) :

In simplified terms, a base weight of 1,000 assigned to a sampled person means that the sampled person “represents” 1,000 persons in the U.S. population. 

DCF (Duplication Control Factors) :

[...] value between 1 and 4 inclusive [...]

NAF (Household Noninterview Adjustment Factors) :

The noninterview adjustment factor is intended to adjust for the presence of Type A noninterview households (households that are not interviewed because the occupants were temporarily absent, no one was home, the occupants refused participation, or the occupants could not be located).