This model predicts the median income per household (over/under $49,000) in US zip codes based on geography and social security payments. That analysis would provide insight on the relationship between certain demographic variables and income.

As a first step in creating this model, in order to get a better understanding of the data, EDA was conducted and a visualization was produced using Matplotlib and Seaborn.

Next, to select the best algorithm for the model, Gridsearch and hyper-parameter tuning were applied to five classifiers: XGBoost, Random Forest, KNN, Logistic Regression, and Decision Tree. Based on comparisons of the accuracy score, XGBoost classifier was chosen for the final model.

Sources of data: US Census Report & Social Security Administration.

I worked with four different datasets, comprising of over 30,000 rows.

Issues to be dealt with included:

-Duplication

-Small number of people in a zip code

-Null/Wrong values

My final dataset before feature engineering was a concatenation of 4 datasets and had 11,260 rows and 12 columns. I chose the number $49,000 as my target over/under income, in order to get relatively balanced numbers in the target value.

One example from my EDA - Household median and mean income per zip code has relatively normal distribution:

After creating dummy variables and dropping columns with multicollinearity, the dataset had 68 columns. Accuracy Score for this dataset was 0.547.

My next steps were as follows:

-Creating polynomials and interactions (total number of columns: 2415)

-Scaling the data with SKLearn StandardScaler

-Selection of best 50 features out of the 2415, using SKLearn SelectKBest (F regression).

Following is en example of the hyper-parameter tuning for the XGBoost classifier:

The model with the best result after Hyper-parameter tuning used the XGBoost Classifier and reached a 0.733 accuracy.

-The State where the zip code is located has big influence, especially when dealing with wealthy or poor states (Connecticut and Puerto Rico for example).

-Some combinations of size of population and isolated places has an influence toward low income prediction.

-High percentage of retired workers is sometime a sign for higher median income (probably because wealthy people move to specific areas after retirement).

-The number of people with disabilities has some influence to predict that the area has relatively lower incomes.

Here is one example of a classifier feature importance( this one is from the logistic regression):

With additional data (for example, tax collection), and with including all US zip codes in the dataset (this model includes roughly 33% of them), the prediction can be further refined.