Adult Income RFClassifier

1. Proposal

• Americans income-to-expense ratio varies by income level.
• What factors will allow us to accurately predict the annual income of adults in the US?
• If we can predict income, we can predict expenditure.

2. Methodology

• Clean the data to be categorical in nature.
• Create a categorical model with a RandomForest algorithm.
• Train a model to correctly categorize the income level of adults.

3. Cleaning the Data

• Binary Income:
  • 0 assigned <= $50k/year,
  • 1 assigned >$50k/year
• Education cleaned into 3 categories
  • Little difference and representation in lower/higher education extremes
• Marital status simplified to binary
  • Little variation in income via technical status (ex: “widowed” vs. “never married” have similar income ratios)
• Race & native country - Not enough data representation to be reliable in categorical model.
• Gender and relationship show strong trends in relation to income.
  • Relationship is in regards to others in the home.
• Age shows strong correlation to income values; with a tail to the outlier ages of 70+.
  • Removed the 1% outliers
• Hours-per-week is focused in on an average of 40; as expected.
  • < 40 hours/week not conducive with making >$50k?

4. Feature Identification

• No feature has a higher score than 30% influence to predicting income.
• As predicted, race had little value in predicting income from this dataset.
• Top 5 features:
  • Age ~ 30%
  • Hours-per-week ~ 15%
  • Marital Status ~12%
  • Education ~ 12%
  • Occupation ~12%

5. Results of Random Forest Model

• Trained on 80% of data, tested on 20%.
  • Training accuracy: 85%
  • Testing Accuracy: 83.1%
• Recall or True Positive Rate
  • (TP/(TP+FN): 83%, better w/ <= $50k data
• Precision
  • (TP/(FP+TN): 82%, better w/ <= $50k data
• F1-score
  • (2PR/(P+R): 82%
  • It is a harmonic mean of precision and recall
• Accuracy
  • ((TP+TN)/(N+P)): Overall ~83%
  • Percentage of total items classified correctly

• Great? Good? Bad?!
• ROC-AUC Score:
  • Likelihood of randomly choosing a positive case & negative case where the positive case outranks the negative case according to the classifier.
  • Our model’s score: 88.4%; pretty good.

6. Conclusions

• We can accurately predict the income level of an adult in the US with an error of < 12%.
  • Allows us to predict expenses, savings, and taxes.
  • Enables tailored marketing to demographics according to expenses.
• For the future:
  • Need more representative data of minorities & immigrants
  • Consider adding more income tiers, changing from binary classification to a new model.