The data used in this analysis can be found in the following zip file: http://thomasdata.s3.amazonaws.com/ds/us_census_full.zip. This US Census dataset contains detailed but anonymized information for approximately 300,000 people.
The archive contains 3 files:
- A large training file (csv)
- Another test file (csv)
- A metadata file (txt) describing the columns of the two csv files (identical for both)
The Jupyter notebook contains all the code along with explanatory comments. Please refer to this notebook for full details of my analysis. In summary, I have performed some EDA, cleaned the data, and created several models to predict whether or not someone earns greater than $50,000 per year.
Initially, I tried Logistic Regression, K-Nearest Neighbors, Random Forest, and XGBoost models on the cleaned data. Class imbalance was quite high so I applied SMOTE rebalancing and re-ran the models without seeing significant improvement; I was using the F1 score to tune the models which is a robust metric with imbalanced datasets.
Next, I applied Principle Component Analysis and determined that 10 components could capture 99.9% of the variance, so used these 10 features in new models. Performance did not improve for the Random Forest or XGBoost, which were already the highest-performing models, so I didn't feel it necessary to use PCA in the final model (training time significantly improved, however, so in a production setting if the models will be frequently retrained, PCA is probably a good idea).
I determined the most influential features to be:
- 8.49% - dividends_from_stocks
- 8.01% - capital_gains
- 7.81% - age
- 6.94% - detailed_occupation_recode
- 3.83% - detailed_industry_recode
- 3.63% - num_persons_worked_for_employer
- 3.18% - weeks_worked_in_year
- 2.93% - sex_ Male
- 2.60% - capital_losses
- 2.36% - major_occupation_code_ Executive admin and managerial
Finally, I used an unseen test data set to measure the performance of the XGBoost model:
- 94.7% - Accuracy
- 59.5% - F1 score
- 74.4% - Precision
- 49.6% - Recall