Predictive modeling next steps

[rebeccajohnson88]

• [x] separate out preprocessing script from modeling script
• [x] read in following new datasets below (csv versions) to preprocessing script: (1) "WHD data" (all states; no TRLA outcomes)- whd_violations.csv; (2) "TRLA data" (7 states; whd and trla outcomes)
• [x] get existing preprocessing code to work with new data, making sure that the ACS variables are newly added as features
• [x] Make sure the acs is included in the mean imputation (separate for train and test) and do min-max scaling separately in each dataset
• [x] Ping me after preprocessing is done for me to look over
• [ ] Then work on the model estimation: (1) for WHD data, this should iterate over two labels: (1) investigations (outcome_is_investigation_overlapsd) and (2) violations (outcome_is_viol_overlapsd); (2) for data combining TRLA, should use the following variable outcome_compare_TRLA_WHD and use np.where to define four outcomes: investigated by either, investigated by TRLA only, investigated by WHD only, investigated by both
• [ ] For model estimation script, write the following outputs for each model and label to the model_outputs folder in Github (so not Dropbox): (1) confusion matrix (can melt to long format where one col is each of the for categories, another col is the count), (2) coefficients/feature importances, (3) continuous predicted probabilities for each row (should have job_row_id attached)

[rebeccajohnson88]

Issues - close to 5x the rate of missingness in the pkl file as in the matched_data_WHD in R:

[rebeccajohnson88]

• [X] Fix issue with labels (high missingness) caused by left_join in original process_data function: (originated with left join after reset_index; didnt seem to be matching up ids correctly)

• [x] ACS not actually imputed --- output of checking which cols in X_train have any missing values

Going back to script 12, issue is somewhere in the imputation function which isnt actually imputing