Model Buildings

[nicklucius]

Discussions so far have included:

1. Model feature importance for each potential feature
2. Use a sigmoid function to separate buildings into two classes
3. Use a linear function to model buildings evenly, so that there is not a glut of buildings at either end of the classification.

[nicklucius]

Will use Ridge regularization. Also will keep in mind Lasso if we eventually have a lot of predictors.

[nicklucius]

• [x] Find R package for Ridge

[nicklucius]

Currently there is one predictor in the prototype of the building.json file: tax delinquencies. Here is a simple linear regression for predicting buildings that need demolition based solely on tax delinquencies. The dots are overplotted so no patterns can be really be seen, but the regression line shows the trend.

Also, here is the breakdown of buildings with a middle range of tax delinquencies, by whether they were recommended for demolition.

Likewise, here is a breakdown of buildings with no tax delinquencies:

Tax delinquency seems to have some explanatory value, but more predictors are needed.

[nicklucius]

Up next:

• [x] Try density graph like this (at bottom)
• [x] Separate into train/test sets and generate confusion matrix / ROC / PR Curve

[nicklucius]

Modeling and analysis using the linked building data can be found here: https://cdn.rawgit.com/Chicago/buildings/24e0ddb4/classify-buildings.html