Capstone of WashU Coursera: Data Science at Scale The detailed code (in R) can be found in RPubs in this notebook
This project aimed at exploring data related to Blight in Detroit, and building a predictive model of blight. We generated 101 features from the provided data, and explored multiple stochastic models. Finally, the Gradient Boosting Model seemed to perform best, with a 78.7% AUC in cross-validation, and 74% AUC in the validation set. Despite not being very accurate, ~68%, the model can help to reduce the number of buildings to assess for bligth.
The data used in the project was the following
- Demolition permits in Detroit, used to define whether a particular building was blighted
The rest of data was used for feature extraction
- 311 call records in Detroit
- Crime records in Detroit
- Blight violation records in Detroit
The following steps were
- Data exploration
- Generating a list of buildings
- Feature engineering
- Data modeling
- Evaluation of best model
- Sliced and diced the data to calculate multiple aggregates and understand the information
- Cleaned data, e.g., plotted top bligth violations and found 21k records that had the same lat/long coordinate, seeming a data entry error
- Built interactive exploratory map to visualize demolitions, 311 calls, blight violations, and crime on the map
- Built a fast algorithm to scan through all the available lat/long coordinates from 311 call, blight violations, and crime. The algorthim groups lat/long coordinates that belong to a 74x74 square feet rectangle around each particular coordinate in all records. Grouped lat/longs are assigned the to the same building ID, which will later be used to mash up the different sources of data.
- Built a lambda function to determine if a building lat/long is within a blight parcel. Blight parcels are defined by a circumference centered at their lat/long and with a radious such as the resulting area is equal to parcel size.
Generated a total of 101 features
Basic features at the building level
- number of blight violations
- number of 311 calls
- minimum, maximum of 311 call rating, and their difference
- number of crime reports
- number of crime reports with determination being "responsible"
- max fine amount for all crime records at a particular building location (from JudgmentAmt)
Surrounding features, around 111m of each building:
- number of blight violations
- number of 311 calls
- average of max rating of surrounding 311 calls
- average of min rating of surrounding 311 calls
- number of crime reports
- average of max fine amounts of surrounding crime reports
- total sum of max fine amounts of surrounding crime reports
Counts of type of crime, violations and 311 calls *
- 23 types of 311 issues
- For violations, reduced from 313 violation codes to 12 groups
Explored different models and sets of features
With number of violations only
- Linear model
- Decision Tree
- Random forest
With combination of features
- Decision tree
- Random forest
With all features and subsets based on importance
- Random forest
- Gradient boosting
The best model is a gradient boosting model with all the features. Selecting the top 20 features yields to similar performance of 78% AUC in cross-validation and 74-75% AUC in the validation set. Below is the confusion matrix for the evaluation set at an optimal operating point of 0.5:
| Ref Not Blighted | Ref Blighted | |
|---|---|---|
| Pred Not Blighted | 913 | 258 |
| Pred Blighted | 476 | 636 |
Despite not being very accurate, ~68%, the model can help to reduce the number of buildings to assess for bligth.
- Include temporal features (e.g., frequency of violations) - some temporal data seem to be noisy (e.g., violation ticked issue time being 01/01/38440)
- Use Detroit Parcel data to define the buildings
- Use Detroit neighboorhood data to calculate local metrics
- Use demographic data from the CENSUS
- Use MLS / Zillow (last time sold, price, # sold houses around)