Evaluating Inglewood Area Robberies

---

Zacharia Schmitz

3 October 2023

Heatmap of Robberies in Cluster 4 (Inglewood area)

---

Project Plan:

(Jump To)

Project Overview

Data Acquisition

Preparation

Exploration

Models

Conclusion

---

From the City of Los Angeles website about the data:

• This dataset reflects incidents of crime in the City of Los Angeles from 2010 - 2019.

• This data is transcribed from original crime reports that are typed on paper and therefore there may be some inaccuracies within the data.

• Some location fields with missing data are noted as (0°, 0°). Address fields are only provided to the nearest hundred block in order to maintain privacy.

• This data is as accurate as the data in the LA database.

---

Overview

Project Goals

• Initially: Explore high areas of gun violence in the Los Angeles area and what to do to protect yourself

• Upon further evaluation: After it was identified in clustering, explore drivers of the high (almost double) robbery rate in the Inglewood area. Identify features and develop a model to encourage public safety and prevent crime.

---

Project Description

• I initially started with 2 datasets from the city of Los Angeles.

• After combing, there was close to 3 million rows (2,943,476).

• Each row was a unique incidence of crime in the Los Angeles area.

• There was 29 different features about each crime. This only included the initial report, and not the final closed case details.

• The target variable was a column created from the crime description, 'is_robbery'.

• The comparison was ultimately Cluster 4 (Inglewood Area) versus all of the other areas.

---

Initial Hypotheses

• Time of day is probably a large driver of robberies. Most likely more at night.

• There are probably not very many robberies during the day.

• Gender and age is most likely a driver of robberies. Most likely people preying on females and the very young and old.

• Most robberies are most likely occuring at gunpoint, or assumed to be gunpoint.

---

Acquire

• 2010 to 2019 Data (crime_data_2010_2019).csv

https://data.lacity.org/Public-Safety/Crime-Data-from-2010-to-2019/63jg-8b9z

• 2020 to 25 Sep 23 (crime_data_2020_2023).csv

https://data.lacity.org/Public-Safety/Crime-Data-from-2020-to-Present/2nrs-mtv8

Both dataframes shared the same features, and were merged together.

• Merged CSV Format (crime_data.csv)

https://drive.google.com/file/d/14FBSb-iADac0jENqeEfceTcmRi2Btaoh/view?usp=sharing

---

Pre-Modeling Data Dictionary:

Column	Definition
is_robbery	Feature engineered from taking robbery and attempted robbery from crime description
is_street	Feature engineered from taking if the robbery occured on the street from premise description
victim_sex	Encoded column for the victim's sex. F = Female, M = Male, X = Non-Binary
victim_descent	The victim's ethnicity

---

Preparation Steps

Logically Rename Columns

• Rename columns to make sense for exploration.

Choose columns to work with

• Based on what we're looking for, we'll start with only looking at some of the features.

Fix Time Occurred

• Change time occurred to 4 digit 24 hour time. We won't move it to index for now, because we may not be doing any time series

Fix Victim Age

• Looks like we have negative values in the victim_age. We will have to drop those, since no fair assumptions can be made.

Fix Victim Descent

• We'll map the victim_descent based on abbreviations and assign nulls 'unknown' for exploration before potentially dropping nulls.

Fix Victim Sex

• We'll have to map victim_sex. We can assume F = Female, M = Male, X = Non binary. Due to large amount of other values, we will make them 'unknown' rather than imputing any assumptions.

Nulls in premise_description

• Very small amount of nulls in premise_description (0.00023). We'll assign them 'unknown'.

Weapon_description

• 66% Nulls in weapon_description, we'll fill them with 'No Weapon'.

Weapon_category

• Make a weapon bin out of all of the weapon_descriptions. No Weapon, Firearm, Melee Object, Threats, Vehicle, Other

Create report_delay column

• Take the date of reporting, subtract the crime_occurred.

Fix Part 1 & 2

• Part 1 is for heinous crimes & 2 is for less heinous.

Fix latitude and longitude

• A total of 3321 values of 0 in both lat/long. We'll sort by the area name, and forward fill so they stay in the correct area.

Filter down DF to guns only

• Due to scope of data and time for project, we will only focus on firearm related crimes.

Bin Robbery and Attempted Robbery

• We made a column named is_robbery and if the crime was robbery or attempted robbery, we marked it true.

Bin Time of Day

• We binned the time of crime to time of day to morning, afternoon, evening, and night.

Make a column for crimes that happened on the street.

• It was identified that crimes that a majority of robberies happened on the street. We'll make a binary column for it name is_street.

---

Explore

• Does age affect the rate of robbery in cluster 4?

• Does time of day affect robbery rate in cluster 4?

• Does the victim's gender affect robbery rate in cluster 4?

• Where are you most likely to be robbed in cluster 4?

• Is the weapon used in cluster 4 different than the rest of Los Angeles?

---

Modeling

• Use drivers in explore to build predictive models of different types

• Evaluate models on train and validate data

• Select the best model based on accuracy

• Evaluate the test data

Best Model on Test

Logistic Regression

C=10
max_iter=1000
penalty="l1"
solver="saga"
random_state=321

Baseline: 53%

Test Set: 60%

---

How to Reproduce:

REQUIRED LIBRARIES:

• numpy

• pandas

• seaborn

• matplotlib

• scipy

• sklearn

• folium

  • pip install folium

• yellowbrick

  • pip install yellowbrick

1. Clone this repo.

2. Acquire data by:

   • Source the two LA Crime Data files through the links

   OR

   • Download my combined CSV (750MB) of crime data from LA here
   
   

3. Save CSVs in same folder as Jupyter Notebook with the correct naming convention

4. Run the notebook

   • Do the required pip installs as they come up

---

Conclusions

Key Findings / Questions Answered:

1. Does age affect the rate of robbery in cluster 4?

• Age does affect the rate of robbert in cluster 4.
  
  

2. Does time of day affect robbery rate in cluster 4?

• People are much more likely to be robbed in the evening.
  
  

3. Does the victim's gender affect robbery rate in cluster 4?

• Gender did not strongly affect the robbery rate.
  
  

4. Where are you most likely to be robbed in cluster 4?

• Most outside areas are very likely to be robbed. Number one being on the street/sidewalk.
  
  

5. Is the weapon used in cluster 4 different than the rest of Los Angeles?

• We found the weapon used, is not different in cluster 4, compared to the rest of LA.

---

Takeaways:

• With only 3 identified features, I developed a model with 60% accuracy that outperformed the baseline accuracy of 53%.

  • is_street

  • victim_sex

  • victim_descent

• The best model between SVC, KNN, LR and DTC was LogisticRegression witha train set accuracy of 58%, validation set accuracy of 62%, and a test set accuracy of 60%.

C=10, max_iter=1000, penalty="l1", solver="saga", random_state=321

Given that this model performed 5% better than baseline on our test set, we would expect it to also perform well on unseen data

---

Recommendations

For Modeling:

• Continue to run feature engineering and potentially test other models with other hyperparameters

• Possibly create models for the other identified clusters

• Potentially include some form of time series analysis to obtain trends

For Population:

• Stay out of Inglewood at night

Next Steps

• When dealing with crime rates and stereotypes, model performance is extremely important. Making policing assumptions from a model 7 points better than baseline would not be appropriate.

• Given more time and access suspect information could drastically improve the model.

• Also with more access, the closed case details could cut down on incorrect, and most importantly biased initial reporting.

• Combine with weather data for more insights

• Combine with poverty data for more insights

• Combine with house values for more insights

Back To Top