Crime Prediction

Logo

View the Project on GitHub cyber-hbliu/Crime-Prediction-In-Chicago

Introduction

Chicago is a city steeped in rich history and cultural heritage. Unfortunately, it is also plagued by crime. Ensuring the safety of its citizens is a persistent challenge that law enforcement agencies are continually trying to tackle.

Social safety is a constant problem that affects the lives of its citizens. Understanding crime, predicting it, and preventing it have become something that police departments want to do. As in the movie Minority Report, a “prophet” predicts crimes before they happen. Nevertheless, reality is not a science fiction movie, and crime prediction is a complex problem because many random human factors are involved. The specifics of how crime is committed vary depending on the type of society and community.

In 2022, Victor Rotaru, Yi Huang, Timmy Li, James Evans, and Ishanu Chattopadhyay developed a new algorithm to predict urban crime at the individual event level a week or more in advance, enabling model-based simulations that reveal both the pattern of reported infractions and the pattern of corresponding police enforcement. Meanwhile, in 2021, Suhong Kim, Param Joshi, Parminder Singh Kalsi, and Pooya Taheri et al. implemented a machine learning prediction model to predict crime in Vancouver.

Like the above studies, in this exercise, I attempted to use machine learning to predict crime rates in Chicago by analyzing data from previous crimes records. The prediction model also includes social factors that may affect crime rates. The overall crime rate of a city is calculated by dividing the total number of crimes submitted by police agencies in the city by the city’s population and multiplying the result by 100,000.

The project is broken down into five parts:

  1. Data wrangling: This involves cleaning, joining, merging, and transforming the crime dataset.
  2. Exploratory Analysis: In this section, the crime trend will be analyzed from a community-based perspective, taking into account the type of crime, location, time series, arrest rate, and domestic violence rate.
  3. Feature Engineering: This involves analyzing the spatial patterns of crime, examining the density of different types of crimes, clustering crime events, using KNN algorithms to capture similarities, and classifying the distance between crime sites and nearby feature points.
  4. Predictive Modeling of Crime Count: This section focuses on predicting the crime rate of the city.
  5. Evaluation: In this final stage, the performance of the model will be evaluated and errors will be validated.

Click here. to see packages used in this project

What Exactly Exists

The raw data used in this project consists of crime records from 2015 to 2019 obtained from the Chicago Data Portal. With a total of 1303648 records, the dataset provides detailed information on the primary type of crime, the crime’s description, its location, and other relevant police information, such as community, time, district, and more.

The second dataset used in this project is the Chicago neighborhood dataset, which contains the names, numbers, and geographic boundaries of 77 communities. To make use of this data, it was preprocessed by reading the JSON files from the API and merging them using the concat function. The formatting was then adjusted for date and time, numeric variables, and categorical variables. Additionally, relevant features were added to the dataset using the merge function, including community information and socioeconomic information. Finally, several indices were calculated, such as the crime count by community, weapon count, month count, week count, arrest index, and domestic index.

The dataset was grouped by communities and sorted to create a new dataset of crime events for each community. An interactive choropleth map was created using the folium function, displaying the cumulative crime events over 5 years for each community. The map reveals that crime events are generally concentrated in a few communities in the city, particularly on the south side.

Community-based Perspective

Plotly Charts

Additionally, the charts provide insight into the most common location and time of occurrence for each type of crime. For example, it appears that theft is most commonly recorded at street locations, while battery and assault are most commonly recorded at residential locations. The charts also show the trend of crime occurrence over time, indicating that crime rates may vary depending on the season and month of the year.

In terms of district, it can be seen that some districts have higher crime rates compared to others. For example, some districts may have a higher rate of theft, while others may have a higher rate of battery. By examining the distribution of crime events across the different districts, police departments can allocate resources more effectively to areas with higher crime rates.

Plotly.express and altair packages have been used here to create corresponding interactive charts. These charts show that theft, battery, criminal damage, assault, other offense, deceptive practice, narcotics, burglary, motor vehicle theft, and robbery topped the list of crimes with criminal records.

The data visualization also provides insight into the relationship between crime and location. Crime rates were similar indoors and outdoors. Streets were the most frequent areas, followed by residences and apartments.

Comparing the different types of crime in different communities shows the distribution and frequency of crime incidents in each community. It can be seen that theft, battery, robbery, and narcotics are widely distributed in each community.

Furthermore, the time series allows us to know the trend of crime occurrence. As can be seen from the following plot, crime is not a cumulative or increasing-decreasing phenomenon. It is a continuous, cyclical, extensive, and random process.

From this, we can see a regular correlation between crime occurrence and months, with May to October having a higher frequency than other months, especially in the summer months (July and August) as the period with high incidences of cases. It shows that the weather (temperature, precipitation, cold) is also a critical factor in crime incidents. This could be because people tend to be out more and socializing during these time, which can create opportunities for criminal activities.

Some crimes have a 100% arrest rate, and we should be concerned about which crimes have lower arrest rates, such as burglary, kidnapping, robbery, criminal sexual assault, and theft, which can become hidden dangers.

Heatmap

A total of 31 Chicago districts are represented in this dataset. Based on the districts heatmap, we can identify the number of specific types of crimes that have occurred in various districts in the last five years. The darker the color, the lower the number of crimes, which shows that some crimes, such as assault, burglary, motor vehicle theft, narcotics, theft, and criminal damage, have a high frequency in all districts.

In contrast, obscenity, stalking, kidnapping, intimidation, human trafficking, criminal sexual assault, etc., occur less frequently. Aditionnally, some crimes are concentrated in specific areas, such as prostitution in district 7 and district 11.

However, when it comes to community security, some communities have been experiencing policing problems for these five years. In the community heatmap, we know that these communities, such as Austin, Auburn Gresham, Loop, Near North Side, Near West Side, North Lawndale, Roseland, South Shore, West Englewood, and West Town, all have a high crime rate. These communities may have some security problems.

Similarly, the day of the week heatmap allows us to understand the crime trend over the week. The chart shows that crimes are more likely to occur in the daytime, such as offenses involving children. some crimes are more likely to occur in the night, especially criminal sexual assault. This information is valuable for both the police and the community, as they can prepare accordingly.

In conclusion, the exploratory analysis provides valuable insights into the nature of crime in Chicago. By visualizing the crime data, it becomes easier to identify patterns, trends, and relationships between different variables. This information can help police departments make data-driven decisions for crime prevention and reduction efforts.

Crime Spatial Patterns

Density

Different types of crime are being distributed unevenly across the complex landscape of modern cities. KDE is a popular tool for analyzing data distributions. In the kernel density estimate, 9 representative types such as ROBBERY, THEFT, BURGLARY, WEAPONS VIOLATION, NARCOTICS, ASSAULT, HOMICIDE, SEX OFFENSE, and KIDNAPPING are displayed.

It is worth mentioning that the KDE gives an unequivocal demonstration of the spatial distribution. We can see that while theft occurs widely throughout the city, it is most concentrated in downtown areas, while narcotics are concentrated in the area on the west side of downtown. For the rest of the several crimes, different peaks can be seen scattered in several city areas.

Clustering

The result is that cluster-0 includes 31 communities with an average crime incident of 10043, cluster-1 includes 17 communities with an average crime incident of 30559, cluster-2 includes 5 communities with an average crime incident of 42966, cluster-3 includes 1 community with an average crime incident of 77134, cluster-4 includes 23 communities with an average crime incident of 7862. Last but not least, the hexagonal grid map shows the clusters in more detailly. Using the h3fy function from the tobler.util module can easily generate a hexgrids covering the face of the Chicago City.

Some features would help predict the crime rate, such as the proximity to depository places, bars, pubs, and liquor retailers, and the location near commercials place, universities, schools, landmarks, shot spotter alerts, or subway stations. Moreover, the crime probably happened in abandoned buildings and vehicles in some areas. Likewise, some socioeconomic features, such as the economic hardship index, can be critical factors in community development and security issues. Specifically, there are some factors related to the number of crimes:

  1. Depository locations
  2. Abandon buildings
  3. Schools
  4. Grocery stores
  5. Subway stations
  6. Retails and commercials
  7. Landmarks
  8. ShotSpotter alerts

Nearest Neighbors

In addition, the sklearn.neighbors.NearestNeighbors provides a regression for data with continuous labels. It identifies the nearest neighbors of a given query point, and the commonly used distance metric is the Euclidean distance.

In this project, the nearest neighbors places of crime occurrence are found. The value of k varies according to different places, for instance, 5 nearest depository locations, 5 nearest abandoned buildings, 3 nearest schools, 3 nearest grocery stores, 5 nearest subway stations, 5 nearest shopping areas, 2 nearest landmarks, and 3 nearest shot spotter alerts. The darker the color, the farther the distance. These maps show that depository locations and subway stations are concentrated in the city, and abandoned buildings are concentrated on the city’s west side and some areas in the middle of the city. Schools are located throughout the city, while grocery stores, retail, and commercials are concentrated in some neighborhoods.

The KNN analysis helps us understand the relationship between crime and amenities. It is used to calculate the log-transformed distance between crime locations and amenity locations. This information can be used to identify areas with high or low levels of access to amenities, which can impact crime patterns.

Hexagon Grid

Based on the dataset, I calculated several indexes, including crime rate, risk rate of people under 18 or over 64, domestic violence rate, and weapon rate. Some socioeconomic features like percent of housing crowded and the hardship index are also presented in hexagon grid maps. The darker the color, the higher the value. These maps show that the crime rate is concentrated in areas with a high incidence of crime because the number of crimes is used by the population to calculate. When I look at the remaining four indicators, I see that the Hardship index and the weapon rate have very similar spatial distribution characteristics. Meanwhile, percent of housing crowded and the domestic violence rate shows complementary spatial distributions. Perhaps in further improvement, I should do some correlation analysis.

Correlation Matrix

Following, I show the correlation matrix using a heatmap, which is a matrix that shows the correlation coefficients of different variables. The correlation coefficient is between -1 and 1, with 0 indicating no linear correlation between the two variables. The farther the correlation coefficient is from zero, the stronger the relationship between the two variables. Also the relationship can be positive or negative.

A New Predictive Try

Model

In this section, the aim is to build a Random Forest regression model to analyze the crime risk factors influencing the number of random crimes. This is achieved by dividing the data into training and testing sets, and using the training set to fit the model and optimize its parameters. The dataset is split into two parts, with 70% of the data constituting the training set and 30% the test set, utilizing the sklearn.model_selection.train_test_split method. The data is then transformed and processed using the sklearn.compose.ColumnTransformer and sklearn.pipeline.make_pipelinefunctions.

A random forest is a meta-estimator that fits several classifying decision trees on various sub-samples of the dataset and uses averaging to improve the predictive accuracy and control over-fitting. Once the model has been trained on previous crime data in Chicago, including factors identified in feature engineering, it can be tested on the testing set to evaluate its performance. The sklearn.model_selection.GridSearchCV method is used to perform a search over specified parameter values for the estimator, resulting in a score of 0.5439, indicating the fitness of the model.

Importance

Feature importance are provided by the fitted attribute feature_importances_ and they are computed as the mean and standard deviation of accumulation of the impurity decrease within each tree. Finally, a bar chart will be generated, which shows the use of a forest of trees to evaluate the importance of features. The added features, such as several values measured by KNN, are of high importance.

Examine Errors

Evaluation

In this section, I use the predict function to get the predicted value of the crime rate, calculate the predicted value minus the observed value, then divide it by the observed value to get the errors and absolute percent of errors, and MAE. As a result, the MAE is shown on the map as a hexagonal grid. Darker colors indicate more significant errors. The model does not perform well in the areas with higher crime rates, such as downtown areas and the west side of downtown. Furthermore, in the southern region, the MAE decreases.

Improvement

The model only achieved a partial prediction, with a final score of 0.54. Because some dataset was unavailable, adding more features to the original dataset was impossible, such as weather, population, average income, eviction rate, poverty level, and so on. The use of demographics and graffiti is problematic because of the potential to introduce racial and socioeconomic bias and have questionable causal value. A proper combination of urban amenities, weather, socioeconomic, traffic, and crime data can increase the model’s fit and yield more accurate crime rate predictions. Moreover, the location of the cases is mostly on the street. Some neighborhoods can be analyzed using osmnx. It is also possible to transform and calculate some factors using correlation.

Reference

  1. Scikit-learn. Retrieved from https://scikit-learn.org/stable/index.html
  2. The dataset of incidents of crime. Retrieved from https://data.cityofchicago.org/Public-Safety/Crimes- 2001-to-Present/ijzp-q8t2 Lee, J., & King, G. (2022). Predicting crime with machine learning. Proceedings of the National Academy of Sciences, 119(34), 21283-21289. doi:10.1073/pnas.2002575117
  3. Rotaru, V., Huang, Y., Li, T. et al. Event-level prediction of urban crime reveals a signature of enforcement bias in US cities. Nat Hum Behav 6, 1056–1068 (2022). https://doi.org/10.1038/s41562-022-01372-0
  4. Tamir, Azwad & Watson, Eric & Willett, Brandon & Hasan, Qutaiba & Yuan, Jiann-Shiun. (2021). Crime Prediction and Forecasting using Machine Learning Algorithms. International Journal of Computer Science and Information Technology. 12. 26-33.
  5. Shah, N., Bhagat, N. & Shah, M. Crime forecasting: a machine learning and computer vision approach to crime prediction and prevention. Vis. Comput. Ind. Biomed. Art 4, 9 (2021). https://doi.org/10.1186/s42492-021-00075-z
  6. S. Kim, P. Joshi, P. S. Kalsi and P. Taheri, “Crime Analysis Through Machine Learning,” 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), 2018, pp. 415-420, doi: 10.1109/IEMCON.2018.8614828.