TheLab @ DC is an organization within DC Mayor Muriel Bowser's office that is responsible for testing and evaluating policies and providing analysis to agencies within the DC government. In a word, The Lab is DC's own data science shop.
Eric Foster-Moore, a data scientist at The Lab, will help us kick off this project on Thursday, June 14th. He will speak about the types of problems that The Lab seeks to address for the city, the specific questions which are formulated and answered by data scientists at The Lab, and the ways in which their findings are communicated to stakeholders.
=================================================================================
Authors James Linek , Eric Plog , Kiros Gebremariam , Awab Idris
===================================================================================
WASHINGTON-DC-AERIAL VIEW
Washington, DC, the nation’s capital, is the place created and planned as the seat of government and policy, a city that evokes the past and the nation’s heritage. The heart of Washington, DC is in its buildings, such as the White House, the Capitol, and the Supreme Court, its memorials, and its many open spaces, including the National Mall. There is so much more to write about the beautiful, clean city of Washington, DC. Today is July 4, independence day, we all are having fun and enjoying the independence day. i am just taking a minute to write something important about the unique public safety issues in the nations capital, just to share some of the presentations we make yesterday to the DC lab project using public data. While celebrating the independence day, just watching the fireworks its astonishing to be here at the nations capital, depending on which quadrant of the city you celebrating independence day; your personal safety is always the priority concern for family and friends . so we used publicly available data to predict crimes in the nations capital using attributes that are commonly used features by the MPD. Contrary to the general belief that crimes happen at night or dark, in the nations capital most offensess or crimes happen during the evening.
The following pie Chart shows total crime recorded from different shifts: Day, Evening and midnight. The specific time range about the three classifications was given in the open dataset, which i used as my source to write this blog. We still can observe a relatively low quantity of crime during midnight time range.
Total Crimes by Shift
The stacked Bar Chart below gives information about different types of crime that usually occur in the Washington DC area grouped using the police districts in the town . A stacked bar plot was chosen to visualise the crimes for each district. This bar plot will count the number of instances for each type of crime, and display the results in vertical bar heights stacked on each other, which gives us a feel for the overall number of crimes for each district and the breakout of crime types.
Crimes by Police District
from the above plot, the following insights may be gleaned:Burglary is greater in district 1&2. Theft F/Auto include stealing motor vehicle parts and anything inside of the motor vehicle.Theft F/Auto is greater in district 3 followed by 4 and 1 respectively. District 7 has less crime overall and many of the crimes are proportional . It also appears to have the greatest number of homicides Robbery is significantly less in district 2 but high in districts 6,7,3 and 4. Assault w/Dw is a defined as an assault with dangerous weapon or caused serious injury.
Distribution of crime based on time using table graph
Crimes are said to escalate during holidays. The definition of holidays for this plot is New Year, Independence Day, Thanksgiving, and Christmas. The result actually shows the total number of crimes occurring during holidays is a relatively minuscule percentage of the total number of yearly crimes.
The chart above was an attempt to see if there was a geographic distribution to the various violent crimes across the district. While there does appear to be a few clumps of dangerous areas, in general it appears to be spread throughout D.C. — with the notable exception of the Northwest area (Police District 2). I am surprised by the density of violent crimes near the downtown DC. if you want read more or download the jupyter notebook please clickhere.
TPOT is a Python Automated Machine Learning tool that optimizes machine learning pipelines by intelligently exploring thousands of possible pipelines to find the best one for your data. TPOT works on every model I have tried, with the exception of Neural Networks.
For a project, I was assigned to predict which type of crime occurred based on various features (i.e., location, month, etc.) in Washington DC. This was the first time I used TPOT, and had great results compared to manually grid searching and tuning hyper parameters over several different types of models. Thus, I wanted to share my results:
First [Install TPOT] (https://epistasislab.github.io/tpot/installing/)
conda install numpy scipy scikit-learn pandas
pip install deap update_checker tqdm stopit
Second For any python version conda install pywin32
pip install xgboost (only required if you plan to use XGBoost)
pip install tpot
Third Import TPOT Classifier or TPOT Regressor depending on the nature of your problem
from tpot import TPOTClassifier
from tpot import TPOTRegressor
Fourth Define X and y variables
features = [‘LONGITUDE’,’LATITUDE’,’PSA’,’AGE’,’SHIFT_EVENING’, ‘SHIFT_MIDNIGHT’,’METHOD_Code’,’Year’,’Month’,’Day’,’hour’,’dayofyear’,’WARD’,’week’,’quarter’]
X = dc3[features]
y = dc3[‘OFFENSE_Code’] #9 different crimes to predict
Fifth Train, test and split your data
X_train, X_test, y_train, y_test = train_test_split(X, y,random_state=42)
Sixths Scale your data
ss = StandardScaler()
X_train = ss.fit_transform(X_train)
X_test = ss.transform(X_test)
Sevenths Decide which parameters and config_dict are ideal for your data and computing power.
- [List of config_dicts from] (https://epistasislab.github.io/tpot/using/#built-in-tpot-configurations)
Default TPOT - TPOT will search over a broad range of preprocessors, feature constructors, feature selectors, models, and parameters to find a series of operators that minimize the error of the model predictions. Some of these operators are complex and may take a long time to run, especially on larger datasets.
TPOT light - TPOT will search over a restricted range of preprocessors, feature constructors, feature selectors, models, and parameters to find a series of operators that minimize the error of the model predictions. Only simpler and fast-running operators will be used in these pipelines, so TPOT light is useful for finding quick and simple pipelines for a classification or regression problem.
This configuration works for both the TPOTClassifier and TPOTRegressor.
TPOT MDRTPOT - TPOT will search over a series of feature selectors and Multifactor Dimensionality Reduction models to find a series of operators that maximize prediction accuracy. The TPOT MDR configuration is specialized for genome-wide association studies (GWAS), and is described in detail online here.
Note that TPOT MDR may be slow to run because the feature selection routines are computationally expensive, especially on large datasets.
TPOT sparse - TPOT uses a configuration dictionary with a one-hot encoder and the operators normally included in TPOT that also support sparse matrices.
This configuration works for both the TPOTClassifier and TPOTRegressor.
- List of TPOT Regression parameters :
- List of TPOT Classifier parameters:
- Keep in mind that TPOT is computationally expensive, so increasing the generations, population size and cv will dramatically increase the time it takes to complete.
- See below for the parameters used on my DC crime model:
As you can see, I used small parameters and the model still took roughly 8 hours to complete. The model performed much better than my initial grid searching, and allowed me to work on more data gathering, EDA and feature engineering while the model ran. In addition, TPOT prints the highest performing model along with the parameters utilized so you can perform additional analysis such as feature importances. See the section below for the feature importances of my model.
Although a score of 59% on the train and test sets is not the greatest, it still out performed my manual models by a large margin. In order to get a better model, I would need to collect additional data such as unemployment rate or housing prices per neighborhood.
Eighths Feature Importance
important_features = pd.Series(data=dt.feature_importances_,index=X.columns)
important_features.sort_values(ascending=False,inplace=True)
pd.DataFrame(important_features).plot(kind = 'barh', grid=False)
In summation, TPOT is a very helpful tool to help identify the optimal model and parameters to achieve the best model possible.
1) Make sure you have `.gitignore` in your local repo by adding this line: `input/csr/*`
2. Create your own local branches
3) Make sure you install and track Git Large File Storage(read more: https://git-lfs.github.com)
- `git lfs install`, `git lfs track "*.csv"`
- this will create `.gitattributes` in your local repository.
Approaches and Deliverables with the DC crime Prediction
Approaches
- Determine which variables provide the highest correlation with crime incidents based on the data of Crime incidents from 2012-2018 datasets
- Visualize crime incidents
- Heatmaps and Visual estimate plots to correlate variables
Feature selection
- Time series analysis on variables across the time period of 2012-2018
- Identify seasonal effects by week, month, weekdays,quarters of the year and annual
- Determine anomalies on the Dataset based on shift,PSA,Ward,Start and end date
Prediction models
- Regression (to determine the evolution of patterns over time,by Age of the crime, length of time to respond) Clustering (to determine common patterns across census tracts, Wards, etc.) Classification (to predict if a given crime is likely to happen in a given time and location)
Deliverables
- A notebook including: Analysis of crime patterns Models for predicting crime
1) Open Data DC
-GIS
-Real-property-tax-assessment-neighborhoods
-Real-property-tax-assessment-sub-neighborhoods
[Time fetched, June, 2018]
Additional from the Project lab
Problem Areas, Potential Questions, and Additional Resources
In consultation with the Lab @ DC, we've organized questions into three main problem areas. You are not limited to addressing these questions, but they are some of the questions that are currently of interest to The Lab!
Problem Areas
-
Crime-Related:
-
Does crime go up when street lights go out? (This is actually an open data science project that The Lab is working on right now, led by a volunteer. You can check out the repo here. You could pick up an existing thread or start from the ground up!)
-
Can we predict how many people are going to be in jail or released from jail? Are there issues associated with this, such as overcrowding? Are there (evidence-based) steps that could be taken to avoid some of these issues?
-
Do wireless hotspots or security cameras reduce crime?
-
If we built similar models to predict gunshots and to predict 311 calls (using the same predictors, say) what might we learn from comparing the model parameters and predictions?
-
Using this data:
-
Can we tell if being charged with one crime is highly predictive of also being charged with certain other crimes, or:
-
Are there other factors that predict a combination of offenses (e.g. maybe arrests in specific locations are more likely to have specific combinations of charges)
-
-
DC makes available crime incident location data. Using that data:
- Can you build a model to predict or explain crime? (Note: you can use continuously update 2018 Data as a real-life test set.)
- Can you merge this dataset with another dataset to explore the relationship between crime and, e.g. house prices, parking tickets, 311 requests, or certain city infrastruture?
-
-Property-Related:
-
Can you identify different features from an aerial image in order to predict whether a house is vacant, or identify the type of property?
- Using Real Property Values:
- Can you estimate or assess housing values based on aerial imagery, voter file data, and other information?
- Can we make inferences about property values or demographic trends given construction and occupancy permits?
-
Forecasting:
- Can we forecast service needs around DC based on 311 service calls and other data?
- Can we forecast traffic accidents, moving violations, and parking tickets? Based on these forecasts, can we make policy recommendations, forecast revenue, or
- Can we build a voter turnout model to predict the likelihood that someone will vote. Data data are available from the Board of Elections for $5 on a CD. If you build the model before the June primary, you’ll have a built-in test set!
Additional Resources
The DC Open Data Portal is your friend.
The DC FOIA Reading Room gives access to documents released under the Freedom of Information Act.
The DC Policy Center releases a lot of reports that may be of interest, including:
- Taking Stock of the District's Housing Stock, with data available.
- Can Neighborhood Characteristics Explain Enrollment in at In-Boundary Schools?, with data from the Deputy Mayor for Education.
The Federal Government's data portal may provide some interesting data to cross refence against.
Case-Shiller Index of US residential house prices. Data comes from S&P Case-Shiller data and includes both the national index and the indices for 20 metropolitan regions. The indices are created using a repeat-sales methodology.
As per the [home page for Indices on S&P website][sp-home]:
The S&P/Case-Shiller U.S. National Home Price Index is a composite of single-family home price indices for the nine U.S. Census divisions and is calculated monthly. It is included in the S&P/Case-Shiller Home Price Index Series which seeks to measure changes in the total value of all existing single-family housing stock.
Documentation of the methodology can be found at:
Key points are (excerpted from methodology):
- The indices use the "repeat sales method" of index calculation which uses data on properties that have sold at least twice, in order to capture the true appreciated value of each specific sales unit.
- The quarterly S&P/Case-Shiller U.S. National Home Price Index aggregates nine quarterly U.S. Census division repeat sales indices using a base period a nd estimates of the aggregate value of single family housing stock for those periods.
- The S&P/Case - Shiller Home Price Indices originated in the 1980s by Case Shiller Weiss's research principals, Karl E. Case and Robert J. Shiller. At the time, Case and Shiller developed the repeat sales pricing technique. This methodology is recognized as the most reliable means to measure housing price movements and is used by other home price ind ex publishers, including the Office of Federal Housing Enterprise Oversight (OFHEO)
To download and process the data do:
python scripts/process.py
Updated data files will then be in data directory.
Note: the URLs and structure of the source data have evolved over time with the source data URLs changing on every release.
Originally (2013) the site provided a table of links but these are not direct file URLs and you have dig around in S&P's javascript to find the actual download locations. As of mid-2014 the data is consolidated in one primary XLS but the HTML you see in your browser and the source HTML are different. In addition, the actual location of the XLS file continues to change on each release.
================================================================
DATA DICTIONARY
Attributes to Remove
The below table provides a description of attributes that will be removed because they provide little to no predictive power for the Washington DC Metropolitant Police Department crime incident report.
===================================================================
The below table provides a description of attributes that will be removed because they provide little to no predictive power.
| Attribute | Description | Removal Reason |
|---|---|---|
| REPORT_DAT | The date/time the offense was reported | When the crime was report is not a predictor of a crime. |
| SHIFT | The duty shift that responded the call. | This is an eight hour period. |
| OFFENSE | The category of crime committed. | This attribute is redundant with OFFENSE_Code. It will be stored off for its labels. |
| METHOD | The category of the method used to commit the crime. | This attribute is redundant with METHOD_Code. It will be stored off for its labels. |
| DISTRICT | The police district. | This attribute was replaced by new feature DistrictID. |
| PSA | Police Service Area | This attribute was replaced to create PSA_ID. |
| WARD | The political Ward identifier. | WARD can be derived from NEIGHBORHOOD_CLUSTER which provides a smaller geographic resolution. |
| CENSUS_TRACT | Land management tract identifier. | This attribute is replaced by other geographical data. |
| VOTING_PRECINCT | Political subdivision | This attribute is replaced by other geographical data. |
| CCN | Criminal Complaint Number - unique to each report | An index for the crime report. |
| XBLOCK | Eastern coordinate of crime scene (meters) | This attribute is replaced by lattitude and longitude. |
| YBLOCK | Northern coordinate of crime scene (meters) | This attribute is replaced by lattitude and longitude. |
| START_DATE | The earliest the crime might have been committed. | END_DATE provides a more accurate time when considering crimes not solved/resolved. |
| DistrictID | The police district. | Redundant with PSA_ID. DistrictID can be found from PSA_ID. |
| SHIFT_Code | The period of duty shift that responded to the call. | This is an eight hour period that is too coarse grain. |
| METHOD_Code | The coded method used to commit the crime. | Over 91% of the crimes are coded OTHER so this attribute is useless. |
| CRIME_TYPE | The code for the type of crime. This attribute can be determined | from OFFENSE_Code. |
| AGE | The difference in END_DATE and START_DATE of the crime. | This attribute is unreliable due to the discrepency in START_DATES. |
| TIME_TO_REPORT | The time it took for the police to report the crime. | This attribue represents action after the crime has happened . |
==============================================================
Attributes to Keep
The below table provides a description of attributes that will be used for predicting crimeincidents in DC.
================================================================
|Attribute |Description |Role|
| :--------------------------------- | :----------------------- | :-------------------: |
|---|---|---|
| ANC | Advisory Neighborhood Commission that is a geo-political grouping. This is our target attribute. | Feature and Target |
|---|---|---|
Attributes Added to the crime incident report from different sources
==================================================================
Our team was able to acquire data on the unemployment rates and housing prices for each of the 8 wards based on year. Of course, no data comes into play without the need for sufficient munging! A separate excel file was created to help fill the gaps with a multitude of INDEX/MATCH functions accross several sheets. The unemployment rates were also incomplete in a sense it was only yearly for the 8 wards, but the housing prices data we were using different informations for cross validation.
Our team felt it was important to bring to light a characteristic about Theft and Robbery which may not be so apparent just by looking at the data set. It is important to take note that while Theft and Robberies may have a higher rate in certain neighborhoods, it is likely that those who are committing the crime most likely live in a completely separate location. It is highly unlikely that a suspect would steal from their neighbors, risking the higher chance of being caught with readily recognizable stolen goods[Here we just taking common sense]. On the other hand, those areas which may have a higher drug use may experience exactly this as the neighborhood may have a different dynamic with respect to a communal mindset which may exist in other neighborhoods[areas with high sex abuse and violent crime areas]. As is usually the case, if there were more data readily available about the suspect[identity,like age,location,status,demographics] as well as the victimology involved[by age and sex], there very well may be the possibility of being able to measure these aspects.
A description of attributes that we have added for predicting crime incidents in Washington DC.Do these features help us predic a class of a crime?
===============================================================
|Attribute|Description|Type|Role|
|:------|:----------------|:---|:----:|
|Housing_Prices| Sales prices of the homes per WARD | Continuous| Feature|
|Unemployment| Unemployment rate per WARD | Continuous| Feature|
|Month|The latest month the crime was comitted.| Ordinal| Feature|
|Day of a week| The latest day the crime was committed.| Ordinal| Feature|
|Hour The latest hour the crime was committed.| Oridinal| Feature|
Dataset Explanations
we have two potential response variables: Crime_Type (nonviolent crime vs. Violent crime), and Offense_Code (The more specific type of offense: Homicide, Robbery, Theft, Arson, etc.). The goal is to provide the Metropolitant police Department with a model that can predict or classify a crime based on the available explanatory variables collected from the opendatadc as well as from the MPD website.
One problem with this data is that the victim profile data is missing (due to privacy concerns, and the fact that non violent crimes are not necessarily because of the owner's profile). The explanatory variables for this dataset focus on time and locations. We believe that the detection/classification of a Violent crime would be based primarily on the victim's characteristics, and not exclusively on the location or time. The other problem with this data is that (fortunately) there are far fewer violent crimes than there are nonviolent crimes (approximately 83% of the 36000+ crime reports are against nonviolent rather than persons), so we have very imbalanced classification.
Our exploration of the variables seem to indicate that time (not necessarily the day, but the time during the day) is one of the more significant factors[this was the visualization and we need to see for our prediction and modeling]. We saw this in the SHIFT variable (which gives the Police duty shift that responded to the call). When we broke the time down into individual hours of the day, we saw a pronounced cyclic effect, where night-time crimes were far more likely than daytime crimes[see feature selection and visualization notebook]. Weekend crimes were slightly more likely than crimes during the work week[visualization by week ], and monthly trends appeared to be opposite intuition (fall crimes were more likely than winter or summer crimes).The quarterly crime shows an increase during the second and third quarters as compared to the first and fourth quarter of the year.
Location also appeared to have some influence, but the way the locations were grouped altered the effect significantly[the visualizationon notebook three by PSA]. Different political areas (Wards and the subordinate Association Neighborhood Committees) showed a different trend than using global locations (Latitude and Longitude). Police districts (and their subordinate Police Service Areas(PSA) showed a different trend than the Ward grouping. This tells us that there are some location effects, but it is difficult to separate them out due to the correlation between geo-physical areas and the different (but overlapping) political mappings[the ward classification based on distance is just not explanatory to justify the crimes]. The top violent crimes show significance across the anacostia river and around the business areas of the district, near the white house and all around the Georgia Avenue and towards the North East side of DC and areas bordering the Prince Georgecounty.