police killings in america
TRANSCRIPT
Hands up, don't shoot!
A Comprehensive Look at Police Shootings in America
By: Maxwell V. Pederson
Introduction and Background• The fatal police shooting of Michael Brown back in 2014 brought about
widespread outrage and debate across the nation
• The Guardian, a British Newspaper, sought about the reporting of police killings, in which they found there is no comprehensive database!
• Only as of recently from all the protests and outrage has the FBI decided to create a voluntary program in which police can choose or choose not to report their justifiable killings.
• From 2005 - 2012 only 1,110 of the 18,000 police departments reported these justifiable killings
• Data for this project was gathered from the Guardian's Open Source Police Homicide database and from FiveThirtyEight’s version of the Guardian data
Problem Statement There has been over 2,100 reported fatalities caused by police since January of 2015. With all the the riots, protests, and uproar caused by certain police shootings such as Michael Brown, what is the general sentiment about police shootings in America? Is there reason to believe that police have an inherent bias towards who they kill? Does the data support the public's sentiment towards police shootings?
Overall Goal: Understand the characteristics of these police shootings to come to a conclusion on whether the general population’s opinion on police fatalities is justified by biases police may have towards who they kill.
Methods• Took the FiveThirtyEight data: Started out with 34 attributes and 467
instances• Only a few attributes consist of the original attributes from the Guardian
database (categorical)
• Most attributes are numerical census data added by FiveThirtyEight
• Very unique instances, show the data is fairly linearly-inseparable and may suffer from the curse of dimensionality
• If columns were used and had bad data/NAs the whole row would be removed• Most if not all imputation methods will not work well when the data is this
linearly-inseparable with many dimensions
• Extensive use of C5.0 , SVMs , RFs , and BBNs are used in this presentation.
What Does the General Population Think?
Rolling Around in the Data I
Rolling Around in the Data II
Florida Police Justified Homicides
Were These People Armed in Florida?
Attribute Selection & Feature Engineering
C5.0 For Classifying if a Person was Armed or Not
Support Vector Machine For if a Person was Armed
Bayesian Belief Network
• Probability that a person is armed given they are in one of the poorest districts, given a certain race: 65 ~ 69%
• Probability a person was killed by a gunshot given they were armed: 90% (also the same if not armed)
• Probability a person is armed given they are a male: 74% and 64% given they are a female
Predictive Modeling Performances• All of the algorithms have
low accuracy and low sensitivity
• The low sensitivity shows that the algorithms are misclassifying unarmed people as being armed
• Just as in the real world, each event classification is extremely different hence the poor accuracies
Conclusion • The general population of twitter shows a negative sentiment towards police shootings,
but it must be kept in mind the population that is on twitter, young adults. This could serve as an “echo-chamber” effect as seen in the election.
• Predictive models don't do well on this data because it's too specific and makes the data linearly inseparable, yet in real life we aren't getting enough of the picture to classify if a person is armed or not.
• The models mostly misclassified people who aren't being armed as being armed, which seems to reflect controversial killings today.
• Police may or may not have biases, but it can be seen that if these advanced algorithms can’t classify properly, imagine being the person in the situation when a police call comes in.
• The next steps would be to get a fuller picture on police killings and in general more data. Coupled with top notch non-biased data scientists, maybe predictive forecasting of crimes could be done.
• As seen a Wall Street Journal however, algorithms aren't biased, the people who work with them are. For me, I definitely tried to get certain results out of the models, which shows my bias.
ReferencesChen, Eugene. "Map on MapInSeconds.com." MapInSeconds.com by Darkhorse Analytics. Darkhorse Analytics, 2016. Web. 15 Dec. 2016. <http://mapinseconds.com/>.
McGinty, Jo Craven. "Algorithms Aren't Biased, But the People Who Write Them May Be." The Wall Street Journal. Dow Jones & Company, 14 Oct. 2016. Web. 15 Dec. 2016. <http://www.wsj.com/articles/algorithms-arent-biased-but-the-people-who-write-them-may-be-1476466555>.
Swaine, Jon. "About The Counted: Why and How the Guardian Is Counting US Police Killings." The Guardian. Guardian News and Media, 2015. Web. 15 Dec. 2016. <https://www.theguardian.com/us-news/ng-interactive/2015/jun/01/about-the-counted>.
Flowers, Andrew. "Fivethirtyeight/data." GitHub. FiveThirtyEight, June 2015. Web. 15 Dec. 2016. <https://github.com/fivethirtyeight/data/tree/master/police-killings>.