I am working with a city's crime data and am trying to classify the type of crime in a city based on various features. Two of the features are latitude and longitude and I have been thinking about what is the best way to use these features in a model? Using them as regular numerical features does not seem right to me intuitively because the numerical variances between different latitude and latitude values are small and are not ordinal (45.0002 vs 45.0003, etc...) so what would be the best approach here? Thank you!! (The one thing I want to add in case though I don't think this should be relevant is the other features I am using in this model I have created dummy variables for due to their categorical nature).
$\begingroup$
$\endgroup$
7
-
1$\begingroup$ Cluster analysis comes to mind. $\endgroup$– CarlMar 13, 2017 at 4:25
-
$\begingroup$ What specifically do you mean in this case? I am trying to understand how to best use these features to create a classification model for identifying what type of crime based on all my other features. How would clustering play a role in here? $\endgroup$– data_science_mathMar 13, 2017 at 4:27
-
$\begingroup$ Probably care more about the distance between two points than the actual numerical value? I.e. how clustered different types of crime are? Being close in latitude probably is only meaningful if also close in longitude... $\endgroup$– Benjamin HorowitzMar 13, 2017 at 4:29
-
$\begingroup$ Yes, that's what i was thinking regarding both latitude and longitude should be close, otherwise its meaningless if one is close, the other is not. So what would be a good approach for incorporating those two into a predictive model? $\endgroup$– data_science_mathMar 13, 2017 at 4:32
-
$\begingroup$ 2D clustering obviously. $\endgroup$– CarlMar 13, 2017 at 4:33
|
Show 2 more comments
1 Answer
$\begingroup$
$\endgroup$
7
Time to read up on cluster analysis and crime.
https://www.ncjrs.gov/html/nij/mapping/ch4_9.html
https://www.icpsr.umich.edu/CrimeStat/files/CrimeStatChapter.6.pdf
http://www.ecostat.unical.it/RePEc/WorkingPapers/WP12_2011.pdf
For lots more search cluster analysis geography and crime type. and similar text.
This is what you asked for: one with cluster analysis that talks about predictive crime models and neural nets.
-
1$\begingroup$ But how would cluster analysis help me create a predictive model? I was referring to what I can do with my features to feed them into a supervised learning model to get predictions for crimes and see my accuracy score. Are you just suggesting that I essentially map graph datapoint on a map as a unsupervised learning problem vs a supervised learning problem where I classify which specific type of crime each crime is (based on variables that include year, month, day, hour and latitude and longitude)... $\endgroup$ Mar 13, 2017 at 5:15
-
1$\begingroup$ Yes, time based training and prediction. Think of it this way, if you know that in the "hood" murder is yearly 10, 20, 30, 40, 50, what will it be next year? Then you discover that the hood is drifting south 1 km per year, what will it be next year? Sure, I suppose you can do this with NN, read the Italian paper first, maybe, and get some idea of what has been done to model these things before. $\endgroup$– CarlMar 13, 2017 at 6:51
-
$\begingroup$ this is a really hazy topic matter to just jump into, I would highly recommend reading up about the systemic bias and issues that come along with predictive policing and applying machine learning to crime data sets. This is an area where improper modeling can have serious implications for peoples lives, please go forward accordingly $\endgroup$– rgalboApr 26, 2018 at 18:47
-
1$\begingroup$ @rgalbo Worse, in the perfect state, there is no crime, no police, and no-one counting. For example, there was no unemployment in some Communist countries. From this, we conclude that police create crime. $\endgroup$– CarlApr 27, 2018 at 18:32
-
1$\begingroup$ @rgalbo Poverty and crime are associated. If someone steals a loaf of bread, he is a criminal, but if someone steals one billion dollars, he gets a knighthood. Despite appearances, the difference between those two situations is not a moral difference, but one of influence and power. Laws are written by the influential and powerful, e.g., in North America at a net worth of approximately 100 million dollars and up. $\endgroup$– CarlDec 21, 2018 at 15:39