SeriesFinder

This is the code for paper "Learning to detect patterns of crime" (http://web.mit.edu/rudin/www/WangRuWaSeECML13.pdf).

The Series Finder model has two components. First, run learn_lambda.m to learn the pattern-general coefficients lambda, and then run find.m to grow series from given seeds.

Step 0: pre-computed similarity of each feature

In our project, we precomputed 11 feature matrix: LocEntry,MnsEntry, dayapart, HB_dis, premises,ransacked,residents, timeframe,dayweek,suspect and victim. For Ncrimes, these are Ncrimes by Ncrimes matrix where an element on i-th row and j-th column represents the similaerity between the i-th crime and j-th crime for that particular feature. Since the similarity matrix is symmetric, we only keep the upper triangular of each matrix and set the lower half to be 0.

Step 1: learn_lambda.m

Learn_lambda.m finds the optimal lambda. It initializes the model with a randomly generated starting point lambda_init, and uses coordinate ascent to gradually improve the objective until it converges. Since this could lead to a local optima, we initialize the search with different randomly generated starting points

• trainSeries: a list of series, where trainSeries(t,:) contains the index of crimes in series t
• NtrainSeries: the number of training series
• stepTable: a table that contains results at different lambdas. For example, stepTable(j,r) stores the results when multiply the j-th feature by scales(r)
• Maxlen: the maximum number of crime that can be discovered in a series
• maxIter:maximum number of steps
• Nchain: the number of times the searching is run from with a randomly generated starting point
• scales: the scales that are used to be multiplied with each feature

Step 2: find.m

find.m grows series of crimes from seeds provided by users. Every time, the model computes the pattern-crime similarity between the current pattern and all candidate crimes, and choose the one that has the highest similarity, adds it to the pattern and update the pattern-specific coefficients eta accordingly. The series keeps growing until the cohesion drops below certain cuttoff or the size of the series is bigger than the maximum length allowed

• Ncrimes: the number of all crimes from which we want to discover series.
• Nseries: the number of series we want to discover
• cutoff: the cutoff for cohesion of a series where if the cohesion is below the cutoff, the series stops growing
• Maxlen:the maximum number of crime that can be discovered in a series