An “Estimate & Score Algorithm” for simultaneous parameter estimation and reconstruction of incomplete data on social networks
 Rachel A Hegemann^{1}Email author,
 Erik A Lewis^{1} and
 Andrea L Bertozzi^{1}
DOI: 10.1186/2190853221
© Hegemann et al.; licensee Springer. 2013
Received: 14 January 2012
Accepted: 27 December 2012
Published: 12 January 2013
Abstract
Abstract
Dynamic activity involving social networks often has distinctive temporal patterns that can be exploited in situations involving incomplete information. Gang rivalry networks, in particular, display a high degree of temporal clustering of activity associated with retaliatory behavior. A recent study of a Los Angeles gang network shows that known gang activity between rivals can be modeled as a selfexciting point process on an edge of the rivalry network. In reallife situations, data is incomplete and lawenforcement agencies may not know which gang is involved. However, even when gang activity is highly stochastic, localized excitations in parts of the known dataset can help identify gangs responsible for unsolved crimes. Previous work successfully incorporated the observed clustering in time of the data to identify gangs responsible for unsolved crimes. However, the authors assumed that the parameters of the model are known, when in reality they have to be estimated from the data itself. We propose an iterative method that simultaneously estimates the parameters in the underlying point process and assigns weights to the unknown events with a directly calculable score function. The results of the estimation, weights, error propagation, convergence and runtime are presented.
Keywords
Inferring incomplete data Social networks Gang rivalries Hawkes process Selfexciting point processesIntroduction
In this work we focus our attention on data sets of events involving rival gangs on a social network. Each event in the data set corresponds to a crime that occurs at a specified time and involves a pair of rival gangs. A subset of these events are unsolved crimes in which one or both of the rival gangs is not known. The method developed in this paper could be broadly applied to any social network involving activities in time between pairs of nodes on the network. However the interest in the problem came about by examining data from the Hollenbeck Division of the Los Angeles Police Department, home to 29 street gangs with a wellknown rivalry network [1][3].
Unlike other methods used to address incomplete data relating to social networks [4, 5], the question at hand is not if a rivalry exists, but rather to which rivalry a violent event belongs. This structure of between gang rivalries can be viewed as a social network [6] often embedded in space [1, 2]. Violent events involving gangs tend to be dyadic, and so we can formulate these events as a realization of a stochastic process occurring on the edges of the rivalry network. For each edge in the network there exists a different stochastic process. In our analysis however, we use identical parameters to generate synthetic data. The method does not assume that the underlying parameters generating each process are identical.
The first step to inferring the affiliation of the violent events is to understand the underlying stochastic process. This requires us to capture the behavior of criminal activity through computational means, much like in [7][9]. Recently methods have been proposed in the literature to mathematically model gang violence. The authors in [10] employ an agentbased model to investigate the geographic influences in the formation of the gang rivalry structure observed in Hollenbeck. These authors consider the longterm structure of the rivalry network embedded in space. In terms of the rivalry violence, a shorter timescale must be considered.
Violence among gangs exhibits retaliatory behavior [11]. In other words, given an event has happened between two gangs, the likelihood that another event will happen shortly after is increased. A problem such as this is modeled naturally by a selfexciting point process. It is interesting to note that these models were first used to analyze earthquakes [12][15]. Since then, they have been used to model financial contagion in credit markets [16, 17], viral videos on the web [18], terrorist activity in Indonesia [19], and the spread of infectious disease [20]. In this analysis we limit the scope of our model to include time only, thus providing a baseline model.
In this Hawkes process, the intensity λ_{ k }(tH_{τ,k}) depends on the history of the process ${H}_{\tau ,k}=\{{t}_{1},{t}_{2},\cdots {t}_{{M}_{k}}\}$, where M_{ k }is the number of events for process, k. In this framework, the window of time, [0,T], observed for each process in the network is the same. However, the number of events in each process, M_{ k }, is stochastic, and therefore varies from process to process. In practice the final time, T, is determined by the end of the data collection period. Further, the edges of the window introduce boundary effects that are adjusted for in the parameter approximation, see Equations 10 and 11.
The background rate of the process is defined by the constant μ_{ k }. In the context of gang rivalries, background events can be thought of as random occurrences between rival gangs that trigger retaliatory events. The expected number of offspring for any event is determined by the constant α_{ k }, and the decay of the intensity back to the background rate is ω_{ k }. Offspring events, in this context, could be interpreted as retaliatory events. Larger values for μ_{ k }and α_{ k }produce more background and offspring events respectively. Larger values of ω_{ k } do not influence the total number of events, but rather the amount of clustering in time.
The authors of [24] produce a mathematical framework to solve the incomplete data problem observed in gang violence data sets. In their work they use an optimization strategy that computes the weights to infer the rivalry affiliation of the incomplete data. In this formulation the authors prove that their optimization has a unique solution under mild constraints. This is substantial contribution in inferring the affiliation of the unknown violent events. However, the authors of [24] assume that the process parameters are known, an assumption that is often not feasible in practice. Further, finding the weights requires solving a computationally expensive optimization problem.
We propose an iterative method that (A) estimates the process parameters assuming the data is generated by the process defined by Equation 1 and (B) infers the process affiliation of simulated data via a direct method of computation. We iterate between (A) and (B) until the estimates for the unknown events converge. We call this the Estimate & Score Algorithm (ESA). The details of the ESA are described in Section “The Estimation & Score Algorithm (ESA)”. The ESA is tested on simulated data in Section “Results”, with analysis of the estimation of the parameters in the presence of incomplete data (see Subsection “Estimation analysis”) and comparison of the proposed score functions with that of the StomakhinShortBertozzi (SSB) method in [24] (see Subsection “Updating Weights analysis”). In Subsection “Runtime Analysis” there is an analysis of the runtime between the StomakhinShortBertozzi and the Forward Backward score functions used to update the weights (see Subsection “Runtime Analysis”). Subsection “Convergence Results” contains an analysis of the convergence of the Estimation & Score Algorithm. This method solves the more realistic problem of estimating the process and the weights. Further, the computation for the weight updates is more direct and therefore avoids performing the costly optimization scheme used in [24]. This is a novel piece of work with many exciting extensions. A final discussion of the results and future work is presented in Section “Discussion and Future Work”. As in [24] we do not use field data in this paper, rather we generated point process data using similar parameters as observed in the field data for Hollenbeck [22]. By using simulated data to test the algorithms we have actual ground truth evaluate the performance of the method.
Problem Formulation
The data is assumed to lie on a known social network containing K processes, where each of the K processes is a pairwise rivalry between two gangs. From this set of events, there are a total of N events where the time is known, but the processes affiliation is not known. These events are referred to as unknown events. Each of the N unknown events are placed into each of the K processes. Since the process affiliation is not known for all of the events in the network, each event is given an associated weight, S_{i,k}. Here S_{i,k}is the i th element of the k th process. If the event is known S_{i,k}=1. If S_{i,k} is unknown then it is assigned a number between 0 and 1 by our algorithm. We enforce the constraint that $\sum _{k=1}^{K}{S}_{i,k}=1$.
As indicated in Figure 1, for each process in the network events e_{i,k} are indexed by increasing time, ${t}_{1}\le {t}_{2}\le {t}_{3}\cdots \le {t}_{{M}_{k}}$. Ordering the events in such a way has the consequence that the first unknown element in time, for example, may have different indexes for different processes. In Figure 1 the triangle index in first process is the third event, e_{1,3}. However the triangle in the K th process is the second event, e_{K,2}. One can easily keep track of the local index of a unknown event for each process.
The Estimation & Score Algorithm (ESA)
Initialization
For this paper, there were two ways of initializing the Estimate & Score Algorithm. The first is used to infer rivalry affiliation given field data. After importing the data, the unknown events are identified and placed into each of the of the K processes. The weights, S_{i,k}, must also be initialized. If the event is known, then S_{i,k}=1. If the event is unknown then ${S}_{i,k}=\frac{1}{K}$.
An alternate initialization utilizes simulated data in order to test the components of the Estimate & Score Algorithm. In this case, data is generated from K independent Hawkes processes with given μ_{ k }, α_{ k }, and ω_{ k }. From these data, choose N events at random from the network to mark as unknown. Place these N unknown events into each of the other processes. Initialize the weights such that for known events S_{i,k}=1 and for unknown events S_{i,k}=1/K. This initialization process is used in this paper to test the method and produce the results in Section “Results”.
Parameter Estimation
In the presence of no unknown events, there are both parametric [12] and nonparametric [25][28] ways to model the underlying stochastic process on each edge of the social network. For this work, we chose a parametric form for the triggering density to validate the model but the results could easily be extended to the nonparametric case. We note that, as is usual with nonparametric estimates, speed would be compromised for the sake of flexibility.
For this paper, the data is assumed to be a realization of Equation 1, where the parameters are estimated using a method similar to the Expectation Maximization (EM) algorithm [29]. An EMlike approach is taken because of the branching structure present in a Hawkes process. In such a process each event can be associated with a background or response event. However, given a realization from this process it is not immediately obvious whether an event is a background or response event. We can view this information as a hidden variable that we must estimate. In this way, every event in each of the K processes is assigned a probability ${P}_{i,j}^{k}$. The probability that event i is a background event is denoted ${P}_{i,i}^{k}$, and probability that event i caused event j is denoted ${P}_{i,j}^{k}$. This assumes that t_{ i }<t_{ j }. From this EM estimation, the approximation for each of the variables is altered to include the weights for the unknown events. In fact, in the case where all the events are known, the estimation formulas are the same. This section derives the EM estimates when in the presence of incomplete data.
for t_{ i }<t_{ j }. The EM algorithm then becomes a matter of iterating between estimating the probabilities and the parameters. It has been proven that this algorithm will converge under mild assumptions [29]. Further, Equation 6 adjusts for boundary effects.
When all of the events are known, i.e. S_{i,k}=1 when unknown event i, k belongs to process k and is zero otherwise, these estimates become identical to the EM parameter estimates.
Updating weights
At the start of the Estimation & Score algorithm all of the weights for the unknown events are S_{i,k}=1/K. Once the parameters are estimated using the altered EM algorithm described in Equation 11, the weights, S_{i,k}, are updated, see Figure 2. Here we present four different score functions and the StomakhinShortBertozzi method [24], used to define, q_{i,k}, the intermediate process affiliation. Each of these score functions synthesize information from different portions of the data set. Given an event early in the data set, a score function that uses future events would be ideal. On the other hand, for later events a score function using previous events is desired. Similar considerations should be made if there are portion of the data with more incomplete data. After all of these intermediate weights, q_{i,k}, have been calculated, they are renormalized as a probability via ${S}_{i,k}=\frac{{q}_{i,k}}{\sum _{k}{q}_{i,k}}$. For simplicity we consider a response function of the form, ${g}_{k}\left(t\right)={\alpha}_{k}{\omega}_{k}{e}^{{\omega}_{k}\left(t\right)}$.
Ratio Score Function
Lambda Score Function
StomakhinShortBertozzi (SSB) method
This method is motivated by the Hawkes process defined in Equation 1.
Probability Score Function
Forward Backward Score Function
Results
The Estimation & Score Algorithm is tested for accuracy on simulated data from the Hawkes process defined in Equation 1. An analysis of the parameter estimation method outlined in Subsection “Parameter Estimation” is conducted in Subsection “Estimation analysis”. A comparison of the score functions when assuming the true parameters is found in Subsection “Updating Weights analysis”. Subsection “Runtime Analysis” provides a comparison of the runtime between the Forward Backward score function and the StomakhinShortBertozzi method. A example of convergence of the Estimate & Score Algorithm is provided in Subsection “Convergence Results”.
Estimation analysis
There are many ways we could allow the unknown events to influence our estimates of the underlying parameters for each process. There are two extremes. On the one hand, we could exclude all of the unknown events from the parameter estimation. This would be equivalent to setting the S_{i,k}=0 for all unknown events. On the other hand, we could include all of the unknown events in the estimation of the parameters for each process. This would be equivalent to letting S_{i,k}=1 for all i and k. Another possible estimation method is some combination of these two. We propose this as a way of allowing the unaffiliated events to play some role in the estimation process. The naive choice is allowing each event to play the same role in each process. This amounts to setting S_{i,k}=1/K for the unknown events. We compare these three choices to the estimations obtained by the Estimate & Score Algorithm (ESA) using the Forward Backward score function. Finally, we want to compare all four of these possible estimation techniques to the best we could possibly do. In this case, that would mean we knew all the affiliations for the events (i.e. there are no unknown events).
Average and standard deviations for μ _{ k } on 100 networks, true value is μ _{ k } = 0.01
# unknown  15  30  45  60  75  

Equal  (Ave)  0.0102  0.0098  0.0100  0.0099  0.0096 
Weights  (StDev)  ±0.0014  ±0.0014  ±0.0017  ±0.0015  ±0.0015 
ESA  (Ave)  0.0099  0.0093  0.0093  0.0091  0.0086 
(StDev)  ±0.0014  ±0.0014  ±0.0017  ±0.0014  ±0.0014  
Unknown  (Ave)  0.0098  0.0091  0.0089  0.0085  0.0079 
Not Included  (StDev)  ±0.0014  ±0.0014  ±0.0017  ±0.0015  ±0.0015 
Unknown  (Ave)  0.0117  0.0129  0.0143  0.0157  0.0167 
Included  (StDev)  ±0.0014  ±0.0017  ±0.0019  ±0.0016  ±0.0019 
No Unknown  (Ave)  0.0100  0.0095  0.0094  0.0093  0.0088 
(StDev)  ±0.0014  ±0.0014  ±0.0017  ±0.0015  ±0.0015 
Average and standard deviations for α _{ k } on 100 networks, true value is α _{ k } = 0.5
# unknown  15  30  45  60  75  

Equal  (Ave)  0.4678  0.4573  0.4340  0.4220  0.3989 
Weights  (StDev)  ±0.0636  ±0.0759  ±0.0686  ±0.0726  ±0.0699 
ESA  (Ave)  0.4853  0.4903  0.4795  0.4786  0.4642 
(StDev)  ±0.0640  ±0.0767  ±0.0712  ±0.0741  ±0.0719  
Unknown  (Ave)  0.4712  0.4646  0.4429  0.4348  0.4132 
Not Included  (StDev)  ±0.0638  ±0.0779  ±0.0700  ±0.0737  ±0.0702 
Unknown  (Ave)  0.4580  0.4364  0.4172  0.4032  0.3855 
Included  (StDev)  ±0.0668  ±0.0822  ±0.0705  ±0.0799  ±0.0818 
No Unknown  (Ave)  0.4820  0.4838  0.4741  0.4750  0.4595 
(StDev)  ±0.0647  ±0.0759  ±0.0726  ±0.0748  ±0.0689 
Average and standard deviations for ω _{ k } on 100 networks, True Estimate is ω _{ k } = 0.1
# unknown  15  30  45  60  75  

Equal  (Ave)  0.1070  0.1041  0.1042  0.1051  0.10364 
Weights  (StDev)  ±0.0264  ±0.0274  ±0.0262  ±0.0248  ±0.0255 
ESA  (Ave)  0.1069  0.1042  0.1039  0.1059  0.1045 
(StDev)  ±0.0263  ±0.0273  ±0.0264  ±0.0255  ±0.0240  
Unknown  (Ave)  0.1075  0.1054  0.1063  0.1060  0.1054 
Not Included  (StDev)  ±0.0264  ±0.0286  ±0.0269  ±0.0246  ±0.0269 
Unknown  (Ave)  0.1048  0.1101  0.0988  0.1022  0.0993 
Included  (StDev)  ±0.0275  ±0.1035  ±0.0273  ±0.0301  ±0.0285 
No Unknown  (Ave)  0.1078  0.1057  0.1055  0.1070  0.1054 
(StDev)  ±0.0265  ±0.0277  ±0.0256  ±0.0241  ±0.0230 
Notice in the estimates for μ_{ k }in Figure 3 and Table 1, the ESA performs the best compared to the true value and has only a slight reduction in accuracy as the number of unknown events increases. On average the other three estimates seem to degrade more rapidly as the number of unknown events increases. When S_{i,k}=1, the estimates for μ_{ k } are far above the true value and growing as the number of unknown events increases. This follows from the fact that letting S_{i,k}=1 means we are effectively adding events to the network. Take the case when K=2. Assume that each process has 1000 events, and there are 100 unknown events from each process. When we estimate the parameters for the first process, we will use the 900 events we know plus the 200 unknown events from the network. We will get the identical number of events in our estimation for process two. This creates 200 new events and thus biases the estimates for μ_{ k }. This motivated the idea of equal weighting for each unknown event, and that choice is validated by the estimates for μ_{ k }. A similar argument shows why S_{i,k}=0 (i.e. ignoring all the unknown events) has the lowest estimate for μ_{ k } at each level of incomplete data.
In the estimates for the branching ratio α_{ k }, the ESA on average yields the best estimates and maintains its accuracy in the presence of unknown events. It is interesting to note that equal weighting performs worse here than if we let S_{i,k}=0 for all unknown events. Using the ESA overcomes this drawback. Again, setting S_{i,k}=1 for all unknown events performs the worst. This could stem from the fact that most of the unknown events are being labeled background and thus this estimation technique underestimates the branching ratio because fewer events are considered offspring. Notice that the estimate for ESA (dashcircle) tracks the best possible estimate (solid) well while the other three start to trail off as more and more information is labeled as unknown.
Finally, in Figure 5 and Table 1, it is shown that the ESA estimate (dashcircle) for ω_{ k }tracks the behavior of the best estimate (solid) closer than the other methods. Including all of the unknown events (dashsquare) provides the poorest estimate for ω_{ k }. For the other three estimation techniques we see that they are all comparable.
Updating Weights analysis
From Figure 6 it is clear that the StomakhinShortBertozzi score function in solid dark blue, and the Forward Backward score function (cyan dashed diamond) perform nearly identically when looking at the Top 1, Top 2, and Top 3 inferences. These functions look both forward and backward in time from the incomplete event, and are therefore able to identify clusters of events in time. The Probability (black dashed asterisk) and Ratio (solid green square) score functions don’t do nearly as well the StomakhinShortBertozzi and Forward Backward score functions, but better than the Lambda (magenta dashed circle) score function. The Lambda score function appears to perform close to chance (dark green solid plus) for the Top 1, Top 2, and Top 3 inferred process affiliation. Due to the success of the Forward Backward score function and the StomakhinShortBertozzi method, only these are used for further analysis.
The analysis comparing the score functions assumed that the true parameters were known. However, when applying this method in practice there will be error in the estimated parameters. This estimation error will propagate through to the score functions. To understand how deviations of the estimated parameters influence the score functions pairwise combinations of the parameters were increased and decreased by 90% from the target values μ_{ k }=0.01, ω_{ k }=0.1, and α_{ k }=0.5 in 10% increments. In particular the Forward Backward and SSB score functions are computed for pairwise combinations of μ in the range of [0.001,0.019], ω in the range of [0.01,0.19], and α in the range of [0.05,0.95]. Further, in these pairwise combinations, the third parameter is kept at the target value. Notice that a 90% change is larger than the errors observed in the parameter estimates in Subsection “Estimation analysis”.
Runtime Analysis
Though the Forward Backward score function and the SSB method produce comparable results in terms of accuracy, there is a sizable difference in the time it takes to update the weights using these methods.
Convergence Results
Discussion and Future Work
In this paper we propose an effective method for simultaneously estimating the parameters and assigning process affiliation in case of incomplete field data from selfexciting point processes on a network. This problem comes from the demand for law enforcement agencies to identify gang affiliation in the case of unsolved crimes in an area of highly complex gang rivalry activity. We present a new framework we name the Estimate & Score Algorithm for possible application to field data. By testing the method on simulated datasets we can understand its performance features and liabilities. The method is an iterative procedure in which process parameters are estimated alternately with the calculation of network affiliation probabilities. We identify several useful ‘score functions’ for calculating the network affiliations. We also compare the use of unknown events in the parameter estimation. One upshot of our analysis is that the inclusion of unknown events may increase the accuracy of the parameter estimation. Several score functions are considered and the Forward Backward score function shows the most promise with comparable results to that of the StomakhinShortBertozzi method of [24] in the parameter regime tested. The score function calculation is a direct method that does not rely on solving a variational problem, and thus is more computationally efficient than [24].
For future work, space often plays a role in understanding criminal activity [8, 30][33]. Further, criminal behavior has nonrandom structure and can often be framed in terms of routine activity theory [34, 35]. In the case of gang violence, there is a strong spatial component [1, 10, 36]. One can extend the Estimate & Score Algorithm to include space. There is a precedence in the earthquake literature of adding space to selfexciting point processes [13, 15], however, in the case of gang violence, the spatial response may be different. Instead of retaliatory events clustering around prior events, it appears that the data is clustered around regions in space. A spatial model similar to that of [37] could be employed, where the triggering density in space is related to their respective gang setspace, or center of activity [38]. Statistically when modeling spatial point processes one needs to tease out the difference between hot spots due to risk heterogeneity versus event dependence. The data given will be one realization of the underlying process, however using techniques such as prototyping [39], one could potentially reformulate the data into multiple realization of the same process and distinguish between these two phenomena.
There are other factors in the data that can be fused into the model, though more analysis would be required. For example, in earthquake modeling the magnitude of the earthquake is often included. To include such a factor to the intensity λ_{ k }(tH_{τ,k}) one would need to determine a numerical metric to define the impact of each event type. This is not a straightforward task and would require further investigation. Extending this model in this way could allow for the inclusion of events involving tagging, or other low level gang crimes, which could be a precursor to more extreme violent interactions between gangs. Including this data is outside of the scope of the current model but has a strong potential to enrich the overall data set allowing for better analysis.
Using a nonparametric method, they are able to approximated the background function μ(t) and the response function g(t) for a broader class of functions. In this paper, the data was assumed to come from a Hawkes process with constant background rate and an exponential response to previous events. There are cases where the background rate is not constant [40]. Further it is conceivable that the response function could be of a form other than an exponential decay. In this circumstances, the model for λ(tH_{ τ }) in Equation 1 would not be appropriate.
Finally, this method has a great potential in the field of policing. Once such a model has been calibrated correctly, the Estimation & Score Algorithm using the quicker Forward Backward score function can be used to infer the gang association in real time, while the investigation is on going. Given an accurate model of the underlying process, such a method could identify rivalries that have heightened activity.
Author’s contributions
RH led in the construction of the Estimation & Score algorithm, coding, analysis of the results and writing of the manuscript. EL contributed to the construction of the Estimation & Score algorithm, coding, analysis of the results, and writing of the manuscript. AB contributed to the analysis of the results, editing of the manuscript, and conception and design of the Estimation & Score algorithm. All authors read and approved the final manuscript.
Abbreviations
 EM:

Expectation Maximization
 SSB:

StomakhinShortBertozzi
 ESA:

Estimation & Score Algorithm
 Avg:

Average.
Declarations
Acknowledgements
This work was supported by NSF grant DMS0968309, ONR grant N000141010221, ARO grants W911NF1010472 and W911NF1110332, and AFOSR MURI grant FA95501010569. Further, we would like to thank Martin Short, Jeff Brantingham, and George Mohler for their helpful comments and conversations.
Authors’ Affiliations
References
 Radilm SM, Flint C, Tita GE: Spatializing Social Networks: Using Social Network Analysis to Investigate Geographies of Gang Rivalry, Territoriality, and Violence in Los Angeles. Annals of the Association of American Geographers 2010, 100(2):307–326. http://www.tandfonline.com/doi/abs/10.1080/00045600903550428View ArticleGoogle Scholar
 Tita G, Radil S: Spatializing the social networks of gangs to explore patterns of violence. Journal of Quantitative Criminology 2011, 27: 1–25. 10.1007/s1094001090955View ArticleGoogle Scholar
 Tita G, Riley JK, Ridgeway G, Abrahamse AF, Greenwood P: Reducing Gun Violence: Results from an Intervention in East Los Angeles. RAND Press, Santa Monica, CA; 2004.Google Scholar
 Hoff P: Multiplicative latent factor models for description and prediction of social networks. Computational & Mathematical Organization Theory 2009, 15: 261–272. http://dx.doi.org/10.1007/s10588–008–9040–4. [10.1007/s10588–008–9040–4]. [10.1007/s1058800890404]. 10.1007/s1058800890404View ArticleGoogle Scholar
 Koskinen JH, Robins GL, Pattison PE: Analysing exponential random graph (pstar) models with missing data using Bayesian data augmentation. Statistical Methodology 2010, 7(3):366–384.MathSciNetView ArticleGoogle Scholar
 Short MB, Mohler GO, Brantingham P, Tita GE: Gang rivalry dynamics via coupled point process networks. 2011.Google Scholar
 Brantingham P, Glässer U, Jackson P, Vajihollahi M: Modeling Criminal Activity in Urban Landscapes. In Mathematical Methods in Counterterrorism. Edited by: Memon N, David Farley J, Hicks DL, Rosenorn T. Springer Vienna; 2009:9–31. http://dx.doi.org/10.1007/978–3211–09442–6\_2. [10.1007/978–3211–09442–6_2]View ArticleGoogle Scholar
 Brantingham P, Glasser U, Kinney B, Singh K, Vajihollahi M: A computational model for simulating spatial aspects of crime in urban environments. In Systems, Man and Cybernetics, 2005 IEEE International Conference on Volume 4. IEEE; 2005:3667–3674.View ArticleGoogle Scholar
 Brantingham P, Brantingham P: Computer simulation as a tool for environmental criminologistsm. Security Journal 2004, 17: 21–30.View ArticleGoogle Scholar
 Hegemann RA, Smith LM, Barbaro AB, Bertozzi AL, Reid SE, Tita GE: Geographical influences of an emerging network of gang rivalries. Physica A: Statistical Mechanics and its Applications 2011, 390(21–22):3894–3914. http://www.sciencedirect.com/science/article/pii/S037843711100447XView ArticleGoogle Scholar
 Decker S: Collective and normative features of gang violence*. Justice Quarterly 1996, 13(2):243–264. http://www.ingentaconnect.com/content/routledg/rjqy/1996/00000013/00000002/art00005View ArticleGoogle Scholar
 Veen A, Schoenberg FP: Estimation of Space–Time Branching Process Models in Seismology Using an EM Type Algorithm. Journal of the American Statistical Association 2008, 103(482):614–624. http://pubs.amstat.org/doi/abs/10.1198/016214508000000148MathSciNetView ArticleGoogle Scholar
 Ogata Y: SpaceTime PointProcess Models for Earthquake Occurrences. Annals of the Institute of Statistical Mathematics 1998, 50: 379–402. http://dx.doi.org/10.1023/A:1003403601725. [10.1023/A:1003403601725]View ArticleGoogle Scholar
 Ogata Y: Statistical models for earthquake occurrences and residual analysis for point processes. J. Amer. Statist. Assoc 1988, 83(401):9–27. 10.1080/01621459.1988.10478560View ArticleGoogle Scholar
 Zhuang J, Ogata Y, VereJones D: Stochastic declustering of spacetime earthquake occurrences. Journal of the American Statistical Association 2002, 97(458):369–380. 10.1198/016214502760046925MathSciNetView ArticleGoogle Scholar
 Errais E, Giesecke K, Goldberg LR: Affine Point Processes and Portfolio Credit Risk. SIAM Journal on Financial Mathematics 2010, 1: 642–665. http://link.aip.org/link/?SJF/1/642/1 http://link.aip.org/link/?SJF/1/642/1 10.1137/090771272MathSciNetView ArticleGoogle Scholar
 AïtSahalia Y, CachoDiaz J, Laeven R: Modeling financial contagion using mutually exciting jump processes, Tech. rep. National Bureau of Economic Research; 2010.View ArticleGoogle Scholar
 Crane R, Sornette D: Robust dynamic classes revealed by measuring the response function of a social system. Proceedings of the National Academy of Sciences 2008, 105(41):15649–15653. http://www.pnas.org/content/105/41/15649.abstract 10.1073/pnas.0803685105View ArticleGoogle Scholar
 Porter M, White G: Selfexciting hurdle models for terrorist activity. The Annals of Applied Statistics 2011, 6: 106–124.MathSciNetView ArticleGoogle Scholar
 Meyer S, Elias J, Höhle M: A Space–Time Conditional Intensity Model for Invasive Meningococcal Disease Occurrence. Biometrics 2011, 68: 607–616.View ArticleGoogle Scholar
 Mohler GO, Short MB, Brantingham PJ, Schoenberg FP, Tita GE: SelfExciting Point Process Modeling of Crime. Journal of the American Statistical Association 2011, 106(493):100–108. http://pubs.amstat.org/doi/abs/10.1198/jasa.2011.ap09546 10.1198/jasa.2011.ap09546MathSciNetView ArticleGoogle Scholar
 Egesdal M, Fathauer C, Louie K, Neuman J: Statistical and Stochastic Modeling of Gang Rivalries in Los Angeles. SIAM Undergraduate Research Online 2010, 3: 72–94.View ArticleGoogle Scholar
 Hawkes AG: Spectra of some selfexciting and mutually exciting point processes. Biometrika 1971, 58: 83–90. http://biomet.oxfordjournals.org/content/58/1/83.abstract 10.1093/biomet/58.1.83MathSciNetView ArticleGoogle Scholar
 Stomakhin A, Short MB, Bertozzi AL: Reconstruction of missing data in social networks based on temporal patterns of interactions. Inverse Problems 2011, 27(11):115013. http://stacks.iop.org/0266–5611/27/i=11/a=115013 10.1088/02665611/27/11/115013MathSciNetView ArticleGoogle Scholar
 Marsan D, Lengline O: Extending earthquakes’ reach through cascading. Science 2008, 319(5866):1076. 10.1126/science.1148783View ArticleGoogle Scholar
 Marsan D, Lengliné O: A new estimation of the decay of aftershock density with distance to the mainshock. Journal of Geophysical Research 2010, 115(B9):B09302.View ArticleGoogle Scholar
 Sornette D, Utkin S: Limits of declustering methods for disentangling exogenous from endogenous events in time series with foreshocks, main shocks, and aftershocks. Physical Review E 2009, 79(6):61110.MathSciNetView ArticleGoogle Scholar
 Lewis E, Mohler G: A Nonparametric EM Algorithm for Multiscale Hawkes Processes. Preprint 2011.Google Scholar
 Dempster AP, Laird NM, Rubin DB: Maximum likelihood from incomplete data via the EM algorithm. J. Roy. Statist. Soc. Ser. B 1977, 39: 1–38. [With discussion] [With discussion]MathSciNetGoogle Scholar
 Brantingham PJ, Brantingham PL: Environmental Criminology. Sage Publications, Inc, Beverly Hills, CA; 1981.Google Scholar
 Herbert DT: Geography of Urban Crime. Longman Inc, London; 1982.Google Scholar
 Eck JE, Chainey S, Leitner JGCM, Wilson RE: Mapping Crime: Understanding Hot Spots, National Institute of Justice. 2005, 1–71. http://www.ojp.usdoj.gov/nijGoogle Scholar
 Brantingham PL, Brantingham PJ: Criminality of place. European Journal on Criminal Policy and Research 1995, 3: 5–26. [10.1007/BF02242925] http://dx.doi.org/10.1007/BF02242925.View ArticleGoogle Scholar
 Brantingham PJ, Brantingham PL: Environment, routine situation: Toward a pattern theory of crime. Advances in Criminological Theory 1993, 5: 259–294.Google Scholar
 Cohen LE, Felson M: Social Change and Crime Rate Trends: A Routine Activity Approach. American Sociological Review 1979, 44(4):588–608. http://www.jstor.org/stable/2094589 10.2307/2094589View ArticleGoogle Scholar
 Block R: Gang Activity and Overall Levels of Crime: A New Mapping Tool for Defining Areas of Gang Activity Using Police Records. Journal of Quantitative Criminology 2000, 16: 369–383. http://dx.doi.org/10.1023/A:1007579007011. [10.1023/A:1007579007011]View ArticleGoogle Scholar
 O’Leary M: Modeling Criminal Distance Decay. Cityscape: A Journal of Policy Development and Research 2011, 13(3):161–198.MathSciNetGoogle Scholar
 Tita G, Cohen J, Engberg J: An Ecological Study of the Location of Gang “Set Space”. Soc. Probl 2005, 52(2):272–299. 10.1525/sp.2005.52.2.272View ArticleGoogle Scholar
 Nichols K, Schoenberg F, Keeley J, Bray A, Diez D: The application of prototype point processes for the summary and description of California wildfires. Journal of Time Series Analysis 2011, 32(4):420–429. 10.1111/j.14679892.2011.00734.xMathSciNetView ArticleGoogle Scholar
 Lewis E, Mohler G, Brantingham PJ, Bertozzi A: Selfexciting point process of Insurgency in Iraq. Security Journal 2011, 25: 0955–1662.Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.