Source-linked AI summary
Once Upon a Crime: Towards Crime Prediction from Demographics and Mobile Data
Andrey Bogomolov, Bruno Lepri, Jacopo Staiano, Nuria Oliver, Fabio Pianesi, Alex Pentland
TL;DR
The paper addresses whether crime hotspots can be predicted without relying solely on historical crime data, using aggregated mobile-network behavior and demographic information. It develops and evaluates a multimodal approach on London data, achieving almost 70% accuracy and reporting a 6% improvement over rich borough statistics.
Problem
The paper investigates whether aggregated human behavioral data from mobile networks, combined with demographics, can predict crime hotspots beyond historical knowledge or offender profiling.
Method
The approach combines anonymized, aggregated mobile-network activity and demographic features to predict whether geographic areas will have high or low crime levels.
Results
Almost 70% accuracy was obtained for predicting high or low crime levels in the next month, with a 6% improvement over rich borough population statistics.
Takeaways & Limitations
Aggregated mobile-network data contain relevant information about geographic areas for predicting crime levels, while daily and monthly behavioral data improve prediction over borough statistics.
Takeaways & Limitations
The approach is evaluated on crime data from London and predicts crime levels for a geographic area in the following month.
Abstract
from arXiv · showhide
In this paper, we present a novel approach to predict crime in a geographic space from multiple data sources, in particular mobile phone and demographic data. The main contribution of the proposed approach lies in using aggregated and anonymized human behavioral data derived from mobile network activity to tackle the crime prediction problem. While previous research efforts have used either background historical knowledge or offenders' profiling, our findings support the hypothesis that aggregated human behavioral data captured from the mobile network infrastructure, in combination with basic demographic information, can be used to predict crime. In our experimental results with real crime data from London we obtain an accuracy of almost 70% when predicting whether a specific area in the city will be a crime hotspot or not. Moreover, we provide a discussion of the implications of our findings for data-driven crime analysis.
I. Introduction
The paper frames crime as a social and economic problem and investigates whether aggregated, anonymized mobile-network behavior combined with demographics can predict neighborhood crime hotspots. It adopts a place-centric, data-driven perspective and evaluates the approach’s predictive power and implications.
- Crime is associated with impacts on quality of life, economic development, and local and national economic growth.
- Prior research links criminal activity to socioeconomic variables including education, ethnicity, income, and unemployment.
- Crime hotspots can cluster at micro-geographic levels without matching broader neighborhood trends.
- The study uses a place-centric formulation to predict whether geographic areas will become crime locations from aggregated mobile activity and demographic information.
- The paper evaluates the proposed model, compares it with an official-statistics approach, and discusses its theoretical and practical implications.
II. Related Work
Related work examines crime through people-centric profiling, place-centric hotspot analysis, and social-media prediction. This paper extends data-driven crime forecasting by using mobile network activity as human behavioral data.
- People-centric research has studied criminal profiling and spatiotemporal patterns in offenders’ behavior and historical crime data.
- Place-centric research has applied quantitative methods to analyze spatial and temporal patterns in crime hotspots.
- Social-media crime prediction combines tweet semantics with spatiotemporal and neighborhood demographic information.
- The paper complements these approaches by leveraging mobile network activity to forecast crime hotspots.
III. Datasets
The study uses datasets from a London-area public competition, combining anonymized mobile-network behavioral data with geo-localized open data. The open data include crime and multiple contextual indicators.
- The datasets were provided through the 2013 Datathon for Social Good competition in London.
- The Smartsteps dataset contains anonymized and aggregated human behavioral data derived from mobile network activity in the London Metropolitan Area.
- Geo-localized open data include reported crimes, property sales, transportation, weather, and borough-profile indicators totaling 68 metrics.
- The paper then describes the specific datasets used to predict crime hotspots.
A. Smartsteps Dataset
Smartsteps divides the London Metropolitan Area into geographically located cells and provides hourly demographic and activity estimates for each cell. The dataset includes population presence, location status, and demographic composition variables.
- Smartsteps divides London into 124119 cells with provided locations and surface areas, although the precise cell shapes are unavailable.
- For each cell, demographic variables were computed hourly across two periods spanning December 9, 2012, to January 5, 2013.
- Footfall estimates the number of people in each cell from aggregated phone-call activity, cell-tower coverage mapping, and network-market-share extrapolation.
- Hourly cell data include estimated population, home/work/visitor percentages, and gender and age splits.
- Table 1 organizes SmartSteps data by one-hour intervals for each Smartsteps cell.
B. Criminal Cases Dataset
The criminal cases dataset records reported UK crimes with geographic and categorical information, but only month and year—not exact dates—are available. Experiments use January 2013 crimes as the Smartsteps-cell ground truth.
- The dataset contains the geolocation of all reported crimes in the UK, but records only the month and year of each crime.The competition data covered December 2012 and January 2013.
- Each crime record includes an ID, month and year, coordinates, address, police department, LSOA identifiers, and one of 11 crime types.Examples include antisocial behavior, burglary, violent crime, and shoplifting.
- LSOAs are small, population-based statistical areas defined by the UK Office for National Statistics after the 2001 census.They have a mean population of 1,500 and a minimum threshold of 1,000.
- The experimental ground truth consists of crimes reported in January 2013 for each Smartsteps cell.
C. London Borough Profiles Dataset
The London borough profiles dataset provides official demographic and socioeconomic information at LSOA-level spatial granularity. It contains 68 metrics describing population, living conditions, economic activity, wellbeing, and civic characteristics.
- The official London borough profiles dataset contains 68 population-related metrics at LSOA spatial granularity.
- Its measures cover population, households, demographics, migration, ethnicity, language, employment, earnings, housing, deprivation, health, happiness, and political participation.
- The profiles are combined with aggregated and anonymized demographic and mobility data to represent fine-grained spatio-temporal dynamics.
IV. Methodology
The paper formulates crime hotspot forecasting as binary classification: for each Smartsteps cell, it predicts whether the cell will be a crime hotspot in the next month.
- Crime hotspot forecasting is cast as a binary classification task over Smartsteps cells.
- For each Smartsteps cell, the model predicts whether that geographic cell will become a crime hotspot in the next month.
A. Data Preprocessing and Feature Extraction
Preprocessing links crime and borough-profile records to Smartsteps cells, while feature extraction summarizes behavioral variables and their temporal structure. Borough profiles retain their original 68 features without preprocessing.
- Data Preprocessing and Feature Extraction: The workflow begins with preprocessing the bulk data described in the preceding section.
- Data Preprocessing and Feature Extraction: Crime events and borough-profile information are linked to Smartsteps cells by matching each record to the closest cell geographically.
- Data Preprocessing and Feature Extraction: For each Smartsteps variable, mathematical functions characterize the distributions of behavioral measurements, including diversity and regularity.
- Data Preprocessing and Feature Extraction: Repeating these computations over 1-hour, 4-hour, and 1-day sliding windows captures temporal relationships while reducing computational complexity and feature-space size.
- Data Preprocessing and Feature Extraction: The London borough profiles require no preprocessing, so all original 68 borough-profile features are used.
A.4. Crime hotspots ground-truth data
The study converts highly skewed crime counts into approximately balanced low- and high-crime classes, then selects a compact feature subset and compares several classifiers. Random Forest performs best among the evaluated models.
- Crime counts were split at the median into low-crime and high-crime classes because the distribution was highly skewed.The distribution had skewness 5.88, kurtosis 72.5, mean 8.2, and median 5.
- 53.15% of cells belonged to the high-crime class, making the two resulting classes approximately balanced.
- The feature-ranking metric was mean decrease in the Gini coefficient, selected because it outperformed mutual information, information gain, and chi-square statistics.The reported average improvements were approximately 28.5%, 19%, and 9.2%, respectively.
- 68 features were retained from an initial pool of about 6000, reducing dimensionality by approximately 90 times.Feature selection used a pipelined ranking and subset-selection approach based only on training data.
- The models used 5-fold cross-validation and included logistic regression, support vector machines, neural networks, decision trees, and tree ensembles.
- Random Forest achieved the best performance among the evaluated classifiers, so subsequent results report this model.The method uses ensembles of randomized trees whose votes determine the predicted class.
V. Experimental Results
Random Forest predicts whether London Smartsteps cells will become crime hotspots, achieving almost 70% accuracy. Behavioral mobile-network features, especially temporal and at-home patterns, outperform borough-profile statistics and improve prediction at finer temporal and spatial resolution.
- Almost 70% accuracy was achieved when predicting whether a Smartsteps cell would be a crime hotspot in the following month.The reported evaluation metrics were accuracy, F1, and AUC, using a test set not used during training.
- Second-order features capturing intertemporal dependencies significantly improved model-performance metrics while making the feature space more compact.
- The majority-class baseline achieved 53.15% accuracy by always predicting the high-crime class.
- 69.54% accuracy was obtained by the Smartsteps+Borough model versus 62.18% for the borough-profiles model using the same number of variables.The combined model improved accuracy by over 7% relative to the borough-only model.
- Mobile behavioral data significantly improved prediction over rich borough statistics while offering finer temporal and spatial resolution despite being less comprehensive.Borough profiles were described as expensive and effort-consuming to collect and updated infrequently.
- Smartsteps features had greater predictive power than official borough statistics, with daily multi-day features outperforming monthly features.The top-20 features contained no borough-profile features.
- At-home features accounted for 11 of the top 20 predictors, while entropy-based features accounted for 8.Shannon entropy was used to capture predictable structure in the types of people present in a place over a day.
VII. Implications and Limitations
The multimodal approach predicts next-month crime levels from aggregated mobile-network behavior and demographic information, with practical implications for narrowly targeted crime monitoring. Its scope is constrained by short mobile-data coverage and monthly crime aggregation, although it improves accuracy over traditional borough statistics.
- Implications: Almost 70% accuracy was achieved when predicting whether a geographical area would have high or low crime levels in the next month.The approach uses a multimodal combination of mobile-network activity and demographic information.
- Implications: Predictions could inform police departments and city governments where to invest efforts and enable quicker responses to criminal events.The approach may also support intervention in narrowly defined geographic areas.
- Implications: Aggregated and anonymized mobile-network activity features combined with demographic information provide relevant information for predicting a geographical area’s crime level.This differs from approaches based mainly on historical crime knowledge, criminal profiling, or broad socio-economic descriptions.
- Implications: Dynamic features describing the demographics and activities of a place may identify similar high-risk areas even when they lack recent reported crimes.The method can also identify novel risk-inducing or risk-reducing geographic features.
- Limitations: Three weeks of Smartsteps data and monthly crime aggregation limited validation at finer time scales such as the week, day, or hour.The authors identify finer-grained crime data as necessary to validate predictions at those temporal resolutions.
- Implications: 6% higher prediction accuracy was reported than when using traditional statistical data about a borough’s population, with a similar number of variables.The traditional statistics were described as rich but expensive to collect.