Suspect matching on spatial and contextual aspects Rob van der Veer, rvdveer@sentient.nl 8th National Crime Mapping Conference, 10 June 2010 Agenda Introduction Work history Matching using associative memory Series detection Match case(s) to offender locations Match case(s) to offenders description Match case(s) to careers Example and results Do it yourself suspect matching 2 Rob van der Veer, CEO Sentient Data mining specialists since 1990 Own software: DataDetective Customers: – – – – – – Fraud analysis (Tax offices) Marketing (Delta Lloyd) Market research (De Telegraaf) Risk analysis (Cordares, KPN) Product advice (Libraries) Crime analysis (Police) Broad co-operation UvA, MapInfo, Hot ITem, Vicar Vision, ParaBots, vtsPN, Police academy, Experian, VU 3 Work history 1990: Company founded. Artificial Intelligence research and application development 1996: First law enforcement application 2002: Amsterdam police force adopts DataDetective 2006: More police forces join KDD2009: Third price ‘best data mining application’ 2010: DataDetective proposed as national standard 2010: Government funds project to pilot DataDetective at city councils 4 Data sources DataDetective Find relations Predict GIS Statistics, charts Cluster Link analysis Fuzzy query Applications / BI systems Query Data+metadata for other tools SPSS Analyts’s Notebook Google earth 5 Automatic reports and dashboards 6 Application: Crime type clustering Problems on queens day 7 Application: link analysis Onderzoek relatie tussen verdachten (rood) van geweld en slachtoffers (wit) 8 Application: Profile Analysis Application: find interacting factors 1.683 794 47,18% Geslacht Man Vrouw 1.116 656 58,78% 567 138 24,34% leeftijdskl 12-17 ja nee ja nee 175 130 74,29% 941 526 55,90% 21 10 47,62% 546 128 23,44% nationaliteit XXXXXXXXXX ja cbs-buurt YYYYYYYYYYY geboorteland NEDERLAND nee leeftijdskl 18-24 ja nee ja nee 390 252 64,62% 551 274 49,73% 135 44 32,59% 411 84 20,44% geboortegemeente Goirle Ooit getrouwd? 22 153 20 110 90,91% 71,90% cbs-wijk Wijk XXXXXXXXX) ja nee nee ja 105 80 76,19% 285 172 60,35% 190 129 67,89% 361 145 40,17% Links/rechtshandig rechts links 29 76 28 52 96,55% 68,42% ja nee 12 123 9 35 75,00% 28,46% burg.staat Gescheiden ja nee 28 333 23 122 82,14% 36,64% 10 Application:prediction models Personal file Dataminingmodel Risk of fire arm (0-100%) Report Domestic violence Dataminingmodel Risk for escalation (0-100%) Neighbourhood, surroundings, infrastructure, households Dataminingmodel What risks to expect Dataminingmodel Risk for an event (0-100%) Exact location Day, time, season, weather prediction 11 The weather report for tomorrow… … for street robbery. Application: Where/when analysis Tue afternoon Using knife Hot spots (location, day of week, time) Th, Fr, Sa Late afternoon Th, Fr Evening We age 12-15 Street robbery Su to Mo Grabbing purse 13 Application: geographic trend analysis 14 Matching using associative memory Input case(s) Matched with processed train data Associated cases Conclusion 15 Matching algorithm Height category Age Series detection Clusters indicate trends and series 17 18 Match case(s) to offender locations ‘geographic profiling’ 19 Match case(s) to offender descriptions Match witness descriptions to known offenders Applications: Selecting foto’s to show Narrow down search Field test 1996: 50% more hits 20 Match case(s) to careers Series Match Similar cases in past Suspect of four similar cases Suspects of similar cases 21 Profile typical suspect What are typical features of the suspects of past cases that are similar to the problem at hand? 22 23 Example Summer 2009: Series: Tilburg area Car break-ins in car parks Laptops stolen Matched to similar incidents in past Printed photos of suspects of those similar incidents Photo of suspect X matched CCTV footage Patrolling offers briefed with photos August 2009: X spotted during patrol and taken in. Was part of international gang. 24 Results Field tests show a 50% increase in search hit Experiments with crime data show a similar gain While initially hesitant, use increases constantly Also applied by non-technical users User surveys report an efficiency gain of factor 20 User surveys report a quicker response 25 Do it yourself suspect matching 1. Select case(s) and possible suspects 2. Score suspects with geographic profile probability 3. Score suspects with similarity to description a) b) c) Find the governing witness description Assign score 100 to exact matching suspects Relax search criteria and assign lower scores for matches 4. Score suspects with matching careers a) b) c) Find the governing MO, location and time plus derived features Match-score past incidents using the method above Select the best matching incidents and score suspects depending on how many best matching past incidents they have 5. Combine the scores to sort and select suspects 26 More information Rob van der Veer rvdveer@sentient.nl +31 20 5300 330 Hedda Roos Amsterdam police force +31 20 5598 495 hedda.roos@amsterdam.politie.nl www.sentient.nl/?crime 27