Predicting When Crime Will Occur: Building a Methodology for Timely Crime Prediction

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K14365
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 25010:Social systems engineering-related
• Research Institution: Tohoku University
• Principal Investigator: Ohyama Tomoya 東北大学, データ駆動科学・AI教育研究センター, 助教 (80893776)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 安全・安心 / 犯罪予測 / 犯罪予防 / 時空間モデリング / 地理的クラスタリング / 異常検知 / 機械学習

Outline of Final Research Achievements

In this study, I developed a method to accurately predict the timing of crimes, an aspect that has been neglected in existing research. Specifically, I aimed to predict the date of occurrence by integrating two factors: (1) the effects of time factors that can be identified in advance (special days, days of the week, weather, events, etc.), and (2) fluctuations due to the increased activity of motivated offenders. I evaluated the prediction performance of the proposed method in Osaka City (vehicle theft and parts theft) and Seoul City (outdoor violence and sex crimes). The results showed that the proposed method outperformed models that randomly select days and models using existing machine learning methods. Furthermore, the prediction performance was higher for sexual crimes than for simple violence, suggesting that the proposed method may be more effective for crimes committed with clear criminal intent.

Free Research Field

犯罪学、社会工学

Academic Significance and Societal Importance of the Research Achievements

犯罪予測は，理論，データ，アプローチ含め，これまで地理的な側面（どこで起こるか）に力点がおかれ，時間的な側面（いつ起こるか）に関しては相対的に関心が低く，予測性能は限定的であった．そのため，犯罪が起こる地域・場所はわかっても，時期がずれて防げないという事態が生じていた．本研究は，こうした問題を解決するために犯罪発生の時間的要素を詳細化し，曜日等の予め把握可能な要因と，犯罪時系列データさえあれば，ある程度の精度で予測が可能であることを示した点で意義がある．また，低頻度な犯罪は発生時期まで予測するのが困難であるが，本研究では予め地域を地理的にグルーピングして予測するアプローチの有効性も示した．