Crime Rate Analysis And Prediction Usingcontext-aware Tensor Decomposition

Crime is one of the most important social problems in the country,affecting economics,children development,public safety and citizen's life quality.Modeling the whole crime situation of a city is critical for policy makers in their efforts to reduce crime.Traditional methods for crime rate analysis and prediction have some drawbacks.First,the time granularity is not accurate enough,traditional methods usually analysis and predict the crime rate for a whole year or even longer a period,it ignores the complex time dependencies among crime incidents and leading the result to be less convincible.Second,traditional methods usually predict the whole crime situation of a place without considering the crime category,resulting the model to be less useful.Based on those weaknesses,in this paper we propose context aware tensor decomposition model to analysis and predict crime rate.The main contributions of this paper are as follow.(1)To our knowledge,this paper is the first work using tensor decomposition to solve crime rate analysis and prediction problem.Crime records recorded by police office usually have information about crime category,crime location and timestamp.The data is actually a result of “crowd sensing”,containing rich information that can help diagnose urban crime situations.Our method is mainly composed of four part,they are tensor construction,tensor decomposition and recovery,crime rate visualization,crime rate prediction,respectively.Among them crime rate visualization can help us found unreported and potential crime incidents,and crime rate prediction can predict the whole crime rate of a place before they happen.(2)Moreover,we optimized traditional tensor decomposition model by introducing POI data.It leveraged the concept of latent factor models and matrix factorization to achieve more accurate prediction result.(3)The method we propose can predict the crime rate of different community for different crime category in different month.Experiment shows that our method outperforms traditional method by several evaluation metrics.