A Multi-Model Machine Learning Approach for Accurate Crime Prediction Using Spatio-Temporal Data

The significance of predicting crime therefore arises from the possibility of making reliable assumptions that serve the intended organizations in preventing crime. Traditional crime models suffer from three main limitations: low density of data, difficulties for quantifying significant information from parameters that probed space-time, and low adaptability of the model to areas or crimes not envisaged in its dataset. To tackle these problems, the paper offers a brand new multi-module crime prediction system based on the modern machine learning methodologies, which operates on multivariate time-space data. The model consists of three sub-models: The system includes the proposed Attention-based Long Short Term Memory (ATTN- LSTM), a temporal spatial, bidirectional LSTM, as well as a combination of spatial-temporal worksheets is a Fusion Learning Framework (FLF) combined with the Dynamic Learning Fusion Tool (DLF). The DLF module refines the model by adding its refinement of the outputs of the various sub-models hence in- creasing on accuracy. Also, the transfer learning method cuts the training time because it uses features from similar datasets. The implemented model is checked on extensive crime datasets from San Francisco and Chicago jurisdictions; the MAE, MSE, R2 and SMAPE which were used in the evaluation of performance show R2 of about 0.92—0.97 which depicts the model performance is accurate. This approach allows for predicting the hourly crime rates for various type of crime and representing these results graphically in a form of pie charts, which can be useful for policemen. This work will be continued in the future to reduce training time and improve the applicability of the model to cases where there is little or no data on certain types of crime. Although MDPIS is quite efficient for MEP training, problems like longer training time and data scarcity are still present, and later versions of this forecasting tool can theoretically solve these problems thereby providing an environment of real-time prediction.