Machine learning approach to identity resolution for criminal profiling

A common dilemma when working on criminal data is that often people manipulate their details to disguise themselves and hide their identities which leads to creating ambiguous and false identities. Deep Neural Network (DNN) is applied to work well on fraudulent and imbalanced data. Two subcategories of DNN, the Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) with Python libraries such as TensorFlow and Keras are applied to the dataset for the detection of suspects with false identities to assist the process of analytical investigation by law enforcement agencies. Upon application of this approach to the anonymized policing dataset from SPIRIT Project funded by European Union’s Horizon, 5 main suspects with false identities were identified out of 23 targets on 39 million records. As working on criminal data is quite sensitive so to avoid data leakage while training and testing the data, K-fold validation has been applied. Furthermore, cascading both MLP and LSTM models on the policing dataset resulted in improved model prediction and accuracy compared to using each model individually. The cascaded model notably reduced false predictions for the recurring criminal patterns such as age-specific trends and most prevalent crime activities.