Feature Engineering and Interpretability of xgboost based on spatial and temporal crime data and Twitter data
The goal of this study is to use data mining techniques to investigate the impact of different levels of feature engineering on the xgboost algorithm on crime prediction. Many studies on the use of trees method to improve crime prediction models have been conducted; however, to the best of our knowledge, many of them lack the appropriate implementation of data mining techniques.
Few of those who mentioned feature engineering were not concerned with estimating the effect of various feature engineering techniques on prediction. None of these studies used xgboost to successfully present the contributions of individual predictors. For these reasons, the study intends to investigate and report on the impact of appropriate feature engineering, as well as to explain the contribution of individual variables on a crime prediction model using a crime dataset from the United Kingdom.
Research questions:
a. At what level of feature engineering does our model have the highest accuracy? b. What is the best type of data splitting for the UK crime dataset? c. What type of resampling is most appropriate for the UK crime dataset? d. What is the best solution for class imbalance on the UK dataset? e. What are the most critical variables?
Research objective
a. Application of data mining techniques to predict crime rate in the UK. b. Estimate the influence of individual predictors on xgboost model. c. Develop a crime prediction shiny App (optional).