Algorithm-Driven Cybercrime Analysis: An Application of Decision Trees, XGBoost and TimeXer

This paper focuses on the use of multiple machine learning algorithms in cybercrime research. In the data processing stage, VCDB and ITU data are cleaned and statistically analysed through outlier checking, missing value processing, date format standardisation and other operations. In order to explore the laws of cybercrime, data visualisation technology is used to draw heat maps of global and regional cybercrime events, visually displaying the characteristics of crime distribution. In the construction of the crime fighting model, XGBoost and decision tree algorithms are used. The decision tree improves model diversity and classification performance through bootstrap sampling, randomly selecting feature subsets, constructing the CART algorithm, and pruning; XGBoost approximates the objective function with the help of a quadratic Taylor expansion and introduces a regularisation term to control the complexity of the model. In comparison, the XGBoost model is more accurate. In addition, the TimeXer time series prediction model based on the Transformer architecture is used to predict cybercrime. The model captures time series dependencies through a self-attention mechanism, extracts features in different subspaces with the help of multi-attention, and injects temporal information using location coding. Experimental results show that the TimeXer model prediction accuracy is significantly improved by adding specific metrics. This study verifies the effectiveness of the algorithm in cybercrime research and provides technical support and decision-making basis for related fields.