The invention discloses a telecommunication fraud detection model fusing spatial-temporal
feature extraction and ensemble classification, and the model comprises the following modules: a
feature selection module which is used for guiding GMGWO to accurately screen an optimal feature subset from
original data through the deep fusion of ST and GMGWO; the
data reconstruction module is used for integrating a deep neural network DCNN and an
elastic network self-encoding ECAE, the deep neural network DCNN extracts local spatial features of data, the
elastic network self-encoding ECAE considers feature sparsity and model
noise immunity through an L1 + L2 regularization strategy, and
data reconstruction is completed based on node similarity; and the integrated depth classification module is used for fusing the BiTCN, the BiGRU and the ST, the BiTCN captures
local space-time characteristics, the BiGRU models a long
time sequence dependency relationship, and the ST mechanism dynamically focuses on key
time sequence fragments and node association information. According to the method, an integrated framework fusing
feature selection and deep classification technology strategies is constructed, and the detection precision and the model generalization are improved.