A two-route CNN model for bank account classification with heterogeneous data.

Area of Science:

• Computer Science
• Data Science
• Financial Technology

Background:

• Classifying bank accounts using transaction data aids in detecting illegal financial activities.
• Existing methods often fail to leverage heterogeneous features embedded within time series data.

Purpose of the Study:

• To propose a novel TRHD-CNN model for bank account classification.
• To effectively utilize heterogeneous features from time series transaction data for improved accuracy.

Main Methods:

• A two-route Convolutional Neural Network (TRHD-CNN) model is developed.
• The model processes two types of heterogeneous feature matrices independently using a divide and conquer strategy.
• A DirectedWalk method is employed to learn network vectors for embedding neighbor relationships.

Main Results:

• TRHD-CNN demonstrates significant advantages over existing methods on a real bank transaction dataset.
• The model achieved recall scores up to 5.15% higher than competing methods in identifying illegal pyramid selling accounts.
• The two-route architecture proved effective in mining complementary classification characteristics.

Conclusions:

• The TRHD-CNN model offers a powerful approach for classifying bank accounts and detecting financial crime.
• The model's architecture is adaptable for multi-route scenarios and other application fields.
• Leveraging heterogeneous features is crucial for enhancing the accuracy of financial transaction analysis.