Feature Encoding For High Cardinality Categorical Variables Using Entity Embeddings: A Case Study in Customs Fraud Detection

Customs authorities nowadays are pressurized by the increasing levels of international trade and insufficient resources to perform physical check on all the trades. Government therefore aims to focus their limited resources on the highly suspicious trade with the help from data mining especially in detecting suspicious trade. In customs fraud detection system, a total of 38,173 records focusing on the Harmonized System code starting with 4011 that shared by the Royal Malaysian Customs Department was analyzed. From the analysis of the datasets provided, one of the main difficulties faced is the issue of high cardinality where there are a lot of unique values in single feature. Thus, Entity Embeddings as a method in solving this issue by providing better representation of the unique values in lower dimensions was introduced. The result of the proposed feature encoding method can achieve higher AUC-ROC and F1-Score in Logistic Regression, Support Vector Machine, and Neural Network models compared to other feature encoding techniques.