Abstract:
Mobile money service is an emerging technology, which has gained rapid growth over
the past years. Especially in developing countries, mobile money service has become a
popular platform where it does not require any formal bank account. Due to its
convenient nature, easy usability and accessibility, mobile money service has opened
doors for fraud activities, such as money laundering and drug trafficking.
In this paper, it is focused on improving the mobile money fraud detection system using
social network analysis. The mobile money transaction data was loaded to Spark
GraphFrame library by forming users as vertexes and transactions as edges. When a
fraudulent transaction occurs, the length of the transaction is usually affected.
Therefore, three user motifs (chains), four user motifs and five user motifs have been
extracted from the mobile money transaction network. Then created new features as
sender and receiver involvement to the motifs. The page rank and degree are calculated
for the vertex in the network. The extracted network features merge with the original
transaction data. In this research, random under-sampling has been implemented to
overcome the class imbalance problem. Gradient boosting, multi-layer perceptron
(MLP) and deep neural network (DNN) classifier were implemented using 70% of data
as train data and tested using 30% of data as the test data on transaction data with and
without network features. Classification models have been implemented using
transaction data with network features showed better performance than the classifiers
used only transaction data. The MLP model outperformed with 99.8% accuracy, 0.64
precision, 0.88 recall and 0.7429 F1-score which was higher than the benchmark study.
