Understanding — Credit card transaction fraud feature extraction using “Parenclitic Networks”
“Parenclitic Network” What is it ? How does it help in Outlier Detection?
The fundamental idea is to use graph network measures as features in addition to the normal transnational features. Also, there is one difference between the traditional graph networks and parenclitic networks. Nodes are associated with the features, but the links or edges are weighted according to the deviation of the transaction value from the values of two features. Below picture explains it a little bit more.
This representation identifies which parts and relationships have deviated by the transaction. Data is projected on to every pair of features and a line is fit using linear regression. When an unlabeled data comes in you calculate the deviation and use it as to create the edge.
Let us dive into the paper
I agree, ordinary features are only useful when the fraud transaction standout on their own. There is always a motivation to figure out new features. [Pg.1]
Parenclitic networks are borrowed from the medical field. They are used to extract new features and augment them with the transaction features. [Pg.1]
Explanation of the parenclitic network representation in the paper. [Pg.2]
A list of the network features [Pg.3]
A list of features from the credit card transactions [Pg.3]
A little bit about the data set. [Pg.3]
A fully connected layers with sigmoid activation activation is used. [Pg.4]
Rest of it the usual stuff normally followed. Results show the ROC with AUC along with various comparisons. Details in the original paper https://arxiv.org/pdf/1706.01953.pdf
At the end I had few questions which I think were not clear from the paper for me. Here they are
