Network Graph AnalysisCategorizing Private Individuals and Private Firms in a Bank

Abstract: This study deals with classifying private individuals and private firmsthrough transaction data of a financial institution. Private firm(Enskild Firma) in Sweden is difficult to distinguish from privateindividual through transactions as it is owned by a sole trader who hasa small business. There is no special kind of bank account required bythe bank for this firm. Bank maintains the status of the users asprivate individuals or private firms, but that information is not up todate in most cases. This creates problems in detecting fraud and targeting customers with appropriate offers and advertisements. Data contains transactions of individuals over the period of one year.To perform appropriate analysis techniques, data is pre-processed andtransformed into features from the transactions. 67 features are beingconsidered for each user. Most of the data is unlabeled. Hence, semisupervised techniques like Label Propagation and Label Spreading areused. Those Algorithms assign labels to the users such as private firmor private individual through a network graph. To validate the labels,Supervised Learning techniques; Logistic Regression, Random Forest and AdaBoost are used. This study shows results of network analysis inclassifying accounts and provides a comparison between LPA and LSA. 15%accuracy increase is observed with AdaBoost after training on additionalsamples labeled by LPA.