Spectral Clustering and Cost-Sensitive Deep Neural Network-Based Undersampling Approach for P2P Lending Data

Peer-to-peer lending, also known as P2P lending, is the new generation loan disbursement process, where lenders and borrowers communicate through online services. Loans through P2P lending platforms are generally unsecured, due to the presence of borrowers with low credit scores. Lendingclub dataset, consisting of quantitative and qualitative information of borrowers from 2007 to 2011, is taken for the research. Machine learning models trained with such imbalanced dataset consists of biasing towards major class samples. The model performs significantly well on major class (safe borrowers) in terms of high precision but does not perform significantly well on minor class (defaulted) borrowers and provides low recall on minor class samples. To deal with the issue, a novel undersampling algorithm based on the combination of spectral clustering and cost sensitive deep neural network (SCCSDNN) is proposed. Experimental results showcased the outstanding performance of the proposed technique, and it outperforms state of the art undersampling, oversampling and ensemble resampling techniques.