A comparison study of semi-supervised SVM algorithms for small business credit prediction

The small companies become increasingly important in bank's lending business. But the challenge is how bank's credit assessment is made in a small amount of time and money. Compare with the big companies, the small companies often need a small amount of cash flow. They may not provide the complete certificates or documents, so that the bank has to collect information of the companies and evaluate their credit rating especially by experts. For the bank, it is worthless to spend time and money to investigate a small company, especially just to lend several hundred thousand dollars. In the real life, credits of most the companies are good, while only small of them cannot repay for some reasons. The few number of small companies' credit data is valuable while considerable unknowing credit data of small companies is within reach. Therefore, the binary classification of the good credit and the bad credit is asymmetry. we choose supervised learning algorithm (Regularized Least Squares Classification and SVM) and semi-supervised learning algorithm (Transductive SVM and Deterministic Annealing Semi-supervised SVM) to predict the credits of small companies. In this paper, we conduct a series of experiments on credit datasets with different proportion classification and the results show that the Deterministic Annealing Semi-supervised SVM (DAS3VM) performance better when the data set is rare and asymmetry.