This exploratory study aims to address the problem that cash loan fraud customers are difficult to detect manually. Cash loan is a new consumption model in the concept of Internet consumer finance(ICF). Manual detection of fraudulent customers requires a lot of manpower and time, and often causes great losses to financial institutions, so our group did the research mentioned above. In this paper, we proposed a Semi-supervised Pre-loan Fraud Detection (SPFD) system via investigating various supervised and unsupervised learning algorithms on basis of 285,771 applicants' desensitized data from MUCFC (a Chinese ICF company). In SPFD, feature selection methods consist of KL Divergence, Wasserstein Distance and Manual Selection, while the clustering algorithms we adopted was K-constrained seed clustering. Final result demonstrates good performance with the Adjusted Rand Index(ARI) reaching 81.7%. Such method would help financial institution to reduce financial losses.