Since the emergence of the World Wide Web, online shopping and online auction have gained increasing popularity. While people are enjoying the benefits of online trading, criminals are also taking advantage of fraudulent activities against honest parties to obtain illegal profits. Therefore, proactive fraud detection restraint systems are commonly applied in practice to detect and prevent such fraudulent and illegal activities. Machine-learned models, especially those that are learned online, are able to detect fraud more efficiently and faster than human-tuned rule-based systems. In this article, we propose an online probit model framework that simultaneously takes into account the selection of online functions, the limits of human knowledge coefficients and multi-instance learning. Through empirical experiments in a real-world online fraud auction, detection data show that this model can detect more fraud and significantly reduce customer complaints compared to several baseline models and system-based human tuned rules .