Is it reasonable to use rules to label positive samples when doing fraud detection in machine learning? 
*

*We use supervised learning algorithm to detect fraud, but we have fewer or even no positive samples, is it reasonable if we use rules to label positive samples? if so, is the supervised learning model is meaningful? 

*If we make use of anomaly detection to catch fraud, we use rules to label data in model evaluation phase, will the model work?

*All above, if we can find rules to label data, does the model we build to detect fraud make any sense? because we've already used rules to find fraud.

 A: If you have no positive cases in your data and instead you will use some expert rules to mark the cases you believe to be frauds, then obviously your model will learn how to classify the cases that were marked by you as frauds. There is no guarantee that it will learn anything at all about the true frauds, since it will simply be learning your rules. If you have data with no positive cases, then you should rather use methods designed for anomaly detection (see anomaly-detection).
A: There is a very interesting work that was published in ICLR 2020:
Learning from Rules Generalizing Labeled Exemplars 
https://openreview.net/pdf?id=SkeuexBtDr
This paper assumes that you have a bunch of rules and some labeled examples supporting each rule. It also assumes that you have access to a large unlabeled data. Then it proceeds to learn a model that can learn from rule exemplars and the unlabeled data jointly to do better than the rule sets themselves.
Given your setup, I felt that this work might be of some use to you.
