As we know, the quality of processes is technically depicted by variation, a product or process with the best quality must naturally require the variation as less as possible. The variation is usually reduced with many ways, say, by adjusting parameters settings under robust design with many turns expensive experiments. So ones are trying to reach the robustness by detecting cheap and simple methods. In this paper, a both practical and simple technique for quality improvement, namely reducing the variation, by data classification is studied. First, all possible system factors are included, which may dominate the variation law. And then we make use of the past observations and their classification as well as boxplot charts to find out the internal rule between the variation and the system factor. Next, adjust the location of the system factor according to the rule so that the variation could, to some extent, be lessened. Finally, two typical quality improvement cases based on data classification are presented.