Abstract: The process data of hot-rolled strips from an iron and steel enterprise were analyzed to find out the inherent relationship between process parameters and production quality by using an improved random forests algorithm. After critical features being extracted， a defect prediction model was built. According to the experiment， balancing operation can improve the prediction accuracy of the imbalanced data sets. Meanwhile， the combination of CART and C4.5 can further improve the prediction accuracy than each single method. Furthermore， in consideration of the characteristics whose features have high or low correlations with the response variable， mutual information was introduced as an evaluation criterion for feature selection. Mutual information makes great contribution to classification effect of random forest algorithm， and recognition rate of defects of hot-rolled strips is obviously improved by using three strategies.

Key words: hot-rolled strip, defect prediction, data driven, feature selection, random forests