To improve the soft sensor performance in chem. processes, a soft sensor modeling method based on improved ensemble learning that not only selected modeling variables with the highest influences on dominant variables by cosine similarity but also established the multi-target soft sensor model was proposed. The prepd. modeling variables were combined via the least squares support vector machine to predict the variable elements of the query sample. Because the variable selection mechanism of modeling used the dynamic information of chem. processes, this method was more suitable for data prediction than the other ones. Real data from a S recovery unit were used for model verification to evaluate the performance of the proposed soft sensor modeling method. The results confirmed its effectiveness.