Induction of classification rules based on rough sets and neural networks has been an active research area in the field of machine learning due to their strength of handing imprecise and nonlinear problems. However, from previous literature, rough set is only a pre-processing tool to eliminate redundant data and neural networks act as a classifier to output classification. Since neural networks operate in black box fashion and lack explanation facilities for resulted knowledge, it is often difficult to extract rules from a trained neural network. In this paper, from a new prospective, a novel hybrid approach based on rough sets and neural networks has been proposed. Our hybrid approach consists of three phrases: Firstly using rough sets to reduce redundant attributes from a decision table, and then a neural network is trained to delete noisy attributes and records in the table. Finally, classification rules are generated from the reduced decision table by variable precision rough sets model. The new approach has been applied to two artificial datasets and two real-world datasets. The empirical results show that the proposed approach is more effective than some available hybrid approaches in generating classification knowledge. Induction of classification rules based on rough sets and neural networks has been an active research area in the field of machine learning due to their strength of handing imprecise and nonlinear problems. However, from previous literature, rough set is only a pre-processing tool to eliminate redundant data and neural networks act as a classifier to output classification. Since neural networks operate in black box fashion and lack of facilities forulting knowledge, it is often difficult to extract rules from a trained neural network. In this paper, from a new prospective, a novel hybrid approach based on rough sets and neural networks has been proposed. Our hybrid approach consists of three phrases: records in the table. Finally, classification rules are generated from the reduced decision table by variable precision rough sets The new results have been applied to two artificial datasets and two real-world datasets. The empirical results show that the proposed approach is more effective than some available hybrid approaches in generating classification knowledge.