Classification using association is a recent data mining approach that integrates association rule discovery and classification. A modified version of the Multi-class Classification based on Association Rule (MCAR) is proposed in this paper. The proposed classifier, known as Modified Multi-class Classification based on Association Rule, MMCAR, employs a new rule production function which resulted only relevant rules are used for prediction. Experiments on UCI data sets using different classification learning algorithms (C4.5, RIPPER, MCAR) is performed in order to evaluate the effectiveness of MMCAR. Results show that the MMCAR produced higher accuracy compared to C4.5 and RIPPER. In addition, the average number of rules generated by MMCAR is less than the one produced by MCAR.