Traditional recommender systems let users provide a single rating indicating their overall preferences toward items. Beside overall rating, multi-criteria recommender systems let users rate on multiple aspects of items with multi-criteria ratings. Most methods in recommender systems are based on collaborative filtering, which makes recommendations by exploiting ratings from neighbor users. However, most of them ignore the fact that rating behaviors of users vary due to their personal preference biases. Thus, exploiting ratings from neighbors directly might result in a poor recommendation. To solve this, the rating conversion techniques have been applied in some single criterion recommendations. For multi-criteria ratings, converting each criterion rating independently might result in loss of relation among criteria ratings. We propose a novel method that simultaneously converts all criteria ratings between users to maintain their implicit relations. The multi-criteria ratings are first normalized by variances of users and principle component analysis is applied to extract user preference patterns. Such patterns are then used for multi-criteria rating conversion. The experiment results show that our method outperforms both current single and multi-criteria rating conversion techniques with RMSEs of 1.1082 and 3.5574 on TripAdvisor and Yahoo datasets respectively, while maintaining considerably high level of prediction coverages.