Recommender Systems (RS) have emerged to help users make good decisions about which products to choose from the vast range of products available on the Internet. Many of the existing recommender systems are developed for simple and frequently purchased products using a collaborative filtering (CF) approach. This approach is not applicable for recommending infrequently purchased products, as no user ratings data or previous user purchase history is available. This paper proposes a new recommender system approach that uses knowledge extracted from user online reviews for recommending infrequently purchased products. Opinion mining and rough set association rule mining are applied to extract knowledge from user online reviews. The extracted knowledge is then used to expand a user's query to retrieve the products that most likely match the user's preferences. The result of the experiment shows that the proposed approach, the Query Expansion Matching-based Search (QEMS), improves the performance of the existing Standard Matching-based Search (SMS) by recommending more products that satisfy the user's needs.