Recommendation techniques play significant role in the design of online applications especially in e-commerce applications. Generally recommendation techniques use filtering methods. Filtering methods fall in the category of content filtering and collaborative filtering. Content filtering requires the matching of user’s profile with product features. Content filtering doesn’t take into account the similarity among users’ profiles. Another filtering method is called collaborative filtering. Recommender systems which use collaborative filtering also consider the similarity among other users’ profiles also for providing recommendations. Such kind of recommender systems compute the ratings about a product feature by considering ratings specified by users with similar profiles. Recommender systems often suffer from the problem of sparse data. If the products to be sold online have several features and all features must be rated by the users and if the product is promoted online and survey is presented to thousands of online users, it may happen that not all users participate in the survey. Even if all users participate in the survey they do not provide ratings on all features of the product. It results in several missing values in user-item matrix. This matrix is sparse in nature. If matrix with sparse data is presented as input to the recommender system, the recommender system may not work correctly on it. Therefore the missing values must be filled before the data is fed to the recommender system. In this paper we propose an approach to handle the problem of sparse data by using user profile similarities in a social net work. Each user’s profile is augmented with an additional attribute called trust. The value of trust represents the degree of trust of social network users on the given user. When a user completes a given survey for a product and he/she skips one or more ratings, then the trust value from his/her profile is retrieved to fill this value. Next this user’s friends’ list is retrieved and the rating specified by these users are also retrieved. On the basis of these rating values and the trust value, missing rating value is computed. Experimental results show that the rating values are computed with reasonably good accuracy