In this research, a revised recommendation system to generate recommended items for each user is constructed, such as "recommended for you" on the e-commerce websites. By using both the purchase and the browsing data, sparseness of matrix derived from the user's behavior history data is reduced. The main purpose is to construct a recommendation system that can recommend new items not browsed by users and appropriate items matching user preferences.As a procedure for generating recommended items, a user-item matrix and must-link constraints are first constructed from user's behavior history data. We add rows and columns to represent various item and user information to the user-item matrix. Next, semi-supervised learning is performed using the user-item matrix and the must-link constraint, and a new user-item matrix is generated. From this matrix on the basis of Pearson similarity, item similarity and user similarity are obtained. Finally, item-based collaborative filtering and user-based collaborative filtering are performed to generate recommended items.Experimental results show that the F-measure to represent the recommendation accuracy increases by generating recommended items with the proposed model using must-link constraints, user information and item information. In addition, it can be seen that the proposed model is more likely to purchase recommended items than the model of existing models.