Considering the limits that the Apriori algorithm produces numerous candidate itemsets during the self-joins of frequent items and scans database time after time, this paper proposed an improved algorithm. This algorithm maps the database to a boolean matrix, and then, deletes those meaningless items and records after the AND operation between matrix columns. This will greatly reduce the time and space complexities. Applying to the friend recommendation algorithm in social networks, this improved algorithm regards the interested users and information as records, takes the concerned users as deal items, builds a transaction database, computes frequent 2-item sets and recommends Top-N users ranked by supporting number as friends. The experiment proves the improved algorithm has higher precision and recall in friend recommendation algorithms of social networks.