Along with the development of online social networks, friend recommendation becomes the favor of the major social networks. It can help people to meet new friends for expanding the scale of social network, which in turn allows people to receive more information from their friends. Therefore, friend recommendation should be focused on expanding the scale of social network and obtaining information form recommended friends. However, existing friend recommendation methods barely consider the people information need, and they are mainly based on the simple and limited user profiles, and are agnostic to users' offline behaviors in the real world. Because human activity in the physical world has a spatial locality, the recommended friends through the existing recommendation methods are limited in geographic space which the target user know. As a result, the recommendation cannot provide more new information on geography to meet the target's need on information. This paper first proposes a new friend recommendation method based on the offline check-in behaviors in the real world instead of the online user profiles, and mines check-in behavior similarity between users in the real world. The essential goal of friend recommendation is to provide users with more new information. In order to meet the requirements of user getting more geographical location information, the recommendation systems can recommend the strangers in broader check-in geography distribution for the target users. Meanwhile, when the recommended friends and the target users have the similar check-in behaviors, it is more probable for the users to accept recommended strangers. Kernel density estimation (KDE) is used to estimate each user's check-in behavior probability distribution and the time entropy to filtering some noise that have side effects on overall check-in behavior similarity, then the recommended strangers who can bring a wider range of new strangers geographic information for the target users are selected. Lastly, a large-scale user check-in data-set of Foursquare is used to validate recommendation precision and the degree of information expanding of this approach. The experimental results show that the proposed approach outperforms the existing friend recommendation methods on the aspect of the information expanding degree while maintaining the recommendation precision of the state-of-the-art stranger recommendation methods.