In the last 20 years, recommendation system has been becoming more and more widely used in many web and mobile applications. It was started when Amazon popularized a recommendation technique called item-based collaborative filtering. This technique is fast, stable, and it performs well in most media sharing contexts. However, we found that there are some serious differences in recommending “stickers”, compared to traditional media items, like movies, songs, and so on. We have tried several approaches to improve recommendations in this context by comparing different similarity measurement methods, comparing personalized and non-personalized recommendations, and altering the time range used for generating the recommendation lists. We found that, in the situation where preference measurement does not have upper-bound, adjusted cosine similarity and cosine similarity methods perform better than Pearson correlation method. Meanwhile, in the situation where items have short-lived popularity period, straightforward personalized recommendations give bad accuracy. Finally, the personalized recommendations show performance improvement when generated using shorter time range.