Fashion is a key defining factor of popular culture, and it changes over time. Each season tons of new products emerge to the market. People who follow fashion wish to discover new and trendy products and quickly catch the most fashionable styles. Traditionally, product trends can be found in fashion magazines and product catalogs, but now the proliferation of the Internet and social networks may have made trend e-discovery possible. This paper explores a novel problem of finding product trends through the posts on Pinterest, a rising social media for sharing interests using uploaded photographs and text comments. A weighted feature subset selection (WFSS) framework is applied to simultaneously group popular products into different types and select the most representative and discriminative terms to describe each product type. We compare WFSS with co-clustering algorithms, non-negative matrix factorization, and unsupervised feature selection methods. Experimental results on a data set collected from Pinterest show the effectiveness of WFSS in both product clustering and keyword selection.