Recent years have witnessed the explosive growth of recommender systems in various exciting application domains such as electronic commerce, social networking, and location-based services. A great many algorithms have been proposed to improve the accuracy of recommendation, but until recently the long tail problem rising from inadequate recommendation of niche items is recognized as a real challenge to a recommender. This is particularly true for ultra-massive online retailers who usually have tremendous niche goods for sale. In light of this, in this paper, we propose a pattern-based method called CORE for flexible recommendation of both popular and niche items. CORE has two notable features compared with various existing recommenders. First, it is superior to previous pattern-based methods by adopting cosine rather than frequent patterns for recommendation. This helps filter out spurious cross-support patterns harmful to recommendation. Second, compared with some benchmark methods such as SVD and LDA, CORE does well in niche item recommendation given particularly heavy tailed data sets. Indeed, the coupled configuration of the support and cosine measures enables CORE to switch freely between recommending popular and niche items. Experimental results on two benchmark data sets demonstrate the effectiveness of CORE especially in long tail recommendation. To our best knowledge, CORE is among the earliest recommenders designed purposefully for flexible recommendation of both head and tail items.
Download Full PDF Version (Non-Commercial Use)