Recommending in the Long Tail of 0−1 Data using Highly Correlated Pairs of Group of Items
Trần Nguyễn Minh Thư, Hồ Tường Vinh, François SEMPE, Jean‐Daniel Zucker · 2010
Among difficulties encountered by modern shopping recommenders one of the most important is the long tail shape of sold items which is also related to cold-start issues. Various approaches including content-based recommendations attempt to overcome this problem which has a serious impact on the accuracy of recommendations especially when new products are continuously added to the catalogue. A hybrid approach that combines recommendations based on highly correlated pairs of concrete items (existing items) with abstract items (based on the item taxonomy) is proposed in this paper (AbstractConcrete RecommendationACReco). The ACReco algorithm is evaluated on an in-house and a benchmark of 0-1 Data (the Ta Feng data). Given-n protocols experimental results show significant improvements in both the recommendation accuracy and the recommendation of products in the long tail. The overall efficiency of the approach relies on the available Product Taxonomy.