Sequences, Items And Latent Links: Recommendation With Consumed Item Packs

In this Zetabyte Era, the abundance of information calls for personalization systems to ease the navigation of users. Among these systems, recommenders are becoming mainstream, and are used by major service providers such as Facebook, Amazon and Netflix. Some recommenders make use of the content of the items: these include popularity-based, knowledge-based or demographic-based schemes [8]. Others are content-agnostic: these are mainly collaborative filtering (CF) [14], [44] schemes, and are predominant today for they achieve good recommendation quality without requiring any prior knowledge of the content of the items recommended. Recommenders typically collect user preferences using explicit feedback [32], such as numerical ratings (star ratings in Imdb, Netflix, Amazon), binary preferences (likes/dislikes in Youtube), or unary preferences (retweets in Twitter). Yet, relying on explicit feedback raises issues regarding feedback sparsity (in systems where the item catalog is large, users tend to give feedback on a trace amount of those items, impacting the quality of recommendations [8]), and limited efficiency for recommending fresh items in reaction to recent user actions [37]. A few implicit recommenders have been proposed to answer those shortcomings.