Recommending Structured Objects: Paths and Sets

Abstract

Recommender systems have been widely adopted in industry to help people find the most appropriate items to purchase or consume from the increasingly large collection of available resources (e.g., books, songs and movies). Conventional recommendation techniques follow the approach of ``ranking all possible options and pick the top'', which can work effectively for single item recommendation but fall short when the item in question has internal structures. For example, a travel trajectory with a sequence of points-of-interest or a music playlist with a set of songs. Such structured objects pose critical challenges to recommender systems due to the intractability of ranking all possible candidates. This thesis study the problem of recommending structured objects, in particular, the recommendation of path (a sequence of unique elements) and set (a collection of distinct elements). We study the problem of recommending travel trajectories in a city, which is a typical instance of path recommendation. We propose methods that combine learning to rank and route planning techniques for efficient trajectory recommendation. Another contribution of this thesis is to develop the structured recommendation approach for path recommendation by substantially modifying the loss function, the learning and inference procedures of structured support vector machines. A novel application of path decoding techniques helps us achieve efficient learning and recommendation. Additionally, we investigate the problem of recommending a set of songs to form a playlist as an example of the set recommendation problem. We propose to jointly learn user representations by employing the multi-task learning paradigm, and a key result of equivalence between bipartite ranking and binary classification enables efficient learning of our set recommendation method. Extensive evaluations on real world datasets demonstrate the effectiveness of our proposed approaches for path and set recommendation.