Recommender System (RS) has been proposed to help users tackle information overload by suggesting potentially interesting items to users. A typical RS usually has a set of users and items with various rating preferences. The key task of RS is to predict an unknown rating or to recommend relevant items to a given user. Many existing recommendation methods such as Collaborative Filtering (CF), Content-based Recommendation, and Hybrid Filtering often assume that users, items and their attributes are identically and independently distributed. In the real world, however, these objects and their attributes are often coupled with each other through explicit or implicit relations. On one hand, users are often connected through social or trust relations, and items interact with linkage or citation relations. On the other hand, the attributes of users or items are also more or less coupled with each other. These dependent relations clearly demonstrate that the users, items, and their attributes in RS are not identically and independently distributed (non-IID), which is rarely considered in most existing recommendation methods. The non-IID RS has emerged with the consideration of non-IID characteristics into RS. The main challenge in non-IID RS is to analyse and model the coupling relations between users and between items.

In this dissertation, we aim to improve recommendation effectiveness by incorporating the coupling relations into RS. The main contributions of the dissertation are summarized as follows:
(1) We propose three novel neighbourhood-based CF methods including coupled user-based CF, coupled item-based CF, and coupled CF. Specifically, we first apply a novel coupled object similarity to compute the coupling relations between users and between items based on their attributes. We then integrate the user and item couplings into the neighbourhood-based CF to produce the proposed methods by inventing new similarity measures.
(2) We propose three novel model-based CF methods including coupled user-based matrix factorization (CUMF), coupled item-based matrix factorization (CIMF), and coupled matrix factorization (CMF). CUMF and CIMF respectively integrate the attribute-based user couplings and item couplings into MF, and CMF incorporates the user couplings, item couplings, and the user-item rating matrix together into MF.
(3) We propose a two-level matrix factorization recommendation model that integrates the textual semantic couplings between items and the user-item rating matrix together.
(4) We conduct experiments to evaluate the effectiveness of incorporating the couplings into non-IID RS.