Recommender systems (RS) are the core software, tools, and techniques that effectively and effciently cope with information overload and locate the information which is really needed. As one of the most widely used artificial intelligence (AI) systems, RS have been integrated with our daily life for a long time. In recent year, the machine learning approach has been dominated AI research in almost all areas. Therefore, modeling advanced RS with the machine learning approach is the basic methodology in this thesis.

Current RS suffer from many issues, e.g. cold start and black sheep, because they fail to consider the non-IIDness, including the heterogeneities and coupled relations within and between users, item, and their interactions. As a result, we propose the non-IID recommender systems by modeling the non-IIDness in recommendation  data with the machine learning approach. More specific, we study the non-IID RS modeling techniques from three perspectives: users, items, and their interactions. This research not only promotes the design of new machine learning models and algorithms in theory but also brings far-reaching influence on the evolution of technology and society.

To construct the non-IID RS from the user perspective, we jointly model two aspects: (1) the heterogeneities of users and (2) the coupling between users. In specific, we study the non-IID user modeling in two representative RS: (1) group-based recommender systems (GBRS) and (2) social network-based recommender systems (SNRS). First, we provide an in-depth analysis of the existing GBRS and shows their deficiencies of modeling the heterogeneity and coupling between group members for making group decisions. A deep neural network is designed to learn a group preference representation which jointly considers all members’ heterogeneous preferences. Second, we model an SNRS by modeling the influential contexts which embed the influence of relevant users and items since a user’s selection is largely influenced by other users with social relationships.

To construct the non-IID RS from the item perspective, we target two modeling aspects: (1) the heterogeneities of items and (2) the coupling between item. In specific, we study the non-IID item modeling in two representative RS: (1) cross-domain recommender systems (CDRS) and (2) session-based recommender systems (SBRS). First, the existing CDRS may fail to conduct cross-domain transfer because of the domain heterogeneity. As a result, we propose an irregular tensor factorization model which can better capture the coupling between heterogeneous domains with learning the domain factors for each domain. Second, we build an effective and efficient personalized SBRS to better capture the couplings between items by modeling intra- and inter-session contexts.

To construct the non-IID RS from the interaction perspective, we target two

modeling aspects: (1) the heterogeneities of interactions and (2) the coupling between interactions. In specific, we study the non-IID interaction modeling in two representative RS: (1) multi-objective recommender systems (MORS) and attraction-based recommender systems (ABRS). First, we study a MORS to tackle with the challenges of recommendation for users and items in the long tail. A coupled regularization model is proposed to jointly optimize two objectives, i.e. credibility and specialty. Second, the existing content-based RS can recommend new content according to similarity, but they are not capable of interpreting the attraction points in user-item interaction. To build interpretable content-based RS, we propose attraction modeling to learn and track user attractiveness.

In the last part, we summarize contributions of our work and present the future

directions that can improve and extend the non-IID RS.