1 Introduction

Recommender systems are indispensable tools to help tackle with the information overload problem [2]. However, with merely relying on user-item interaction data, traditional recommender systems inherently suffer from the data sparsity (i.e., most users only rate a small portion of items) and cold start (i.e., new users merely rate few items) issues [12]. To address such issues, a number of recommendation algorithms have been designed by leveraging the valuable side information of users, items and their interactions to compensate for the insufficiency of rating information [4, 5, 7, 13, 17].

This tutorial provides a comprehensive analysis of state-of-the-art recommendation approaches with side information in a principle way from two perspectives: representation and methodology. By the end of this tutorial, the audiences would know how the recommendation approaches evolve with more complicated representations and methodologies for using various kinds of side information.

Various Representations. This section introduces two common ways of organizing side information (or features) in both flat and hierarchical representations.

  • Flat features (FF) are mainly explored by early studies, where the features are organized independently in the same layer [7, 10, 16]. Such kind of side information can be utilized for better modeling the characteristics of users or items. Many content-based recommender systems have been developed by extensively using FF.

  • Feature hierarchy (FH) is a natural yet powerful structure to human knowledge, providing a machine- and human-readable description of features and the affiliatedTo relations among them. FH has been proven to be more effective to achieve high-quality recommendation than FF [5, 6, 13].

  • Knowledge graph (KG) connects various types of features in a unified global representation space. KG helps with the inference of subtler user or item relations from different angles, which are difficult to uncover with the homogeneous information [11, 17, 18].

Various Methodologies. Early approaches with side information are mainly memory-based, which are ineffective due to the time-consuming search in the user or item space [8, 10]. Thus, many prevalent recommendation algorithms with side information are devised based on more advanced models, including association rule, clustering, topic model, regression, factorization, representation learning and deep neural network. Factorization, representation learning and deep learning based models have been widely investigated in recent research:

  • Factorization model (FM) can incorporate different kinds of side information [12], such as matrix factorization (MF), tensor factorization (TF) and SVD++, etc. They model users’ behavior patterns by employing global statistical information of user-item interaction data. The basic idea is that both users and items can be characterized by a few latent vectors.

  • Representation learning (RL) based recommendation methods have proven to be effective in capturing local item relations by modeling item co-occurrence in individual user’s interaction record [1]. Some researchers attempt to integrate side information into RL models that help learn better user and item embeddings, thus achieving further performance enhancements for recommendation [3, 16].

  • Deep neural network (DNN) has recently attracted major research interests from the recommendation community. In contrast to factorization and RL based methods, DNN based recommendation models can learn nonlinear latent representations. Both structural (e.g., social networks, knowledge graph) and non-structural side information (visual, textual content) can be exploited by DNN models [9, 15].

2 Tutorial Outlines

As this tutorial covers a variety of research topics about employing side information for recommendation, it should be organized in a structural way as below:

  • Introduction (10 min).

    • A brief introduction to the history of recommender systems.

    • What are the traditional recommendation algorithms?

    • What are the limitations of traditional recommendation algorithms?

    • What is side information and why does it can help create more effective recommender system?

  • Various representations (60 min).

    • How are flat features utilized?

    • How are feature hierarchy utilized?

    • An introduction to our works [4, 14].

    • An introduction to other representation methods for side information.

  • Break (30 min).

  • Various methodologies (75 min).

    • An introduction to memory-based models and their limitations.

    • How does factorization model use side information?

    • How does representation learning use side information?

    • How does deep neural network use side information.

    • An introduction to our works [13, 15].

  • Conclusion and future directions (5 min).

3 Biographies and Expertise of Tutors

Qing Guo is a part-time Ph.D. student in Wee Kim Wee School of Communication and Information at Nanyang Technological University. He focuses on Point-of-Interest (POI) recommendation by exploiting the heterogeneous information in location-based social networks. He obtained his M.Sc. in The University of Hong Kong in 2014 and B.E. from University of Electronic Science and Technology of China in 2013. While doing Ph.D. study, he was also a research associate in SAP Innovation Center network from 2015 to 2018, where he participated in machine learning products development in SAP products. Now, he is a data scientist in Shopee and continue to work on recommendation research and applications.

Zhu Sun is a data scientist in Shopee, Singapore. She obtained her Ph.D. degree from Nanyang Technological Univeristy in 2018. She received M.E. in 2016 and B.E. in 2013 from Yanshan University, China. Her research is highly related to artificial intelligence. Specifically, she mainly focus on applying data mining and machine learning techniques to design effective recommender systems. She is interested in leveraging side information to address data sparsity and cold start problems of recommender systems.

Yin-Leng Theng is Professor and Director of the Centre of Healthy and Sustainable Cities (CHESS) at Wee Kim Wee School of Communication and Information, and Research Director at the Research Strategy and Coordination Unit (President’s Office) at Nanyang Technological University. Her research interests are mainly in user-centred design, interaction design and usability engineering. She has participated in varying capacities as principal investigator, co-investigator and collaborator in numerous research projects in the United Kingdom and Singapore since 1998.