Keywords

1 Motivation and Overview

Similar to a typical recommendation task, the focus of sequential recommendation is to model user’s preferences on items. The difference is that items are listed in a ordered sequence, for example in e-commerce, sequence is commonly called as a session [5]. Specifically, sequential recommendation focuses more on sequential information such as behavioral patterns and sequential dependencies among items [3]. It takes the user’s past behaviors as input, detects behavior patterns, models sequential dependencies, and uses them to develop recommendations that match the preference of individual users. Thus, traditional sequential recommender systems are particularly interested in employing appropriate and effective machine learning approaches to model sequential information, such as Markov Chain [2] and session-based KNN [1].

In recent years, deep learning (DL) techniques obtain tremendous achievements in nature language processing [4], demonstrating their effectiveness in processing sequence data. Thus, they have attracted an increasing interest in sequential recommender systems, and many DL-based models achieved state-of-the-art performance [6].

In this tutorial, we focus on the DL-based recommender systems. The main highlights of this tutorial are as follows:

  • We will give a comprehensive overview of sequential recommender systems based on DL techniques.

  • We will show an original classification framework for sequential recommendation, corresponding to three different recommendation scenarios, and then introduce the representative algorithms under each scenario.

  • We will summarize the influential factors for typical DL-based sequential recommendation and demonstrate their effects by empirical studies, which can serve as a guidance for sequential recommendation research and practices.

2 Target Audience and Learning Objectives

This tutorial will cover topics related to DL and sequential recommendation. The target audience of this tutorial are researchers and practitioners who are interested in sequential recommendation or deep learning. Through this tutorial, the participants will get a comprehensive overview of deep learning based sequential recommender systems and understand how different DL techniques model deal with different types of behavioral sequences, what are the main influential factors of DL-based sequential recommenders and how to take advantage of these factors to improve model performances.

3 Outline of the Tutorial

3.1 Introduction

  1. 1.

    How to define a sequential recommendation problem? How do they distinguish from other recommendation tasks such as rating prediction or top-k item recommendation models?

  2. 2.

    What are the major types of sequential recommender systems in terms of the different input data (e.g., behavior trajectories) and what are the prediction goals for each type?

  3. 3.

    What are the typical DL techniques applied to sequential recommendation? What are their strengths?

  4. 4.

    What are the advantages of DL-based algorithms in contrast with traditional sequential recommendation algorithms? What are the suitable scenarios for both the traditional recommendation approaches and DL techniques?

3.2 Categorization of the Sequential Recommendation Systems

We design a categorization of the sequential recommendation algorithms in the literature from two perspectives: action objects (i.e., different sequential information) and user representation (i.e., whether to distinguish different users).

  1. 1.

    Categorization based on action objects. What forms of action object exist in the sequence data? How does the model handle different types of action objects? We analyze the next-item recommendation and the next-basket recommendation.

  2. 2.

    Categorization based on user representation. Does the model include user modeling modules? How does the model represent the relationship between users and sequences? Does it model user representation based on behavior sequence or other ancillary information such as user identification? We analyze the implicit and explicit user representations.

3.3 DL-Based Algorithms

In terms of the aforementioned categorization, we specifically present the definition and representative DL algorithms for each recommendation scenario:

  1. 1.

    Experience-based sequential recommendation. The input is user’s behavior trajectory with the same item. The goal is to predict next item under a target action type.

  2. 2.

    Transaction-based sequential recommendation. The input is user’s behavior trajectory with same action type but different action objects. The goal is to predict next action object.

  3. 3.

    Interaction-based sequential recommendation. The input is user’s behavior trajectory with multiple action types and different action objects. The goal is to predict both the action type and the action object.

3.4 Empirical Evaluations

In this section, we summarize the influential factors for typical DL-based sequential recommendation, and show how we validate them through experiments.

  1. 1.

    Influential Factors. What are the major factors that possibly influence the performance of DL-based recommender systems? How and to what extent do they improve the performance?

  2. 2.

    Empirical Experiment. Through a specially designed experimental framework, we strive to answer: how does the performance change with the change of these factors? How should we utilize each factor to improve DL models?

3.5 Challenges and Future Directions

In this section, we conclude our tutorial by summarizing challenges and research directions based on our categorization and evaluation.

4 Similar Tutorials

Sequence-aware Recommender Systems (https://www.um.org/umap2018/tutorial/index.html, UMAP 2018, Singapore) - This tutorial focuses on sequence-aware recommender systems [3], which contains a large number of non-deep-learning methods. The main difference is that our tutorial will specifically focus on DL methods, and demonstrate their advantages and disadvantages.

5 Presenters

Dr. Hui Fang is an Assistant Professor in Research Institute for Interdisciplinary Sciences. She received her Ph.D. from the School of Computer Engineering, Nanyang Technological University, Singapore. Her main research interests include trust prediction in online communities, personalized product/social recommendation, link prediction and data-driven decision making. Her research has been published in top conferences like IJCAI, AAA and AAMAS. Presently she is the Associate Editor of the ECRA journal, and invited referees of other journals like TKDE, Cybernetics, and DSS. She also serves as a Program Committee Member for international conferences, including UMAP and IJCAI, etc.

Dr. Guibing Guo is an Associate Professor in the Software College, Northeastern University, Shenyang, China. He received a Ph.D degree in computer science at Nanyang Technological University, and did a post-doc at Singapore Management University. His research interests include recommender systems, deep learning, social network analysis and data mining. The main research focus is to resolve the challenges of recommender systems, including data sparsity, cold start, diversity and so on. He is the original author (and now the team leader) of the open-source library for recommender systems, LibRec, which implements a large number (over 90) of recommender algorithms.