序列推荐系统：挑战、进展和展望

序列推荐系统

需积分: 46 159 浏览量 2020-04-25 15:47:00 上传评论收藏 517KB PDF 举报

资源推荐

资源详情

资源评论

Sequential Recommender Systems: Challenges, Progress and Prospects

∗

Shoujin Wang

, Liang Hu

, Yan Wang

, Longbing Cao

, Quan Z. Sheng

, Mehmet Orgun

Department of Computing, Macquarie University

Advanced Analytics Institute, University of Technology Sydney

{shoujin.wang, yan.wang}@mq.edu.au, rainmilk@gmail.com,

Abstract

The emerging topic of sequential recommender

systems (SRSs) has attracted increasing attention

in recent years. Different from the conventional

recommender systems (RSs) including

collaborative filtering and content-based filtering,

SRSs try to understand and model the sequential

user behaviors, the interactions between users

and items, and the evolution of users’ preferences

and item popularity over time. SRSs involve the

above aspects for more precise characterization of

user contexts, intent and goals, and item

consumption trend, leading to more accurate,

customized and dynamic recommendations. In

this paper, we provide a systematic review on SRSs.

We first present the characteristics of SRSs, and

then summarize and categorize the key challenges

in this research area, followed by the

corresponding research progress consisting of the

most recent and representative developments on

this topic. Finally, we discuss the important

research directions in this vibrant area.

1 Introduction

Sequential recommender systems (SRSs) suggest items

which may be of interest to a user by mainly modelling the

sequential dependencies over the user-item interactions

(e.g., view or purchase items on an online shopping platform)

in a sequence [27]. The traditional recommender systems

(RSs), including the content-based and collaborative filtering

RSs, model the user-item interactions in a static way and can

only capture the users’ general preferences. In contrast,

SRSs treat the user-item interactions as a dynamic sequence

and take the sequential dependencies into account to

capture the current and recent preference of a user for more

accurate recommendation [1]. In order to enhance the

understanding of SRSs, next we present the motivation and

formalization of SRSs.

A comprehensive survy on session-based recommender systems

can be found: https://arxiv.org/abs/1902.04864.

Figure 1: Two examples of SRSs: (1) After Jimmy has booked a flight,

a hotel and rented a car, what will be his next action? (2) After Tina

has bought an iPhone, an iWatch and a pair of AirPods, what would

she buy next?

Motivation: Why Sequential Recommender Systems?

The user-item interactions are essentially sequentially

dependent. In the real world, users’ shopping behaviours

usually happen successively in a sequence, rather than in an

isolated manner. Taking the shopping events of Jimmy

depicted in Figure 1 as an example, before Jimmy started

holiday, he booked a flight and a hotel and rented a car

successively, and his next action may be visiting a tourist

attraction via selfdriving. In such a case, the hotel may be

close to the destination airport of the flight and the location

for picking up the rented car may be not far away from the

hotel. In this scenario, each of Jimmy’s next actions depends

on the prior ones and thus all the four consumption actions

are sequentially dependent. Likewise, we can see the

sequential dependencies in Tina’s case. Such kind of

sequential dependencies commonly exist in transaction data

but cannot be well captured by the conventional content-

based RSs or collaborative filtering RSs [12], which

essentially motivates the development of SRSs.

Both the users’ preference and items’ popularity are

dynamic rather than static over time. In fact, a user’s

preference and taste may change over time. For instance,

many young people who used to be iPhone fans now have

switched to become fans of the phones manufactured by

Huawei or Samsung and the popularity of iPhone has been

dropping in recent years. Such dynamics are of great

significance for precisely profiling a user or an item for more

accurate recommendations and they can only be captured

by SRSs.

Tina

Jimmy

User-item interactions usually happen under a certain

sequential context. Different contexts usually lead to

different users’ interactions with items, which is, however,

often ignored by traditional RSs like collaborative filtering. In

contrast, an SRS takes the prior sequential interactions as a

context to predict which items would be interacted in the

near future. As a result, it is much easier to diversify the

recommendation results by avoiding repeatedly

recommending those items identical or similar to those

already chosen.

Formalization: What are Sequential Recommender

Systems?

Generally, an SRS takes a sequence of user-item interactions

as the input and tries to predict the subsequent user-item

interactions that may happen in the near future through

modelling the complex sequential dependencies embedded

in the sequence of user-item interactions. More specifically,

given a sequence of user-item interactions, a

recommendation list consisting of top ranked candidate

items are generated by maximizing a utility function value

(e.g., the likelihood):

R = arg max f(S) (1)

where f is a utility function to output a ranking score for the

candidate items, and it could be of various forms, like a

conditional probability [19], or an interaction score [11]. S =

,...,i

|S|

} is a sequence of user-item interactions where

each interaction i

=< u,a,v > is a triple consisting of a user u,

the user’s action a, and the corresponding item v. In some

cases, users and items are associated with some meta data

(e.g., the demographics or the features), while the actions

may have different types (e.g., click, add to the cart,

purchase) and happen under various contexts (e.g., the time,

location, weather). The output R is a list of items ordered by

the ranking score.

Different from the general sequence modelling in which

the sequence structure is much simpler since a sequence is

often composed of atomic elements (e.g., real values, genes),

the learning task in SRSs is much more challenging because

of the more complex sequence structure (e.g., each element

is a triple). This motivates us to systematically analyze the

challenges in SRSs and summarize the corresponding

progress. Contributions. The main contributions of this work

are summarized below:

• We systematically analyze a number of key challenges

caused by different data characteristics in SRSs and

categorize them from a data driven perspective, which

provides a new view to deeply understand the

characteristics of SRSs.

• We summarize the current research progress in SRSs by

systematically categorizing the state-of-the-art works

from a technical perspective.

• We share and discuss some prospects of SRSs for the

reference of the community.

2 Data Characteristics and Challenges

Due to the diversity and complexity of the customers’

shopping behaviours, item characteristics and the specific

shopping contexts in the real world, the generated user-item

interaction data often has different characteristics. Different

data characteristics essentially bring different challenges for

SRSs, which require different solutions, as presented in

Table 1. In the following five subsections, we specifically

discuss five key challenges respectively in SRSs caused by

different data characteristics. In each subsection, we first

introduce the particular data characteristics and then

illustrate the corresponding challenges.

2.1 Handling Long User-Item Interaction Sequences

A long user-item interaction sequence consists of a relatively

large number of user-item interactions. As a result, it has a

much higher chance to have more complex and

comprehensive dependencies over the multiple interactions

inside it, which makes the sequential recommendations

much more challenging. Specifically, two most critical

challenges in long user-item interaction sequences are

learning higher-order sequential dependencies and learning

long-term sequential dependencies, which will be presented

respectively below. Learning higher-order sequential

dependencies. Hig herorder sequential dependencies

commonly exist in the useritem interaction sequences,

especially in long ones. Compared to the lower-order

sequential dependencies, which are relatively simple and

can be easily modeled by Markov chain models [3] or

factorization machines [14; 10], higher-order sequential

dependencies are much more complex and harder to be

captured because of their complicated multi-level cascading

dependencies crossing multiple user-item interactions. So

far, there have been mainly two basic approaches reported

that can address this challenge in SRSs to some extent:

higher-order Markov-chain models [6] and recurrent neural

networks (RNN) [7], as shown in Table 1. However, each

approach has its own limitations, for example, the historical

states that can be involved in a higher-order Markov-chain

model are quite limited as the number of the model

parameters to be estimated grows exponentially with the

order, while the overly strong order assumption employed

in RNN limits the application of RNN in sequences with a

flexible order. The technical progress achieved in both

approaches will be presented in Sections 3.1 and 3.3

respectively in more details. Learning long-term sequential

dependencies. Long-term sequential dependencies refer to

the dependencies between interactions that are far from

each other in a sequence. For instance, given a shopping

sequence S

={a rose, eggs, bread, a bottle of milk, a vase},

which consists of a basket of items that are purchased

successively by a user Janet. Obviously, the vase and the rose

are highly dependent even though they are far from each

other. Such cases are not uncommon in the real world as

users’ behaviours are usually highly uncertain and thus they

may put any items into the cart. To address such a critical

剩余7页未读，继续阅读

评论收藏

内容反馈

syp_net

粉丝: 158
资源: 1187

序列推荐系统：挑战、进展和展望

推荐系统研究进展及展望

AAAI 2021上与【序列推荐（SR）】相关的论文（五篇）

对话推荐系统的进展和挑战：综述论文

2016年以来推荐系统前沿论文进展 2018.11.02 方建勇1

IJCAI-2021 会议报告PPT：推荐系统的进展

达芬奇机器人系统在妇科领域的应用进展及展望.pdf

深度学习在视频目标跟踪中的应用进展与展望.pdf

基于深度学习推荐系统的研究与展望.pdf

最佳离散信号之序列概述

风力和光伏发电短期功率预测研究进展与展望.pdf

地球大数据促进土地退化零增长目标实现：实践与展望.docx

5G移动通信系统关键技术与标准进展及6G展望.pdf

5G移动通信系统关键技术与标准进展及6G展望.zip

任务型对话系统最新的进展和挑战（2020年04月最新进展）

基于注意力机制的深度学习推荐研究进展.pdf

深度学习在视频目标跟踪中的应用进展与展望_深度学习在视频目标跟踪中的应用进展与展望_

时间序列聚类——十年回顾

智能视频监控系统的研究现状及展望.pdf

人工免疫系统进展与展望

2009自然科学进展 个性化推荐系统的研究进展

电子商务协同过滤推荐系统的研究与进展

智能推荐系统研究综述.pdf

基于数据挖掘的生物序列聚类研究

最新资源

2009自然科学进展个性化推荐系统的研究进展