Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

Sequential Recommender Systems: Challenges, Progress and Prospects ∗

Shoujin Wang1 , Liang Hu2,3 , Yan Wang1 , Longbing Cao2 , Quan Z. Sheng1 and Mehmet Orgun1
1
Department of Computing, Macquarie University
2
Advanced Analytics Institute, University of Technology Sydney
3
University of Shanghai for Science and Technology
{shoujin.wang, yan.wang}@mq.edu.au, [email protected],
[email protected],{michael.sheng, mehmet.orgun}@mq.edu.au

Abstract
The emerging topic of sequential recommender
Jimmy
systems (SRSs) has attracted increasing attention
in recent years. Different from the conventional
recommender systems (RSs) including collabora-
tive filtering and content-based filtering, SRSs try
to understand and model the sequential user behav- Tina
iors, the interactions between users and items, and Figure 1: Two examples of SRSs: (1) After Jimmy has booked a
the evolution of users’ preferences and item popu- flight, a hotel and rented a car, what will be his next action? (2)
larity over time. SRSs involve the above aspects for After Tina has bought an iPhone, an iWatch and a pair of AirPods,
more precise characterization of user contexts, in- what would she buy next?
tent and goals, and item consumption trend, leading
to more accurate, customized and dynamic recom- happen successively in a sequence, rather than in an isolated
mendations. In this paper, we provide a systematic manner. Taking the shopping events of Jimmy depicted in
review on SRSs. We first present the characteris- Figure 1 as an example, before Jimmy started holiday, he
tics of SRSs, and then summarize and categorize booked a flight and a hotel and rented a car successively, and
the key challenges in this research area, followed his next action may be visiting a tourist attraction via self-
by the corresponding research progress consisting driving. In such a case, the hotel may be close to the desti-
of the most recent and representative developments nation airport of the flight and the location for picking up the
on this topic. Finally, we discuss the important re- rented car may be not far away from the hotel. In this sce-
search directions in this vibrant area. nario, each of Jimmy’s next actions depends on the prior ones
and thus all the four consumption actions are sequentially de-
pendent. Likewise, we can see the sequential dependencies
1 Introduction in Tina’s case. Such kind of sequential dependencies com-
Sequential recommender systems (SRSs) suggest items monly exist in transaction data but cannot be well captured
which may be of interest to a user by mainly modelling the by the conventional content-based RSs or collaborative filter-
sequential dependencies over the user-item interactions (e.g., ing RSs [Kang et al., 2018], which essentially motivates the
view or purchase items on an online shopping platform) in development of SRSs.
a sequence. The traditional recommender systems (RSs), Both the users’ preference and items’ popularity are dy-
including the content-based and collaborative filtering RSs, namic rather than static over time. In fact, a user’s prefer-
model the user-item interactions in a static way and can only ence and taste may change over time. For instance, many
capture the users’ general preferences. In contrast, SRSs treat young people who used to be iPhone fans now have switched
the user-item interactions as a dynamic sequence and take the to become fans of the phones manufactured by Huawei or
sequential dependencies into account to capture the current Samsung and the popularity of iPhone has been dropping in
and recent preference of a user for more accurate recommen- recent years. Such dynamics are of great significance for pre-
dation [Chen et al., 2018]. In order to enhance the under- cisely profiling a user or an item for more accurate recom-
standing of SRSs, next we present the motivation and formal- mendations and they can only be captured by SRSs.
ization of SRSs. User-item interactions usually happen under a certain se-
quential context. Different contexts usually lead to different
Motivation: Why Sequential Recommender Systems? users’ interactions with items, which is, however, often ig-
The user-item interactions are essentially sequentially depen- nored by traditional RSs like collaborative filtering. In con-
dent. In the real world, users’ shopping behaviours usually trast, an SRS takes the prior sequential interactions as a con-
∗ text to predict which items would be interacted in the near
A comprehensive survy on session-based recommender systems
can be found: https://arxiv.org/abs/1902.04864. future. As a result, it is much easier to diversify the rec-

6332
 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

ommendation results by avoiding repeatedly recommending 2.1 Handling Long User-Item Interaction
those items identical or similar to those already chosen. Sequences
Formalization: What are Sequential Recommender A long user-item interaction sequence consists of a relatively
Systems? large number of user-item interactions. As a result, it has
Generally, an SRS takes a sequence of user-item interactions a much higher chance to have more complex and compre-
as the input and tries to predict the subsequent user-item in- hensive dependencies over the multiple interactions inside
teractions that may happen in the near future through mod- it, which makes the sequential recommendations much more
elling the complex sequential dependencies embedded in the challenging. Specifically, two most critical challenges in long
sequence of user-item interactions. More specifically, given user-item interaction sequences are learning higher-order se-
a sequence of user-item interactions, a recommendation list quential dependencies and learning long-term sequential de-
consisting of top ranked candidate items are generated by pendencies, which will be presented respectively below.
maximizing a utility function value (e.g., the likelihood):
Learning higher-order sequential dependencies. Hig her-
R = arg max f (S) (1) order sequential dependencies commonly exist in the user-
where f is a utility function to output a ranking score for the item interaction sequences, especially in long ones. Com-
candidate items, and it could be of various forms, like a condi- pared to the lower-order sequential dependencies, which are
tional probability [Wang et al., 2018], or an interaction score relatively simple and can be easily modeled by Markov chain
[Huang et al., 2018]. S = {i1 , i2 , ..., i|S| } is a sequence of models [Garcin et al., 2013] or factorization machines [Ren-
user-item interactions where each interaction ij =< u, a, v > dle et al., 2010; Hidasi and Tikk, 2016], higher-order sequen-
is a triple consisting of a user u, the user’s action a, and the tial dependencies are much more complex and harder to be
corresponding item v. In some cases, users and items are as- captured because of their complicated multi-level cascading
sociated with some meta data (e.g., the demographics or the dependencies crossing multiple user-item interactions. So far,
features), while the actions may have different types (e.g., there have been mainly two basic approaches reported that
click, add to the cart, purchase) and happen under various can address this challenge in SRSs to some extent: higher-
contexts (e.g., the time, location, weather). The output R is a order Markov-chain models [He and McAuley, 2016] and
list of items ordered by the ranking score. recurrent neural networks (RNN) [Hidasi et al., 2016a], as
Different from the general sequence modelling in which shown in Table 1. However, each approach has its own limi-
the sequence structure is much simpler since a sequence is tations, for example, the historical states that can be involved
often composed of atomic elements (e.g., real values, genes), in a higher-order Markov-chain model are quite limited as
the learning task in SRSs is much more challenging because the number of the model parameters to be estimated grows
of the more complex sequence structure (e.g., each element is exponentially with the order, while the overly strong order
a triple). This motivates us to systematically analyze the chal- assumption employed in RNN limits the application of RNN
lenges in SRSs and summarize the corresponding progress. in sequences with a flexible order. The technical progress
Contributions. The main contributions of this work are sum- achieved in both approaches will be presented in Sections 3.1
marized below: and 3.3 respectively in more details.
• We systematically analyze a number of key challenges Learning long-term sequential dependencies. Long-term
caused by different data characteristics in SRSs and cat- sequential dependencies refer to the dependencies between
egorize them from a data driven perspective, which pro- interactions that are far from each other in a sequence. For
vides a new view to deeply understand the characteris- instance, given a shopping sequence S1 ={a rose, eggs, bread,
tics of SRSs. a bottle of milk, a vase}, which consists of a basket of items
• We summarize the current research progress in SRSs that are purchased successively by a user Janet. Obviously,
by systematically categorizing the state-of-the-art works the vase and the rose are highly dependent even though they
from a technical perspective. are far from each other. Such cases are not uncommon in the
real world as users’ behaviours are usually highly uncertain
• We share and discuss some prospects of SRSs for the and thus they may put any items into the cart. To address
reference of the community. such a critical issue, Long Short Term Memory (LSTM)-
based [Wu et al., 2017] and Gated Recurrent Unit (GRU)-
2 Data Characteristics and Challenges based [Hidasi et al., 2016a] RNN have been applied in SRSs
Due to the diversity and complexity of the customers’ shop- to capture the long-term dependencies among the user-item
ping behaviours, item characteristics and the specific shop- interactions in a sequence. However, it is easy for RNN mod-
ping contexts in the real world, the generated user-item in- els to generate false dependencies by overly assuming any
teraction data often has different characteristics. Different adjacent items in a sequence are highly dependent. In the
data characteristics essentially bring different challenges for above example of Janet’s shopping sequence, an RNN usu-
SRSs, which require different solutions, as presented in Table ally models S1 by assuming the milk and vase are dependent
1. In the following five subsections, we specifically discuss due to the close distance between them, but actually they are
five key challenges respectively in SRSs caused by different not. Some other efforts have been made to solve this issue by
data characteristics. In each subsection, we first introduce utilizing the advantage of mixture models to combine multi-
the particular data characteristics and then illustrate the cor- ple sub-models with different temporal ranges to capture both
responding challenges. short- and long-term dependencies in a unified model [Tang et

6333
 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

Data characteristics Challenges Existing solutions

Long user-item interac- Learning higher-order se- Higher-order Markov chain [He and McAuley, 2016], RNN
tion sequences quential dependencies [Hidasi et al., 2016a]
Learning long-term sequen- LSTM- [Wu et al., 2017] and GRU-based [Hidasi et al.,
tial dependencies 2016a] RNN, mixture models [Tang et al., 2019]
User-item interaction Learning collective sequen- CNN [Tang and Wang, 2018; Yuan et al., 2019]
sequences with a tial dependencies under the
flexible order assumption of flexible order
User-item interaction Learning sequential depen- Attention models [Wang et al., 2018], memory networks
sequences with noise dencies attentively and dis- [Chen et al., 2018]
craminatively
User-item interaction Learning heterogeneous re- Mixture models [Kang et al., 2018; Tang et al., 2019; Wang
sequences with hetero- lations discriminatively and et al., 2019]
geneous relations integrating them effectively
User-item interac- Learning hierarchical de- Feature-enriched RNN [Hidasi et al., 2016b], hierarchical
tion sequences with pendencies embedding [Wang et al., 2015], hierarchical RNN [Quad-
hierarchical structures rana et al., 2017], hierarchical attention [Ying et al., 2018]

Table 1: A summary of challenges driven by data characteristics in SRSs

al., 2019]. Overall, the works that are able to tackle this chal- that they may contain some noisy and irrelevant interactions
lenge are quite limited and more investigations are required that generate interference for the next interaction prediction.
to bridge this gap. The technical progress achieved in RNN In practice, in a user-item interaction sequence, some histor-
and mixture models will be presented in Section 3.3. ical interactions are strongly relevant to the next interaction,
while others may be weakly relevant or even irrelevant. For
2.2 Handling User-Item Interaction Sequences example, in another shopping sequence S3 = {bacon, a rose,
with a Flexible Order eggs, bread}, the item “rose” may be a noisy item as it is
In the real world, some user-item interaction sequences are quite different from others and has no correlation to them.
strictly ordered while others may not be, namely, not all adja- The next item may be a bottle of milk with a high probability
cent interactions are sequentially dependent in a sequence. and it only sequentially depends on bacon, eggs and bread
For instance, in a shopping sequence S2 = {milk, butter, while has nothing to to with the rose. Therefore, another
flour}, it does not matter whether to buy milk or butter first, key challenge in SRSs is to learn sequential dependencies
but the purchase of both items leads to a higher probability attentively and discriminatively over user-item interaction se-
of buying flour next; namely, there is no strict order between quences with noise.
milk and butter, but flour sequentially depends on the union Quite a few works have attempted to solve such a typical is-
of them. Therefore, for a sequence with a flexible order, it is sue by employing the attention models [Wang et al., 2018] or
much better to capture the collective sequential dependencies, memory networks [Chen et al., 2018] to selectively retain and
rather than the point-wise ones as the former is fuzzy and does utilize information from those interactions that are truly rele-
not assume a strict order over user-item interactions. As a re- vant to the next interaction prediction. The technical progress
sult, how to capture collective sequential dependencies under achieved in these solutions will be presented in Section 3.3.
an assumption of flexible order becomes the key challenge in
handling sequences with a flexible order in SRSs. 2.4 Handling User-Item Interaction Sequences
Although common and important, reported studies in SRSs with Heterogeneous Relations
have not paid much attention to this issue yet. Exist- Heterogeneous relations refer to different types of relations
ing SRSs built on Markov-chains, factorization machines or which deliver different kinds of information and should be
RNN can only handle the point-wise dependencies but are modelled differently in SRSs. For instance, in a user-item
not good at modelling and capturing collective dependen- interaction sequence, except for the widespread occurrence-
cies. Only quite few works like [Tang and Wang, 2018; based sequential dependencies over user-item interactions,
Yuan et al., 2019] have attempted to address such a challenge there are also similarity-based relations between the inter-
by employing the strength of convulotional neural networks acted items in terms of their features. Furthermore, even
(CNN) to model the local and global dependencies between though both are sequential dependencies, long-term sequen-
different areas in an “image”, i.e., the embedding matrix of a tial dependencies are quite different from short-term ones
sequence of interactions. The technical progress achieved in and they cannot be modelled in the same way. Therefore,
CNN-based SRSs will be presented in Section 3.3. another key challenge in SRSs is how to effectively capture
these heterogeneous relations embedded in the user-item in-
2.3 Handling User-Item Interaction Sequences teraction sequences respectively and to make them work col-
with Noise laboratively for the sequential recommendations when han-
Due to the uncertainty of user shopping behaviours, most of dling user-item interaction sequences associated with hetero-
the user-item interaction sequences are not clean, meaning geneous relations.

6334
 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

There are quite limited works reported in the literature to first categorized into 11 atomic classes ( e.g., the sequential
solve this challenge in SRSs. Mixture models [Kang et al., pattern mining, factorization machine, and recurrent neural
2018; Tang et al., 2019; Wang et al., 2019] are the only so- networks) from the technical perspective. All these atomic
lution to address such challenge so far. A mixture model classes are then further categorized into three taxonomies,
integrates different types of relations modelled by different including traditional sequence models, latent representation
sub-models to collaboratively generate sequential recommen- models, and deep neural network models. Generally speak-
dations. The specific technical progress will be presented in ing, these three taxonomies change from simple to complex
Section 3.3. and are reported successively. Next we summarize the re-
search progress in each of the three taxonomies.
2.5 Handling User-Item Interaction Sequences
with Hierarchical Structures 3.1 Traditional Sequence Models for SRSs
Generally, there are mainly two kinds of hierarchical struc- Traditional sequence models including sequential pattern
tures that may be associated with a user-item interaction se- mining and Markov chain models are intuitive solutions to
quence: (1) the hierarchical structure between meta data SRSs by taking advantage of their natural strength in mod-
and user-item interactions. To be specific, the users’ demo- elling sequential dependencies among the user-item interac-
graphics can determine the users’ preferences in some de- tions in a sequence.
gree and can further affect their interactions with the items. Sequential pattern mining. Sequential pattern-based RSs
Similarly, the features of items often have some effects on first mine frequent patterns on sequence data and then uti-
whether they will be liked and interacted by users [Hidasi et lize the mined patterns to guide the subsequent recommenda-
al., 2016b]; and (2) the hierarchical structure between sub- tions. Although simple and straightforward, sequential pat-
sequences and user-item interactions. More specifically, in tern mining usually generates a large number of redundant
some SRSs, one user-item interaction sequence includes mul- patterns, which increases unnecessary cost w.r.t. time and
tiple sub-sequences (also called sessions). In such a case, space. Another obvious shortcoming is that it often loses
in addition to the prior interactions within the current sub- those infrequent patterns and items due to the frequency con-
sequence, the historical sub-sequences may also influence the straint, which limits the recommendation results to those pop-
next user-item interaction to be predicted in the current sub- ular items. Therefore, quite few works have been reported in
sequence [Ying et al., 2018]. Therefore, one more key chal- this class, except a representative one [Yap et al., 2012].
lenge in SRSs is how to incorporate the hierarchical depen- Markov chain models. Markov chain-based RSs adopt
dencies embedded in these two kinds of hierarchical struc- Markov chain models to model the transitions over user-item
tures into sequential dependency learning to generate more interactions in a sequence, for the prediction of the next in-
accurate sequential recommendations. teraction. According to the specific technique used, Markov
Although quite a few works have attempted to address this chain-based RSs are divided into basic Markov Chain-based
challenge from certain aspects, some other aspects have been approaches and latent Markov embedding-based approaches.
less studied. On the one hand, to take the influences of items’ The former one directly calculates the transition probability
features on the user-item interactions into account, a series based on the explicit observations [Garcin et al., 2013], while
of feature-enriched neural models including [Hidasi et al., the latter first embeds the Markov chains into an Euclidean
2016b] have been proposed for SRSs. In comparison, the in- space and then calculates the transition probabilities between
fluences of users’ demographics have been rarely considered interactions based on their Euclidean distance [Feng et al.,
in existing SRSs and more efforts should be devoted into this 2015]. The shortcomings of Markov chain-based RSs are ob-
direction. On the other hand, some hierarchical models in- vious, namely, on the one hand, they can only capture the
cluding hierarchical embedding models [Wang et al., 2015], short-term dependencies while ignoring long-term ones due
hierarchical RNN [Quadrana et al., 2017] and hierarchical at- to the Markov property which assumes that the current in-
tention networks [Ying et al., 2018] have been devised to in- teraction depends on one or several most recent interactions
corporate the historical sub-sequences into sequential depen- only; on the other hand, they can only capture the point-wise
dency learning to build more powerful SRSs. Particularly, the dependencies while ignoring the collective dependencies over
technical progress achieved to address this challenge will be user-item interactions. Consequently, they are less and less
presented in Sections 3.2 and 3.3. employed in SRSs in recent years.

3 Research Progress 3.2 Latent Representation Models for SRSs

To provide an overview of the technical progress in SRSs and Latent representation models first learn a latent representation
to give more technical details of the solutions to the afore- of each user or item, and then predict the subsequent user-
mentioned challenges, we summarize and briefly discuss the item interactions by utilizing the learned representations. As
research progress in SRSs from a technical perspective in this a result, more implicit and complex dependencies are cap-
section. Particularly, we first present a categorization of all tured in a latent space, which greatly benefits the recommen-
the approaches for SRSs from the technical perspective and dations. Next, we introduce two representative models falling
then briefly highlight the recent progress in each category. into this taxonomy.
The categorization of SRS approaches is presented in Fig- Factorization machines. Factorization machine-based SRSs
ure 2. We observe that the various approaches for SRSs are usually utilize the matrix factorization or tensor factorization

6335
 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

Sequential pattern mining

Traditional
sequence models
Markov chain models

Factorization machines
Sequential recommender Latent representation
system approaches models
Embedding
Recurrent neural networks

Basic deep neural Convolutional neural networks

networks

Graph neural networks

Deep neural network
models
Attention models

Advanced models Memory networks

Mixture models

Figure 2: A categorization of SRS approaches from the technical perspective

to factorize the observed user-item interactions into latent fac- convolutional neural networks (CNN) and graph neural net-
tors of users and items for recommendations [Rendle et al., works (GNN) have also been applied in SRSs to make up the
2010; Hidasi and Tikk, 2016]. Different from collaborative defects of RNN. Next, we introduce the SRSs built on top of
filtering (CF), the matrix or tensor to be factorized is com- these three deep neural networks respectively.
posed of interactions rather than the ratings in CF. Such a RNN-based SRSs. Given a sequence of historical user-item
model is easily affected by the sparsity of the observed data interactions, an RNN-based SRS tries to predict the next pos-
and thus cannot achieve ideal recommendations. sible interaction by modelling the sequential dependencies
Embedding. Embedding-based SRSs learn a latent repre- over the given interactions. Except for the basic RNN, long-
sentations for each user and item for the subsequent rec- short-term-memory (LSTM)- [Wu et al., 2017] and gated re-
ommendations by encoding all the user-item interactions in current unit (GRU)-based [Hidasi et al., 2016a] RNN have
a sequence into a latent space. Specifically, some works also been developed to capture the long-term dependencies
take the learned latent representations as the input of a net- in a sequence. Recent years have witnessed the prosperity of
work to further calculate an interaction score between users RNN-based SRSs and they dominate the research on the deep
and items, or successive users’ actions [Wang et al., 2015; learning-based SRSs or even the whole SRSs. Besides the
2018], while other works directly utilize them to calculate a basic RNN structure, some variants are proposed to capture
metric like the Euclidean distance as the interaction score [He more complex dependencies in a sequence, like hierarchical
et al., 2018]. This model has shown great potential in recent RNN [Quadrana et al., 2017]. However, RNN is not flaw-
years due to its simplicity, efficiency and efficacy. less for SRSs, with the shortcomings in two aspects: (1) it is
easy to generate fake dependencies due to the overly strong
3.3 Deep Neural Network Models for SRSs assumption that any adjacent interactions in a sequence must
Deep neural networks have natural strength to model and cap- be dependent, which may not be the cases in the real world
ture the comprehensive relations over different entities (e.g., because there are usually irrelevant or noisy interactions in-
users, items, interactions) in a sequence, and thus they nearly side a sequence; and (2) it is likely to capture the point-wise
dominate SRSs in the past few years. The latest progress dependencies only while ignoring the collective dependencies
achieved in SRSs also belongs to this taxonomy. Generally, (e.g., several interactions collaboratively affect the next one).
this taxonomy can be divided into two sub classes: SRSs built CNN-based SRSs. Different from RNN, given a sequence
on basic deep neural networks and SRSs built on deep neural of user-item interactions, a CNN first puts all the embeddings
networks with some advanced models incorporated. of these interactions into a matrix, and then treats such a ma-
trix as an “image” in the time and latent spaces. Finally, a
Basic Deep Neural Networks CNN learns sequential patterns as local features of the image
The most commonly used deep neural networks for SRSs are using convolutional filters for the subsequent recommenda-
recurrent neural networks (RNN) due to their natural strength tions. Since a CNN does not have strong order assumptions
in sequence modelling, but they also have defects. Recently, over the interactions in a sequence, and they learn patterns

6336
 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

between the areas in an “image” rather than over interactions, models are in their early stages.
therefore, CNN-based SRSs can make up the aforementioned
drawbacks of RNN-based SRSs to some degree. However, 4 Open Research Directions
CNN-based SRSs cannot effectively capture long-term de- Recent years, particularly the recent three years, have wit-
pendencies due to the limited sizes of the filters used in CNN, nessed the fast development of sequential recommender sys-
which limits their applications. The typical works include tems, along with the prosperity of deep learning, especially
[Tang and Wang, 2018; Yuan et al., 2019]. that of the recurrent neural networks. While categorizing and
GNN-based SRSs. Recently, with the fast development of summarizing the research practices in this filed, we have iden-
GNN, GNN-based SRSs have been devised to leverage GNN tified further open research directions discussed below.
to model and capture the complex transitions over user-item Context-aware sequential recommender systems. The cur-
interactions in a sequence. Typically a directed graph is first rent context in which a user or an item is could greatly in-
built on the sequence data by taking each interaction as a node fluence the user’s choice on the item, which should be con-
in the graph while each sequence is mapped to a path. Then, sidered when conducting recommendations. This is even
the embeddings of users or items are learned on the graph to more necessary in SRSs as the context may change over time.
embed more complex relations over the whole graph [Wu et However, most existing SRSs ignore such significant aspect.
al., 2019]. Such an approach makes full use of the advan- Therefore, context-aware SRSs would be an important direc-
tage of GNN to capture the complex relations in structured tion for future work.
relation datasets. GNN-based SRSs have shown a great po- Social-aware sequential recommender systems. Users live
tential to provide explainable recommendations by revealing in a society and are connected with various people both online
the complex relations between the recommended items and and offline. Others’ behaviors or opinions often affect the
the corresponding sequential context. Such kind of SRSs are users’ choices greatly. Therefore, the social influence needs
still in their early stages. to be taken into account in SRSs, which is usually ignored in
Advanced Models the existing works.
To address the limitations of SRSs built on basic neural net- Interactive sequential recommender systems. Most of
work structures, some advanced models are usually combined shopping behaviours in the real-world are continuous rather
together with a certain kind of basic deep neural networks than isolated events. In other words, there are actually se-
(e.g., RNN, CNN) to build more powerful SRSs which are quential interactions between a user and the shopping plat-
able to address particular challenges. Next, we introduce form (e.g., Amazon). However, the existing SRSs often ne-
three advanced models that are commonly used in SRSs. glect such interactions and only generate recommendations
Attention models. Attention models are commonly em- for one action at a single time step. How to incorporate the
ployed in SRSs to emphasize those really relevant and im- user-seller interactions and thus generate multi-time step rec-
portant interactions in a sequence while downplaying those ommendations is a promising research direction.
ones irrelevant to the next interaction. They are widely incor- Cross-domain sequential recommender systems. In the
porated into shallow networks [Wang et al., 2018] and RNN real world, items purchased by a user during a certain time
[Ying et al., 2018] to handle interaction sequences with noise. period are often from multi-domains rather than one domain.
Essentially, there are some sequential dependencies between
Memory networks. Memory networks are introduced into
items from different domains, such as the purchase of a car
SRSs to capture the dependencies between any historical
insurance after the purchase of a car. Such cross-domain se-
user-item interaction and the next one directly by incorpo-
quential dependencies are ignored in most SRSs. Therefore,
rating an external memory matrix. Such matrix enables it
cross-domain SRS is another promising research direction to
possible to store and update the historical interactions in a
generate more accurate recommendations by leveraging in-
sequence more explicitly and dynamically to improve the ex-
formation from other domains and more diverse recommen-
pressiveness of the model and reduce the interference of those
dations from different domains.
irrelevant interactions [Chen et al., 2018]. Furthermore, some
works incorporate a key-value memory network to store and
update the corresponding knowledge base information of the 5 Conclusions
interacted items in a sequence to learn the attribute level pref- Recommender systems (RS) is one of the most direct and
erence for the enhancement of recommendations [Huang et practical applications of artificial intelligence in our daily
al., 2018]. Generally, memory networks have shown their lives. Sequential recommender systems (SRSs) have been at
potential in SRSs, but are not sufficiently studied yet. the core of the RS field in the past three to five years as they
Mixture models. A mixture model-based SRS combines dif- provide more intelligent and favorable recommendations to
ferent models that excel at capturing different kinds of de- satisfy our daily requirements. It is our hope that this sum-
pendencies to enhance the capability of the whole model in mary provides an overview of the challenges and the recent
capturing various dependencies for better recommendations. progress as well as some future directions in SRSs to the RS
A typical example is [Tang et al., 2019], which combines dif- research community.
ferent kinds of encoders that are suitable for short- and long-
term dependencies respectively to learn a more precise se- Acknowledgements
quence representation for the subsequent recommendations This work was partially supported by Australian Research
and has demonstrated to be quite effective. However, such Council Discovery Project DP180102378.

6337
 Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI-19)

References [Rendle et al., 2010] Steffen Rendle, Christoph Freuden-

[Chen et al., 2018] Xu Chen, Hongteng Xu, Yongfeng thaler, and Lars Schmidt-Thieme. Factorizing person-
Zhang, and et al. Sequential recommendation with user alized markov chains for next-basket recommendation.
memory networks. In Proceedings of the 11th ACM In- In Proceedings of the 19th International Conference on
ternational Conference on Web Search and Data Mining, World Wide Web, pages 811–820, 2010.
pages 108–116, 2018. [Tang and Wang, 2018] Jiaxi Tang and Ke Wang. Person-
[Feng et al., 2015] Shanshan Feng, Xutao Li, Yifeng Zeng, alized top-n sequential recommendation via convolutional
Gao Cong, and Yeow Meng Chee. Personalized ranking sequence embedding. In Proceedings of the 11th ACM In-
metric embedding for next new poi recommendation. In ternational Conference on Web Search and Data Mining,
Proceedings of the 24th International Joint Conference on pages 565–573, 2018.
Artificial Intelligence, pages 2069–2075, 2015. [Tang et al., 2019] Jiaxi Tang, Francois Belletti, Sagar Jain,
[Garcin et al., 2013] Florent Garcin, Christos Dimitrakakis, Minmin Chen, Alex Beutel, and et al. Towards neural
and Boi Faltings. Personalized news recommendation with mixture recommender for long range dependent user se-
context trees. In Proceedings of the 7th ACM Conference quences. In Proceedings of the 28th International Confer-
on Recommender Systems, pages 105–112, 2013. ence on World Wide Web, pages 811–820, 2019.
[He and McAuley, 2016] Ruining He and Julian McAuley. [Wang et al., 2015] Pengfei Wang, Jiafeng Guo, Yanyan
Fusing similarity models with markov chains for sparse Lan, Jun Xu, Shengxian Wan, and Xueqi Cheng. Learning
sequential recommendation. In Proceedings of the 16th hierarchical representation model for next basket recom-
IEEE International Conference on Data Mining, pages mendation. In Proceedings of the 38th ACM SIGIR Con-
191–200, 2016. ference on Research and Development in Information Re-
trieval, pages 403–412, 2015.
[He et al., 2018] Ruining He, Wang-Cheng Kang, and Julian
McAuley. Translation-based recommendation: A scalable [Wang et al., 2018] Shoujin Wang, Liang Hu, Longbing
method for modeling sequential behavior. In Proceedings Cao, Xiaoshui Huang, Defu Lian, and Wei Liu. Attention-
of the 27th International Joint Conference on Artificial In- based transactional context embedding for next-item rec-
telligence, pages 5264–5268, 2018. ommendation. In Proceedings of the 32nd AAAI Confer-
ence on Artificial Intelligence, pages 2532–2539, 2018.
[Hidasi and Tikk, 2016] Balázs Hidasi and Domonkos Tikk.
General factorization framework for context-aware rec- [Wang et al., 2019] Shoujin Wang, Liang Hu, Yang Wang,
ommendations. Data Mining and Knowledge Discovery, and et al. Modeling multi-purpose sessions for next-
30(2):342–371, 2016. item recommendations via mixture-channel purpose rout-
ing networks. In Proceedings of the 28th International
[Hidasi et al., 2016a] Balázs Hidasi, Alexandros Karat-
Joint Conference on Artificial Intelligence, pages 1–7,
zoglou, and et al. Session-based recommendations with 2019.
recurrent neural networks. In Proceedings of the 4th Inter-
national Conference on Learning Representations, pages [Wu et al., 2017] Chao-Yuan Wu, Amr Ahmed, Alex Beu-
1–10, 2016. tel, Alexander J Smola, and How Jing. Recurrent recom-
mender networks. In Proceedings of the 10th ACM In-
[Hidasi et al., 2016b] Balázs Hidasi, Massimo Quadrana,
ternational Conference on Web Search and Data Mining,
Alexandros Karatzoglou, and Domonkos Tikk. Paral- pages 495–503, 2017.
lel recurrent neural network architectures for feature-rich
session-based recommendations. In Proceedings of the [Wu et al., 2019] Shu Wu, Yuyuan Tang, and et al. Session-
10th ACM Conference on Recommender Systems, pages based recommendation with graph neural networks. In
241–248, 2016. Proceedings of the 33rd AAAI Conference on Artificial In-
telligence, pages 1–9, 2019.
[Huang et al., 2018] Jin Huang, Wayne Xin Zhao, Hongjian
Dou, Ji-Rong Wen, and Edward Y Chang. Improv- [Yap et al., 2012] Ghim-Eng Yap, Xiao-Li Li, and Philip Yu.
ing sequential recommendation with knowledge-enhanced Effective next-items recommendation via personalized se-
memory networks. In Proceedings of the 41st ACM SIGIR quential pattern mining. In Database Systems for Ad-
Conference on Research & Development in Information vanced Applications, pages 48–64, 2012.
Retrieval, pages 505–514, 2018. [Ying et al., 2018] Haochao Ying, Fuzhen Zhuang, Fuzheng
[Kang et al., 2018] Wang-Cheng Kang, Mengting Wan, and Zhang, Yanchi Liu, Guandong Xu, and Xing Xie. Sequen-
Julian McAuley. Recommendation through mixtures of tial recommender system based on hierarchical attention
heterogeneous item relationships. In Proceedings of the networks. In Proceedings of the 27th International Joint
27th ACM International Conference on Information and Conference on Artificial Intelligence, 2018.
Knowledge Management, pages 1143–1152, 2018. [Yuan et al., 2019] Fajie Yuan, Alexandros Karatzoglou,
[Quadrana et al., 2017] Massimo Quadrana, Alexandros Ioannis Arapakis, Joemon M Jose, and Xiangnan He. A
Karatzoglou, and et al. Personalizing session-based simple convolutional generative network for next item rec-
recommendations with hierarchical recurrent neural ommendation. In Proceedings of the 12th ACM Interna-
networks. In Proceedings of the 11th ACM Conference on tional Conference on Web Search and Data Mining, pages
Recommender Systems, pages 130–137, 2017. 582–590, 2019.

6338