Controllable Multi-Interest Framework for Recommendation
Yukuo Cen
†
, Jianwei Zhang
‡
, Xu Zou
†
, Chang Zhou
‡
, Hongxia Yang
‡∗
, Jie Tang
†∗
†
Department of Computer Science and Technology, Tsinghua University
‡
DAMO Academy, Alibaba Group
{cyk18,zoux18}@mails.tsinghua.edu.cn
{zhangjianwei.zjw,ericzhou.zc,yang.yhx}@alibaba-inc.com
jietang@tsinghua.edu.cn
ABSTRACT
Recently, neural networks have been widely used in e-commerce
recommender systems, owing to the rapid development of deep
learning. We formalize the recommender system as a sequential rec-
ommendation problem, intending to predict the next items that the
user might be interacted with. Recent works usually give an overall
embedding from a user’s behavior sequence. However, a unied
user embedding cannot reect the user’s multiple interests during a
period. In this paper, we propose a novel
co
ntrollable
m
ulti-
i
nterest
framework for the sequential
rec
ommendation, called ComiRec.
Our multi-interest module captures multiple interests from user
behavior sequences, which can be exploited for retrieving candidate
items from the large-scale item pool. These items are then fed into
an aggregation module to obtain the overall recommendation. The
aggregation module leverages a controllable factor to balance the
recommendation accuracy and diversity. We conduct experiments
for the sequential recommendation on two real-world datasets,
Amazon and Taobao. Experimental results demonstrate that our
framework achieves signicant improvements over state-of-the-art
models
1
. Our framework has also been successfully deployed on
the oine Alibaba distributed cloud platform.
CCS CONCEPTS
• Information systems → Recommender systems
;
• Com-
puting methodologies → Neural networks.
KEYWORDS
recommender system; sequential recommendation; multi-interest
framework
ACM Reference Format:
Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang and Jie
Tang. 2020. Controllable Multi-Interest Framework for Recommendation. In
Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and
Data Mining (KDD ’20), August 23–27, 2020, Virtual Event, CA, USA. ACM,
New York, NY, USA, 10 pages. https://doi.org/10.1145/3394486.3403344
∗
Hongxia Yang and Jie Tang are the corresponding authors.
1
Code is available at https://github.com/THUDM/ComiRec
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
KDD ’20, August 23–27, 2020, Virtual Event, CA, USA
© 2020 Association for Computing Machinery.
ACM ISBN 978-1-4503-7998-4/20/08.. . $15.00
https://doi.org/10.1145/3394486.3403344
1 INTRODUCTION
The development of e-commerce revolutionized our shopping styles
in recent years. Recommender systems play a fundamental role
in e-commerce companies. Traditional recommendation methods
mainly use collaborative ltering methods [
47
,
48
] to predict scores
between users and items. Recently, neural networks have been
widely used in e-commerce recommender systems, owing to the
rapid development of deep learning. Neural recommender systems
generate representations for users and items and outperform tradi-
tional recommendation methods. However, due to the large-scale
e-commerce users and items, it is hard to use deep models to di-
rectly give the click-through rate (CTR) prediction between each
pair of users and items. Current industrial practice is to use fast K
nearest neighbors (e.g., Faiss [
25
]) to generate the candidate items
and then use a deep model (e.g., xDeepFM [
33
]) to integrate the
attributes of users and items to optimize the business metrics such
as CTR.
Some recent works leverage graph embedding methods to obtain
representations for users and items, which can be used for down-
stream applications. For example, PinSage [
56
] builds on Graph-
SAGE [
15
] and has applied graph convolutional network based
methods to production-scale data with billions of nodes and edges.
GATNE [
6
] considers dierent user behavior types and leverages a
heterogeneous graph embedding method to learn representations
for users and items. However, this kind of method ignores the se-
quential information in the user behaviors and cannot capture the
correlations between adjacent user behaviors.
Recent researches [
7
,
27
,
36
] formalize the recommender system
as a sequential recommendation problem. With a user’s behavior
history, the sequential recommendation task is to predict the next
item he/she might be interested in. This task reects the real-world
recommendation situation. Many recent models can give an overall
embedding for each user from his/her behavior sequence. However,
a unied user embedding is hard to represent multiple interests.
For example, in Figure 1, the click sequence shows three dierent
interests of Emma. As a modern girl, Emma is interested in jewelry,
handbags, and make-ups. Therefore, she may click items of the
three categories during this period of time.
In this paper, we propose a novel controllable multi-interest
framework, called ComiRec. Our multi-interest module can capture
the multiple interests of users, which can be exploited for retriev-
ing candidate items. Our aggregation module combines these items
from dierent interests and outputs the overall recommendation.
Figure 1 shows a motivating example of our multi-interest frame-
work. We conduct experiments for the sequential recommendation,
which is similar to our online situation. The experimental results
arXiv:2005.09347v2 [cs.IR] 3 Aug 2020
