Real-time Aention Based Look-alike Model for Recommender
System
Yudan Liu
WeiXin Group, Tencent Inc.
Beijing, China
danydliu@tencent.com
Kaikai Ge
WeiXin Group, Tencent Inc.
Beijing, China
kavinge@tencent.com
Xu Zhang
WeiXin Group, Tencent Inc.
Beijing, China
xuonezhang@tencent.com
Leyu Lin
WeiXin Group, Tencent Inc.
Beijing, China
goshawklin@tencent.com
ABSTRACT
Recently, deep learning models play more and more important roles
in contents recommender systems. However, although the perfor-
mance of recommendations is greatly improved, the "Matthew ef-
fect" becomes increasingly evident. While the head contents get
more and more popular, many competitive long-tail contents are
dicult to achieve timely exposure because of lacking behavior
features. This issue has badly impacted the quality and diversity of
recommendations. To solve this problem, look-alike algorithm is a
good choice to extend audience for high quality long-tail contents.
But the traditional look-alike models which widely used in online
advertising are not suitable for recommender systems because of
the strict requirement of both real-time and eectiveness. This pa-
per introduces a real-time attention based look-alike model (RALM)
for recommender systems, which tackles the challenge of conict
between real-time and eectiveness. RALM realizes real-time look-
alike audience extension beneting from seeds-to-user similarity
prediction and improves the eectiveness through optimizing user
representation learning and look-alike learning modeling. For user
representation learning, we propose a novel neural network struc-
ture named attention merge layer to replace the concatenation
layer, which signicantly improves the expressive ability of multi-
elds feature learning. On the other hand, considering the various
members of seeds, we design global attention unit and local atten-
tion unit to learn robust and adaptive seeds representation with
respect to a certain target user. At last, we introduce seeds clus-
tering mechanism which not only reduces the time complexity of
attention units prediction but also minimizes the loss of seeds in-
formation at the same time. According to our experiments, RALM
shows superior eectiveness and performance than popular look-
alike models. RALM has been successfully deployed in "Top Stories"
Recommender System of WeChat, leading to great improvement
on diversity and quality of recommendations. As far as we know,
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 ’19, August 4–8, 2019, Anchorage, AK, USA
© 2019 Association for Computing Machinery.
ACM ISBN 978-1-4503-6201-6/19/08.. . $15.00
https://doi.org/10.1145/3292500.3330707
this is the rst real-time look-alike model applied in recommender
systems.
CCS CONCEPTS
• Information systems → Recommender systems.
KEYWORDS
recommender system; look-alike; audience extension; deep learn-
ing; attention model; user representation learning
ACM Reference Format:
Yudan Liu, Kaikai Ge, Xu Zhang, and Leyu Lin. 2019. Real-time Attention
Based Look-alike Model for Recommender System. In The 25th ACM SIGKDD
Conference on Knowledge Discovery and Data Mining (KDD ’19), August 4–
8, 2019, Anchorage, AK, USA. ACM, New York, NY, USA, 9 pages. https:
//doi.org/10.1145/3292500.3330707
1 INTRODUCTION
As for contents recommender systems, the traditional recommenda-
tion algorithms such as collaborative ltering[
15
] and content based
algorithms[
12
] have been applied widely. Recently, deep learning
models such as deep neural networks (DNNs) and recurrent neural
networks (RNNs) are more and more popular on recommendation
task[
4
][
3
][
2
]. These deep learning based methods eectively cap-
ture the user preferences, item features and non-liner relationship
between user and item, which show better performance compared
with traditional algorithms on recommendation in most situations.
However, these end-to-end models like deep learning networks
aim at improving the performance of recommendation, and tend
to predict higher CTR (click-through rate) for the head contents
than the long-tail ones. It means that popular articles and those
target user have clicked are always preferred. At the same time,
there are many competitive long-tail contents including contents
from manual pushing, novelties and latest news. These long-tail
contents are usually short of behavior features, which are essen-
tial for recommendation models. As a result, they are dicult to
achieve wide and timely exposure. We call it the "Matthew eect"
in recommender systems, leading to low quality and poor diversity
of recommended contents. Apart from the performance, improving
the quality and diversity of recommendation results have become
common problems faced with many recommender systems.
arXiv:1906.05022v1 [cs.IR] 12 Jun 2019