Collaborative Memor y Network for Recommendation Systems
Travis Ebesu
Santa Clara University
Department of Computer Engineering
Santa Clara, CA, USA
tebesu@scu.edu
Bin Shen
Google
New York, NY, USA
bshen@google.com
Yi Fang
Santa Clara University
Department of Computer Engineering
Santa Clara, CA, USA
yfang@scu.edu
ABSTRACT
Recommendation systems play a vital role to keep users engaged
with personalized content in modern online platforms. Deep learn-
ing has revolutionized many research elds and there is a recent
surge of interest in applying it to collaborative ltering (CF). How-
ever, existing methods compose deep learning architectures with
the latent factor model ignoring a major class of CF models, neigh-
borhood or memory-based approaches. We propose Collaborative
Memory Networks (CMN), a deep architecture to unify the two
classes of CF models capitalizing on the strengths of the global struc-
ture of latent factor model and local neighborhood-based structure
in a nonlinear fashion. Motivated by the success of Memory Net-
works, we fuse a memory component and neural attention mecha-
nism as the neighborhood component. The associative addressing
scheme with the user and item memories in the memory module
encodes complex user-item relations coupled with the neural atten-
tion mechanism to learn a user-item specic neighborhood. Finally,
the output module jointly exploits the neighborhood with the user
and item memories to produce the ranking score. Stacking multiple
memory modules together yield deeper architectures capturing
increasingly complex user-item relations. Furthermore, we show
strong connections between CMN components, memory networks
and the three classes of CF models. Comprehensive experimen-
tal results demonstrate the eectiveness of CMN on three public
datasets outperforming competitive baselines. Qualitative visual-
ization of the attention weights provide insight into the model’s
recommendation process and suggest the presence of higher order
interactions.
CCS CONCEPTS
• Information systems → Recommender systems
;
• Comput-
ing methodologies → Neural networks;
KEYWORDS
deep learning; memory networks; collaborative ltering
ACM Reference Format:
Travis Ebesu, Bin Shen, and Yi Fang. 2018. Collaborative Memory Network
for Recommendation Systems. In SIGIR ’18: The 41st International ACM
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.
SIGIR ’18, July 8–12, 2018, Ann Arbor, MI, USA
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-5657-2/18/07.. . $15.00
https://doi.org/10.1145/3209978.3209991
SIGIR Conference on Research & Development in Information Retrieval, July
8–12, 2018, Ann Arbor, MI, USA. ACM, New York, NY, USA, 10 pages. https:
//doi.org/10.1145/3209978.3209991
1 INTRODUCTION
Recommendation systems are vital to keeping users engaged and
satised with personalized recommendations in the age of infor-
mation explosion. Users expect personalized content in modern
E-commerce, entertainment and social media platforms but the ef-
fectiveness of recommendations are restricted by existing user-item
interactions and model capacity. The ability to leverage higher or-
der reasoning may help alleviate the problem of sparsity. A popular
and successful technique, collaborative ltering (CF), establishes
the relevance between users and items from past interactions (e.g.,
clicks, ratings, purchases) by assuming similar users will consume
similar items.
CF can generally be grouped in three categories: memory or
neighborhood-based approaches, latent factor models and hybrid
models [
17
,
26
]. Memory or neighborhood-based methods form
recommendations by identifying groups or neighborhoods of sim-
ilar users or items based on the previous interaction history. The
simplicity of these models such as item
K
nearest neighbor (KNN)
have shown success in production systems at Amazon [
21
]. Latent
factor models such as matrix factorization project each user and
item into a common low dimensional space capturing latent rela-
tions. Neighborhood methods capture local structure but typically
ignore the mass majority of ratings available due to selecting at
most
K
observations from the intersection of feedback between two
users or items [
17
]. On the other hand, latent factor models capture
the overall global structure of the user and item relationships but
often ignore the presence of a few strong associations. The follow-
ing weaknesses between the local neighborhood-based and global
latent factor models lead to the development of hybrid models such
as SVD++ [
17
] and generalizations such as Factorization Machines
[
24
] which integrate both neighborhood-based approaches and
latent factor models to enrich predictive capabilities.
Recently, deep learning has made massive strides in many re-
search areas obtaining state of the art performance in computer
vision [
9
], question answering [
18
,
30
,
35
,
39
], learning programs
[
8
], machine translation [
1
] and many other domains. The suc-
cessful integration of deep learning methods in recommendation
systems have demonstrated the noticeable advantages of com-
plex nonlinear transformations over traditional linear models [
40
].
However, existing composite architectures incorporate the latent
factor model ignoring the integration of neighborhood-based ap-
proaches in a nonlinear fashion. Hence, we propose to represent the
neighborhood-based component with a Memory Network [
30
,
35
]
to capture higher order complex relations between users and items.
arXiv:1804.10862v2 [cs.IR] 21 Jun 2018