Multiplex Memory Network for Collaborative Filtering_BUPT, Ant Group, SMU_2020_SDM.pdf

tinge

149

9页

1次

2024-01-15

免费下载

Multiplex Memory Network for Collaborative Filtering

Xunqiang Jiang

∗

Binbin Hu

†

Yuan Fang

‡

Chuan Shi

∗§

Abstract

Recommender systems play an important role in helping

users discover items of interest from a large resource collec-

tion in various online services. Although current deep neural

network-based collaborative ﬁltering methods have achieved

state-of-the-art performance in recommender systems, they

still face a few major weaknesses. Most importantly, such

deep methods usually focus on the direct interaction be-

tween users and items only, without explicitly modeling

high-order co-occurrence contexts. Furthermore, they treat

the observed data uniformly, without ﬁne-grained diﬀerenti-

ation of importance or relevance in the user-item interactions

and high-order co-occurrence contexts. Inspired by recent

progress in memory networks, we propose a novel multiplex

memory network for collaborative ﬁltering (MMCF). More

speciﬁcally, MMCF leverages a multiplex memory layer con-

sisting of an interaction memory and two co-occurrence con-

text memories simultaneously, in order to jointly capture and

locate important and relevant information in both user-item

interactions and co-occurrence contexts. Lastly, we conduct

extensive experiments on four datasets, and the results show

the superior performance of our model in comparison with

a suite of state-of-the-art methods.

1 Introduction

In the era of information overload, recommender sys-

tems have been playing an increasingly important role in

various online services [6], including E-commerce, online

news and social media. Current recommender systems

evolve around learning an eﬀective preference predic-

tion function based on historical user-item interaction

records, known as collaborative ﬁltering (CF) [17]. In

particular, matrix factorization (MF) [7] is one of the

most successful methods among CF techniques, which

models the preference as the inner product of user and

item latent factors. Since the interactions between users

and items are often complex and may involve vastly dif-

∗

Beijing University of Posts and Telecommunications, China.

Email: {skd621, shichuan}@bupt.edu.cn

†

Ant Financial Services Group, China. Email:

bin.hbb@antﬁn.com

‡

Singapore Management University, Singapore. Email:

yfang@smu.edu.sg

Corresponding author

ferent underlying intentions, such a shallow representa-

tion could be inadequate in expressing the preferences.

Due to the ability of modeling non-linear functions,

deep neural networks have yielded state-of-the-art per-

formance in many research areas such as computer vi-

sion and natural language processing. The recent inte-

gration of deep models into recommender systems has

also revealed the remarkable strength of complex non-

linear transformations of user-item interactions. Exist-

ing neural network-based recommendation fall into two

broad categories. The ﬁrst category replaces the tradi-

tional inner product with nonlinear neural networks to

model more complex prediction functions [2, 18]. Recent

eﬀorts [15, 27] further utilize random walks to augment

user-item interactions, which ultimately train a more

eﬀective deep model. The second category extends the

matrix factorization by incorporating deep representa-

tions learned from additional side information such as

review texts and videos [25, 28].

Unfortunately, these deep recommendation models

still suﬀer from several limitations. Consider the sce-

nario in Fig. 1(a1)—given that the user u

purchased

items such as camera (i

), pen (i

) and book (i

) in

the past, the recommender system needs to determine

whether u

is likely to purchase a memory card (i

One limitation of the previous deep models, as

sketched in Fig. 1(b), is that they mainly focus on the

direct interaction between users and items, without ex-

plicitly accounting for the high-order co-occurrence con-

texts between users and items, respectively. Examples

are the co-purchase contexts between users such as u

and u

both purchasing the same item camera (i

) in

Fig. 1(a2), as well as the co-purchased context between

items such as the memory card (i

) and camera (i

) both

purchased by the same user u

in Fig. 1(a3). Such high-

order contexts are especially crucial in sparse datasets

lacking enough user-item interactions. While some ex-

isting works [1, 4, 27] have attempted to utilize higher-

order information, they focus on enriching user-item in-

teractions, rather than explicitly modeling user-user and

item-item contexts.

Second, most previous works treat the observed

data (i.e., user-item interactions and co-occurrence con-

texts) uniformly, without ﬁne-grained diﬀerentiation of

their underpinning preferences or importance. Intu-

Unauthorized reproduction of this article is prohibited

Downloaded 01/12/24 to 27.19.3.156 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

User Item

(b) Previous Models

𝒖

𝟑

𝒊

𝟑

Interaction

Embedding

Co-occurrence

User Embedding

Co-occurrence

Item Embedding

Prediction

(a) User-Item Interactions

and Co-occurrence context

𝒊

𝟏

𝒊

𝟐

𝒊

𝟑

𝒊

𝟒

𝒖

𝟐

𝒖

𝟏

𝒖

𝟑

𝒖

𝟐

𝒖

𝟏

𝒖

𝟑

𝒊

𝟐

𝒖

𝟐

𝒖

𝟏

𝒊

𝟏

𝒊

𝟐

𝒊

𝟑

𝒖

𝟐

𝒖

𝟏

𝒊

𝟏

𝒊

𝟐

𝒊

𝟑

User-Item

Interaction

0.7 0.3 0.8

0.3 0.4 0.7

Prediction

0.7 0.3 0.8

0.3 0.4 0.7

𝒖

𝟑

𝒊

𝟏

𝒖

𝟑

𝒊

𝟏

(a1) User-Item Interactions

(a2) User Co-occurrence

(a3) Item Co-occurrence

Figure 1: Illustration of our problem. (a) User-item

interactions and co-occurrence contexts. (b) Previous

models focusing on user-item interactions. (c) Our

model with ﬁne-grained interactions and co-occurrence

contexts.

itively, a user-item interaction could be inﬂuenced by

diverse preferences, given users from diﬀerent demo-

graphic, socioeconomic or cultural groups. Generally,

such diverse preferences may be weighed diﬀerently in

diﬀerent interactions. While such diﬀerentiation has

been explored to some extent for user-item interactions

[1, 23, 30], it remains an open question for co-occurrence

contexts, which are likewise not uniformly important

to the recommendation. As shown in Fig. 1(c), given

a memory card (i

), the co-occurring camera (i

) is a

more important context than a pen (i

), as memory card

and camera are more complementary in nature. Similar

scenarios exist in user co-occurrence contexts. Thus, in

both user-item interactions and co-occurrence contexts,

we need a ﬁne-grained model that allows diﬀerent pref-

erences and contexts to receive varying attention.

To deal with the above considerations, two critical

questions must be addressed. First, how to integrate

user-item interactions and co-occurrence contexts in a

uniﬁed and end-to-end manner? Second, how to enable

ﬁne-grained diﬀerentiation for both user-item interac-

tions and co-occurrence contexts? Inspired by the re-

cent progress in memory networks [14, 20], a few studies

[1, 23, 24] have employed memory networks in recom-

mender systems, which have the advantage of modeling

ﬁne-grained preferences with a memory module. How-

ever, their architecture only operates with a single type

of memory centering around user-item interactions. Dif-

ferent from them, we ﬁrstly propose a novel Multiplex

Memory Network for Collaborative Filtering (MMCF),

which leverages multiplex memory layers to jointly cap-

ture user-item interactions and the co-occurrence con-

texts simultaneously in one framework. More speciﬁ-

cally, the multiplex memory layer is designed to accom-

modate multiple types of memory concurrently, consist-

ing of an interaction memory to model user-item in-

teractions and two co-occurrence context memories to

model user-user and item-item contexts, respectively.

These memories are not independent and their inter-

play is also captured by our model. Furthermore, to-

wards ﬁne-grained interactions and contexts, for each

type of memory, an attention mechanism is employed

to locate the important and relevant information in the

memory slots.

In summary, we make the following contributions.

(1) We highlight the importance of explicitly modeling

the user and item co-occurrence contexts in deep mod-

els especially when the user-item interactions are sparse,

which is also observed empirically in our experiments.

(2) We propose a novel model MMCF, a deep multiplex

memory network to jointly capture ﬁne-grained user-

item interactions and co-occurrence contexts through

multiple types of memory in an uniﬁed, end-to-end

framework. To our best knowledge, MMCF is the ﬁrst

memory network concurrently employing multiple types

of memory for collaborative ﬁltering. (3) We conduct

extensive experiments on four real-world benchmarks,

validating the eﬀectiveness of MMCF and its assump-

tions.

2 Related Work

Collaborative ﬁltering (CF) aims to recommend a suit-

able list of items based on historical user-item inter-

action records. In particular, the matrix factorization

(MF) method [5, 7] has shown its eﬀectiveness in many

applications, which decomposes the user-item interac-

tion matrix to learn low-rank latent user and item fac-

tors. Furthermore, many studies [4, 10, 13, 19] have

been proposed to incorporate additional side informa-

tion. In spite of the success of existing CF-based rec-

ommendation methods, they still suﬀer from the limited

ability of modeling more complex user-item interactions.

Due to the ability of modeling arbitrary non-linear

functions, deep neural networks endow recommender

system with the potential of capturing more complex

and intricate user-item interactions [2, 8, 9, 12, 18].

Similar to the traditional MF-based model, various

attempts [4, 22] have been made to integrate diﬀerent

side information into the deep models. However, most

of these deep models only concern with coarse-grained

user-item interactions.

To enable ﬁne-grained modeling, several studies at-

Unauthorized reproduction of this article is prohibited

Downloaded 01/12/24 to 27.19.3.156 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

Multiplex Memory Network for Collaborative Filtering

Xunqiang Jiang

∗

Binbin Hu

†

Yuan Fang

‡

Chuan Shi

∗§

Abstract

Recommender systems play an important role in helping

users discover items of interest from a large resource collec-

tion in various online services. Although current deep neural

network-based collaborative ﬁltering methods have achieved

state-of-the-art performance in recommender systems, they

still face a few major weaknesses. Most importantly, such

deep methods usually focus on the direct interaction be-

tween users and items only, without explicitly modeling

high-order co-occurrence contexts. Furthermore, they treat

the observed data uniformly, without ﬁne-grained diﬀerenti-

ation of importance or relevance in the user-item interactions

and high-order co-occurrence contexts. Inspired by recent

progress in memory networks, we propose a novel multiplex

memory network for collaborative ﬁltering (MMCF). More

speciﬁcally, MMCF leverages a multiplex memory layer con-

sisting of an interaction memory and two co-occurrence con-

text memories simultaneously, in order to jointly capture and

locate important and relevant information in both user-item

interactions and co-occurrence contexts. Lastly, we conduct

extensive experiments on four datasets, and the results show

the superior performance of our model in comparison with

a suite of state-of-the-art methods.

1 Introduction

In the era of information overload, recommender sys-

tems have been playing an increasingly important role in

various online services [6], including E-commerce, online

news and social media. Current recommender systems

evolve around learning an eﬀective preference predic-

tion function based on historical user-item interaction

records, known as collaborative ﬁltering (CF) [17]. In

particular, matrix factorization (MF) [7] is one of the

most successful methods among CF techniques, which

models the preference as the inner product of user and

item latent factors. Since the interactions between users

and items are often complex and may involve vastly dif-

∗

Beijing University of Posts and Telecommunications, China.

Email: {skd621, shichuan}@bupt.edu.cn

†

Ant Financial Services Group, China. Email:

bin.hbb@antﬁn.com

‡

Singapore Management University, Singapore. Email:

yfang@smu.edu.sg

Corresponding author

ferent underlying intentions, such a shallow representa-

tion could be inadequate in expressing the preferences.

Due to the ability of modeling non-linear functions,

deep neural networks have yielded state-of-the-art per-

formance in many research areas such as computer vi-

sion and natural language processing. The recent inte-

gration of deep models into recommender systems has

also revealed the remarkable strength of complex non-

linear transformations of user-item interactions. Exist-

ing neural network-based recommendation fall into two

broad categories. The ﬁrst category replaces the tradi-

tional inner product with nonlinear neural networks to

model more complex prediction functions [2, 18]. Recent

eﬀorts [15, 27] further utilize random walks to augment

user-item interactions, which ultimately train a more

eﬀective deep model. The second category extends the

matrix factorization by incorporating deep representa-

tions learned from additional side information such as

review texts and videos [25, 28].

Unfortunately, these deep recommendation models

still suﬀer from several limitations. Consider the sce-

nario in Fig. 1(a1)—given that the user u

purchased

items such as camera (i

), pen (i

) and book (i

) in

the past, the recommender system needs to determine

whether u

is likely to purchase a memory card (i

One limitation of the previous deep models, as

sketched in Fig. 1(b), is that they mainly focus on the

direct interaction between users and items, without ex-

plicitly accounting for the high-order co-occurrence con-

texts between users and items, respectively. Examples

are the co-purchase contexts between users such as u

and u

both purchasing the same item camera (i

) in

Fig. 1(a2), as well as the co-purchased context between

items such as the memory card (i

) and camera (i

) both

purchased by the same user u

in Fig. 1(a3). Such high-

order contexts are especially crucial in sparse datasets

lacking enough user-item interactions. While some ex-

isting works [1, 4, 27] have attempted to utilize higher-

order information, they focus on enriching user-item in-

teractions, rather than explicitly modeling user-user and

item-item contexts.

Second, most previous works treat the observed

data (i.e., user-item interactions and co-occurrence con-

texts) uniformly, without ﬁne-grained diﬀerentiation of

their underpinning preferences or importance. Intu-

Unauthorized reproduction of this article is prohibited

Downloaded 01/12/24 to 27.19.3.156 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

User Item

(b) Previous Models

𝒖

𝟑

𝒊

𝟑

Interaction

Embedding

Co-occurrence

User Embedding

Co-occurrence

Item Embedding

Prediction

(a) User-Item Interactions

and Co-occurrence context

𝒊

𝟏

𝒊

𝟐

𝒊

𝟑

𝒊

𝟒

𝒖

𝟐

𝒖

𝟏

𝒖

𝟑

𝒖

𝟐

𝒖

𝟏

𝒖

𝟑

𝒊

𝟐

𝒖

𝟐

𝒖

𝟏

𝒊

𝟏

𝒊

𝟐

𝒊

𝟑

𝒖

𝟐

𝒖

𝟏

𝒊

𝟏

𝒊

𝟐

𝒊

𝟑

User-Item

Interaction

0.7 0.3 0.8

0.3 0.4 0.7

Prediction

0.7 0.3 0.8

0.3 0.4 0.7

𝒖

𝟑

𝒊

𝟏

𝒖

𝟑

𝒊

𝟏

(a1) User-Item Interactions

(a2) User Co-occurrence

(a3) Item Co-occurrence

Figure 1: Illustration of our problem. (a) User-item

interactions and co-occurrence contexts. (b) Previous

models focusing on user-item interactions. (c) Our

model with ﬁne-grained interactions and co-occurrence

contexts.

itively, a user-item interaction could be inﬂuenced by

diverse preferences, given users from diﬀerent demo-

graphic, socioeconomic or cultural groups. Generally,

such diverse preferences may be weighed diﬀerently in

diﬀerent interactions. While such diﬀerentiation has

been explored to some extent for user-item interactions

[1, 23, 30], it remains an open question for co-occurrence

contexts, which are likewise not uniformly important

to the recommendation. As shown in Fig. 1(c), given

a memory card (i

), the co-occurring camera (i

) is a

more important context than a pen (i

), as memory card

and camera are more complementary in nature. Similar

scenarios exist in user co-occurrence contexts. Thus, in

both user-item interactions and co-occurrence contexts,

we need a ﬁne-grained model that allows diﬀerent pref-

erences and contexts to receive varying attention.

To deal with the above considerations, two critical

questions must be addressed. First, how to integrate

user-item interactions and co-occurrence contexts in a

uniﬁed and end-to-end manner? Second, how to enable

ﬁne-grained diﬀerentiation for both user-item interac-

tions and co-occurrence contexts? Inspired by the re-

cent progress in memory networks [14, 20], a few studies

[1, 23, 24] have employed memory networks in recom-

mender systems, which have the advantage of modeling

ﬁne-grained preferences with a memory module. How-

ever, their architecture only operates with a single type

of memory centering around user-item interactions. Dif-

ferent from them, we ﬁrstly propose a novel Multiplex

Memory Network for Collaborative Filtering (MMCF),

which leverages multiplex memory layers to jointly cap-

ture user-item interactions and the co-occurrence con-

texts simultaneously in one framework. More speciﬁ-

cally, the multiplex memory layer is designed to accom-

modate multiple types of memory concurrently, consist-

ing of an interaction memory to model user-item in-

teractions and two co-occurrence context memories to

model user-user and item-item contexts, respectively.

These memories are not independent and their inter-

play is also captured by our model. Furthermore, to-

wards ﬁne-grained interactions and contexts, for each

type of memory, an attention mechanism is employed

to locate the important and relevant information in the

memory slots.

In summary, we make the following contributions.

(1) We highlight the importance of explicitly modeling

the user and item co-occurrence contexts in deep mod-

els especially when the user-item interactions are sparse,

which is also observed empirically in our experiments.

(2) We propose a novel model MMCF, a deep multiplex

memory network to jointly capture ﬁne-grained user-

item interactions and co-occurrence contexts through

multiple types of memory in an uniﬁed, end-to-end

framework. To our best knowledge, MMCF is the ﬁrst

memory network concurrently employing multiple types

of memory for collaborative ﬁltering. (3) We conduct

extensive experiments on four real-world benchmarks,

validating the eﬀectiveness of MMCF and its assump-

tions.

2 Related Work

Collaborative ﬁltering (CF) aims to recommend a suit-

able list of items based on historical user-item inter-

action records. In particular, the matrix factorization

(MF) method [5, 7] has shown its eﬀectiveness in many

applications, which decomposes the user-item interac-

tion matrix to learn low-rank latent user and item fac-

tors. Furthermore, many studies [4, 10, 13, 19] have

been proposed to incorporate additional side informa-

tion. In spite of the success of existing CF-based rec-

ommendation methods, they still suﬀer from the limited

ability of modeling more complex user-item interactions.

Due to the ability of modeling arbitrary non-linear

functions, deep neural networks endow recommender

system with the potential of capturing more complex

and intricate user-item interactions [2, 8, 9, 12, 18].

Similar to the traditional MF-based model, various

attempts [4, 22] have been made to integrate diﬀerent

side information into the deep models. However, most

of these deep models only concern with coarse-grained

user-item interactions.

To enable ﬁne-grained modeling, several studies at-

Unauthorized reproduction of this article is prohibited

Downloaded 01/12/24 to 27.19.3.156 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

of 9

免费下载

tugraph paper

关注

评论