Accurate Bundle Matching and Generation via Multitask Learning with Partially Shared Parameters Hyunsik Jeon

2025-04-30 0 0 4.56MB 10 页 10玖币

侵权投诉

Accurate Bundle Matching and Generation via

Multitask Learning with Partially Shared Parameters

Hyunsik Jeon

Seoul National University

Seoul, South Korea

jeon185@snu.ac.kr

Jun-Gi Jang

Seoul National University

Seoul, South Korea

elnino4@snu.ac.kr

Taehun Kim

Seoul National University

Seoul, South Korea

kbiglight@snu.ac.kr

U Kang

Seoul National University

Seoul, South Korea

ukang@snu.ac.kr

Abstract—How can we recommend existing bundles to users

accurately? How can we generate new tailored bundles for

users? Recommending a bundle, or a group of various items,

has attracted widespread attention in e-commerce owing to

the increased satisfaction of both users and providers. Bundle

matching and bundle generation are two representative tasks in

bundle recommendation. The bundle matching task is to correctly

match existing bundles to users while the bundle generation is to

generate new bundles that users would prefer. Although many

recent works have developed bundle recommendation models,

they fail to achieve high accuracy since they do not handle

heterogeneous data effectively and do not learn a method for

customized bundle generation.

In this paper, we propose BUNDLEMAGE, an accurate ap-

proach for bundle matching and generation. BUNDLEMAGE

effectively mixes user preferences of items and bundles using

an adaptive gate technique to achieve high accuracy for the

bundle matching. BUNDLEMAGE also generates a personalized

bundle by learning a generation module that exploits a user

preference and the characteristic of a given incomplete bundle to

be completed. BUNDLEMAGE further improves its performance

using multi-task learning with partially shared parameters.

Through extensive experiments, we show that BUNDLEMAGE

achieves up to 6.6% higher nDCG in bundle matching and 6.3×

higher nDCG in bundle generation than the best competitors. We

also provide qualitative analysis that BUNDLEMAGE effectively

generates bundles considering both the tastes of users and the

characteristics of target bundles.

Index Terms—bundle recommendation, bundle matching, bun-

dle generation, multi-task learning

I. INTRODUCTION

Given item and bundle purchase histories of users, how

can we match existing bundles to the users and generate

new bundles for them? Recommending a bundle, or a group

of various items, instead of individual items has attracted

widespread attention in e-commerce since 1) it recommends

items that users would prefer at once and 2) it increases

the chances of unpopular items being exposed to users. Bun-

dle recommendation is divided into two different but highly

related tasks, bundle matching and bundle generation, both

of which play important roles in bundle recommendation.

Bundle matching, which is to accurately match pre-constructed

bundles to users, is crucial because it reduces the cost of

manually constructing a bundle every time. Bundle generation,

which automatically generates personalized bundles for users,

is necessary because it enables us to construct a new bundle

that better reﬂects user preferences than the pre-constructed

bundles in a long-term perspective.

Bundle recommendation, however, is challenging due to

the following reasons. First, bundle matching requires careful

handling of heterogeneous types of data (i.e., user-item in-

teractions and user-bundle interactions) to extract meaningful

preferences of users. Previous works [1], [2], [3], [4], [5] fail

to achieve high accuracy for bundle matching since they do

not establish a relationship between the heterogeneous data.

Second, bundle generation is a demanding task since the search

space of possible bundles is burdensome to cope with; ﬁnding

all possible bundles requires exponential computational costs

to the number of items. Existing methods [1], [5] do not learn

any generation mechanism from the observable data. Instead,

they heuristically generate new personalized bundles based

on a learned bundle matching and show poor performances

on bundle generation as a result. Third, it requires careful

design of architecture to achieve high accuracy in both bundle

matching and generation since they are highly related but

different tasks. Previous works [1], [2], [3], [4], [5] have

not studied architectures that perform both tasks concurrently

since they have focused only on the bundle matching model.

In this paper, we propose BUNDLEMAGE (Accurate Bundle

Matching and Generation via Multitask Learning with Partially

Shared Parameters), an accurate method for bundle recom-

mendation. To achieve high accuracy for the bundle match-

ing, BUNDLEMAGE carefully aggregates information of user-

bundle and user-item interactions by exploiting an adaptive

gate technique which adaptively balances the contribution

of heterogeneous information. BUNDLEMAGE also learns a

generation mechanism to provide a new tailored bundle for

users. We train a generation module of BUNDLEMAGE by

reconstructing given incomplete bundles, exploiting the prefer-

ences of users who have interacted with them. BUNDLEMAGE

further improves its performance via multi-task learning with

partially shared parameters to address the bundle matching and

bundle generation problems simultaneously. With these ideas,

BUNDLEMAGE accurately recommends existing bundles to

users, and successfully generates new bundles that users would

prefer.

Our contributions are summarized as follows:

•Method. We propose BUNDLEMAGE, an accurate

method for personalized bundle matching and generation.

arXiv:2210.15460v2 [cs.IR] 28 Oct 2022

Method

0.0

0.2

0.4

0.6

0.8

nDCG@5

Method

0.2

0.4

0.6

0.8

1.0

nDCG@5

Bundle Matching

on Steam dataset

Bundle Generation

on Netease dataset

𝟔. 𝟑×

𝟔. 𝟔%

BR Random

Best

0 2 4 6 8 10

BundleMage (proposed)

VAE-CF

GRAM-SMOT

BGCN

NCF

EFM

DAM

BPR

POP

Random

(a) Bundle Matching and Generation

User A

Preference:

RPG

Characteristics:

Shooting

Shooting &

RPG

Shooting &

Adventure

Shooting &

Simulation

Given

bundle

Target

users

BundleMage's Top-1

recommendations

Simultaneous consideration of bundle

characteristics and user preferences

Neon

Chrome

King

Oddball

Sparkle 2

Crimsonland

Deadly Sin

Dinosaur

Hunt

The

Odyssey

User B

Preference:

Adventure

User C

Preference:

Simulation

(b) Case Study

Fig. 1: [Best viewed in color] (a) Evaluation of BUNDLEMAGE and competitors for bundle matching on Steam dataset

and bundle generation on Netease dataset with respect to nDCG@5. BUNDLEMAGE outperforms the competitors for both

bundle matching and generation. (b) Top-1recommendations of BUNDLEMAGE for different target users in bundle generation.

BUNDLEMAGE considers the characteristics of a given bundle and the preferences of target users for bundle generation. For

instance, BUNDLEMAGE recommends a shooting and RPG game (e.g., Deadly Sin) for a bundle of shooting games and a

target user A who prefers RPG.

TABLE I: Table of frequently used symbols.

Symbol Description

vuuser u’s item interaction vector (∈Ni)

ruuser u’s bundle interaction vector (∈Nb)

ruuser u’s partially masked bundle interaction vector (∈Nb)

xbbundle b’s item afﬁliation vector (∈Ni)

xbbundle b’s partially masked item afﬁliation vector (∈Ni)

Nu, Ni, Nbnumbers of users, items, and bundles, respectively

Ω(vu),Ω(ru),Ω(xb)indices of observable entries in vu,ru, and xb, respectively

U,I,Bsets of users, items, and bundles, respectively

BUNDLEMAGE accurately matches users to bundle using

their past item and bundle interactions. BUNDLEMAGE

also effectively generates personalized bundles using tar-

get users’ preferences.

•Experiments. Extensive experiments on real-world

datasets show that BUNDLEMAGE provides the state-of-

the-art performance with up to 6.6% higher nDCG in

bundle matching, and up to 6.3×higher nDCG in bundle

generation compared to the best competitors (see Fig. 1a,

and Tables III and IV).

•Case studies. We show in case studies that

BUNDLEMAGE successfully generates personalized

bundles even with unpopular items which would

otherwise be rarely exposed (see Fig. 1b and 7).

The code and datasets are available at https://github.com/

BundleRecommender/BundleMage.

II. PROBLEM DEFINITION AND RELATED WORKS

In this section, we deﬁne the problem of bundle recommen-

dation and summarize related works. Symbols used frequently

in this paper are summarized in Table I.

A. Problem Deﬁnition

Bundle recommendation [6] aims to predict bundles, instead

of items, that a user would prefer. For each user u, we observe

item interaction vector vu∈RNiand bundle interaction vector

ru∈RNb, where Niand Nbare the numbers of items and

bundles, respectively. vuand ruare binary vectors, where each

nonzero entry indicates the interaction with the corresponding

item or bundle. We have a binary bundle-item afﬁliation matrix

X∈RNi×Nbwhere each nonzero entry indicates the inclusion

of an item to a bundle; xb∈RNi, which indicates bth column

of X, is the item afﬁliation vector of bundle b. We denote

the sets of indices of observable entries in vu,ru, and xbas

Ω(vu) = {i:i∈ I},Ω(ru) = {b:b∈ B}, and Ω(xb) = {i:

i∈ I}, respectively; U,I, and Bare the sets of users, items,

and bundles, respectively. We describe the formal deﬁnition of

bundle matching and bundle generation as follows.

Problem 1(Bundle matching):

Given: a user u’s item interaction vector vuand bundle

interaction vector ru,

Predict: the user u’s next interacted bundle b0, where b0∈ B

and b06∈ Ω(ru).

Problem 2(Bundle generation):

Given: a user u’s item interaction vector vu, bundle inter-

action vector ru, and an incomplete bundle ˜

G={i:i∈ I}

to be completed,

Construct: a personalized bundle G(u, ˜

G) = {i0:i0∈

I, i06∈ ˜

G} of size k |I|, which denotes a small set of

items to complete ˜

Gfor user u, to be recommended to user

uas the complete set ˜

G ∪ G(u, ˜

G).

B. Collaborative Filtering

Collaborative ﬁltering is the most extensively used recom-

mendation approach due to its powerful performance in real

world services. Collaborative ﬁltering predicts items a user

would prefer by capturing similar patterns across users and

items. On early works, matrix factorization approaches [7],

[8], [9] learn latent factors of users and items while predicting

interactions by a linear way. They still largely prevail rec-

ommender system community because of their simplicity and

effectiveness. Recent collaborative ﬁltering-based approaches

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AccurateBundleMatchingandGenerationviaMultitaskLearningwithPartiallySharedParametersHyunsikJeonSeoulNationalUniversitySeoul,SouthKoreajeon185@snu.ac.krJun-GiJangSeoulNationalUniversitySeoul,SouthKoreaelnino4@snu.ac.krTaehunKimSeoulNationalUniversitySeoul,SouthKoreakbiglight@snu.ac.krUKangSeoulNation...

展开>> 收起<<

Accurate Bundle Matching and Generation via Multitask Learning with Partially Shared Parameters Hyunsik Jeon.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Accurate Bundle Matching and Generation via Multitask Learning with Partially Shared Parameters Hyunsik Jeon

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: