MMGA Multimodal Learning with Graph Alignment Xuan Yang Zhejiang University

2025-05-02 0 0 405.28KB 3 页 10玖币

侵权投诉

MMGA: Multimodal Learning with Graph Alignment

Xuan Yang

Zhejiang University

xuany@zju.edu.cn

Quanjin Tao

Zhejiang University

taoquanjin@zju.edu.cn

Xiao Feng

Zhejiang University

3200104919@zju.edu.cn

Donghong Cai

Zhejiang University

donghongcai@zju.edu.cn

Xiang Ren

University of Southern California

xiangren@usc.edu

Yang Yang∗

Zhejiang University

yangya@zju.edu.cn

ABSTRACT

Multimodal pre-training breaks down the modal barriers and allows

the individual modalities to be mutually augmented with informa-

tion, resulting in signicant advances in representation learning.

However, graph modality, as a very general and important form of

data, cannot be easily interacted with other modalities because of its

non-regular nature. In this paper, we propose MMGA (Multimodal

learning with Graph Alignment), a novel multimodal pre-training

framework to incorporate information from graph (social network),

image and text modalities on social media to enhance user represen-

tation learning. In MMGA, a multi-step graph alignment mechanism

is proposed to add the self-supervision from graph modality to op-

timize the image and text encoders, while using the information

from the image and text modalities to guide the graph encoder

learning. We conduct experiments on the dataset crawled from

Instagram. The experimental results show that MMGA works well

on the dataset and improves the fans prediction task’s performance.

We release our dataset, the rst social media multimodal dataset

with graph, of 60,000 users labeled with specic topics based on 2

million posts to facilitate future research.

CCS CONCEPTS

•Computing methodologies →Articial intelligence.

KEYWORDS

social media, multimodal learning, graph pre-training, user repre-

sentation

ACM Reference Format:

Xuan Yang, Quanjin Tao, Xiao Feng, Donghong Cai, Xiang Ren, and Yang

Yang. 2022. MMGA: Multimodal Learning with Graph Alignment. In Pro-

ceedings of ACM Conference (Conference’17). ACM, New York, NY, USA,

3 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn

∗Corresponding author.

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 prot 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 specic permission and/or a

fee. Request permissions from permissions@acm.org.

Conference’17, July 2017, Washington, DC, USA

ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00

https://doi.org/10.1145/nnnnnnn.nnnnnnn

1 INTRODUCTION

Multimodal learning has gained increasing attention in recent years

as the heterogeneous data become ubiquitous in the real world.

Nowadays, many multimodal works study the question of how to

utilize the relationships between the multi modalities and enhance

the cross-modal representation learning, such as CLIP [

], BLIP [

]

and FLAVA [

]. However, these works mostly focus on the vision-

language modalities, few works pay attention to the graph modality.

The graph modality is commonly exists in the real world and is

very important in many applications. For example, on the social

media, besides the texts and the images that posted by users, there

is abundant information in graph data (such as interaction graph

and following graph), which can help us better capture the interest

of user and then improve the user recommendation or advertising

strategy. However, it is dicult to deeply incorporate the graph

modality with image and text modalities: compared with the im-

age and text data, the graph is naturally non-euclidean structure

data, which makes the existing modality fusion method unworkable.

Thus, rather than the natural alignment learning method that com-

monly used in image-text modalities, how to incorporate the graph

data into the multimodal learning is a key challenge. Though there

are works incorporate the graph data into the multimodal learning,

they mainly treat the graph data as an individual modality, which

only use the graph structure for information aggregation after the

multi-modal (text-image) fusion process [

][

] and ignoring the

cross-modality information between graph and other.

To address the above problem, we propose the Multimodal learn-

ing with Graph Alignment (MMGA) framework to better leverage

the information in multi modalities and enlarge their mutual in-

formation. Specically, we conduct our study on the social media

dataset. On social media, there are user social networks which

contains rich user social information and user posts which simul-

taneously contain image and text. We propose to use the graph

structure in the user social network to supervise the learnt space of

image and text modality. Meanwhile, we utilize the image and text

representations of users to infer the relationships between users,

such as the closeness between users, and then use it to support

the graph learning process. In other word, we align the explicit

graph structure with the implicit image semantic structure and text

semantic structure respectively. In this way, we align the graph

modality’s space with the semantic spaces of image and text modal-

ities; and then, we use the distance between two nodes in image

and text spaces to adjust the graph weight. In this way, we use

graph modality’s information to add an additional supervision on

text and image modality representation learning and use the text

arXiv:2210.09946v2 [cs.MM] 31 Oct 2022

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：

MMGA:MultimodalLearningwithGraphAlignmentXuanYangZhejiangUniversityxuany@zju.edu.cnQuanjinTaoZhejiangUniversitytaoquanjin@zju.edu.cnXiaoFengZhejiangUniversity3200104919@zju.edu.cnDonghongCaiZhejiangUniversitydonghongcai@zju.edu.cnXiangRenUniversityofSouthernCaliforniaxiangren@usc.eduYangYang∗Zhejian...

展开>> 收起<<

MMGA Multimodal Learning with Graph Alignment Xuan Yang Zhejiang University.pdf

共3页,预览1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！

MMGA Multimodal Learning with Graph Alignment Xuan Yang Zhejiang University

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: