VIT-CAT PARALLEL VISION TRANSFORMERS WITH CROSS ATTENTION FUSION FOR POPULARITY PREDICTION IN MEC NETWORKS Zohreh HajiAkhondi-Meybodiy Arash Mohammadiz Ming Houyz

2025-05-06 0 0 454.22KB 5 页 10玖币

侵权投诉

VIT-CAT: PARALLEL VISION TRANSFORMERS WITH CROSS ATTENTION FUSION FOR

POPULARITY PREDICTION IN MEC NETWORKS

Zohreh HajiAkhondi-Meybodi†, Arash Mohammadi‡, Ming Hou†‡,

Jamshid Abouei††, Konstantinos N. Plataniotis‡‡

†Electrical and Computer Engineering, Concordia University, Montreal, Canada

‡Concordia Institute of Information Systems Engineering (CIISE), Montreal, Canada

†‡ Defence Research and Development Canada (DRDC), Toronto, Canada

†† Department of Electrical Engineering, Yazd University, Yazd, Iran

‡‡ Electrical and Computer Engineering, University of Toronto, Toronto, Canada

ABSTRACT

Mobile Edge Caching (MEC) is a revolutionary technology for the

Sixth Generation (6G) of wireless networks with the promise to sig-

niﬁcantly reduce users’ latency via offering storage capacities at the

edge of the network. The efﬁciency of the MEC network, however,

critically depends on its ability to dynamically predict/update the

storage of caching nodes with the top-Kpopular contents. Conven-

tional statistical caching schemes are not robust to the time-variant

nature of the underlying pattern of content requests, resulting in a

surge of interest in using Deep Neural Networks (DNNs) for time-

series popularity prediction in MEC networks. However, existing

DNN models within the context of MEC fail to simultaneously cap-

ture both temporal correlations of historical request patterns and the

dependencies between multiple contents. This necessitates an ur-

gent quest to develop and design a new and innovative popularity

prediction architecture to tackle this critical challenge. The paper

addresses this gap by proposing a novel hybrid caching framework

based on the attention mechanism. Referred to as the parallel Vision

Transformers with Cross Attention (ViT-CAT) Fusion, the proposed

architecture consists of two parallel ViT networks, one for collecting

temporal correlation, and the other for capturing dependencies be-

tween different contents. Followed by a Cross Attention (CA) mod-

ule as the Fusion Center (FC), the proposed ViT-CAT is capable of

learning the mutual information between temporal and spatial cor-

relations, as well, resulting in improving the classiﬁcation accuracy,

and decreasing the model’s complexity about 8times. Based on the

simulation results, the proposed ViT-CAT architecture outperforms

its counterparts across the classiﬁcation accuracy, complexity, and

cache-hit ratio.

Index Terms—Mobile Edge Caching, Popularity Prediction,

Deep Neural Networks, Vision Transformer, Cross-Attention.

1. INTRODUCTION

The phenomenal growth in demand for mobile wireless data ser-

vices, together with the emergence of advanced Internet of Things

(IoT) applications bring new technical challenges to wireless com-

munications. According to Ericsson’s mobility report [1], global

mobile data trafﬁc is projected to exponentially grow from 67 ex-

abytes/month in 2021 to 282 exabytes/month in 2027. To accommo-

date the huge amount of mobile data trafﬁc, Mobile Edge Caching

This Project was partially supported by Department of National De-

fence’s Innovation for Defence Excellence & Security (IDEaS), Canada.

(MEC) [2–5] has emerged as a promising solution for potential de-

ployment in the Sixth Generation (6G) of communication networks.

MEC networks provide low-latency communication for IoT devices

by storing multimedia contents in the storage of nearby caching

nodes [6, 7]. The limited storage of caching nodes, however, makes

it impossible to preserve all contents on nearby devices. To tackle

this challenge, predicting the most popular content is of paramount

importance, as it can signiﬁcantly inﬂuence the content availability

in the storage of caching nodes and reduce users’ latency.

Existing popularity prediction solutions are typically developed

based on statistical models [6–10], Machine Learning (ML)-based

architectures [11–14], and Deep Neural Networks (DNNs) [15–24],

among which the latter is the most efﬁcient one for popularity pre-

diction. This is mainly due to the fact that DNN-based models can

capture users’ interests from raw historical request patterns without

any feature engineering or pre-processing. In addition, DNN-based

popularity prediction models are not prone to sparsity and cold-start

problems with new mobile user/multimedia contents. As a result, re-

cent research has shifted its primary attention to DNN-based frame-

works to monitor and forecast the popularity of content using its his-

torical request pattern. A critical aspect of a DNN-based popularity

prediction architecture is its ability to accurately capture both tempo-

ral and spatial correlations within the time-variant request patterns of

multiple contents. While the temporal correlation illustrates the vari-

ation of users’ preferences over time, spatial correlation reﬂects the

dependency between different multimedia contents. The majority of

works in this ﬁeld [18–21], however, are not appropriately designed

to simultaneously capture both dependencies. This necessitates an

urgent quest to develop and design a new and innovative popularity

prediction architecture, which is the focus of this paper.

Literature Review: Recently, a variety of promising strategies have

been designed to forecast the popularity of multimedia contents with

the application to MEC networks. In [25], an auto-encoder architec-

ture was proposed to improve content popularity prediction by learn-

ing the latent representation of historical request patterns of contents.

To boost the decision-making capabilities of caching strategies, Re-

inforcement Learning (RL) [26, 27] and Convolutional Neural Net-

work (CNN) [28]-based caching frameworks were introduced to ex-

ploit the contextual information of users. Despite all the beneﬁts

that come from the aforementioned works, they relied on a common

assumption that the content popularity/historical request patterns of

contents would remain unchanged over time, which is not applicable

in highly dynamic practical systems.

arXiv:2210.15125v1 [cs.LG] 27 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：

VIT-CAT:PARALLELVISIONTRANSFORMERSWITHCROSSATTENTIONFUSIONFORPOPULARITYPREDICTIONINMECNETWORKSZohrehHajiAkhondi-Meybodiy,ArashMohammadiz,MingHouyz,JamshidAboueiyy,KonstantinosN.PlataniotiszzyElectricalandComputerEngineering,ConcordiaUniversity,Montreal,CanadazConcordiaInstituteofInformationSystemsEn...

展开>> 收起<<

VIT-CAT PARALLEL VISION TRANSFORMERS WITH CROSS ATTENTION FUSION FOR POPULARITY PREDICTION IN MEC NETWORKS Zohreh HajiAkhondi-Meybodiy Arash Mohammadiz Ming Houyz.pdf

共5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！

VIT-CAT PARALLEL VISION TRANSFORMERS WITH CROSS ATTENTION FUSION FOR POPULARITY PREDICTION IN MEC NETWORKS Zohreh HajiAkhondi-Meybodiy Arash Mohammadiz Ming Houyz

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: