Citation Trajectory Prediction via Publication Inuence Representation Using Temporal Knowledge Graph Chang ZongYueting ZhuangWeiming LuJian ShaoSiliang Tang

2025-04-27 0 0 1.01MB 9 页 10玖币

侵权投诉

Citation Trajectory Prediction via Publication Inﬂuence Representation Using

Temporal Knowledge Graph

Chang Zong∗Yueting Zhuang∗†Weiming Lu∗Jian Shao∗Siliang Tang∗

Abstract

Predicting the impact of publications in science and technol-

ogy has become an important research area, which is useful

in various real world scenarios such as technology invest-

ment, research direction selection, and technology policy-

making. Citation trajectory prediction is one of the most

popular tasks in this area. Existing approaches mainly rely

on mining temporal and graph data from academic articles.

Some recent methods are capable of handling cold-start pre-

diction by aggregating metadata features of new publica-

tions. However, the implicit factors causing citations and

the richer information from handling temporal and attribute

features still need to be explored. In this paper, we propose

CTPIR, a new citation trajectory prediction framework that

is able to represent the inﬂuence (the momentum of citation)

of either new or existing publications using the history infor-

mation of all their attributes. Our framework is composed of

three modules: diﬀerence-preserved graph embedding, ﬁne-

grained inﬂuence representation, and learning-based trajec-

tory calculation. To test the eﬀectiveness of our framework

in more situations, we collect and construct a new tempo-

ral knowledge graph dataset from the real world, named

AIPatent, which stems from global patents in the ﬁeld of

artiﬁcial intelligence. Experiments are conducted on both

the APS academic dataset and our contributed AIPatent

dataset. The results demonstrate the strengths of our ap-

proach in the citation trajectory prediction task.

1 Introduction

Distinguishing high-impact publications is crucial to

making decisions in business and research activities,

such as investment in technology ﬁelds, selection of re-

search topics, and development policymaking. Citations

of a publication are usually applied to evaluate its po-

tential impact. The question of how to predict citations

has attracted more attention in recent years. With the

development of knowledge graph technologies, the key

issue is how to utilize the information provided by a

∗Zhejiang University, Hangzhou. zongchang@zju.edu.cn,

yzhuang@zju.edu.cn, luwm@zju.edu.cn, jshao@zju.edu.cn, sil-

iang@zju.edu.cn

†Corresponding author.

knowledge graph to predict a publication’s future cita-

tion trend.

Existing methods on this problem can be summa-

rized in three ways. The ﬁrst approach [33, 27] tries

to make use of prior knowledge and network techniques

by assuming that citation trajectories obey the Power

Law or log-normal functions. Traditional statistical

methods are applied to make predictions. Another way

[14, 1] focuses on taking advantage of text features of

abstracts and reviews. Features are fed into recurrent

neural network (RNN) models for time-series predic-

tions. With the increasing popularity of graph neural

networks (GNNs), recent works attempt to apply var-

ious structural learning models to exploit information

from attributes of publications [6, 31, 9, 13, 8].

authors keywords institutes

citation citation

Figure 1: A diagram illustrating that citations of

publications can be aﬀected by their attributes and

relation types in diﬀerent levels.

However, citations are aﬀected by many potential

factors, and there is a lot of implicit information that

must be considered in practice. For example, attributes

of a publication, such as authors and keywords, should

be treated signiﬁcantly. The reputation of a scholar

and the popularity of a ﬁeld can greatly aﬀect future

citations of a publication. In addition, each attribute

contributes to a publication at diﬀerent levels (Figure

1). The approaches in previous works simply apply

GNNs to aggregate attribute features, which leads to

a lack of ﬁne-grained inﬂuence expression. On the basis

of the above knowledge, a more powerful framework

for predicting citation trajectories with the inﬂuence of

arXiv:2210.00450v1 [cs.AI] 2 Oct 2022

publications derived from a temporal knowledge graph

is needed to handle the problem: How to represent

and calculate the inﬂuence of a publication using

as much of its information as possible?

Current studies on temporal knowledge graphs try

to manage changes in two adjacent snapshots, assuming

that nodes should update smoothly or evolve dramati-

cally [7, 26]. However, these assumptions require one to

manually set a change rate to limit the evolution, which

is not ﬂexible. Existing works still mainly focus on han-

dling structural and temporal features in separate steps,

which leads to a lack of expression to treat dynamic

graphs as a whole. Furthermore, accumulative citations

are usually modeled as log-normal or cumulative dis-

tribution functions [9, 2]. It is still worth trying some

alternatives to perform a further analysis. The poten-

tial enhancement mentioned above should be studied to

handle another problem: How can we improve the

expressiveness of the framework for prediction

tasks using temporal knowledge graphs?

With the observations above, we propose CT-

PIR (Citation Trajectory Prediction via Inﬂuence

Representation), a new framework to predict citation

trajectories with inﬂuence representation using tempo-

ral knowledge graphs. First, we optimize the R-GCN

mechanism [22] to automatically learn the gaps between

two adjacent snapshots. Second, we implement a ﬁne-

grained inﬂuence (citation momentum) representation

module to make use of all historical information from a

publication’s attributes. Third, a learnable general lo-

gistic function is applied to ﬁt the trajectories using the

inﬂuence representation from the previous module.

We experiment our framework with two real world

datasets. One is APS1, a public dataset of academic

papers. Another, named AIPatent, is a new dataset

that we construct with global patents in the ﬁeld of

artiﬁcial intelligence. Compared to some baselines, the

results show that CTPIR outperforms those methods in

all cases.

Our key contributions are summarized in the fol-

lowing points:

•Novel framework: We propose a new framework,

named CTPIR, which implements a ﬁne-grained in-

ﬂuence representation approach using a more ex-

pressive temporal graph learning process and opti-

mizes existing methods to bring prediction results

much closer to observations.

•Improved evaluation: We construct a new tem-

poral knowledge graph dataset named AIPatent for

the task, which is also a strong supplement for the

1https://journals.aps.org/datasets

community to carry out various temporal graph

studies. We also design and implement multiple

subtasks to evaluate approaches from a more com-

prehensive view.

•Multifaceted analysis: We analyze the experi-

mental performance from multiple aspects. Expla-

nations on how CTPIR performs better compared

to other recent approaches are discussed. Some

weaknesses and further eﬀorts are also mentioned

to guide future studies.

The dataset we use in this work, including our

AIPatent contributed dataset, and the code to re-

produce are available in our GitHub repository:

https://github.com/changzong/CTPIR

2 Related Work

2.1 Citation and Popularity Prediction. Mod-

ern approaches to citation count prediction (CCP) aim

to combine attribute information with temporal fea-

tures. GNNs are commonly used to capture topolog-

ical features of citation networks. The encoded nodes

are sent to RNNs or attention models for time-series

forecasting. A previous work [8] follows this simple

encoder-decoder architecture. Some previous studies

[30, 13, 25, 32] put emphasis on cascade graphs for pop-

ularity prediction, using a kernel method to estimate

structural similarities. These works are based on sim-

ply combining graph embedding with time-series meth-

ods. In contrast, we introduce a method to fully utilize

all past characteristics of a publication’s attributes. A

recent work called HINTS [9] adds an imputation mod-

ule to aggregate the information from each snapshot of

graphs. Another work proposes a heterogeneous dynam-

ical graph neural network (HDGNN) [29] to predict the

cumulative impact of articles and authors. The latest

work [24] uses an attention mechanism to represent the

sequence of content from citation relations. Although

these works can take advantage of richer information,

their lack of ﬁne-grained design to represent the inﬂu-

ence of a publication is not conducive to achieving good

prediction performance.

2.2 Temporal Graph Embedding. We focus on

deep learning-based temporal graph embedding ap-

proaches. Several previous works implement a straight-

forward way to combine GCN and RNN models to ex-

tract structural and temporal features [18, 23, 5]. RNN

variants are applied as the temporal module to perform

downstream tasks such as anomaly detection. Mean-

while, temporal attention models can be a substitute

for GCN to extract topological features [28, 15, 17]. A

recent paper [3] tries to represent global structural in-

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：

CitationTrajectoryPredictionviaPublicationInuenceRepresentationUsingTemporalKnowledgeGraphChangZong*YuetingZhuangWeimingLuJianShaoSiliangTangAbstractPredictingtheimpactofpublicationsinscienceandtechnol-ogyhasbecomeanimportantresearcharea,whichisusefulinvariousrealworldscenariossuchastechnologyi...

展开>> 收起<<

Citation Trajectory Prediction via Publication Inuence Representation Using Temporal Knowledge Graph Chang ZongYueting ZhuangWeiming LuJian ShaoSiliang Tang.pdf

共9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！

Citation Trajectory Prediction via Publication Inuence Representation Using Temporal Knowledge Graph Chang ZongYueting ZhuangWeiming LuJian ShaoSiliang Tang

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: