Corresponding author Yawen Li warmly0716126.com. Adaptive Dual Channel Convolution H yper graph Representation Learning for Technological Intellectual Property

2025-04-26 0 0 652.81KB 8 页 10玖币

*Corresponding author: Yawen Li (warmly0716@126.com). 
Adaptive Dual Channel Convolution Hypergraph 
Representation Learning for Technological Intellectual 
Property 
Yuxin Liu1, Yawen Li2*, Yingxia Shao1, Zeli Guan1 
1Beijing Key Laboratory of Intelligent Communication Software and Multimedia, School of Computer Science 
(National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876 
2School of Economics and Management, Beijing University of Posts and Telecommunications, Beijing 100876
Abstract: In the age of big data, the demand for hidden 
information mining in technological intellectual property 
is  increasing  in  discrete  countries.  Definitely,  a 
considerable  number  of  graph  learning  algorithms  for 
technological intellectual property have been proposed. 
The  goal  is  to  model  the  technological  intellectual 
property entities and their relationships through the graph 
structure and use the neural network algorithm to extract 
the hidden structure information in the graph. However, 
most  of  the  existing  graph  learning  algorithms  merely 
focus  on  the  information  mining  of  binary relations  in 
technological intellectual property, ignoring the higher-
order  information  hidden  in  non-binary  relations. 
Therefore, a hypergraph neural network model based on 
dual channel convolution is proposed. For the hypergraph 
constructed from technological intellectual property data, 
the  hypergraph  channel  and  the  line  expanded  graph 
channel  of  the  hypergraph  are  used  to  learn  the 
hypergraph, and the attention mechanism is introduced to 
adaptively  fuse  the  output  representations  of  the  two 
channels. The proposed model outperforms the existing 
approaches on a variety of datasets. 
Keywords:  Graph  Convolution;  Hypergraph;  Line 
Expansion; Attention; Representation Fusion 
1 Introduction 
Technological  intellectual  property  data  is  generally 
composed of complex information like creative entities, 
resource  entities,  inter-entity  relationships,  and  entity 
attributes[33].  Mainstream  models  usually  need  to  map 
data  into  a  vector  space  that  is  convenient  for 
calculation[20][23],[27], filter and fuse data in the limited 
space[39]-[43],  and  cannot  ignore  the  correlation 
information of the data itself. The purpose of this paper is 
to propose a hypergraph learning algorithm to represent 
the hypergraph modeled the complex information about 
technological  intellectual  property  data  for  further 
analysis and mining. 
In  terms  of  the  modelling  technological  intellectual 
property data, existing methods include a homogeneous 
[30],[32]  and  a  heterogeneous  graph  [25][26],[31]. 
However,  homogeneous  and  heterogeneous  graphs 
merely model binary relationships, which are not able to 
model  non-binary  relationships  in  technological 
intellectual property data. Consequently, graph learning 
algorithms  based  on  a  homogenous  graph  and  a 
heterogeneous  graph  are  unable  to  represent  nonlinear 
higher-order correlation among technological intellectual 
property entities. Some researchers have proposed to use 
multi-agent to learn multiple features and solve the above 
problems through feature fusion [12-15], but this method 
requires  a  lot  of  artificial  construction  work.  Instead, 
hypergraph [29] has been used as a modeling method of 
non-binary relations due to its structural characteristics, 
such  as  the  research  [10],  [46]  on natural  language 
processing,  the  research  [12],[34]-[36]  on 
recommendation systems, the research [11],[45] on image 
classification, and the research on various social networks 
[38],[44].  Accordingly,  using  hypergraph  to  model  the 
entities  and  relationships  of  technological  intellectual 
property data is able to fully express its complexity. 
For  hypergraph  learning,  traditional  algorithms  are 
typically separated into two categories: one is based on 
incidence matrix decomposition, and the other is based on 
hypergraph  expansion  [28],[37],  which  converts 
hypergraphs  into  weighted  simple  graphs.  Neither  the 
performance nor the efficiency of traditional methods can 
be compared with neural networks, so this paper mainly 
discusses  hypergraph  neural  network  algorithms.  The 
neural  network  algorithms  include  hypergraph  neural 
networks based on hypergraph expansion and hypergraph 
neural  networks  based  on  non-hypergraph  expansion. 
Methods  [1][3]  based  on  hypergraph  expansion 
commonly use simple graph learning methods such as the 
Graph  Convolution  Network  (GCN)  to  represent  the 
converted simple weighted graph from the hypergraph. 
However, it would cause the loss of some information for 
the same reason that it simplifies the hypergraph into a 
simple weighted graph. The methods [5][7] based on non-
hypergraph  expansion  are  to  carry  out  hypergraph 
learning through hypergraph convolution neural networks 
or hypergraph attention neural networks. Some of these 
methods are  merely appropriate for learning k-uniform 
graphs with k hypernodes on each hyperedge but not for 
the  technological  intellectual  property  data  that 
hypergraph structure is a k-uniform graph. The others [8] 
integrate  the  features  of  hypergraph  and  simple  graph 
expanded from hypergraph. Nevertheless, the method of 
fusion is the average of the representation of two graphs. 
In  view  of  the  problems  with  the  above  methods,  we 
propose  an  adaptive  dual-channel  hypergraph  node 

 
representation  algorithm  based  on  hypergraph 
convolution and line expansion (LE) of the hypergraph. 
Furthermore, it uses the node classification task to verify 
the method. Firstly, we construct a hypergraph based on 
some technological intellectual property data by taking 
the author as the hyperedge and his publications as the 
hypernodes. Secondly, using the method proposed by [3] 
expands  the  hypergraph  to  a  weighted  simple  graph. 
Thirdly, the hypergraph and the graph of LE pass through 
the  hypergraph  convolution  channel  and  the  graph 
convolution channel, respectively. Finally, the final node 
representation  is  obtained  by  using  the  attention 
mechanism to adaptively fuse the two representations of 
dual  channels.  Furthermore,  we  use  the  node 
classification  task  to  determine  the  effect  of  node 
representation. 
The main contributions of this paper are as follows: 
1)  We  propose  an  adaptive  hypergraph  dual  channel 
convolution  node  representation  learning  method 
(ADHCN). The hypergraph line expansion convolution 
channel and the hypergraph convolution channel are used 
to represent nodes, so as to avoid the information loss in 
the  simple  graph,  which  is  line  expanded  from  the 
hypergraph. 
2)  We  introduce  the  attention  mechanism  to  fuse  the 
representations of the two channels to avoid the deviation 
caused by direct averaging or splicing.   
3)  The  experimental  results  demonstrate  that  ADHCN 
exceeds the compared models in terms of accuracy, F1, 
and recall. Moreover, the experiment shows that the effect 
of the model is significantly improved after introducing 
the attention mechanism for adaptive fusion. 
2 Related Work 
A  hypergraph     is  a  generalized  graph, 
where    is the set of hypernodes,  
   is  the  set  of  hyperedges,  and  
  is the weights set on each hyperedge. In 
particular,  each  hyperedge     connects  more  than  two 
hypernodes.  Moreover,  the  incidence  matrix  of  the 
hypergraph  is   ,  where 
  only takes 0 or 1. In particular, there are  more than 
two 1 in the column of  , which means one hyperedge 
connects several vertices. 
For the methods in hypergraph neural networks based on 
expansion, Yadati et al. [1] proposed the hyperGCN model, 
whose  principle  is  to  use  the  GCN  training  graph 
transformed  from  the  hypergraph  through  the  spectral 
theory and introduce the  mediator to prevent the loss of 
the  hypergraph’s  information.  Bandyopadhyay  [2] 
simplified  the  hypergraph  into  a  normal  graph  named 
“line graph” by taking the hyperedge as the vertex and the 
Jaccard similarity between the hyperedges as the weight, 
and  then  trained  the  graph  with  GCN.  Yang  et  al.  [3] 
proposed  the  line  expansion  (LE)  of  hypergraphs  to 
transform  the  hypergraph  into  a  weighted simple  graph 
. The    is the node set which is composed 
of  hyperedge-hypernode  pairs   
from the original hypergraph. The edge of the edge set   
connects  two  pairs  if  both  pairs  have  a  common 
hypernode  or  hyperedge.  The  adjacency  matrix  of  the 
weighted  graph     shown in  Eq.1 
is defined by the pairwise relation    and  . 








   (1) 
The  above  algorithms  reduce  the  hypergraph  learning 
problem  with  simple  graph  learning,  which  brings 
problems  of  hypergraph  information  loss.  For  the 
methods  in  hypergraph  neural  networks  based  on  non-
expansion,  Feng  et  al.  [5]  proposed  a  hypergraph 
convolution  neural  network  by  extending  graph 
convolution to hypergraph. And the proposed method is 
used  to  learn  the  hypergraph  constructed  by  k-nearest 
neighbor  (KNN).  Owing  to  KNN  being  the  method  of 
constructing  hypergraphs,  this  model  is  merely 
appropriate  for  k-uniform  hypergraphs.  In  order  to 
address  the  issue  that  the  hypergraph  structure  in  [5] 
changes as a result of the alteration in node representation. 
Jiang  et  al.  [6]  proposed  a  dynamic  hypergraph  neural 
network  (DHGNN)  that  contains  dynamic  hypergraph 
reconstruction  that  reconstructs  the  hypergraph  at  each 
layer  and  dynamic  graph  convolution  that  gathers  the 
information of nodes and edges. However, the method is 
incapable of solving the k-uniform graph problem. Bai et 
al.  [7]  proposed  a  hypergraph  convolution  theory  from 
another  perspective,  which  is  consistent  with  the 
hypergraph  convolution  proposed  by  [5].  Besides,  it 
introduced  hypergraph  attention  to  enhance  the 
expression ability. Xia et al. [8] proposed a dual channel 
hypergraph convolution network to solve the problem of 
session-based  recommendation.  In particular, the  model 
uses hypergraph convolution and line graph convolution 
to learn hypergraph and fuse the learning result. In spite 
of that, there are still some problems in the model, such 
as  information  loss  in  the  line  graph  channel  and  the 
simplicity of the fusion method. Yu et al. [9] proposed a 
multi-channel  hypergraph  convolutional  network  for 
social  recommendation.  The  model  builds  three 
hypergraphs  and  uses  three  hypergraph  convolutions  to 
represent  these  graphs.  However,  the  method  can’t  be 
used in the technological intellectual property dataset. 
3  ADHCN:  Adaptive  Dual  Channel 
Hypergraph Convolution Network  
Figure 1 shows the model structure proposed in this paper. 
It  mainly  includes  the  hypergraph  construction  and  the 
line  expansion  (HCLE)  module,  the  dual  channel 
convolution  (DHC)  module,  the adaptive  representation 
fusion (ARF) module, and the model verification module. 
Firstly, the dataset is constructed into a hypergraph H and 
a  line-expanded  simple  graph     through  the  HCLE 
module.  Moreover,  the  initial  representation  of  the  two 

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：

*Correspondingauthor:YawenLi(warmly0716@126.com).AdaptiveDualChannelConvolutionHypergraphRepresentationLearningforTechnologicalIntellectualPropertyYuxinLiu1,YawenLi2*,YingxiaShao1,ZeliGuan11BeijingKeyLaboratoryofIntelligentCommunicationSoftwareandMultimedia,SchoolofComputerScience(NationalPilotSoftw...

展开>> 收起<<

Corresponding author Yawen Li warmly0716126.com. Adaptive Dual Channel Convolution H yper graph Representation Learning for Technological Intellectual Property.pdf

共8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！

Corresponding author Yawen Li warmly0716126.com. Adaptive Dual Channel Convolution H yper graph Representation Learning for Technological Intellectual Property

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: