tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
best result is in bold and the second best is underlined
Type Method Cloth Sports Economics Mathematics Geology
LM Embedings GNNBERTGNN 06362 06727 01822 02693 03667
BERTFTGNN 06457 06804 01826 02861 03633
GIANTGNN 06471 07201 01651 02421 03120
PattonGNN 06831 07015 01664 02599 03375
OPTGNN 06484 07255 01747 02713 03660
LM GNN Iter Training GLEM 06743 07413 02286 02834 04100
LM FinetuningOPT 06541 06005 02324 02889 04206
OPTNConcat 06798 06728 02372 02903 04218
Ours HiComOPT 06882 07524 02463 03062 04464
node text using different LMs Then we feed these LMencoded fea
tures into a GraphSAGE model 14 for node classification The LMs
we consider include nonfinetuned BERT 11 finetuned BERT on
the same node classification task 19 nonfinetuned OPT 50 GI
ANT 7 an XRTransformer 47 after selfsupervised finetuning
with graph information and Patton 20 For Patton we use its
available finetuned model checkpoints on the MAPLE and Amazon
Products These models are denoted as BERTGNN BERTFTGNN
OPTGNN GIANTGNN and PattonGNN Second we consider
theGLEM 53 method that iteratively trains a GNN and an LM
Third we directly finetune LLMs on the node text for classifica
tion We also leverage graph context by dynamically sampling and
concatenating neighbor text to the node text which we refer to as
LLM neighbor concatenation NConcat The LLM we use for this
type of approach is the pretrained OPT model which is the same
LLM used in our proposed HiCom framework and the baselines
are denoted as OPT and OPTNConcat
53 Experiment Settings
Data processing Our main goal is to evaluate the challenging
cases where nodes have rich neighborhood information analogous
to long documents in NLP research Therefore we consider two
different experiment settings of where the nodes can be sampled
from dense regions andall regions For the dense regions we mimic
the challenging long document setting to focus on nodes from the
dense region of a graph where nodes possess rich neighborhood
information In particular we consider nodes belonging to the k
core graph In the literature the Amazon product graphs are usually
evaluated by taking their 5core To make the case more challenging
we consider 8core for the dense region setting For all regions
nodes are sampled uniformly randomly which is more inclusive
but not the best setting to test models ability to handle rich inputs
Data splits We randomly pick 20 nodes per class as the training
set for each graph and sample 1000 nodes from the rest as the
validation set and sample up to 10000 nodes from the rest as the
test set We evaluate experiments with the F1 score considering
the imbalance We set the number of summary vectors 𝑘 to be 50
Given the difference in graph characteristics we experiment with a
few different hyperparameters eg fanouts to make sure the modelproduces the best result We do such hyperparameter tuning for
our method as well as the baselines In Table 3 we report the test
results with the best hyperparameter selected on the validation set
The full results of all methods with different hyperparameters can
be found in Table 6 in Appendix A4 Details of hyperparameters
and implementation are shown in Appendix A3
54 Main Results
The node classification results for the dense regions are shown in
Table 3 We see that our HiComOPT outperforms all baselines on
all datasets We also make some observations and provide related
discussions below
GNN as Backbone vs LM as Backbone Our first observation is
based on the comparison of models from two different categories
eg GNN as the backbone the first multirow LM embeddings
GNN and LM as the backbone the last two multirows LM
finetuning and ours We exclude GLEM from this discussion as it
involves in finetuning both models for predictions We observe that
all three LMbased methods can provide reasonable performance on
these graph datasets outperforming most GNNs on most datasets
especially for the MAPLE graphs This observation connects to the
reasons we discussed in Section 43 First LMs can directly take raw
text as input and model tokenlevel interactions through attention
whereas GNNs can only work with encoded embeddings and only
model nodelevel interactions Secondly The LMpreserved knowl
edge from pretraining can be utilized for context understanding
and prediction whereas GNNs do not have preserved knowledge
The second point is more significant on the MAPLE graphs as these
graphs contain academic papers as nodes which are likely included
in the pretrain data of the LM In contrast the Amazon data which
consists of diverse and less standardized item descriptions is less
likely to be included in the training data might be partially included
via the Common Crawl data As a result finetuning only on the
text of the target nodes OPT has good performance that is close
to using neighbor text OPTNConcat
LM Finetuning vs HiCom Our second observation is the mono
tone performance improvement from finetuning pure LM OPT
to LM Neighbors OPTNConcat to HiCom HiComOPT This
trend is observed on all datasets for both the dense regions Table 3Conference acronym XX June 0305 2018 Woodstock NY Zhang et al
and all regions Table 7 The improvement from LM to LM Neigh
bors is intuitive because more information are added for the LM
to process to help with the prediction However doing neighbor
concatenation also has two bottlenecks One is that the number of
neighbors that can be included is low 4 for a 2048 window size
of OPT with 512 tokens per node The other is that the hierarchy
in the graph structure is not modeled due to simple concatena
tion These are exactly the two bottlenecks that HiCom overcomes
With many more multihop incorporated following a hierarchy
the HiCom performance further improves over finetuning LM
Neighbors
Dense Regions vs All Regions Our results in Table 3 demonstrate
that HiComOPT can outperform baselines for nodes in dense re
gions especially on graphs with high average degrees We also

best result is in bold and the second best is underlined

Type Method Cloth Sports Economics Mathematics Geology

LM Embedings GNNBERTGNN 06362 06727 01822 02693 03667

BERTFTGNN 06457 06804 01826 02861 03633

GIANTGNN 06471 07201 01651 02421 03120

PattonGNN 06831 07015 01664 02599 03375

OPTGNN 06484 07255 01747 02713 03660

LM GNN Iter Training GLEM 06743 07413 02286 02834 04100

LM FinetuningOPT 06541 06005 02324 02889 04206

OPTNConcat 06798 06728 02372 02903 04218

Ours HiComOPT 06882 07524 02463 03062 04464

node text using different LMs Then we feed these LMencoded fea

tures into a GraphSAGE model 14 for node classification The LMs

we consider include nonfinetuned BERT 11 finetuned BERT on

the same node classification task 19 nonfinetuned OPT 50 GI

ANT 7 an XRTransformer 47 after selfsupervised finetuning

with graph information and Patton 20 For Patton we use its

available finetuned model checkpoints on the MAPLE and Amazon

Products These models are denoted as BERTGNN BERTFTGNN

OPTGNN GIANTGNN and PattonGNN Second we consider

theGLEM 53 method that iteratively trains a GNN and an LM

Third we directly finetune LLMs on the node text for classifica

tion We also leverage graph context by dynamically sampling and

concatenating neighbor text to the node text which we refer to as

LLM neighbor concatenation NConcat The LLM we use for this

type of approach is the pretrained OPT model which is the same

LLM used in our proposed HiCom framework and the baselines

are denoted as OPT and OPTNConcat

53 Experiment Settings

Data processing Our main goal is to evaluate the challenging

cases where nodes have rich neighborhood information analogous

to long documents in NLP research Therefore we consider two

different experiment settings of where the nodes can be sampled

from dense regions andall regions For the dense regions we mimic

the challenging long document setting to focus on nodes from the

dense region of a graph where nodes possess rich neighborhood

information In particular we consider nodes belonging to the k

core graph In the literature the Amazon product graphs are usually

evaluated by taking their 5core To make the case more challenging

we consider 8core for the dense region setting For all regions

nodes are sampled uniformly randomly which is more inclusive

but not the best setting to test models ability to handle rich inputs

Data splits We randomly pick 20 nodes per class as the training

set for each graph and sample 1000 nodes from the rest as the

validation set and sample up to 10000 nodes from the rest as the

test set We evaluate experiments with the F1 score considering

the imbalance We set the number of summary vectors 𝑘 to be 50

Given the difference in graph characteristics we experiment with a

few different hyperparameters eg fanouts to make sure the modelproduces the best result We do such hyperparameter tuning for

our method as well as the baselines In Table 3 we report the test

results with the best hyperparameter selected on the validation set

The full results of all methods with different hyperparameters can

be found in Table 6 in Appendix A4 Details of hyperparameters

and implementation are shown in Appendix A3

54 Main Results

The node classification results for the dense regions are shown in

Table 3 We see that our HiComOPT outperforms all baselines on

all datasets We also make some observations and provide related

discussions below

GNN as Backbone vs LM as Backbone Our first observation is

based on the comparison of models from two different categories

eg GNN as the backbone the first multirow LM embeddings

GNN and LM as the backbone the last two multirows LM

finetuning and ours We exclude GLEM from this discussion as it

involves in finetuning both models for predictions We observe that

all three LMbased methods can provide reasonable performance on

these graph datasets outperforming most GNNs on most datasets

especially for the MAPLE graphs This observation connects to the

reasons we discussed in Section 43 First LMs can directly take raw

text as input and model tokenlevel interactions through attention

whereas GNNs can only work with encoded embeddings and only

model nodelevel interactions Secondly The LMpreserved knowl

edge from pretraining can be utilized for context understanding

and prediction whereas GNNs do not have preserved knowledge

The second point is more significant on the MAPLE graphs as these

graphs contain academic papers as nodes which are likely included

in the pretrain data of the LM In contrast the Amazon data which

consists of diverse and less standardized item descriptions is less

likely to be included in the training data might be partially included

via the Common Crawl data As a result finetuning only on the

text of the target nodes OPT has good performance that is close

to using neighbor text OPTNConcat

LM Finetuning vs HiCom Our second observation is the mono

tone performance improvement from finetuning pure LM OPT

to LM Neighbors OPTNConcat to HiCom HiComOPT This

trend is observed on all datasets for both the dense regions Table 3Conference acronym XX June 0305 2018 Woodstock NY Zhang et al

and all regions Table 7 The improvement from LM to LM Neigh

bors is intuitive because more information are added for the LM

to process to help with the prediction However doing neighbor

concatenation also has two bottlenecks One is that the number of

neighbors that can be included is low 4 for a 2048 window size

of OPT with 512 tokens per node The other is that the hierarchy

in the graph structure is not modeled due to simple concatena

tion These are exactly the two bottlenecks that HiCom overcomes

With many more multihop incorporated following a hierarchy

the HiCom performance further improves over finetuning LM

Neighbors

Dense Regions vs All Regions Our results in Table 3 demonstrate

that HiComOPT can outperform baselines for nodes in dense re

gions especially on graphs with high average degrees We also