tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
testing We conduct experiments with fanouts equal to 44 for all
the GNNs and HiComOPT Notice that the absolute performance
results are not directly comparable to results in the dense regions
as the trainingvalidationtesting nodes are all different We show
the relative performance improvement by HiComOPT for nodes in
dense regions vs all regions in Figure 3 Here we report the detailed
performance results on nodes in all regions in Table 7
A6 Experiments for Increasing Training Data
In Section 55 we discussed the results of an ablation study of in
creasing the training data We now describe the detailed experiment
setting Given the imbalanced label distribution in these real graphs
some classes have limited labeled samples making it unfeasible to
increase the number of training samples per class Consequently
we have adopted a strategy of uniformly random sampling from
the unused data to augment the training set By unused data we
mean the data was not included in either train valid or test set
in the main experiment The label distribution of the additional
training data thus should roughly follow the label distribution over
the entire graph We conduct experiments with fanouts equal to
44 for the OPTGNN and HiComOPTConference acronym XX June 0305 2018 Woodstock NY Zhang et al
Table 6 Node classification F1 score on Amazon product graphs and MAPLE citation graphs detailed
Method Cloth Sports Economics Mathematics Geology
BERTGNN 82 06136 06727 01822 02693 03667
BERTGNN 44 06362 06546 01730 02685 03448
BERTGNN 28 06262 06503 01705 02674 03405
BERTFTGNN 82 06432 06804 01656 02861 03528
BERTFTGNN 44 06457 06801 01826 02677 03633
BERTFTGNN 28 06455 06804 01763 02712 03529
GIANTGNN 82 06267 07201 01643 02421 03074
GIANTGNN 44 06159 07065 01651 02366 03120
GIANTGNN 28 06471 06997 01599 02334 03086
OPTGNN 82 06484 07255 01694 02634 03330
OPTGNN 44 06378 06796 01747 02713 03497
OPTGNN 28 06417 06607 01593 02647 03660
OPT 06541 06005 02324 02889 04206
OPTNConcat 06798 06728 02372 02903 04218
HiComOPT 82 06882 06812 02463 03060 04464
HiComOPT 44 06622 07524 02341 02993 04339
HiComOPT 28 06009 07354 02331 03062 04185
Table 7 Node classification F1 score on Amazon product graphs and MAPLE citation graphs all regions
Method Cloth Sports Economics Mathematics Geology
BERTGNN 05807 06054 01866 02353 03890
GIANTGNN 06549 06461 01726 02326 03662
OPTGNN 06438 06869 01707 02794 03949
OPT 06537 06553 03055 03072 04654
OPTNConcat 06817 07154 03062 03131 04731
HiComOPT 06849 07183 03118 03243 04737
Figure 5 Method performance on the Sports dataset with the
training set in different sizes
We have shown the F1score vs training set size plot for the
MAPLE graph Geology in Section 55 Here we show a similar plot
for an Amazon graph eg Sports in Figure 5A7 Future Work
The HiCom framework presents several intriguing avenues for fu
ture research and enhancement A natural progression is to explore
the use of more advanced LLMs such as LLaMA 36 which offer
superior textunderstanding capabilities However these models
typically feature an increased number of parameters generally
7𝐵 necessitating substantial hardware support and optimization
for efficient use within the HiCom framework Further exploration
in more tasks like link prediction or content generation based on
graph structures is also exciting Especially the generation would
involve leveraging the compression and contextual understand
ing capabilities of HiCom to generate coherent and contextually
relevant textual content opening new possibilities in automated
content creation The adaptability and effectiveness of HiCom in
handling textrich graph data promise to unlock new potentials in
the realm of graphbased learning and LLM applications
Received 20 February 2007 revised 12 March 2009 accepted 5 June 2009MultiStage Balanced Distillation Addressing LongTail Challenges in
SequenceLevel Knowledge Distillation
Yuhang Zhou1Jing Zhu2Paiheng Xu1Xiaoyu Liu1Xiyao Wang1
Danai Koutra2Wei Ai1Furong Huang1
1University of Maryland College Park
2University of Michigan Ann Arbor
tonyzhou paiheng xliu1231 xywang aiwei furonghumdedu
jingzhuu dkoutraumichedu
Abstract
Large language models LLMs have signifi
cantly advanced various natural language pro
cessing tasks but deploying them remains com
putationally expensive Knowledge distillation
KD is a promising solution enabling the trans
fer of capabilities from larger teacher LLMs
to more compact student models Particularly
sequencelevel KD which distills rationale
based reasoning processes instead of merely
final outcomes shows great potential in en
hancing students reasoning capabilities How
ever current methods struggle with sequence
level KD under longtailed data distributions
adversely affecting generalization on sparsely
represented domains We introduce the Multi
Stage Balanced Distillation BalDistill frame
work which iteratively balances training data
within a fixed computational budget By dy
namically selecting representative head domain
examples and synthesizing tail domain exam
ples BalDistill achieves stateoftheart perfor
mance across diverse longtailed datasets en
hancing both the efficiency and efficacy of the
distilled models

testing We conduct experiments with fanouts equal to 44 for all

the GNNs and HiComOPT Notice that the absolute performance

results are not directly comparable to results in the dense regions

as the trainingvalidationtesting nodes are all different We show

the relative performance improvement by HiComOPT for nodes in

dense regions vs all regions in Figure 3 Here we report the detailed

performance results on nodes in all regions in Table 7

A6 Experiments for Increasing Training Data

In Section 55 we discussed the results of an ablation study of in

creasing the training data We now describe the detailed experiment

setting Given the imbalanced label distribution in these real graphs

some classes have limited labeled samples making it unfeasible to

increase the number of training samples per class Consequently

we have adopted a strategy of uniformly random sampling from

the unused data to augment the training set By unused data we

mean the data was not included in either train valid or test set

in the main experiment The label distribution of the additional

training data thus should roughly follow the label distribution over

the entire graph We conduct experiments with fanouts equal to

44 for the OPTGNN and HiComOPTConference acronym XX June 0305 2018 Woodstock NY Zhang et al

Table 6 Node classification F1 score on Amazon product graphs and MAPLE citation graphs detailed

Method Cloth Sports Economics Mathematics Geology

BERTGNN 82 06136 06727 01822 02693 03667

BERTGNN 44 06362 06546 01730 02685 03448

BERTGNN 28 06262 06503 01705 02674 03405

BERTFTGNN 82 06432 06804 01656 02861 03528

BERTFTGNN 44 06457 06801 01826 02677 03633

BERTFTGNN 28 06455 06804 01763 02712 03529

GIANTGNN 82 06267 07201 01643 02421 03074

GIANTGNN 44 06159 07065 01651 02366 03120

GIANTGNN 28 06471 06997 01599 02334 03086

OPTGNN 82 06484 07255 01694 02634 03330

OPTGNN 44 06378 06796 01747 02713 03497

OPTGNN 28 06417 06607 01593 02647 03660

OPT 06541 06005 02324 02889 04206

OPTNConcat 06798 06728 02372 02903 04218

HiComOPT 82 06882 06812 02463 03060 04464

HiComOPT 44 06622 07524 02341 02993 04339

HiComOPT 28 06009 07354 02331 03062 04185

Table 7 Node classification F1 score on Amazon product graphs and MAPLE citation graphs all regions

Method Cloth Sports Economics Mathematics Geology

BERTGNN 05807 06054 01866 02353 03890

GIANTGNN 06549 06461 01726 02326 03662

OPTGNN 06438 06869 01707 02794 03949

OPT 06537 06553 03055 03072 04654

OPTNConcat 06817 07154 03062 03131 04731

HiComOPT 06849 07183 03118 03243 04737

Figure 5 Method performance on the Sports dataset with the

training set in different sizes

We have shown the F1score vs training set size plot for the

MAPLE graph Geology in Section 55 Here we show a similar plot

for an Amazon graph eg Sports in Figure 5A7 Future Work

The HiCom framework presents several intriguing avenues for fu

ture research and enhancement A natural progression is to explore

the use of more advanced LLMs such as LLaMA 36 which offer

superior textunderstanding capabilities However these models

typically feature an increased number of parameters generally

7𝐵 necessitating substantial hardware support and optimization

for efficient use within the HiCom framework Further exploration

in more tasks like link prediction or content generation based on

graph structures is also exciting Especially the generation would

involve leveraging the compression and contextual understand

ing capabilities of HiCom to generate coherent and contextually

relevant textual content opening new possibilities in automated

content creation The adaptability and effectiveness of HiCom in

handling textrich graph data promise to unlock new potentials in

the realm of graphbased learning and LLM applications

Received 20 February 2007 revised 12 March 2009 accepted 5 June 2009MultiStage Balanced Distillation Addressing LongTail Challenges in

SequenceLevel Knowledge Distillation

Yuhang Zhou1Jing Zhu2Paiheng Xu1Xiaoyu Liu1Xiyao Wang1

Danai Koutra2Wei Ai1Furong Huang1

1University of Maryland College Park

2University of Michigan Ann Arbor

tonyzhou paiheng xliu1231 xywang aiwei furonghumdedu

jingzhuu dkoutraumichedu

Abstract

Large language models LLMs have signifi

cantly advanced various natural language pro

cessing tasks but deploying them remains com

putationally expensive Knowledge distillation

KD is a promising solution enabling the trans

fer of capabilities from larger teacher LLMs

to more compact student models Particularly

sequencelevel KD which distills rationale

based reasoning processes instead of merely

final outcomes shows great potential in en

hancing students reasoning capabilities How

ever current methods struggle with sequence

level KD under longtailed data distributions

adversely affecting generalization on sparsely

represented domains We introduce the Multi

Stage Balanced Distillation BalDistill frame

work which iteratively balances training data

within a fixed computational budget By dy

namically selecting representative head domain

examples and synthesizing tail domain exam

ples BalDistill achieves stateoftheart perfor

mance across diverse longtailed datasets en

hancing both the efficiency and efficacy of the

distilled models