tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
evaluate the performance of HiComOPT vs baselines on randomly
selected nodes from all regions To compare these two cases we
consider the relative performance improvement for HiComOPT
over the secondbest method Results for nodes in the dense region
vs nodes in all regions are shown in Figure 3 We observe that the
performance enhancement of HiComOPT is less significant when
applied to nodes in all regions This observation is expected as
when selecting nodes from all regions many of them can have only
a few neighbors and hierarchical compression becomes unneces
sary In reality we observe that a large portion of nodes belong to
the dense region and our method will perform better when the data
gets richer with more connections The detailed experiment setting
and performance results on nodes in all regions can be found in
Appendix A5
HiCom Performance vs Node Degrees We found the HiCom per
formance improvements are quite consistent with the average node
degrees of the graphs When we consider the performance improve
ment of HiComOPT over the secondbest model on each dataset in
terms of percentage we get 075 149 383 548 583 im
provement in the order of columns presented in the table This gives
a 348 average performance improvement Also these numbers
are positively correlated with the average node degree reported in
Table 2 with a Pearson correlation coefficient equal to 0667 This
is expected as we discussed above
55 Ablation Studies
We conduct a series of ablation studies to understand the effective
ness of each module of the proposed HiCom method and quantify
the advantage of HiCom under different situations
Hierarchy and Summary Accumulation We study the signifi
cance of two key designs of HiCom the hierarchy constructed fol
lowing the graph structure and the summary accumulation We saw
in the main results that HiComOPT can outperform OPTNConcat
by significant margins The main reason is that HiComOPT can
compress much more neighbor information than a vanilla OPT with
limited input length Here we further show that the hierarchy ie
how the graph structure information is used to determine compres
sion order is also important for HiCom performance enhancement
We demonstrate this point by considering a case of compressing
nodes sampled from the neighborhood in a random order as if these
nodes form a set without any graph structure This ablation can
Figure 3 HiCom always gains Relative performance im
provement for HiComOPT over the secondbest method
on nodes in dense regions vs all regions
also be seen as applying the OPTbased AutoCompressor directly
to a long sequence generated by neighbor concatenation The dif
ference between this setting and OPTNConcat is that the number
of neighbors in the sequence can be much larger The result of this
ablation study is shown in Table 4 We see that the performance
of HiComOPT without a hierarchy drops significantly compared
to the complete version demonstrating the significance of the com
pression hierarchy Nonetheless the ablation performance is still
better than OPTNConcat which shows the power of compression
Another key design of HiCom is summary accumulation which as
we discussed above acts similarly as skip connections or multihop
filter matrices to enhance model performance We also perform an
ablation study on it and show results in Table 4 We observe that
removing summary accumulation causes a slight performance drop
smaller than removing hierarchy HiCom remains the best method
on four out of six datasets even with these two ablations
Increasing Training Data In our main experiment we used a
training set comprising 20 samples per class adhering to the prin
ciple that foundational models like LLMs should demonstrate effec
tiveness with minimal finetuning To further explore the models
robustness and performance we undertake an ablation study to
gradually increase the training set size Specifically we expand
the training set to be 2 5 and 10 times larger than the original
training set of 20 samples per class Considering the timeintensive
nature of tuning the LLM with a dataset that is 10 times larger
than before we select the Geology graph for this study The re
sults of these expanded training sets are illustrated in Figure 4
Our observations indicate that even with the enlarged training set
HiComOPT maintains superior performance compared to the two
strong baselines OPTNConcat and OPTGNN This outcome un
derscores the efficacy of HiCom on larger training datasets while
still achieving notable performance improvements The detailed
experiment setting and a similar figure for the Sports graph can be
found in Appendix A6Hierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY
Table 4 Node classification F1 score Ablation on hierarchy graph structure and summary accumulation For each ablation
results withmean that HiComOPT with ablation still performs the best on that dataset
Method Cloth Sports Economics Mathematics Geology
OPTNConcat 06798 06728 02372 02903 04218
HiComOPT 06882 07524 02463 03062 04464
HiComOPT wo hierarchy 0686107015 023740294204115
HiComOPT wo sum accumulation 06777 0743602413 0305104309
Figure 4 HiCom wins Method performance on the Geology
dataset with the training set in different sizes
Table 5 Run time comparison For each setting the training
time for one epoch is shown as well as the performance in
F1 score for the final model
OPTNConcat HiComOPT
Neighbors Fanouts 4 22 82 44 28
Train Time s 5546 3124 8002 8254 8755
Test F1 06250 06716 06835 06885 07001
56 Run Time Comparison
We discussed that one advantage of HiCom is its efficiency To
illustrate this we compare its run time to the backbone OPT with
neighbor concatenation OPTNConcat We show their training
time for one epoch in Table 5 Notably for a comparable number of
neighbors HiComOPT with fanouts 22 only takes 31245546
56of the time required by OPTNConcat with 4 neighbors and the

evaluate the performance of HiComOPT vs baselines on randomly

selected nodes from all regions To compare these two cases we

consider the relative performance improvement for HiComOPT

over the secondbest method Results for nodes in the dense region

vs nodes in all regions are shown in Figure 3 We observe that the

performance enhancement of HiComOPT is less significant when

applied to nodes in all regions This observation is expected as

when selecting nodes from all regions many of them can have only

a few neighbors and hierarchical compression becomes unneces

sary In reality we observe that a large portion of nodes belong to

the dense region and our method will perform better when the data

gets richer with more connections The detailed experiment setting

and performance results on nodes in all regions can be found in

Appendix A5

HiCom Performance vs Node Degrees We found the HiCom per

formance improvements are quite consistent with the average node

degrees of the graphs When we consider the performance improve

ment of HiComOPT over the secondbest model on each dataset in

terms of percentage we get 075 149 383 548 583 im

provement in the order of columns presented in the table This gives

a 348 average performance improvement Also these numbers

are positively correlated with the average node degree reported in

Table 2 with a Pearson correlation coefficient equal to 0667 This

is expected as we discussed above

55 Ablation Studies

We conduct a series of ablation studies to understand the effective

ness of each module of the proposed HiCom method and quantify

the advantage of HiCom under different situations

Hierarchy and Summary Accumulation We study the signifi

cance of two key designs of HiCom the hierarchy constructed fol

lowing the graph structure and the summary accumulation We saw

in the main results that HiComOPT can outperform OPTNConcat

by significant margins The main reason is that HiComOPT can

compress much more neighbor information than a vanilla OPT with

limited input length Here we further show that the hierarchy ie

how the graph structure information is used to determine compres

sion order is also important for HiCom performance enhancement

We demonstrate this point by considering a case of compressing

nodes sampled from the neighborhood in a random order as if these

nodes form a set without any graph structure This ablation can

Figure 3 HiCom always gains Relative performance im

provement for HiComOPT over the secondbest method

on nodes in dense regions vs all regions

also be seen as applying the OPTbased AutoCompressor directly

to a long sequence generated by neighbor concatenation The dif

ference between this setting and OPTNConcat is that the number

of neighbors in the sequence can be much larger The result of this

ablation study is shown in Table 4 We see that the performance

of HiComOPT without a hierarchy drops significantly compared

to the complete version demonstrating the significance of the com

pression hierarchy Nonetheless the ablation performance is still

better than OPTNConcat which shows the power of compression

Another key design of HiCom is summary accumulation which as

we discussed above acts similarly as skip connections or multihop

filter matrices to enhance model performance We also perform an

ablation study on it and show results in Table 4 We observe that

removing summary accumulation causes a slight performance drop

smaller than removing hierarchy HiCom remains the best method

on four out of six datasets even with these two ablations

Increasing Training Data In our main experiment we used a

training set comprising 20 samples per class adhering to the prin

ciple that foundational models like LLMs should demonstrate effec

tiveness with minimal finetuning To further explore the models

robustness and performance we undertake an ablation study to

gradually increase the training set size Specifically we expand

the training set to be 2 5 and 10 times larger than the original

training set of 20 samples per class Considering the timeintensive

nature of tuning the LLM with a dataset that is 10 times larger

than before we select the Geology graph for this study The re

sults of these expanded training sets are illustrated in Figure 4

Our observations indicate that even with the enlarged training set

HiComOPT maintains superior performance compared to the two

strong baselines OPTNConcat and OPTGNN This outcome un

derscores the efficacy of HiCom on larger training datasets while

still achieving notable performance improvements The detailed

experiment setting and a similar figure for the Sports graph can be

found in Appendix A6Hierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY

Table 4 Node classification F1 score Ablation on hierarchy graph structure and summary accumulation For each ablation

results withmean that HiComOPT with ablation still performs the best on that dataset

Method Cloth Sports Economics Mathematics Geology

OPTNConcat 06798 06728 02372 02903 04218

HiComOPT 06882 07524 02463 03062 04464

HiComOPT wo hierarchy 0686107015 023740294204115

HiComOPT wo sum accumulation 06777 0743602413 0305104309

Figure 4 HiCom wins Method performance on the Geology

dataset with the training set in different sizes

Table 5 Run time comparison For each setting the training

time for one epoch is shown as well as the performance in

F1 score for the final model

OPTNConcat HiComOPT

Neighbors Fanouts 4 22 82 44 28

Train Time s 5546 3124 8002 8254 8755

Test F1 06250 06716 06835 06885 07001

56 Run Time Comparison

We discussed that one advantage of HiCom is its efficiency To

illustrate this we compare its run time to the backbone OPT with

neighbor concatenation OPTNConcat We show their training

time for one epoch in Table 5 Notably for a comparable number of

neighbors HiComOPT with fanouts 22 only takes 31245546

56of the time required by OPTNConcat with 4 neighbors and the