tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
423 Summary Accumulation To include more explicit context
information during both compression and prediction summary
vectors from the previous levels in the hierarchy can be accumulated
to enrich the input For the same example of predicting 𝑣0 the input
can become𝒔1𝒔2𝒔0𝑥0by accumulating 𝒔1𝒔2 and 𝒔0together
as an enriched summary Similarly for compression if the hierarchy
has three levels summaries from level one can be accumulated to
enrich compression in level three Such operation is referred to as
summary accumulation in the AutoCompressor model We discuss
its relationship to skip connection and multihop filter matrix in
GNN message passing MP in Section 43 and demonstrate its
usefulness in enhancing the model performance in Section 55
43 Connection to Message Passing
This section compares HiCom with LLMs to GNN MP high
lighting their similarities and distinctions For similarities both
HiCom and MP focus on aggregating contextual information from
a neighborhood and they both operate across multihop neigh
bors to gather information level by level However there are three
critical differences between them
Firstly the situations where these two methods can apply are
different GNN MP is not designed to handle raw text and the mes
sages passed along have to be feature vectors In textrich graphs
this necessitates an additional step of feature extraction from raw
text which can potentially be a bottleneck if the semantic meanings
cannot be well preserved by the encoder Using LLM as encoders or
even cotraining the encoder with GNNs was shown to be subopti
mal 41 and we will also verify this in our experiments HiCom
on the other hand can be trained endtoend with raw text and
graph as inputs to the final labels as outputs which is more aligned
with the common practice of deep learning models 22
Secondly the ways the information is aggregated are differ
ent MP involves an aggregation function to combine all messages
passed to a node usually mean or sum 1421 These aggrega
tion functions treat all messages uniformly without considering
their potential influences on each other Although some GNNs
like GAT 38 and GaAN 48 employ the attention mechanism to
assign varying weights during message aggregation the weight
is assigned at the message level rather than the token level mak
ing the aggregation still limited in capturing semantic meanings
In contrast HiCom concatenates text sequences on several nodes
from a local neighborhood and employs an LLMbased compres
sor to summarize them which allows all the words that appear in
the neighborhood to interact with each other through the model
forward pass and potentially preserves more semantic information
Thirdly the third distinction lies in the utilization of the model
preserved knowledge GNNs are often trained from scratch and thus
have no access to prior knowledge beyond the downstream dataset
Even though there are some GNN pretraining methods the pre
trained GNN focuses on learning graph structural knowledge 27 orTable 2 Dataset Statistics Avg D means the average node
degree Avg T means the average number of tokens on each
node estimated by 1 token 34words
Dataset Nodes Edges Classes Avg D Avg T
Cloth 469274 2578746 18 550 17375
Sports 293712 2390076 23 814 13093
Economics 161727 1530898 40 545 16848
Mathematics 437685 2385322 20 947 19858
Geology 376503 7622754 18 2025 34389
domainspecific knowledge 1718 So far we have not seen GNN
pretraining methods that can be as effective as LLM pretraining
for memorizing general knowledge extracted from a large corpus
The general knowledge encapsulated in LLMs can be leveraged in
HiCom for generating summary vectors and making predictions
thereby enriching the models ability to process text data
Furthermore the summary accumulation operation used in HiCom
shares the same idea of skip connection 42 and the combination of
filter matrices 13 used in MP The summary accumulation brings
summary from a lower level in the hierarchy to a higher level to
make better predictions which creates a shortcut for information
flow across levels like the skip connection Similarly studies of
GNN architectures found that combining different filter matrices
to pass messages can improve GNN performance where each filter
is responsible for passing messages from a different hop From this
perspective summary accumulation achieves the same result
5 EXPERIMENTS
51 Datasets
We evaluate our method on two types of textrich graphs from dif
ferent domains the MAPLE academic citation graphs and the graph
of Amazon products Dataset statistics are presented in Table 2
MAPLE 51 MetadataAware Paper colLEction MAPLE is a
benchmarking dataset for scientific literature tagging constructed
from the Microsoft Academic Graph 33 MAPLE contains graphs
where papers serve as nodes and citations as edges Each graph
is specific to a specific scientific field such as mathematics with
nodes representing papers and their corresponding labels indicat
ing subfields like algebra This structure sets up a multilabel node
classification task In our experiments we focus on three specific
subfield graphs namely Mathematics Economics and Geology
Amazon Products 26 The Amazon products dataset comprises
items available on Amazon with detailed descriptions These items
are structured into a graph where each item represents a node
Edges are established between nodes items that are coviewed by
users Furthermore items are labeled with various product cate
gories setting the definition for a multilabel node classification
task Given the extensive size of the full graph which contains over
nine million items our analysis focuses on product graphs from
three subdomains ie Sports and Cloth
52 Baselines
We consider three types of methods First we evaluate GNN models
that incorporate LMencoded features Specifically we encode theHierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY
Table 3 HiCom wins Node classification F1 score on Amazon product graphs and MAPLE citation graphs For each dataset the

423 Summary Accumulation To include more explicit context

information during both compression and prediction summary

vectors from the previous levels in the hierarchy can be accumulated

to enrich the input For the same example of predicting 𝑣0 the input

can become𝒔1𝒔2𝒔0𝑥0by accumulating 𝒔1𝒔2 and 𝒔0together

as an enriched summary Similarly for compression if the hierarchy

has three levels summaries from level one can be accumulated to

enrich compression in level three Such operation is referred to as

summary accumulation in the AutoCompressor model We discuss

its relationship to skip connection and multihop filter matrix in

GNN message passing MP in Section 43 and demonstrate its

usefulness in enhancing the model performance in Section 55

43 Connection to Message Passing

This section compares HiCom with LLMs to GNN MP high

lighting their similarities and distinctions For similarities both

HiCom and MP focus on aggregating contextual information from

a neighborhood and they both operate across multihop neigh

bors to gather information level by level However there are three

critical differences between them

Firstly the situations where these two methods can apply are

different GNN MP is not designed to handle raw text and the mes

sages passed along have to be feature vectors In textrich graphs

this necessitates an additional step of feature extraction from raw

text which can potentially be a bottleneck if the semantic meanings

cannot be well preserved by the encoder Using LLM as encoders or

even cotraining the encoder with GNNs was shown to be subopti

mal 41 and we will also verify this in our experiments HiCom

on the other hand can be trained endtoend with raw text and

graph as inputs to the final labels as outputs which is more aligned

with the common practice of deep learning models 22

Secondly the ways the information is aggregated are differ

ent MP involves an aggregation function to combine all messages

passed to a node usually mean or sum 1421 These aggrega

tion functions treat all messages uniformly without considering

their potential influences on each other Although some GNNs

like GAT 38 and GaAN 48 employ the attention mechanism to

assign varying weights during message aggregation the weight

is assigned at the message level rather than the token level mak

ing the aggregation still limited in capturing semantic meanings

In contrast HiCom concatenates text sequences on several nodes

from a local neighborhood and employs an LLMbased compres

sor to summarize them which allows all the words that appear in

the neighborhood to interact with each other through the model

forward pass and potentially preserves more semantic information

Thirdly the third distinction lies in the utilization of the model

preserved knowledge GNNs are often trained from scratch and thus

have no access to prior knowledge beyond the downstream dataset

Even though there are some GNN pretraining methods the pre

trained GNN focuses on learning graph structural knowledge 27 orTable 2 Dataset Statistics Avg D means the average node

degree Avg T means the average number of tokens on each

node estimated by 1 token 34words

Dataset Nodes Edges Classes Avg D Avg T

Cloth 469274 2578746 18 550 17375

Sports 293712 2390076 23 814 13093

Economics 161727 1530898 40 545 16848

Mathematics 437685 2385322 20 947 19858

Geology 376503 7622754 18 2025 34389

domainspecific knowledge 1718 So far we have not seen GNN

pretraining methods that can be as effective as LLM pretraining

for memorizing general knowledge extracted from a large corpus

The general knowledge encapsulated in LLMs can be leveraged in

HiCom for generating summary vectors and making predictions

thereby enriching the models ability to process text data

Furthermore the summary accumulation operation used in HiCom

shares the same idea of skip connection 42 and the combination of

filter matrices 13 used in MP The summary accumulation brings

summary from a lower level in the hierarchy to a higher level to

make better predictions which creates a shortcut for information

flow across levels like the skip connection Similarly studies of

GNN architectures found that combining different filter matrices

to pass messages can improve GNN performance where each filter

is responsible for passing messages from a different hop From this

perspective summary accumulation achieves the same result

5 EXPERIMENTS

51 Datasets

We evaluate our method on two types of textrich graphs from dif

ferent domains the MAPLE academic citation graphs and the graph

of Amazon products Dataset statistics are presented in Table 2

MAPLE 51 MetadataAware Paper colLEction MAPLE is a

benchmarking dataset for scientific literature tagging constructed

from the Microsoft Academic Graph 33 MAPLE contains graphs

where papers serve as nodes and citations as edges Each graph

is specific to a specific scientific field such as mathematics with

nodes representing papers and their corresponding labels indicat

ing subfields like algebra This structure sets up a multilabel node

classification task In our experiments we focus on three specific

subfield graphs namely Mathematics Economics and Geology

Amazon Products 26 The Amazon products dataset comprises

items available on Amazon with detailed descriptions These items

are structured into a graph where each item represents a node

Edges are established between nodes items that are coviewed by

users Furthermore items are labeled with various product cate

gories setting the definition for a multilabel node classification

task Given the extensive size of the full graph which contains over

nine million items our analysis focuses on product graphs from

three subdomains ie Sports and Cloth

52 Baselines

We consider three types of methods First we evaluate GNN models

that incorporate LMencoded features Specifically we encode theHierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY

Table 3 HiCom wins Node classification F1 score on Amazon product graphs and MAPLE citation graphs For each dataset the