IRES

Introduction

This repository contains the implementation of our approach to entity summarization using Relational Graph Convolutional Networks (RGCN). Our method aims to leverage the structure and semantics of knowledge graphs by detecting the anti-communities in the given graph.

Installation

To get started with IRES, clone this repository and install the required dependencies:

bash git clone https://github.com/atefemoradan/IRES cd IRES curl -sSL https://install.python-poetry.org | python3 - poetry install

Running the Code

After setting up your environment and choosing your configuration, you can run the main experiment with the main.py script. Here's a generic command to start the training and evaluation process:

bash python3 main.py [-h] [-b {esbm,esbm_plus}] [-d {dbpedia,lmdb}] [-f {rdf,json}] [-s SEED] [-p DATA_PATH] [-O OUTPUT_PATH] [-tf {transe_nodes,transe_node_relations,node_freq,node_relation_freq}] [-e {RGCN,GCN}] [-o {ASGD,AdamW,Adam}] [-sch {StepLR,OneCycleLR,LinearLR,ReduceLROnPlateau,CosineAnnealingLR,CosineAnnealingWarmRestarts}] [-lr LEARNING_RATE] [-wd WEIGHT_DECAY] [-ss STEP_SIZE] [-ne NUM_EPOCHS] [-nc NUM_CNN_LAYERS] [-he HIDDEN_EMBEDDINGS] [-af {ELU,RReLU,ReLU,LeakyReLU}] [-sl SKIP_LAYER] [-bn BATCH_NORM] [-do DROP_OUT] [-la LOSS_ALPHA] [-dt {cuda,cpu}] [-n NUM_EXECUTIONS] [-bm BEST_MODEL] `

Note: Ensure your dataset is prepared and located in the path specified by -p or --data_path. The output, including models and results, will be saved to the directory specified by -O or --output_path.

Configuration

You can adjust parameters in the config.py file or directly via command line arguments.

Parameters for Datasets:

Using node_freq + relation embedding:

ESBM:

bash python3 main.py -b esbm -d lmdb -tf node_relation_freq -lr 0.01 -wd 0.1 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0 python3 main.py -b esbm -d dbpedia -tf node_relation_freq -lr 0.01 -wd 0.1 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0

ESBM_Plus

bash python3 main.py -b esbm_plus -d lmdb -tf node_relation_freq -lr 0.01 -wd 0.1 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0 python3 main.py -b esbm_plus -d dbpedia -tf node_relation_freq -lr 0.01 -wd 0.1 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0

Using transe embedding:

ESBM:

bash python3 main.py -b esbm -d lmdb -tf transe_nodes -lr 0.01 -wd 0.01 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0 python3 main.py -b esbm -d dbpedia -tf transe_nodes -lr 0.01 -wd 0.01 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0

ESBM_Plus:

bash python3 main.py -b esbm_plus -d lmdb -tf transe_nodes -lr 0.01 -wd 0.01 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0 python3 main.py -b esbm_plus -d dbpedia -tf transe_nodes -lr 0.01 -wd 0.01 -ne 50 -ss 7 -he 64 -bn True -do 0.25 -la 0.0

These command-line snippets allow for straightforward model execution tailored to your specific dataset and feature preference. Adjust the parameters as necessary for your experiments.

Evaluation

The effectiveness of the Indisum model is evaluated using F1 scores among other metrics, which are automatically calculated during the testing phase of the script execution. Here's how you can interpret the results and understand the model's performance:

• F1 Score: This metric provides a balance between precision and recall. A higher F1 score indicates better summarization quality.

• Output Directory: All evaluation results, including detailed F1 scores and other performance metrics, are saved in the output directory specified by the --output_path option. Within this directory, look for files results_{configurations}.csv for aggregated results and additional files for more granular insights.

To further analyze the model's performance or compare it against different configurations or benchmarks, you can manually inspect the output files.

Related Work and Comparative Results

Our research critically evaluates the performance of the INDISUM model against four notable related works within the entity summarization domain. We conducted this comparison using the F1 Score metric across the ESBM and ESBM_Plus datasets. The results underscore our model's significant advancements in entity summarization.

F1-Score on ESBM Dataset

| Method | DBpedia (top5) | DBpedia (top10) | LinkedMDB (top5) | LinkedMDB (top10) | All (top5) | All (top10) | |------------------|------------------|------------------|------------------|-------------------|--------------|------------------| | RELIN | 0.265 | 0.459 | 0.285 | 0.345 | 0.270 | 0.426 | | MPSUM | 0.306 | 0.504 | 0.265 | 0.424 | 0.293 | 0.480 | | BAFREC | 0.321 | 0.459 | 0.335 | 0.307 | 0.325 | 0.416 | | KAFCA | 0.318 | 0.509 | 0.229 | 0.402 | 0.273 | 0.455 | | IRES (TransE) | 0.306 | 0.535 | 0.322 | 0.430 | 0.310 | 0.505 | | IRES (frequency) | 0.327 | 0.540 | 0.350 | 0.445 | 0.334 | 0.513 |

F1-Score on ESBM_Plus Dataset

| Method | DBpedia (top5) | DBpedia (top10) | LinkedMDB (top5) | LinkedMDB (top10) | All (top5) | All (top10) | |------------------|------------------|------------------|------------------|-------------------|------------------|------------------| | RELIN | 0.211 | 0.441 | 0.20 | 0.312 | 0.208 | 0.404 | | MPSUM | 0.235 | 0.450 | 0.174 | 0.353 | 0.217 | 0.422 | | BAFREC | 0.143 | 0.180 | 0.155 | 0.128 | 0.146 | 0.165 | | KAFCA | 0.183 | 0.369 | 0.108 | 0.249 | 0.162 | 0.334 | | IRES (TransE) | 0.267 | 0.467 | 0.23 | 0.38 | 0.258 | 0.436 | | IRES (frequency) | 0.334 | 0.534 | 0.30 | 0.433 | 0.324 | 0.505 |

MAP on ESBM Dataset.

| Method | DBpedia (top5) | DBpedia (top10) | LinkedMDB (top5) | LinkedMDB (top10) | All (top5) | All (top10) | |------------------|------------------|------------------|------------------|-------------------|------------------|------------------| | RELIN | 0.14 | 0.29 | 0.153 | 0.22 | 0.144 | 0.27 | | MPSUM | 0.181 | 0.328 | 0.152 | 0.254 | 0.17 | 0.306 | | BAFREC | 0.179 | 0.273 | 0.156 | 0.13 | 0.172 | 0.232 | | KAFCA | 0.199 | 0.337 | 0.137 | 0.24 | 0.181 | 0.309 | | IRES (TransE) | 0.19 | 0.352 | 0.22 | 0.267 | 0.198 | 0.327 | | IRES (frequency) | 0.207 | 0.367 | 0.226 | 0.272 | 0.212 | 0.34 |

MAP on ESBM_Plus Dataset

| Method | DBpedia (top5) | DBpedia (top10) | LinkedMDB (top5) | LinkedMDB (top10) | All (top5) | All (top10) | |------------------|------------------|-----------------|------------------|-------------------|------------------|------------------| | RELIN | 0.123 | 0.274 | 0.11 | 0.19 | 0.12 | 0.25 | | MPSUM | 0.137 | 0.287 | 0.088 | 0.18 | 0.126 | 0.256 | | BAFREC | 0.073 | 0.091 | 0.081 | 0.04 | 0.075 | 0.76 | | KAFCA | 0.104 | 0.21 | 0.052 | 0.117 | 0.089 | 0.048 | | IRES (TransE) | 0.153 | 0.3 | 0.133 | 0.224 | 0.147 | 0.278 | | IRES (frequency) | 0.214 | 0.366 | 0.168 | 0.25 | 0.2 | 0.332 |

For detailed insights into these related works and their approaches to entity summarization, please refer to their respective publications:

• MPSUM
• BAFREC
• KAFKA

The comparative analysis demonstrates the value of employing RGCN for entity summarization, particularly for its ability to capture and utilize the rich semantic relationships and community structures inherent in knowledge graphs.

Citation

bibtex @inproceedings{moradan2025untapping, title={Untapping the Power of Indirect Relationships in Entity Summarization}, author={Moradan, Atefeh and Sorkhpar, Mohammad and Miyauchi, Atsushi and Mottin, Davide and Assent, Ira}, booktitle={WSDM '25: Proceedings of The Eighteenth ACM International Conference on Web Search and Data Mining}, year={2025}, doi={10.1145/3701551.3703566}, url={https://github.com/atefemoradan/IRES}, version={0.1.0} }

License

This project is licensed under the MIT License. For more details, see the LICENSE file in the repository. Utilizing the IRES code in your work implies agreement with the license terms. The MIT License is permissive and allows for considerable freedom in usage and redistribution, but it requires acknowledgment of the original source.