TeleGraph

TeleGraph: A Benchmark Dataset for Hierarchical Link Prediction.

• Link prediction is the problem of detecting the presence of a connection between two entities in a network. Research fields, ranging from network science to machine learning and data mining, have taken a great interest in link prediction task. Given that hierarchical patterns found in many real-world applications while the available research datasets are inadequate, in this work, we present a new real-world dataset TeleGraph, which is a medium sized and heterogeneous telecommunication network with a rich set of attributes. Our descriptive analysishas demonstrated it is highly hierarchical and sparse, which makes the heuristic measures fail to work. We verified this precognition by a series of experiments. Our findings show that most of the available algorithms fail to produce the satisfactory performance on this tree-like dataset except the subgraph-based GNN-models. More specifically, the results of a series heuristic measures are even close to random guesses, which calls for special attention in practice. We believe TeleGraph can serve as an useful benchmark to assess and foster novel link prediction and node embedding techniques.
• The repository contains the code and dataset for paper "TeleGraph: A Benchmark Dataset for Hierarchical Link Prediction" accepted by WebConf GLB 2022 and the manuscript is avaible at https://arxiv.org/abs/2204.07703.

Requirements:

• torch
• numpy
• torch_geometric
• sklearn
• scipy

Data

• TeleGraph.gpickle is an attributed telecom network with three type of devices and 240 types of alarms as sketeched in below:

• unzip the TELECOM.zip to get telecom_graph.pt in which each node is associated with an 1x 240 feature vector regarding the alarm occoured or not.

Runs:

Run GAE models:

cd gaes
python gae.py --dataset Telecom --encoder GCN -epochs 4001 --lr 0.0001 --val_ratio 0.05 --test_ratio 0.10 --patience 200

Run heuristics methods:

cd heuristics
python heuristics.py --dataset Telecom --batch_size 32 --use_heuristic CN

Run SEAL:

cd SEAL
python seal.py --dataset Telecom --embedding DRNL --epoches 4001 --lr 0.0001 --weight_deccay 5e-4 --val_ratio 0.05 --test_ratio 0.10 --batch_size 32 --patience 20

Contact

• We have open positions for internships and full-time jobs. If you are interested in research and practice in graph learning and mining, please send your CV to [email protected].