Logo:Dataset Platter [ FUJITSU's Dataset Hub for AI ]



Bermuda triangles consist of two datasets: one for the task of determining the presence or absence of a triangle loop, and the other for the task of determining whether the distance between two creeks is above a certain threshold.


Synthetic datasets for Graph benchmark.




In this task we generate relatively complex graphs which contain either exactly 1 or exactly 0 cycles of length 3 (triangle structures).


In this task we generate a BA graph with two cliques attached. We assign class 1 to graphs where the distance between cliques is larger than a threshold and class 0 othrewise.

Data Format

Both datasets are stored in our internal format. It consists of two files: sp (adjacency matrix) and cl (graph classification labels). We provide a pair of these files for each split.

SP file

SP file contains adjacency matrices for each data set. It is a tab-separated file where each line correspond to an edge in a graph. The SP file has the following columns:

Graphs corresponding to each ID must be stored successively without interleaving with other graphs, but the order of edges can be arbitrary.

CL file

It is also a tab-separated file with two columns:

The order of entries must be the same with the SP file.

Citation (Bib Text)

title={Bermuda Triangles: {GNN}s Fail to Detect Simple Topological Structures},
author={Arseny Tolmachev and Akira Sakai and Masaru Todoriki and Koji Maruhashi},
booktitle={ICLR 2021 Workshop on Geometrical and Topological Representation Learning},


All data is licensed under CC0

CC0 To the extent possible under law, Akira Sakai has waived all copyright and related or neighboring rights to Bermudatriangles. This work is published from: 日本.