Embedding Heterogeneous Networks into Hyperbolic Space Without Meta-path

Authors: Lili Wang, Chongyang Gao, Chenghan Huang, Ruibo Liu, Weicheng Ma, Soroush Vosoughi10147-10155

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct thorough experiments for the tasks of network reconstruction and link prediction on two public datasets, showing that our model outperforms a variety of well-known baselines across all tasks.
Researcher Affiliation Collaboration Lili Wang,1 Chongyang Gao, 1 Chenghan Huang, 2 Ruibo Liu, 1 Weicheng Ma, 1 Soroush Vosoughi 1 1 Department of Computer Science, Dartmouth College 2 Millennium Management LLC lili.wang.gr@dartmouth.edu, soroush.vosoughi@dartmouth.edu
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The code and data for this paper will be made available upon request.
Open Datasets Yes DBLP: This is a subset network of DBLP, which contains three node types: 14,475 authors (A), 14,376 papers (P), and 20 venues (V). The network also contains the following edge types: 41,794 paper authorship relations (P-A) and 14,376 publish relations (P-V). Movie Lens: This is a subset network of Movie Lens, which contains three node types: 11,718 actors (A), 9,160 movies (M), and 3,510 directors (D). The network also contains the following edge types: 64,051 act in relations (M-A) and 9,160 direct relations (M-D).
Dataset Splits No The paper describes the creation of the test set by randomly removing 20% of edges, but does not explicitly mention a separate validation set or its split details.
Hardware Specification Yes All the experiments are run on an Amazon AWS p2.8xlarge instance running a Linux OS with 488GB of RAM, and the random seeds are set to 0 at the beginning.
Software Dependencies No The paper mentions 'Linux OS' but does not provide specific version numbers for software dependencies or libraries used for the implementation (e.g., PyTorch, TensorFlow, scikit-learn versions).
Experiment Setup Yes For our method, we use the following parameters: We do 10 random walks from each node in the training set with the length of 80, and use a sliding window of size 5 to generate positive samples. For the hyperboloid embedding learning, we generate 20 negative samples for each positive one, and use the learning rate of 0.3 and a batch size of 512 to train 5 epochs.