This is the official implementation of the AttrE2vec (Attributed Edge 2 vector) method, which is designed to obtain low-dimensional representation vectors for edges in graphs.

We observe that for each edge $(u, v)$ we can distinguish two neighborhoods: one "behind" node $u$ and one behind $v$. Each of those neighborhoods is sampled using uniform random walks over edges, and the information of the a single random walk (features of encountered edges) is aggregated using one of the following methods:

• Avg – average aggregation
• Exp – exponential aggregation (more distant edges are less important)
• GRU - recurrent neural network taking only edge features into account
• ConcatGRU - recurrent neural network taking into account both node and edge features

The aggregated information of a single random walk is the averaged over all walks in a single neighborhood to obtain the neighborhood summary $S_u$ (and $S_v$, respectively).

The neighborhood summaries and the features of the current edge are passed into the encoder module, which consists of a custom self-attention mechanism and fully connected layers.

Our proposed model relies on a hybrid cost function that includes:

• a contrastive learning objective based on the cosine similarity of edge representations
• an edge feature reconstruction loss

As the model learns the parameters of a transformation function, the model is fully inductive.

We use Python 3.7 to implement our model and experiments. Please refer to the requirements.txt file to install the required packages.

Our experimental pipeline is based on DVC (Data Version Control). You can get all pipeline artifacts using the dvc pull command. If needed, pipeline stages can be reproduced using the dvc repro <stage-name> command.

If you make use of AttrE2vec in your research, please cite it using the following entry:

@article{BIELAK2022,
    title = {AttrE2vec: Unsupervised Attributed Edge Representation Learning},
    journal = {Information Sciences},
    year = {2022},
    issn = {0020-0255},
    doi = {https://doi.org/10.1016/j.ins.2022.01.048},
    url = {https://www.sciencedirect.com/science/article/pii/S0020025522000779},
    author = {Piotr Bielak and Tomasz Kajdanowicz and Nitesh V. Chawla},
}              

MIT