Deep-Walk-Transformers

Node embedding technique based on Masked Language Model.

Transductive Embeddings

You can use the provided additional public method get_transductive_embeddings() to obtain the embeddings of the nodes used for training.

from deep_walk_transformers import DeepWalkTransformers

dwt = DeepWalkTransformers(
    num_walks,
    walk_len,
    mask_rate,
    embed_dim
)
dwt.fit(G, epochs=epochs, lr=lr)
node_embeddings = dwt.get_transductive_embeddings()

Inductive Embeddings

Through the get_inductive_embeddings() public method you can get the embeddings of nodes unseen on the training data. The required arguments are the full graph G_full ( $G_{train}$ + $G_{unseen}$ ) and the unseen nodes starting_nodes.

inductive_node_embeddings = dwt.get_inductive_embeddings(G_full, starting_nodes)

Feature Information

Additionally you can pass the node features as optional argument to the fit() and get_inductive_embeddings() functions. For the functions:

• fit(): features argument must be of type np.ndarray.
• get_inductive_embeddings(): features argument must be dict(np.array) where the key is the node id and the value is the feature as np.array.

See notebooks directory for more complete examples.