Testing graph2vec with similarity of inferred embeddings.

[nikosandriotis]

Since there is no way of printing the loss function to understand why I am not getting expected results, I wanted to open this issue.

I am training graph2vec with 4000 random erdos_renyi graphs:

Graph2Vec generic example

graphs = []

Define ranges for n and p

n_range = range(400, 800, 50) # n from 400 to 800 in steps of 50
p_range = [round(p, 2) for p in list([i/100 for i in range(30, 71, 10)])] # p from 0.3 to 0.7

Create graphs

for _ in range(4000):
n = random.choice(n_range)
p = random.choice(p_range)
G = nx.erdos_renyi_graph(n, p)
graphs.append(G)

And training:

model = Graph2Vec(dimensions=1024, epochs=200)

Train the model

model.fit(graphs)

Save the full pipeline

os.makedirs("saved_model", exist_ok=True)
joblib.dump(model, "saved_model/graph2vec_diverseplus_pipeline_1024_200epochs_narrow.pkl")

similarity of generated embeddings

Then I am using the trained model to infer embeddings for 2 sets of similar structure graphs:

new_graphs_1 = [nx.erdos_renyi_graph(500, 0.6) for _ in range(10)]
new_graphs_2 = [nx.erdos_renyi_graph(500, 0.6) for _ in range(10)]

new_embeddings_1 = model.infer(new_graphs_1)
new_embeddings_2 = model.infer(new_graphs_2)

Shouldn't the similarity be very close of these 2 embeddings? Instead I get values of 0.6 similarity.