Does it really work？Is it better than a randomly generated vector

[freedomRen]

def get_subgraph_embedding(embeddings_model_file):
    embedding_word_size = 0
    embedding_dims = 0
    word_embeddings = {}
    print("Loading embedding model %s" %(embeddings_model_file))
    with open(embeddings_model_file) as fh:
        # skip first line
        embedding_word_size, embedding_dims = fh.readline().rstrip().split()
        embedding_word_size = int(embedding_word_size)
        embedding_dims = int(embedding_dims)
        print('sub2Vec Info ： %d X %d' %(embedding_word_size, embedding_dims))

        for line in fh.readlines():
            values = line.rstrip().split()
            word = values[0]
            # embedding = np.asarray(values[1:], dtype=np.float64)
            rand_seed = np.uint32(hash(word))
            np.random.seed(rand_seed)
            embedding = np.random.randn(embedding_dims)
            word_embeddings[word] = embedding

    return embedding_word_size, embedding_dims, word_embeddings

I change it to py3， and using randomly generated vectors，and got better result。
So I'm trying to keep thinking scientifically.
Use subgraphs to encode the full graph,
Does the vector of the subgraph have to be important，that we got it from skipgram or randomly generated vector