RandomLinkSplit and LinkNeighborLoader seem to sample ordinary subgraph or graph with Isolated node in directed heterogeneous graph

[0zero000zero0]

It seems that RandomLinkSplit and LinkNeighborLoader often sample ordinary subgraph in directed heterogeneous graph.
I don't it is a bug or something wrong with my code.
Here I have a directed heterogeneous graph , and use RandomLinkSplit to split data as train data and test data:

transform = RandomLinkSplit(
    num_val=0,
    num_test=0.2,
    disjoint_train_ratio=0.3,
    neg_sampling_ratio=1.0,
    add_negative_train_samples=False,
    edge_types=prevent_edge,
    # rev_edge_types=rev_prevent_edge,
)

The train data as follow:

HeteroData(
  pesticide={
    num_nodes=6110,
    x=[6110, 768],
  },
  disease={
    num_nodes=279,
    x=[279, 768],
  },
  plant={
    num_nodes=241,
    x=[241, 768],
  },
  (disease, impact, plant)={ edge_index=[2, 833] },
  (pesticide, prevent, disease)={
    edge_index=[2, 6482],
    edge_label=[2778],
    edge_label_index=[2, 2778],
  },
  (pesticide, cure, plant)={ edge_index=[2, 11577] }
)

Then I use LinkNeighborLoader to sample and load data to train my model:

num_neighbors=[3]
batch_size=1
neg_sampling_ratio=1
train_loader = LinkNeighborLoader(
    data=train_data,
    num_neighbors=num_neighbors,
    batch_size=batch_size,
    neg_sampling_ratio=neg_sampling_ratio,
    edge_label=edge_label,
    edge_label_index=(prevent_edge,edge_label_index),
)

the reason why batchsize is one is that I use sampled for visualization
I find that the sampled subgraph is often ordinary grpah:

HeteroData(
  pesticide={
    num_nodes=2,
    x=[2, 768],
    n_id=[2],
  },
  disease={
    num_nodes=2,
    x=[2, 768],
    n_id=[2],
  },
  plant={
    num_nodes=0,
    x=[0, 768],
    n_id=[0],
  },
  (disease, impact, plant)={
    edge_index=[2, 0],
    e_id=[0],
  },
  (pesticide, prevent, disease)={
    edge_index=[2, 0],
    edge_label=[2],
    edge_label_index=[2, 2],
    e_id=[0],
    input_id=[1],
  },
  (pesticide, cure, plant)={
    edge_index=[2, 0],
    e_id=[0],
  }
)

As you can see, there are some nodes but aren not edges in the subgraph . When I use a larger num_neighbors such as num_neighbors=[5,2] , the number of edge is not zero , but there is isolated node:

HeteroData(
  pesticide={
    num_nodes=3,
    x=[3, 768],
    n_id=[3],
  },
  disease={
    num_nodes=2,
    x=[2, 768],
    n_id=[2],
  },
  plant={
    num_nodes=0,
    x=[0, 768],
    n_id=[0],
  },
  (disease, impact, plant)={
    edge_index=[2, 0],
    e_id=[0],
  },
  (pesticide, prevent, disease)={
    edge_index=[2, 1],
    edge_label=[2],
    edge_label_index=[2, 2],
    e_id=[1],
    input_id=[1],
  },
  (pesticide, cure, plant)={
    edge_index=[2, 0],
    e_id=[0],
  }
)

I tried different parameter combination (such as add_negative_train_samples ,disjoint_train_ratio,neg_sampling_ratio,num_neighbors) , but still don't work.

I am curious about the process and principles of sampling

the offical website says that Negative sampling is currently implemented in an approximate way, i.e. negative edges may contain false negatives.

I gauss the it is a bug because of approximate way of sampling.

Here are my equipment information:
operating system:Window11 23H2
python : 3.12.3
related packes verson:
torch=2.2.1+cu118
torch_cluster=1.6.3+pt22cu118
torch_geometric = 2.5.3
torch_scatter =2.1.2+pt22cu118
torch_sparse =0.6.18+pt22cu118
torch_spline_conv =1.2.2+pt22cu118