How to create OnDiskDataset for Link Prediction

[lyndonkl]

I have a dataset that I have converted to a heterogenous graph in DGL. The following are the size of my train, test and val graphs when I save them using dgl.save_graph

• Train ( Approx 150 GB)
• Test (Approx 30GB)
• Val (Approx 12 GB)

I want to instead create an ondisk dataset as specified in the documentation so that I can perform stochastic training using GraphBolt. I am following the documentation below:

https://docs.dgl.ai/en/2.2.x/stochastic_training/ondisk_dataset_heterograph.html

Couple of things that are unclear to me, if create the dataset for link prediction, does the metadata.yaml file need to include every single edge type under the train_set, test_set and val_set types. In the documentation, I see the following under the link prediction task

- name: link_prediction
    num_classes: 10
    train_set:
      - type: "user:like:item"
        data:
          - name: seeds
            format: numpy
            path: {os.path.basename(lp_train_like_seeds_path)}
      - type: "user:follow:user"
        data:
          - name: seeds
            format: numpy
            path: {os.path.basename(lp_train_follow_seeds_path)}

It is not clear to me whether this is because this is specific to the use case where the link prediction task is occuring on both user:like:item and user:follow:user edges. To be specific, if I only wanted to predict links for user:like:item, do I need to include user:follow:user as well?

Secondly, in the documentation above, I see the negative edges included for link prediction in the indexes within validation set and test set, is there a way to include them in the training set as well? Can I just make the seeds for user:like:item include both positive and negative edges so that I do not need to perform negative sampling in the pipeline.

  - name: link_prediction
    num_classes: 10
    train_set:
      - type: "user:like:item"
        data:
          - name: seeds
            format: numpy
            path: {os.path.basename(lp_train_like_seeds_path)}

[BarclayII]

@lyndonkl

do I need to include user:follow:user as well?

No you don’t have to. You only need to include the edges you want to predict.

@lyndonkl

is there a way to include them in the training set as well?

There is no restriction on how the training/validation/test data should look like. So you only need to make the training set the same format as the validation/test set, and remove the sample_uniform_negative call in the training data loader definition

[lyndonkl]

I have a follow up question. I created the following scenario to understand how ondiskdataset loading works:

Imagine the yellow nodes are person nodes, the blue nodes are item nodes and the red nodes are food nodes in the image below:

I recreate this and save as shown in the code below:

import os
import dgl
import torch
import numpy as np
import pandas as pd

# Define the graph structure
# Define directories
base_dir = './sample_dataset'
os.makedirs(base_dir, exist_ok=True)
personNodes = [0, 1, 2, 3, 4]
personNodesWithLinks = [0, 1, 3, 4]
personNodesNp = np.array(personNodesWithLinks)
itemNodes = [0, 1, 2, 3]
foodNodes = [0, 1, 2, 3]
personItemEdges = [[0, 0], [1, 1], [3, 2], [4, 1]]
personItemTotalEdges = [[[0, 0], [0,1]], [[1, 1], [1,2]], [[3, 2], [3, 1]], [[4, 1], [4,3]]]
personItemEdgesNp = np.array(personItemEdges)
personItemTotalEdgesNp = np.array(personItemTotalEdges)
# df = pd.DataFrame(personItemEdgesNp)
# df.to_csv(os.path.join(base_dir, "person_item_edges.csv"), index=False, header=False)
np.save(os.path.join(base_dir, "person_item_edges.npy"), personItemEdgesNp)
np.save(os.path.join(base_dir, "person_item_edges_t.npy"), personItemEdgesNp.T)
np.save(os.path.join(base_dir, "person_nodes.npy"), personNodesNp)
np.save(os.path.join(base_dir, "person_item_edges_total.npy"), personItemTotalEdgesNp)

personFoodEdges = [[0, 0], [1, 2], [2, 1], [4, 3]]
personFoodEdgesNp = np.array(personFoodEdges)
# df = pd.DataFrame(personFoodEdgesNp)
# df.to_csv(os.path.join(base_dir, "person_food_edges.csv"), index=False, header=False)
np.save(os.path.join(base_dir, "person_food_edges.npy"), personFoodEdgesNp)
np.save(os.path.join(base_dir, "person_food_edges_t.npy"), personFoodEdgesNp.T)

# Add features to nodes (dummy features for this example)
features = torch.eye(len(personNodes))  # Identity matrix as features

# Save node features
torch.save(features, os.path.join(base_dir, "node_features.pt"))

# Create the metadata.yaml file
yaml_content = f"""
dataset_name: sample_graph
graph:
  nodes:
    - type: person
      num: {len(personNodes)}
    - type: item
      num: {len(itemNodes)}
    - type: food
      num: {len(foodNodes)}
  edges:
    - type: "person:like:item"
      format: numpy
      path: ./person_item_edges_t.npy
    - type: "person:like:food"
      format: numpy
      path: ./person_food_edges_t.npy
feature_data:
  - domain: node
    type: person
    name: feature
    format: torch
    in_memory: false
    path: node_features.pt
tasks:
  - name: link_prediction
    train_set:
      - type: "person:like:item"
        data:
          - name: seeds
            format: numpy
            in_memory: false
            path: ./person_item_edges.npy
          - name: seeds_with_negative
            format: numpy
            in_memory: false
            path: ./person_item_edges_total.npy
        
"""

# Save metadata.yaml
with open(os.path.join(base_dir, "metadata.yaml"), "w") as f:
    f.write(yaml_content)

I have saved negative pairs as seeds_with_negative. Now I am loading the dataset as in the documentation for link prediction but removing the negative sampling as suggested. The code is shown below:

base_dir = './sample_dataset'
print(base_dir)
dataset = gb.OnDiskDataset(base_dir).load()

itemset = dataset.tasks[0].train_set
graph = dataset.graph
print(itemset)
print(graph)

sampler = gb.ItemSampler(
        itemset,
        batch_size=1,
        # minibatcher=minibatcher,
        shuffle=True,
    )

sampler = sampler.sample_neighbor(
        graph, [-1]
    )

for minibatch in sampler:
    print(minibatch)

I am seeing the following output when I print the minibatch after sample_neighbours. As you can see, the person:like:food edges are showing as 0 everywhere. Why is this the case? Given that I use a fanout of -1, I should be seeing some edges in at least one of the minibatches for this edge type.

MiniBatch(seeds_with_negative={'person:like:item': tensor([[[4, 1],
                                       [4, 3]]])},
          seeds={'person:like:item': tensor([[4, 1]], dtype=torch.int32)},
          sampled_subgraphs=[SampledSubgraphImpl(sampled_csc={'person:like:food': CSCFormatBase(indptr=tensor([0], dtype=torch.int32),
                                                                         indices=tensor([], dtype=torch.int32),
                                                           ), 'person:like:item': CSCFormatBase(indptr=tensor([0, 2], dtype=torch.int32),
                                                                         indices=tensor([1, 0], dtype=torch.int32),
                                                           )},
                                               original_row_node_ids={'person': tensor([4, 1], dtype=torch.int32)},
                                               original_edge_ids=None,
                                               original_column_node_ids={'food': tensor([], dtype=torch.int32), 'item': tensor([1], dtype=torch.int32), 'person': tensor([4], dtype=torch.int32)},
                            )],
          node_features=None,
          labels=None,
          input_nodes={'person': tensor([4, 1], dtype=torch.int32)},
          indexes=None,
          edge_features=None,
          compacted_seeds={'person:like:item': tensor([[0, 0]], dtype=torch.int32)},
          blocks=[Block(num_src_nodes={'food': 0, 'item': 0, 'person': 2},
                       num_dst_nodes={'food': 0, 'item': 1, 'person': 1},
                       num_edges={('person', 'like', 'food'): 0, ('person', 'like', 'item'): 2},
                       metagraph=[('person', 'food', 'like'), ('person', 'item', 'like')])],
       )
MiniBatch(seeds_with_negative={'person:like:item': tensor([[[1, 1],
                                       [1, 2]]])},
          seeds={'person:like:item': tensor([[1, 1]], dtype=torch.int32)},
          sampled_subgraphs=[SampledSubgraphImpl(sampled_csc={'person:like:food': CSCFormatBase(indptr=tensor([0], dtype=torch.int32),
                                                                         indices=tensor([], dtype=torch.int32),
                                                           ), 'person:like:item': CSCFormatBase(indptr=tensor([0, 2], dtype=torch.int32),
                                                                         indices=tensor([0, 1], dtype=torch.int32),
                                                           )},
                                               original_row_node_ids={'person': tensor([1, 4], dtype=torch.int32)},
                                               original_edge_ids=None,
                                               original_column_node_ids={'food': tensor([], dtype=torch.int32), 'item': tensor([1], dtype=torch.int32), 'person': tensor([1], dtype=torch.int32)},
                            )],
          node_features=None,
          labels=None,
          input_nodes={'person': tensor([1, 4], dtype=torch.int32)},
          indexes=None,
          edge_features=None,
          compacted_seeds={'person:like:item': tensor([[0, 0]], dtype=torch.int32)},
          blocks=[Block(num_src_nodes={'food': 0, 'item': 0, 'person': 2},
                       num_dst_nodes={'food': 0, 'item': 1, 'person': 1},
                       num_edges={('person', 'like', 'food'): 0, ('person', 'like', 'item'): 2},
                       metagraph=[('person', 'food', 'like'), ('person', 'item', 'like')])],
       )
MiniBatch(seeds_with_negative={'person:like:item': tensor([[[3, 2],
                                       [3, 1]]])},
          seeds={'person:like:item': tensor([[3, 2]], dtype=torch.int32)},
          sampled_subgraphs=[SampledSubgraphImpl(sampled_csc={'person:like:food': CSCFormatBase(indptr=tensor([0], dtype=torch.int32),
                                                                         indices=tensor([], dtype=torch.int32),
                                                           ), 'person:like:item': CSCFormatBase(indptr=tensor([0, 1], dtype=torch.int32),
                                                                         indices=tensor([0], dtype=torch.int32),
                                                           )},
                                               original_row_node_ids={'person': tensor([3], dtype=torch.int32)},
                                               original_edge_ids=None,
                                               original_column_node_ids={'food': tensor([], dtype=torch.int32), 'item': tensor([2], dtype=torch.int32), 'person': tensor([3], dtype=torch.int32)},
                            )],
          node_features=None,
          labels=None,
          input_nodes={'person': tensor([3], dtype=torch.int32)},
          indexes=None,
          edge_features=None,
          compacted_seeds={'person:like:item': tensor([[0, 0]], dtype=torch.int32)},
          blocks=[Block(num_src_nodes={'food': 0, 'item': 0, 'person': 1},
                       num_dst_nodes={'food': 0, 'item': 1, 'person': 1},
                       num_edges={('person', 'like', 'food'): 0, ('person', 'like', 'item'): 1},
                       metagraph=[('person', 'food', 'like'), ('person', 'item', 'like')])],
       )
MiniBatch(seeds_with_negative={'person:like:item': tensor([[[0, 0],
                                       [0, 1]]])},
          seeds={'person:like:item': tensor([[0, 0]], dtype=torch.int32)},
          sampled_subgraphs=[SampledSubgraphImpl(sampled_csc={'person:like:food': CSCFormatBase(indptr=tensor([0], dtype=torch.int32),
                                                                         indices=tensor([], dtype=torch.int32),
                                                           ), 'person:like:item': CSCFormatBase(indptr=tensor([0, 1], dtype=torch.int32),
                                                                         indices=tensor([0], dtype=torch.int32),
                                                           )},
                                               original_row_node_ids={'person': tensor([0], dtype=torch.int32)},
                                               original_edge_ids=None,
                                               original_column_node_ids={'food': tensor([], dtype=torch.int32), 'item': tensor([0], dtype=torch.int32), 'person': tensor([0], dtype=torch.int32)},
                            )],
          node_features=None,
          labels=None,
          input_nodes={'person': tensor([0], dtype=torch.int32)},
          indexes=None,
          edge_features=None,
          compacted_seeds={'person:like:item': tensor([[0, 0]], dtype=torch.int32)},
          blocks=[Block(num_src_nodes={'food': 0, 'item': 0, 'person': 1},
                       num_dst_nodes={'food': 0, 'item': 1, 'person': 1},
                       num_edges={('person', 'like', 'food'): 0, ('person', 'like', 'item'): 1},
                       metagraph=[('person', 'food', 'like'), ('person', 'item', 'like')])],
       )

[lyndonkl]

I see this in the documentation

Neighbor sampler is responsible for sampling a subgraph from given data. It
returns an induced subgraph along with compacted information. In the
context of a node classification task, the neighbor sampler directly
utilizes the nodes provided as seed nodes. However, in scenarios involving
link prediction, the process needs another pre-peocess operation. That is,
gathering unique nodes from the given node pairs, encompassing both
positive and negative node pairs, and employs these nodes as the seed nodes
for subsequent steps. When the graph is hetero, sampled subgraphs in
minibatch will contain every edge type even though it is empty after
sampling.

I am not sure if the last line is related to what I am seeing?