Questions about DCRec

[SONGLEI-arch]

For a user, why compute the edge of the same item pair twice?

[yuh-yang]

Thanks for the question.

This is intentional behavior, not a bug. The code is constructing a bidirectional (undirected) item–item graph from each user’s interaction sequence, following the design of DCRec.

Concretely:

• For a sequence [..., item_{i-1}, item_i, item_{i+1}, ...],
  the code adds edges between adjacent items in both directions:

  • (item_i → item_{i-1}) and (item_{i-1} → item_i)
  • (item_i → item_{i+1}) and (item_{i+1} → item_i)

This is why you may observe the same item pair appearing twice:

1. One direction is added explicitly.
2. The reverse direction is also added to make the graph symmetric.

Such a construction is common in graph-based recommendation models, especially when:

• Message passing (e.g., GNN / DGL) assumes directed edges, and
• An undirected item–item transition graph needs to be represented.

In addition:

• item_adj_dict is used to build the global item–item adjacency, and
• item_edges_of_user stores per-user edge lists for later graph construction or batching.

So although the same item pair appears twice at the code level, it corresponds to two opposite directed edges, not duplicated semantics.

Hope this clarifies the design choice.