Graph-Based Models

Graph-Based Models#

Models that treat users and items as nodes in a bipartite interaction graph.

Production models (CI-tested)#

Model

Import

Tutorial

LightGCN

from corerec.engines.collaborative import LightGCN

LightGCN

GNNRec#

Graph neural message passing on the user–item bipartite graph. Uses BCE loss — pass 0/1 labels (or normalize explicit ratings to [0, 1]).

from corerec.engines.gnnrec import GNNRec

model = GNNRec(
    embedding_dim=64,
    num_gnn_layers=3,
    epochs=20,
    batch_size=256,
    verbose=True,
)
model.fit(user_ids=user_ids, item_ids=item_ids, ratings=binary_ratings)
recs = model.recommend(user_id=1, top_k=10)

LightGCN#

Simplified graph convolution for collaborative filtering.

from corerec.engines.collaborative import LightGCN

model = LightGCN(n_factors=64, n_layers=3, epochs=100, verbose=True)
model.fit(user_ids, item_ids, ratings)  # triplet implicit feedback
recs = model.recommend(user_id=1, top_k=10)

Sandbox models (experimental)#

Model

Import

Tutorial

GeoIMC

sandbox graph_based

GeoIMC

LightGCN-Base

corerec.sandbox.collaborative_full.graph_based_base.lightgcn_base

LightGCN Base

GNN-Base

corerec.sandbox.collaborative_full.graph_based_base.GNN_base

GNN Base

When to use#

  • Rich user–item interaction graphs

  • Social or knowledge-graph extensions (sandbox)

  • When CF matrix methods underperform on link structure

See also#

Contents