Ranking Models

Ranking models predict CTR or preference scores to re-rank retrieved candidates. Torch-RecHub includes classic and modern architectures covering feature crosses, attention, and sequence modeling.

Model Overview (what to pick)

Model	When to use	Highlights
WideDeep	Baseline, memory + generalization	Linear (Wide) + DNN (Deep) joint training
DeepFM	Feature interactions matter	FM for 2nd-order + DNN for higher-order, shared embeddings
DCN / DCNv2	Explicit feature crosses	Efficient cross network; v2 adds feature selection & scaling
EDCN	Stronger crosses	Enhanced cross + deep with importance weights
AFM	Need interaction importance	Attention-weighted FM interactions
FiBiNET	Uneven feature importance	SE-style importance + bilinear interactions
DeepFFM / FatDeepFFM	Field-aware interactions	FFM + deep network for higher-order field-aware crosses
BST	Sequence-focused CTR	Transformer over behavior sequences
DIN	Dynamic interest	Attention over user history for target-aware interest
DIEN	Evolving interest	GRU + attention with interest evolution
AutoInt	Automatic interaction	Multi-head self-attention for feature crosses

Quick Usage (example: DeepFM)

from torch_rechub.models.ranking import DeepFM
from torch_rechub.trainers import CTRTrainer
from torch_rechub.basic.features import SparseFeature, DenseFeature

sparse_features = [
    SparseFeature("user_id", vocab_size=10000, embed_dim=32),
    SparseFeature("item_id", vocab_size=50000, embed_dim=32),
]
dense_features = [DenseFeature("age", embed_dim=1)]

model = DeepFM(
    deep_features=sparse_features + dense_features,
    fm_features=sparse_features,
    mlp_params={"dims": [256, 128, 64], "dropout": 0.2, "activation": "relu"},
)

trainer = CTRTrainer(
    model,
    optimizer_params={"lr": 1e-3, "weight_decay": 1e-5},
    n_epoch=10,
)
trainer.fit(train_dl, val_dl)
auc = trainer.evaluate(trainer.model, test_dl)

Parameter Notes (common fields)

• deep_features: feature list for deep/MLP branches.
• fm_features / cross_features: feature list for FM / cross layers.
• mlp_params: dict with dims, dropout, activation.
• cross_num_layers: number of cross layers (DCN/DCNv2/EDCN).
• attention_params: attention dim/dropout (AFM).

Tips

• Need explicit feature crosses → DCN/DCNv2/EDCN.
• Need interpretable interaction weights → AFM / FiBiNET.
• Strong field-aware interactions → DeepFFM / FatDeepFFM.
• User behavior sequences → DIN (target-aware), DIEN (evolving), BST (Transformer).
• Fast baseline with good coverage → WideDeep or DeepFM.