Multi-Task Models

Multi-task learning jointly optimizes related objectives (e.g., CTR, CVR, retention) to improve generalization.

Model Overview (what to pick)

Model	When to use	Highlights
SharedBottom	Tasks highly related	Shared bottom + task-specific towers
MMOE	Task conflict exists	Multi-gate mixture-of-experts per task
PLE	Complex multi-task	Progressive layered extraction to reduce negative transfer
ESMM	Sample selection bias	Full-space modeling for CVR/CTR with post-click modeling
AITM	Task dependency	Adaptive information transfer between tasks

Quick Usage (example: SharedBottom)

from torch_rechub.models.multi_task import SharedBottom
from torch_rechub.basic.features import SparseFeature, DenseFeature

common_features = [
    SparseFeature("user_id", vocab_size=10000, embed_dim=32),
    DenseFeature("age", embed_dim=1),
]

model = SharedBottom(
    features=common_features,
    task_types=["classification", "classification"],
    bottom_params={"dims": [256, 128], "dropout": 0.2, "activation": "relu"},
    tower_params_list=[
        {"dims": [64, 32], "dropout": 0.2, "activation": "relu"},
        {"dims": [64, 32], "dropout": 0.2, "activation": "relu"},
    ],
)

Parameter Notes (common fields)

• features: shared feature list.
• task_types: list of task types (classification, regression).
• bottom_params: shared bottom MLP config.
• tower_params_list: per-task tower MLP configs.
• (MMOE/PLE) expert/tower configs follow the same dict style (dims, dropout, activation).

Tips

• Strongly related tasks, simple setup → SharedBottom.
• Task competition / conflicts → MMOE.
• Need finer separation and reduced negative transfer → PLE.
• CVR/CTR with post-click bias → ESMM.
• Explicit task dependency → AITM.