MMoE Tutorial: Multi-gate Mixture-of-Experts

MMoE Tutorial: Multi-gate Mixture-of-Experts#

Sandbox Model — Experimental

This model is part of the CoreRec Sandbox — an experimental collection of algorithms for research and exploration. Sandbox models are not production-tested and may have incomplete implementations. For production-ready models, see the Production Models.

Introduction#

MMoE uses a Mixture-of-Experts (MoE) architecture with task-specific gating networks to learn task relationships automatically.

Paper: Ma et al. 2018 - Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts (Google)

How MMoE Works#

Architecture#

Multi-Task Learning with Gates:

  1. Shared Experts: Multiple expert networks share input

  2. Task-Specific Gates: Learn how to weigh experts for each task

  3. Task Towers: Specific networks for each objective (e.g., CTR, CVR)

Key Innovation: Solves “seesaw” problem where improving one task hurts another

Architecture:

Input -> [Expert 1] [Expert 2] ... [Expert N]
            |          |              |
         [Gate A]   [Gate B]       [Gate C]
            ↓          ↓              ↓
         [Tower A]  [Tower B]      [Tower C]

Mathematical Foundation#

Mixture of Experts:

f(x) = Σ_{i=1}^n g(x)_i · E_i(x)

Gating Network (Task k):

g^k(x) = softmax(W_g^k x)

Final Output (Task k):

y_k = h^k(f^k(x))

where h^k is the task tower.

Reference Implementation#

Implementation Note

The code examples below are reference implementations for learning purposes. This is a sandbox model and has not been production-tested. Verify behavior thoroughly before using in production.

Step 1: Import and Load Data#

from corerec.sandbox.collaborative_full.nn_base.MMoE_base import MMoE_base as MMoE
from cr_learn import ml_1m
from sklearn.model_selection import train_test_split
import numpy as np

# Load dataset with CORRECT API
data = ml_1m.load()  # Returns dict with 'ratings', 'users', 'movies'
ratings_df = data['ratings']  # DataFrame with user_id, movie_id, rating, timestamp

print(f"Loaded {len(ratings_df)} ratings")

# Split data
train_df, test_df = train_test_split(ratings_df, test_size=0.2, random_state=42)

# Extract arrays for model
train_users = train_df['user_id'].values
train_items = train_df['movie_id'].values
train_ratings = train_df['rating'].values

test_users = test_df['user_id'].values
test_items = test_df['movie_id'].values
test_ratings = test_df['rating'].values

Step 2: Initialize Model#

model = MMoE(
    name="MMoE_Model",
    num_experts=4,
    expert_dim=64,
    task_names=['income', 'marital'],
    epochs=20,
    batch_size=256,
    learning_rate=0.001,
    verbose=True
)

print(f"Initialized {model.name}")

Step 3: Train#

model.fit(
    user_ids=train_users,
    item_ids=train_items,
    ratings=train_ratings
)

print("Training complete!")

Step 4: Predict#

# Single prediction
score = model.predict(user_id=1, item_id=100)
print(f"Predicted score: {score:.3f}")

# Batch predictions
test_predictions = model.batch_predict(list(zip(test_users[:100], test_items[:100])))

Step 5: Recommend#

# Get top-10 recommendations for user
user_id = 1
recommendations = model.recommend(
    user_id=user_id,
    top_k=10
)

print(f"Top-10 recommendations for User {user_id}:")
for rank, item_id in enumerate(recommendations, 1):
    print(f"  {rank}. Item {item_id}")

Step 6: Evaluate#

from sklearn.metrics import mean_squared_error
import numpy as np

# Predict all test ratings
test_pred = [model.predict(u, i) for u, i in zip(test_users, test_items)]
rmse = np.sqrt(mean_squared_error(test_ratings, test_pred))
print(f"Test RMSE: {rmse:.4f}")

Step 7: Save & Load#

# Save model
model.save('artifacts/mmoe_model')

# Load model
loaded = MMoE.load('artifacts/mmoe_model')
test_score = loaded.predict(1, 100)
print(f"Loaded model prediction: {test_score:.3f}")

Key Takeaways#

When to Use MMoE#

✅ Perfect For:

  • Multi-objective optimization (e.g., Clicks & Conversions)

  • When tasks have complex relationships (correlated or conflicting)

  • Large-scale production systems

  • Reducing parameter count vs separate models

❌ Not For:

  • Single task learning

  • Small datasets

  • When tasks are completely unrelated

Best Practices#

  1. Number of Experts: 4-8 usually sufficient

  2. Expert Size: Smaller than single model

  3. Gate Bias: Initialize to uniform

  4. Task Weights: Tune loss weights (e.g., 1.0 for CTR, 2.0 for CVR)

Scaling & Production Considerations#

If you plan to move this sandbox model toward production use, consider:

  • Data Volume: Test with realistic dataset sizes; sandbox implementations may not be optimized for large-scale data.

  • Distributed Training: For datasets exceeding single-machine memory, consider wrapping the model in distributed training frameworks (e.g., PyTorch DDP, Horovod).

  • Hyperparameter Tuning: Use systematic search (Optuna, Ray Tune) rather than manual tuning. The defaults in sandbox models are not tuned for any specific domain.

  • Serving Latency: Profile inference time. Some architectures (e.g., attention-based) may need quantization or ONNX export for low-latency serving.

  • Monitoring: Implement A/B testing and online metrics (CTR, engagement, diversity) before replacing an existing system.

  • Validation: Write comprehensive unit tests and integration tests before deploying. Sandbox models do not have CI coverage.

Further Reading#

  • Paper: Ma et al. 2018 - Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts (Google)

  • GitHub Repository

Contents