DIEN Tutorial: Deep Interest Evolution Network

DIEN Tutorial: Deep Interest Evolution Network#

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#

DIEN models the evolution of user interests over time using AUGRU (Attention-based GRU) to capture interest dynamics.

Paper: Zhou et al. 2019 - Deep Interest Evolution Network

How DIEN Works#

Architecture#

Interest Evolution with Attention:

  1. Interest Extractor Layer:

    • GRU over user behavior sequence

    • Extracts interest states at each time step

  2. Interest Evolving Layer:

    • AUGRU (Attention-based GRU) with attention scores

    • Models how interests evolve toward target item

    • Attention controls update gate based on relevance

  3. Final Layer: Combines evolved interests with other features

Key Innovation: Uses attention to model interest evolution, not just representation

Architecture:

Behavior Seq → GRU (Interest Extract) → AUGRU (Evolution) → Final Interest → Predict
                                          ↑
                                    Attention from Target

Mathematical Foundation#

Interest Extractor (GRU):

h_t = GRU(e_t, h_{t-1})

Attention Score:

a_t = softmax(e_target^T · W · h_t)

AUGRU Update:

u_t' = a_t · u_t               # attention-weighted update gate
h_t' = (1 - u_t') ⊙ h_{t-1} + u_t' ⊙ h̃_t

where h̃_t is candidate state

Final Representation:

interest = Σ_t a_t · h_t'

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.DIEN_base import DIEN_base as DIEN
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 = DIEN(
    name="DIEN_Model",
    gru_hidden_dim=36,
    attention_hidden_dim=36,
    seq_len=50,
    use_auxiliary_loss=True,
    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/dien_model')

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

Key Takeaways#

When to Use DIEN#

✅ Ideal For:

  • E-commerce with browsing sequences

  • Interest drift is important (fashion, trends)

  • Long user sequences (20-100 items)

  • When you need to understand WHY user clicked

  • Evolving user preferences

❌ Not For:

  • Static preferences

  • Short sequences (<10 items)

  • Real-time constraints (AUGRU is slower)

  • Simple collaborative filtering tasks

Best Practices#

  1. Sequence Length: 20-50 items (longer than DIN)

  2. GRU Hidden Dim: 36-64 typical

  3. Attention Type: Scaled dot-product works well

  4. Auxiliary Loss: Add on intermediate interest states

  5. Negative Sampling: Sample from non-clicked items in session

  6. Feature Fusion: Combine with user/item features

  7. Learning Rate: 0.001, decay after epochs

  8. Training: Expensive - use GPU acceleration

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#

Contents