FuxiCTR

GDCN (Geometric Deep Cross Network) is a CTR (Click-Through Rate) prediction model designed to learn explicit and bounded-degree cross features. It is based on the paper Towards Deeper, Lighter and Interpretable Cross Network for CTR Prediction.

Key components include:

• CrossNet: Provides explicit feature crossing with bounded degree using the formula: $x_{l+1} = x_0x_l^Tw + b + x_l$.
• Dynamic embedding size: Automatically computes the embedding size for each feature field using the formula: $emb_dim = 6\times(vocab_size)^{1/4}$.

The QNN model is designed for Click-through rate (CTR) prediction. It analyzes feature interaction mechanisms through the lens of quadratic neural networks. Key innovations include:

• Multi-head Khatri–Rao products: An efficient alternative to the standard Hadamard product for feature interaction.
• Self-Ensemble Loss: A method to improve model performance without increasing inference latency.
• Interpretive Framework: Provides theoretical and empirical analysis of Hadamard product-based feature interaction.

FinalMLP is an enhanced two-stream MLP model designed for Click-Through Rate (CTR) prediction. It builds upon the concept of 'DualMLP' (two well-tuned MLP networks) by adding layers to better capture high-order feature interactions.

Key architectural components include:

• Stream-specific feature gating: Produces feature importance weights (global, user-specific, or item-specific) by conditioning on learnable parameters, user features, or item features.
• Stream-level interaction aggregation: Fuses the outputs of the two streams using a multi-head bilinear fusion mechanism to capture complementary feature interactions.

DCN is a CTR (Click-Through Rate) prediction model designed to learn explicit and bounded-degree cross features. It is composed of two key components:

1. CrossNet: Provides explicit feature crossing with bounded degree using the formula: $x_{l+1} = x_0x_l^Tw + b + x_l$.
2. Dynamic embedding size: Automatically computes the embedding size for each feature field using the formula: $emb_dim = 6\times(vocab_size)^{1/4}$.

DCN is highly efficient, performing similarly to a standard DNN (MLP) while providing explicit feature interactions.

FinalNet implements FINAL (Factorized Interaction Layer), a layer designed for Click-Through Rate (CTR) prediction. It extends standard linear layers by learning quadratic feature interactions.

Key characteristics include:

• Factorized Interaction Layer: Capable of learning quadratic feature interactions.
• FINAL Blocks: Multiple FINAL layers can be stacked into a 'FINAL block' (similar to MLPs) to achieve exponential growth in feature interaction degrees.
• Unified Architecture: It unifies feature interactions and MLPs into a single block, which can serve as an effective replacement for traditional MLP blocks in recommendation models.

The MMoE implementation follows this directory structure:

• config/: Contains configuration files.
  • dataset_config.yaml: Dataset configuration.
  • model_config.yaml: Model hyper-parameters.
• src/: Contains model source code.
  • MMoE.py: The MMoE model implementation.
• run_expid.py: The main execution script.
• fuxictr_version.py: Version checking and loading utilities.