DeepCTR-Torch is a deep learning library for building click-through rate prediction models. It provides a modular framework for assembling custom prediction architectures from pre-built core, interaction, and sequence layers, enabling the construction of deep neural networks that estimate click probability from user behavior data.
The library specializes in feature interaction modeling, offering components for learning low-order, high-order, and adaptive-order feature crosses. It supports multi-task learning for predicting multiple objectives simultaneously, such as click and conversion rates, using shared representations. Sequence-aware user modeling is also a core capability, with attention and recurrent structures for processing user behavior sequences to capture evolving preferences.
Additional capabilities include dynamic feature selection through squeeze-excitation networks, multi-head self-attention for feature relationships, and gated expert routing for task-specific information processing. The framework handles data preparation tasks like defining sparse and dense features, normalizing numerical inputs, and preprocessing categorical variables.
Models are trained through a simple fit and predict interface without requiring manual training loops, with support for checkpoint saving and early stopping.