About the DLRM Model (Machine Learning Model)

Deep Learning Recommendation Model (DLRM) is a neural network-based model designed for personalization and recommendation systems. It uniquely processes categorical data through embeddings and dense features using a multilayer perceptron (MLP), addressing the challenges of handling categorical features in recommendation tasks. DLRM includes a specialized parallelization scheme for optimizing memory usage and computational efficiency in large-scale deployments.

Overview

• Use Case: Personalization and recommendation systems, ad click-through rate prediction, content ranking
• Creator: Meta (Facebook)
• Architecture: Neural network using embeddings for categorical data and MLP for dense features with bottom and top MLPs
• Release Date: 2019
• License: MIT

Training Data

Proprietary click-through rate data; public implementations use Criteo or synthetic datasets

Evaluation Benchmarks

• Click-through rate prediction accuracy
• AUC (Area Under ROC Curve)
• Log loss

References

• https://arxiv.org/abs/1906.00091
• https://github.com/facebookresearch/dlrm

Notes

• Parameter count and memory requirements vary significantly based on embedding table sizes
• Open-source implementations available in PyTorch and Caffe2
• Focuses on memory efficiency and computational scalability
• Embedding tables can dominate memory requirements in production deployments
• Source code licensed under MIT; pre-trained models under CC-BY-NC license