PAPERS-11-007 · Paper — DLRM — Deep Learning Recommendation Model

The paper

The paper “Deep Learning Recommendation Model (DLRM)” was published in 2020 by Facebook AI researchers and presented at the RecSys conference. It is considered a landmark because it offered the first widely‑adopted, open‑source architecture and benchmark for web‑scale recommendation systems.

Before this paper

Before DLRM, large‑scale recommendation systems struggled to combine massive categorical embedding tables with dense numerical features in a single, scalable model. Engineers had no standard way to distribute huge embedding tables across hardware, and benchmarks were fragmented. This made it hard to compare approaches or reproduce results across platforms. The field needed a unified, reproducible baseline that could handle web‑scale data.

The key idea

The key insight of DLRM is to treat categorical features as learned embeddings, process dense features with a bottom‑up neural network, explicitly compute all pairwise dot‑product interactions between these vectors, and then feed the combined information into a top‑level network for the final prediction. This design cleanly separates embedding handling from interaction modeling, making the system both expressive and hardware‑friendly.

How it works

First, each categorical feature is looked up in a large embedding table, producing a dense vector. Second, the dense numerical features pass through a bottom multilayer perceptron to create another vector. Third, DLRM forms explicit dot‑product interactions between every pair of embedding vectors and the dense vector, creating a rich interaction tensor. Fourth, this interaction tensor is concatenated with the original vectors and sent to a top multilayer perceptron that produces the final recommendation score. Throughout, the architecture is engineered so that the massive embedding tables can be sharded across multiple devices, while the interaction and top networks run efficiently on accelerators.

See it — embeddings meet a top MLP

flowchart LR
  A["Categorical:<br/>embeddings"] --> C["Dot-product<br/>interactions"]
  B["Dense: bottom MLP"] --> C
  C --> D["Top MLP"]
  D --> E["Click probability"]

What they showed

DLRM demonstrated a strong, open, reproducible model and benchmark for industrial‑scale recommendation, showing that the proposed architecture could handle web‑scale workloads and serve as a reference point for future research.

What it changed

The paper turned DLRM into a reference architecture and benchmark that is now used in MLPerf and many large‑scale recommender deployments. It gave the community a common baseline, accelerating progress and enabling fair comparisons across systems. By exposing how to distribute enormous embedding tables, it also guided hardware and software co‑design for recommendation workloads. Next: SASRec.

The one-line takeaway

DLRM showed that a simple, explicit interaction model combined with scalable embeddings can become the standard backbone for large‑scale recommendation systems.

Read it / meet it

Reading the DLRM paper gives you a clear view of how modern recommenders are built and why they work at web scale. Its legacy lives on in today’s production systems and benchmark suites like MLPerf, making it a valuable read for anyone entering the field.