Research map / Architectures and efficiency / RNNs and normalization
The trade-offs of model size in large recommendation models : 100GB to 10MB Criteo-tb DLRM model
Aditya Desai, Anshumali Shrivastava
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compress embeddings 10000x with PSS | DLRM · Criteo-TB | AUC target 0.8025 | No clear effect |
| compress embeddings 100000x with PSS (ablation) | DLRM · Criteo-TB | target AUC reached | Hurt |
| compress embeddings 1000x with PSS (ablation) | DLRM · Criteo-Kaggle | AUC: 0.8031 -> 0.8048 | No clear effect |
| compress embeddings 10000x with PSS (ablation) | DLRM · Criteo-Kaggle | AUC: 0.8031 -> 0.8001 | Hurt |
| compress embeddings 10000x with PSS (ablation) | XDEEPFM · Criteo-Kaggle | AUC: 0.8007 -> 0.795 | Hurt |
| compress embeddings 10000x with PSS (ablation) | FIBINET · Criteo-Kaggle | AUC: 0.8016 -> 0.7963 | Hurt |
| increase compression from original to 10000x (ablation) | DLRM · Criteo-TB | epochs to target AUC: 1.00 -> 4.55 | Hurt |
| increase compression from original to 10000x (ablation) | DLRM · Criteo-TB | time per 1000 iterations: 50.6 -> 11.6 | Improved |
| increase compression from original to 10000x (ablation) | DLRM · Criteo-TB | relative total training time: 1 -> 1.04 | No clear effect |
| replace multi-GPU original model with PSS on single GPU (ablation) | DLRM · Criteo-TB inference 89M samples | inference time: 203 seconds -> 97 seconds | Improved |
| use dense instead of sparse gradients at high compression (ablation) | PSS · embedding lookup benchmark | backward pass latency: up to 8× faster | Improved |
| hash chunks of size 32 instead of individual elements (ablation) | PSS · embedding lookup benchmark | latency | Improved |
Closest papers by meaning
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.87)
- On the Embedding Collapse when Scaling up Recommendation Models (ICML 2024 · similarity 0.86)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.85)
- Wukong: Towards a Scaling Law for Large-Scale Recommendation (ICML 2024 · similarity 0.85)
- FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.85)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
