Research map / Learning theory and algorithms / Clustering
Clustering the Sketch: Dynamic Compression for Embedding Tables
Henry Tsang, Thomas Dybdahl Ahle
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 |
|---|---|---|---|
| use CCE instead of the hashing trick | CCE · Criteo Kaggle up to 10 epochs | compression at baseline BCE: 4,600× -> 8,500× | Improved |
| use CCE instead of CE with concatenation | CCE · Criteo Kaggle up to 10 epochs | compression at baseline BCE: 3,800× -> 8,500× | Improved |
| use CCE instead of deep hash embeddings | CCE · Criteo Kaggle up to 10 epochs | compression at baseline BCE: 1,300× -> 8,500× | Improved |
| use CCE instead of the hashing trick | CCE · Criteo Kaggle 1 epoch | compression at baseline BCE: 78 −122× -> 212× | Improved |
| use CCE instead of the hashing trick | CCE · Criteo Terabyte 1 epoch | compression at baseline BCE: 23 −32× -> 101× | Improved |
| use CCE instead of CE with concatenation | CCE · Criteo Terabyte 1 epoch | compression at baseline BCE: 25 −48× -> 101× | Improved |
| train CCE with optimal parameters to convergence vs full table baseline | CCE · Criteo Kaggle | test BCE: 0.66% lower BCE | Improved |
| replace full embedding table by compressed embeddings when training multiple epochs | DLRM · Criteo Kaggle | test BCE | Improved |
| fine-tune tables after product quantization (ablation) | Product Quantization · Criteo Kaggle | test BCE | Hurt |
| increase number of clusterings (ct1 to ct6, once per epoch) (ablation) | CCE · Criteo Kaggle 10 epochs | test BCE | Improved |
| cluster once per epoch for first 6 epochs (ct6 cf300000) (ablation) | CCE · Criteo Kaggle 10 epochs | test BCE | Improved |
| finish all clusterings by 2/3 of an epoch (Strategy 2) (ablation) | CCE · Criteo Kaggle 1 epoch | test BCE | Hurt |
Closest papers by meaning
- Probabilistic Hash Embeddings for Online Learning of Categorical Features (AAAI 2026 · similarity 0.86)
- Objective-Based Hierarchical Clustering of Deep Embedding Vectors (AAAI 2021 · similarity 0.84)
- Compressing Word Embeddings via Deep Compositional Code Learning (ICLR 2018 · similarity 0.84)
- End-To-End Efficient Representation Learning via Cascading Combinatorial Optimization (CVPR 2019 · similarity 0.84)
- Hierarchically Clustered Representation Learning (AAAI 2020 · similarity 0.84)
- Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.84)
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.
