Research map / Learning theory and algorithms / Prediction and decision losses
Representer Point Selection for Explaining Regularized High-dimensional Models
Che-Ping Tsai, Jiong Zhang, Hsiang-Fu Yu, Eli Chien, Cho-Jui Hsieh, Pradeep Kumar Ravikumar
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 high-dimensional representer (l1-projected local similarity) instead of l2 representer / influence function for explaining l1 logistic regression | l1-regularized logistic regression · 20 newsgroups, Gisette, Rcv1 | AUC-DEL+ and AUC-DEL-: e.g. -3.733 vs -2.472 (l2) on 20 newsgroups | Improved |
| Use high-dimensional representer instead of FIA for MF with reparameterized embeddings | MF and YoutubeNet on MovieLens-1M and Amazon video games · MovieLens-1M, Amazon reviews 2018 | AUC-DEL+ / AUC-DEL-: e.g. AUC-DEL+ -0.196 vs -0.101 (FIA) on MovieLens-1M MF | Improved |
| Use high-dimensional representer vs TracInCP checkpoint ensemble on vanilla MF | MF · MovieLens-1M, Amazon reviews 2018 | AUC-DEL+: -0.196 vs -0.217 (MovieLens-1M); -0.184 vs -0.312 (Amazon) | Hurt |
| Use high-dimensional representer vs TracInCP on YoutubeNet | YoutubeNet · MovieLens-1M, Amazon reviews 2018 | AUC-DEL+ / AUC-DEL-: -0.227 vs -0.161 (MovieLens-1M AUC-DEL+) | Improved |
| Use nuclear-norm high-dimensional representer for MF with nuclear norm regularization (vs random deletion) | MF w. nuclear norm · MovieLens-1M | AUC-DEL+ / AUC-DEL-: -0.225 vs -0.002 (random) for AUC-DEL+ | Improved |
| Use high-dimensional representer in negative sampling false-negative detection (aggregated importance) vs random selection | MF with BCE, binarized MovieLens-100k · MovieLens-100k implicit CF | false negatives found and recall@20 improvement: qualitative: outperforms random on both metrics; slightly fewer false negatives than loss/FIA but more influential ones | Improved |
| Random deletion of training samples as explanation baseline | all models · MovieLens-1M, Amazon reviews 2018 | AUC-DEL+: roughly zero, e.g. -0.002 (MovieLens-1M MF) | No clear effect |
| Use high-dimensional representer instead of l2 representer for l1 logistic regression runtime | l1-regularized logistic regression · 20 newsgroups, Gisette, Rcv1 | runtime (ms): 61.35 vs 59.47 ms (20NG, l2 faster); 87.34 vs 130.16 ms (Gisette, high-dim faster than l2) | Mixed |
Closest papers by meaning
- ImplicitSLIM and How it Improves Embedding-based Collaborative Filtering (ICLR 2024 · similarity 0.85)
- FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.85)
- Language Representations Can be What Recommenders Need: Findings and Potentials (ICLR 2025 · similarity 0.85)
- On the Embedding Collapse when Scaling up Recommendation Models (ICML 2024 · similarity 0.84)
- Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering (ICML 2025 · similarity 0.83)
- Recommendations with Sparse Comparison Data: Provably Fast Convergence for Nonconvex Matrix Factorization (ICML 2025 · similarity 0.83)
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.
