Research map / Learning theory and algorithms / Prediction and decision losses
Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a user-controllable recommendation paradigm where users indicate per-interaction willingness to exclude items from training, formulated as a multiplayer game and solved efficiently with influence-function-based approximation (IFRQE, IFRQE++ with multiple anchor vectors).
Problem
Recommender models train on all user interactions, but users may not want some behaviors used in training; existing methods lack user-indicated willingness and retraining per candidate data selection is too costly.
Approach
Each user acts as a player choosing a binary selection vector over their interactions; rewards trade off validation loss against willingness violation, and influence functions approximate validation loss for different selections without retraining, with multiple anchor selection vectors improving approximation accuracy.
Key results
- IFRQE++ improves the base model's reward by about 12.4% (simulation), 11.1% (Diginetica), 14.6% (Steam), and 21.3% (Amazon Video) on average, with an average F1 drop of about 8.02% relative to the base model.
- Increasing anchor vectors reduces validation-loss approximation error from 19.2% to 15.2% on Diginetica and from 9.4% to 1.2% on Amazon Video (IFRQE vs IFRQE++).
- IFRQE++ achieves comparable reward to Proactive while substantially reducing training time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IFRQE++ for user-controlled training instead of training on all interactions | MF · Simulation | reward: -2.25(.032) -> -1.92(.012) | Helped |
| Use IFRQE++ for user-controlled training instead of training on all interactions | MF · Diginetica | reward: -2.65(.072) -> -1.92(.022) | Helped |
| Use IFRQE++ for user-controlled training instead of training on all interactions | MF · Steam | reward: -2.99(.093) -> -2.42(.014) | Helped |
| Use IFRQE++ for user-controlled training instead of training on all interactions | MF · Amazon Video | reward: -2.43(.022) -> -1.98(.056) | Helped |
| Use multiple anchor selection vectors rather than one (ablation) | IFRQE validation-loss approximation · Diginetica | approximation error: 19.2% to 15.2% | Helped |
| Use multiple anchor selection vectors rather than one (ablation) | IFRQE validation-loss approximation · Amazon Video | approximation error: 9.4% to 1.2% | Helped |
Surprising findings
- On Diginetica and Amazon Video with MF, the proposed method sometimes improves recommendation F1 over the full-data base model, suggesting removing willing-to-omit items can help.
Limitations
- Does not model the temporal/dynamic nature of user willingness or online settings.
- Assumes overall user utility is a linear combination of recommendation quality and willingness violation.
- More anchor vectors require additional recommender retraining, reducing efficiency.
Experimental setup
Datasets: RecSim synthetic Diginetica Steam Amazon Video
Benchmarks: RecSim synthetic Diginetica Steam Amazon Video
Metrics: F1 willingness violation (wv) reward
Baselines: Random Threshold Proactive MF NeuMF LightGCN
Builds on: influence functions (Koh and Liang 2017) Proactive (Chen et al. 2022b)
Closest papers by meaning
- KuaiSim: A Comprehensive Simulator for Recommender Systems (NeurIPS 2023 · similarity 0.88)
- Infinite Recommendation Networks: A Data-Centric Approach (NeurIPS 2022 · similarity 0.88)
- LLM-Powered User Simulator for Recommender System (AAAI 2025 · similarity 0.88)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.88)
- Influential Exemplar Replay for Incremental Learning in Recommender Systems (AAAI 2024 · similarity 0.86)
- Identify Then Recommend: Towards Unsupervised Group Recommendation (NeurIPS 2024 · similarity 0.86)
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.