Research map / Label-efficient and robust learning / Recognition and long-tail
TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems
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 TranSUN, a jointly trained auxiliary bias branch that intrinsically removes retransformation bias from transformed-MSE regression models, generalized into the GTS model family, and validates it offline and in a Taobao online A/B test.
Problem
Transformed MSE regression models used in recommender systems suffer from retransformation bias when predictions are back-transformed to the original scale, causing systematic under- or overestimation; prior fixes are post-hoc corrections external to the model.
Approach
Add an auxiliary branch trained jointly with a stop-gradient multiplicative bias target (y divided by the back-transformed prediction) so that the product of the branch output and the back-transformed prediction is an unbiased estimate of E[y]; generalize to GTS, where unbiasedness comes from conditional linearity of a target-scaling slope kappa(Q) fed by a conditional point estimate Q.
Key results
- On CIKM16 test, LogSUN reduces TRE to 0.0123 from 0.3451 for LogMSE and achieves the best TRE and MRE (0.0439) among compared methods.
- On synthetic data, T-MSE with ln(y+1) shows |SRE| up to 0.53 while TranSUN keeps |SRE| below 0.7% across all distributions.
- In an online A/B test on Taobao, LogSUN gave +1.06% GMV in product recommendation vs LogMAE and +1.83% GMV in short video recommendation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TranSUN joint bias learning to transformed MSE | LogMSE · CIKM16 test | TRE: 0.3451 -> 0.0123 | Helped |
| Add TranSUN joint bias learning to transformed MSE | SqrtMSE · CIKM16 test | TRE: 0.1339 -> 0.0347 | Helped |
| Add TranSUN to the regression model | DTMart test · DTMart test | NRMSE: 0.5232 -> 0.4151 | Helped |
| Add GTS debiasing | Indus test · Indus test | TRE: 0.0917 -> 0.0197 | Helped |
| Use LogSUN instead of LogMAE | Product recommendation · Product recommendation | GMV: +1.06% | Helped |
| Use LogSUN instead of LogMSE | Product recommendation · Product recommendation | GMV: +1.02% | Helped |
| Use LogSUN instead of LogMAE | Short video recommendation · Short video recommendation | GMV: +1.83% | Helped |
| Replace the inverse-ground-truth ratio scheme with additive bias modeling (ablation) | LogSUN · CIKM16 test | TRE: 0.0123 -> 0.0689 | Hurt |
| Learn biased prediction divided by ground truth instead of ground truth divided by biased prediction (ablation) | LogSUN · CIKM16 test | TRE: 0.0123 -> 0.1288 | Hurt |
| Replace MSE point loss with MAE (ablation) | GTS · CIKM16 test | TRE: 0.0123 -> 0.0145 | Hurt |
| Replace MSE point loss with MSPE (ablation) | GTS · CIKM16 test | TRE: 0.0123 -> 0.1077 | Hurt |
| Replace κ(u)=1/(|T⁻¹(u)|+ε) with κ(u)=|u| (ablation) | GTS with MSE point loss · CIKM16 test | TRE: 0.0123 -> 0.1810 | Hurt |
Surprising findings
- Supervising the auxiliary branch with the ratio of the biased prediction to ground truth (the reverse ratio) yields a biased estimate, showing that the specific bias-learning target, not explicit bias modeling alone, determines unbiasedness.
- Debiasing reduces XAUC (overall ranking metric), a trade-off the authors attribute to overestimation of hard low-value samples.
Limitations
- Evaluated only on right-skewed target distributions
- Debiasing causes a decline in XAUC (overall ranking accuracy)
- No general method yet to identify the optimal conditional point Q and slope kappa for arbitrary data
Experimental setup
Datasets: CIKM16 DTMart Indus synthetic distributions
Benchmarks: CIKM16 DTMart Indus synthetic data
Metrics: TRE MRE NRMSE NMAE XAUC NDCG SRE
Baselines: MSE MAE LogMSE SqrtMSE WLR ZILN MDME TPM CREAD OptDist NTE Smearing
Builds on: transformed MSE Smearing Normal Theory Estimates conditional transformation models ZILN
Closest papers by meaning
- Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.85)
- Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems (NeurIPS 2026 · similarity 0.84)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.84)
- Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random (NeurIPS 2024 · similarity 0.84)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.84)
- Backdoor Adjustment via Group Adaptation for Debiased Coupon Recommendations (AAAI 2024 · 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.