Research map / Learning theory and algorithms / Prediction and decision losses
Boosting Methods for Interval-censored Data with Regression and Classification
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 L2Boost-CUT and L2Boost-IMP, nonparametric boosting methods that use censoring unbiased transformations and an ICRF survivor estimate to handle interval-censored responses for regression and classification.
Problem
Boosting assumes fully observed data, but interval-censored survival times (known only to lie in an interval) are common and existing boosting extensions mainly handle right censoring or missing responses.
Approach
Adjust the L2 loss via a censoring unbiased transformation (CUT) so its expectation matches the uncensored risk, estimate the conditional survivor function with interval-censored recursive forests, and run functional gradient L2 boosting with smoothing-spline base learners; a companion imputation variant (IMP) replaces the response with the CUT transform.
Key results
- On synthetic interval-censored AFT data, CUT and IMP yield errors and Kendall's tau comparable to the complete-data reference method and outperform the naive midpoint surrogate, which has the largest SMaxAE and SMSqE.
- For survival-status classification, CUT and IMP achieve significantly higher sensitivity than the naive method at similar specificity.
- Theory: with smoothing-spline base learners, the first iterations improve on the unboosted learner and the number of iterations t_n = O(n^{2r/(2v+1)}) attains the minimax MSE rate O(n^{-2v/(2v+1)}).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the naive midpoint response with L2Boost-CUT or L2Boost-IMP | interval-censored survival-time prediction · synthetic lognormal AFT and loglogistic AFT | SMaxAE | Helped |
| replace the naive midpoint response with L2Boost-CUT or L2Boost-IMP | interval-censored survival-time prediction · synthetic lognormal AFT and loglogistic AFT | SMSqE | Helped |
| replace the naive midpoint response with L2Boost-CUT or L2Boost-IMP | interval-censored survival-time prediction · synthetic lognormal AFT and loglogistic AFT | SKDT | Helped |
| replace the naive midpoint response with L2Boost-CUT or L2Boost-IMP | interval-censored survival-status prediction · synthetic data | sensitivity | Helped |
| use L2Boost-CUT instead of L2Boost-IMP | L2Boost for interval-censored data · synthetic survival-status prediction | sensitivity | No clear effect |
| use L2Boost-CUT instead of L2Boost-IMP | L2Boost for interval-censored data · synthetic survival-status prediction | specificity | No clear effect |
| use L2Boost-CUT instead of L2Boost-IMP | L2Boost for interval-censored data · synthetic experiments | performance | No clear effect |
Surprising findings
- L2Boost-CUT and L2Boost-IMP produce nearly identical results, differing mainly in stopping criterion.
- Running boosting indefinitely does not reduce MSE below the noise variance, so excess iterations overfit.
Limitations
- Requires consistent estimation of the conditional survivor function; boosting quality depends on ICRF estimation and its extra computational cost.
- Optimal rates are stated for one-dimensional features; extension to multivariate features is only via componentwise selection heuristic.
- Theoretical guarantees assume regularity conditions (e.g., weak base learners, condition C4) that may not hold for strong learners.
Experimental setup
Datasets: synthetic AFT data (lognormal and loglogistic) Signal Tandmobiel data Bangkok HIV data
Benchmarks: synthetic AFT simulations Signal Tandmobiel Bangkok HIV
Metrics: SMaxAE SMSqE SKDT (sample Kendall's tau) sensitivity specificity
Baselines: oracle method (true phi) reference method (complete data) naive surrogate response (midpoint) COX model
Builds on: L2Boost L2WCBoost interval censored recursive forests censoring unbiased transformation
Closest papers by meaning
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.86)
- Sample-Optimal Agnostic Boosting with Unlabeled Data (ICML 2025 · similarity 0.85)
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.85)
- Learning Survival Models with Right-Censored Reporting Delays (NeurIPS 2026 · similarity 0.84)
- Doubly Robust Conformalized Survival Analysis with Right-Censored Data (ICML 2025 · similarity 0.84)
- Optimal Aggregation of Prediction Intervals under Unsupervised Domain Shift (NeurIPS 2024 · 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.