Research map / Bandits and online learning / Contextual bandits and Thompson sampling
Dynamic Assortment Selection and Pricing with Censored Preference Feedback
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 censored multinomial logit (C-MNL) model for joint multi-product selection and pricing where buyers only buy products priced below their valuation, and two bandit algorithms (UCB and TS assortment selection with LCB pricing) with sublinear regret, validated on synthetic simulations.
Problem
Dynamic multi-product assortment selection and pricing when buyers filter out overpriced items and the seller never observes which items were censored, only the purchase choice.
Approach
A censored MNL choice model with price-threshold activation; an LCB pricing rule sets low prices to ensure exploration and avoid censorship, paired with UCB or Thompson sampling for assortment selection, using online mirror descent estimation.
Key results
- UCBA-LCBP achieves regret bound of order d^{3/2} sqrt(T/kappa) and TSA-LCBP achieves d^2 sqrt(T/kappa) (plus lower-order terms).
- In simulations with K=5 and d=4, the proposed algorithms show sublinear regret while DASP-MNL, ONM and ETC do not work properly in the censored setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UCBA-LCBP or TSA-LCBP instead of DASP-MNL, ONM, or ETC | Synthetic C-MNL dynamic assortment and pricing · Synthetic datasets with varying product counts | regret | Helped |
Surprising findings
- Existing MNL dynamic pricing methods that maximize revenue at each step fail to learn valuations under censorship.
Limitations
- Regret scales as d^{3/2} (UCB) and d^2 (TS), and whether the dependence on d can be improved is open.
- Experiments are limited to synthetic data with small d and K.
- Randomness-in-activation extension only achieves the main bound when noise scale c=O(1/sqrt(T)).
Experimental setup
Datasets: synthetic
Benchmarks: synthetic C-MNL simulations
Metrics: cumulative regret
Baselines: DASP-MNL ONM ETC
Builds on: multinomial logit bandit dynamic pricing with MNL and arbitrary features (DASP-MNL) online Newton method for MNL
Closest papers by meaning
- Dynamic pricing and assortment under a contextual MNL demand (NeurIPS 2022 · similarity 0.89)
- Near-Optimal Policies for Dynamic Multinomial Logit Assortment Selection Models (NeurIPS 2018 · similarity 0.88)
- Online Learning of Delayed Choices (NeurIPS 2024 · similarity 0.87)
- Contextual Dynamic Pricing with Heterogeneous Buyers (NeurIPS 2025 · similarity 0.85)
- Transfer Learning for Nonparametric Contextual Dynamic Pricing (ICML 2025 · similarity 0.84)
- Improved Algorithms for Contextual Dynamic Pricing (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.