Research map / Label-efficient and robust learning / Recognition and long-tail

Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning

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.

Diagnoses popularity bias in SID-based generative recommenders as arising from gradient starvation of tail tokens under MLE and undifferentiated tokenization, and proposes Ghost with asymmetric unlikelihood optimization and skeleton-founded tokenization.

Problem

Generative recommenders built on semantic IDs and LLM backbones inherit popularity bias, over-recommending head items while tail items are rarely surfaced; existing debiasing transfers (re-weighting, re-ranking) give marginal gains and the root causes are under-explored.

Approach

Attributes popularity bias to (i) token-level gradient starvation of tail-item SID tokens under MLE and (ii) undifferentiated SID branching; proposes AUO, an unlikelihood loss penalizing similar-but-divergent head item SIDs for each tail item, and SKT, which makes tail items inherit head-item SID prefixes and append extra tail-specific tokens.

Key results

• On Instruments, Ghost raises tail HR@10 to 0.0173 vs 0.0040 for LC-Rec and cuts ARP@10 to 209.9 vs 338.5 for LC-Rec.
• Ghost lowers MGU by about 55.76% on average versus standard GRs, and reports an average CNS gain of 16.81%.
• Overall HR and NDCG drop only 7.46% and 6.34% on average relative to LC-Rec, per the paper.

Surprising findings

• Scaling the LC-Rec backbone from 0.6B to 8B does not spontaneously mitigate popularity bias.

Limitations

• Overall utility degrades slightly relative to standard GRs and LC-Rec.
• Hyperparameters (SID lengths, alpha, undesired set size) involve trade-offs that require tuning.

Experimental setup

Datasets: Amazon Musical Instruments Amazon Arts, Crafts and Sewing Amazon Video Games

Benchmarks: Amazon Musical Instruments Amazon Arts, Crafts and Sewing Amazon Video Games

Metrics: HR@5 HR@10 NDCG@5 NDCG@10 MGU ARP CNS

Baselines: LETTER LC-Rec ED^2 IFairLRS-RW IFairLRS-RR Augmentation Substitution

Builds on: LC-Rec LETTER ED^2 IFairLRS unlikelihood training RQ-KMeans

Closest papers by meaning

• Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.88)
• Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.87)
• On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.87)
• Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.86)
• Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.86)
• Developing Generative Recommender Systems for Government Subsidy Pro-Grams with a New RQ-VAE Model: Wello & the Korean Government Case (AAAI 2025 · similarity 0.85)