Research map / Probabilistic and scientific modeling / Molecules and proteins
Developing Generative Recommender Systems for Government Subsidy Pro-Grams with a New RQ-VAE Model: Wello & the Korean Government Case
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.
Builds a generative retrieval recommender for Korean government subsidy programs by modifying RQ-VAE (encoder-decoder between codebooks) to produce semantic IDs for users and documents, then trains a transformer to predict next semantic IDs, deployed at Wello.
Problem
Citizens miss government subsidy programs they qualify for because the programs are scattered and hard to search and match; existing keyword recommenders perform poorly, and generative recommenders may struggle to distinguish semantically similar subsidy documents.
Approach
A modified RQ-VAE inserts encoder-decoder modules between codebooks so each quantization layer learns from residual representations, yielding semantic IDs for both users and subsidy documents; a scaled-down TIGER-style transformer then generates the next semantic IDs.
Key results
- On the Wello dataset, the 256/4 base model reaches Recall@5 0.4084, NDCG@5 0.2964, Recall@10 0.5017, NDCG@10 0.3264; removing user IDs lowers performance by about 3-6%.
- On Amazon Sports and Outdoors, the model reaches Recall@5 0.0267 and NDCG@5 0.0182, slightly above TIGER (0.0264, 0.0181) and similar to TIGER on Recall@10 (0.0395 vs 0.0400).
- In an A/B test on Wello, CTR rose from 41.4% (Q3 2024) to 69.6% (Q4 2024), a reported 68% improvement; the authors estimate $7.8 million in otherwise unused subsidy funds could be saved.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 256 cardinalities and 4 codebooks instead of 32 cardinalities and 6 codebooks (ablation) | Generative recommender system on Wello · Wello dataset | Recall@5: 0.3989 -> 0.4084 | Helped |
| Use 256 cardinalities and 4 codebooks instead of 32 cardinalities and 6 codebooks (ablation) | Generative recommender system on Wello · Wello dataset | NDCG@5: 0.2911 -> 0.2964 | Helped |
| Use 256 cardinalities and 4 codebooks instead of 32 cardinalities and 6 codebooks (ablation) | Generative recommender system on Wello · Wello dataset | Recall@10: 0.4844 -> 0.5017 | Helped |
| Use 256 cardinalities and 4 codebooks instead of 32 cardinalities and 6 codebooks (ablation) | Generative recommender system on Wello · Wello dataset | NDCG@10: 0.3190 -> 0.3264 | Helped |
| Add user IDs to semantic ID sequences (ablation) | Generative recommender system on Wello · Wello dataset | Recall@5: 0.3863 (-5.4%) -> 0.4084 | Helped |
| Add user IDs to semantic ID sequences (ablation) | Generative recommender system on Wello · Wello dataset | NDCG@5: 0.2868 (-3.2%) -> 0.2964 | Helped |
| Add user IDs to semantic ID sequences (ablation) | Generative recommender system on Wello · Wello dataset | Recall@10: 0.4730 (-5.7%) -> 0.5017 | Helped |
| Add user IDs to semantic ID sequences (ablation) | Generative recommender system on Wello · Wello dataset | NDCG@10: 0.3144 (-3.7%) -> 0.3264 | Helped |
| Use the generative recommender instead of keyword-based recommendations | Wello platform · Wello platform, Q3 to Q4 2024 | CTR: 41.4% to 69.6%; 68% improvement | Helped |
| Use the generative recommender instead of keyword-based recommendations | Wello platform · Wello platform, Q3 to Q4 2024 | Average time spent per user: 420 seconds to 480 seconds; 14.3% | Helped |
| Use the generative recommender instead of keyword-based recommendations | Wello platform · Wello platform, Q3 to Q4 2024 | Service satisfaction: 19.8% | Helped |
| Deploy the generative recommender system | Wello platform · Government subsidy applications on Wello | Applications: 36% | Helped |
Surprising findings
- The authors report that the existing RQ-VAE tends to concentrate information in the first codebook, motivating their modification, yet the 32/6 configuration showed the best average codebook usage while 256/4 had the lowest validation loss.
- Performance on the Amazon benchmark is only marginally different from TIGER, and the authors acknowledge they did not compare against traditional collaborative or content-based filtering.
Limitations
- Evaluated only on the proprietary Wello dataset and one Amazon dataset; unclear generalization to other datasets and model sizes.
- No comparison with traditional recommenders such as content-based or collaborative filtering.
- Business impact figures (savings, market share scaling) rely on estimates and assumptions by the authors.
Experimental setup
Datasets: Wello dataset (185,610 user click-stream histories, ~240,000 subsidy documents) Amazon Sports and Outdoors
Benchmarks: Wello dataset Amazon Sports and Outdoors
Metrics: Recall@5 NDCG@5 Recall@10 NDCG@10 click-through rate (A/B test) average time spent per user user satisfaction survey
Baselines: P5 GRU4Rec BERT4Rec SASRec TIGER Wello keyword-based recommender
Builds on: RQ-VAE TIGER VQ-VAE SentenceT5
Closest papers by meaning
- Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems (NeurIPS 2026 · similarity 0.85)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.84)
- Multimodal Quantitative Language for Generative Recommendation (ICLR 2025 · similarity 0.83)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.83)
- Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations (NeurIPS 2025 · similarity 0.83)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.82)
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.