Abstract
Government subsidy programs often fail to reach their intended recipients due to information asymmetry and complex eligibility criteria. This paper presents a generative recommender system deployed in South Korea that utilizes a novel Residual Quantization Variational Auto-Encoder (RQ-VAE) to match citizens with relevant support programs. By treating both documents and user profiles as semantic IDs, the system achieved a 68% increase in click-through rates and significant economic savings.
Introduction
In many nations, substantial portions of government budgets allocated for social welfare go unspent. In South Korea, approximately $2.7 billion in annual subsidies remains unclaimed because eligible citizens are either unaware of the programs or overwhelmed by the application process. Traditional recommendation systems struggle with the high semantic similarity of policy documents and the complexity of multi-faceted user demographic data.
The Challenge of Subsidy Matching
Standard retrieval models often fail to distinguish between nuanced policies, such as different types of family support or regional financial aid. Furthermore, existing generative retrieval systems typically focus on encoding item information, neglecting the rich demographic context—such as income, education, and household status—that determines subsidy eligibility.
Methodology: Generative Retrieval with Modified RQ-VAE
The proposed system utilizes Generative Retrieval, which directly generates the identifier of a relevant item rather than ranking a pre-filtered list. The core innovation is a modified Residual Quantization Variational Auto-Encoder (RQ-VAE).
To address the common "codebook imbalance" problem in standard RQ-VAEs—where the first codebook dominates the representation—the researchers introduced encoder-decoder modules between successive codebooks. This architecture forces the model to learn deeper residual representations, capturing the subtle differences between similar government policies.
Data and User Profiling
The system was trained and validated using two primary sources: a proprietary dataset from the Korean startup Wello (containing 185,610 user histories and 240,000 documents) and the public Amazon Sports dataset for benchmarking.
A key insight of this work is the conversion of user profiles into semantic IDs. By integrating demographic data (e.g., marriage status, job description) directly into the sequence generation process, the model's performance improved by approximately 4.5%.
Results and Real-World Impact
The system was deployed on the Wello platform in 2024. A/B testing revealed a dramatic increase in the Click-Through Ratio (CTR) from 41.4% to 69.6%. User satisfaction surveys indicated a 19.8% improvement, and subsidy applications rose by 36%. Economically, the system is estimated to save the Korean government $7.8 million in unspent funds and $27.4 million in advertising costs. On the Amazon benchmark, the "Gen-Rec" model consistently outperformed state-of-the-art baselines like TIGER and SASRec.
Conclusion
This research demonstrates the efficacy of generative retrieval in complex, high-stakes domains like government services. By improving the semantic representation of both items and users through an enhanced RQ-VAE, the system bridges the gap between citizens and the support they need. Future work will focus on dynamic retraining triggers and broader cross-domain validation.
Summary updated