From Prediction to Persuasion: Agentic Recommendation Reason Generation for Regulatory-Compliant Financial AI
Fuente:
Zenodo
Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866902092358090752 |
|---|---|
| author | Jeong, Seonkyu Sim, Euncheol Kim, Youngchan |
| author_facet | Jeong, Seonkyu Sim, Euncheol Kim, Youngchan |
| contents | <p>Financial product recommendation systems must explain why a product is recommended - first and foremost because persuasion requires narrative, not probability. A customer who asks "Why was this card recommended to me?" needs a reason they can accept, not a score they must trust. Regulators (Korean FSS, EU AI Act) are one important audience for such explanations, but the primary driver is the human act of persuasion itself.</p> <p>We present a multi-stage pipeline that bridges the gap between model prediction and human persuasion:</p> <p>(1) Adaptive knowledge distillation from a heterogeneous-expert PLE teacher to per-task LGBM students, with three-layer fallback (distill / direct hard-label / rule-based) and teacher threshold gating that ensures service continuity when teacher quality varies across tasks, enabling CPU-only inference.</p> <p>(2) A multi-agent recommendation reason generation pipeline where three specialized serving agents (Feature Selector, Reason Generator, Safety Gate) collaboratively produce natural-language explanations grounded in business-mapped feature attributions.</p> <p>(3) Two operational agents (OpsAgent and AuditAgent) that interpret monitoring outputs and compliance reports in natural language, enabling regulation-compliant MLOps for small teams without dedicated MLOps staff. Environment-adaptive consensus: AWS uses 3-agent independent parallel voting (Claude Sonnet); on-premises uses a 2-Round hybrid (5 Qwen 14B Q4 agents in Round 1 independent vote + 2 agents in Round 2 sequential deliberation) with structural minority-opinion preservation.</p> <p>(4) Regulatory compliance by design, with built-in drift monitoring, fairness auditing, and governance reporting aligned to Korean FSS guidelines, the EU AI Act, and the Korean AI Basic Act.</p> <p>We evaluate distillation quality (AUC gap < 3.6 percentage points across 7 binary tasks, mean 2.6 pp), reason generation quality via automated compliance validation, and Safety Gate reliability. The system achieves 120ms warm latency on AWS Lambda (L1 predict + 13 tasks). The system targets low-risk products (check cards, deposits); investment and insurance recommendations are excluded from the deployment scope.</p> <p>Companion paper: "Heterogeneous Expert PLE: An Explainable Multi-Task Architecture for Financial Product Recommendation".</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19622052 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | From Prediction to Persuasion: Agentic Recommendation Reason Generation for Regulatory-Compliant Financial AI Jeong, Seonkyu Sim, Euncheol Kim, Youngchan recommendation reason generation knowledge distillation LLM agents regulatory compliance financial AI EU AI Act Korean FSS MLOps audit agent environment-adaptive consensus <p>Financial product recommendation systems must explain why a product is recommended - first and foremost because persuasion requires narrative, not probability. A customer who asks "Why was this card recommended to me?" needs a reason they can accept, not a score they must trust. Regulators (Korean FSS, EU AI Act) are one important audience for such explanations, but the primary driver is the human act of persuasion itself.</p> <p>We present a multi-stage pipeline that bridges the gap between model prediction and human persuasion:</p> <p>(1) Adaptive knowledge distillation from a heterogeneous-expert PLE teacher to per-task LGBM students, with three-layer fallback (distill / direct hard-label / rule-based) and teacher threshold gating that ensures service continuity when teacher quality varies across tasks, enabling CPU-only inference.</p> <p>(2) A multi-agent recommendation reason generation pipeline where three specialized serving agents (Feature Selector, Reason Generator, Safety Gate) collaboratively produce natural-language explanations grounded in business-mapped feature attributions.</p> <p>(3) Two operational agents (OpsAgent and AuditAgent) that interpret monitoring outputs and compliance reports in natural language, enabling regulation-compliant MLOps for small teams without dedicated MLOps staff. Environment-adaptive consensus: AWS uses 3-agent independent parallel voting (Claude Sonnet); on-premises uses a 2-Round hybrid (5 Qwen 14B Q4 agents in Round 1 independent vote + 2 agents in Round 2 sequential deliberation) with structural minority-opinion preservation.</p> <p>(4) Regulatory compliance by design, with built-in drift monitoring, fairness auditing, and governance reporting aligned to Korean FSS guidelines, the EU AI Act, and the Korean AI Basic Act.</p> <p>We evaluate distillation quality (AUC gap < 3.6 percentage points across 7 binary tasks, mean 2.6 pp), reason generation quality via automated compliance validation, and Safety Gate reliability. The system achieves 120ms warm latency on AWS Lambda (L1 predict + 13 tasks). The system targets low-risk products (check cards, deposits); investment and insurance recommendations are excluded from the deployment scope.</p> <p>Companion paper: "Heterogeneous Expert PLE: An Explainable Multi-Task Architecture for Financial Product Recommendation".</p> |
| title | From Prediction to Persuasion: Agentic Recommendation Reason Generation for Regulatory-Compliant Financial AI |
| topic | recommendation reason generation knowledge distillation LLM agents regulatory compliance financial AI EU AI Act Korean FSS MLOps audit agent environment-adaptive consensus |
| url | https://doi.org/10.5281/zenodo.19622052 |