{
  "paper": {
    "arxiv_id": "2509.05292",
    "title": "Deep Reinforcement Learning for Ranking Utility Tuning in the Ad Recommender System at Pinterest",
    "url": "https://arxiv.org/abs/2509.05292"
  },
  "dataset": {
    "name": "MovieLens-1M",
    "users": 420,
    "items": 640,
    "full_catalog_evaluation": true
  },
  "setup": {
    "seed": 42,
    "mechanism": "policy_utility_tuning"
  },
  "baseline": {
    "name": "fixed ranking utility weights",
    "hit_at_10": 0.14047619047619048,
    "ndcg_at_10": 0.07513562832974019,
    "mrr_at_10": 0.05541099773242629,
    "head_share_at_10": 0.21761904761904763
  },
  "method": {
    "name": "DRL-PUT contextual policy over ranking utility weights",
    "hit_at_10": 0.15714285714285714,
    "ndcg_at_10": 0.0895118697352302,
    "mrr_at_10": 0.06898526077097504,
    "head_share_at_10": 0.11047619047619048
  },
  "relative": {
    "hit_at_10_percent": 11.864406779661014,
    "ndcg_at_10_percent": 19.133720879259098,
    "mrr_at_10_percent": 24.497416747659727,
    "head_share_at_10_percent": -49.23413566739606
  },
  "diagnostics": {
    "learned_action": [
      0.75,
      0.15,
      0.1
    ],
    "logged_policy_updates": 18,
    "validation_blend_weight": 0.5,
    "validation_ndcg_candidates": [
      0.067197,
      0.07709,
      0.069012,
      0.06934
    ]
  },
  "paper_results": {
    "platform_revenue_percent": 0.27,
    "ctr_percent": 1.62,
    "cvr_percent": 0.67
  },
  "scope": "从公开交互构造带 propensity 的 logged contextual bandit，使用 REINFORCE 学习相关性/新颖性/收益权重策略；未复刻 Pinterest 广告拍卖、真实收入和反事实校准。",
  "schema_version": 2,
  "manifest_ref": "reproduction:drl-put",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2509.05292-drl-put/metrics/movielens-1m-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}