模型评估总结生成器

模型性能总结

• 验证集表现

  • AUC = 0.89：总体区分能力较强。
  • F1 = 0.78：查准与查全相对均衡。
  • 召回 = 0.83：覆盖目标样本能力较高。
  • 说明：若F1与召回来源于同一阈值，则对应精确率约为0.74（由F1与召回推算）。

• 业务效果（绝对百分点，pp）

  • 转化率：+2.4pp。
  • 流失率：-1.1pp。
  • 说明：体现为正向业务增益。

• 分群表现

  • 新客 AUC = 0.86。
  • 老客 AUC = 0.91。
  • 差异 = 0.05，模型在老客群体的判别力更强。

• 数据窗口与样本量

  • 数据窗：上周。
  • 样本量：约3万。

• 结论

  • 离线判别力稳健，召回较高且F1表现良好；线上业务指标呈正向改善。分群结果显示老客优于新客，后续可结合阈值与特征策略针对新客优化。

Model Performance Summary

• Online (gray release)

  • Hit rate: +15% versus baseline, indicating improved detection/coverage in production canary.
  • False block rate: 0.7%, suggesting low collateral impact during gray testing.

• Offline evaluation

  • AUC: 0.92, demonstrating strong ranking/discriminative power.
  • F1: 0.81, reflecting a good precision–recall balance at the chosen threshold.

• Latency

  • P95: 120 ms, suitable for near–real-time inference in most interactive settings.

• Stability

  • Drift index: 0.08. If this follows a PSI-like scale, this indicates small distribution shift and stable feature/score behavior.

• Fairness

  • ΔFPR: 1.2 percentage points across groups, indicating a modest but non-zero disparity in false positives that should be monitored.

Overall assessment: The model shows strong offline discrimination and favorable online uplift with low false positives, meets practical latency requirements, and exhibits low drift. Fairness shows a small FPR gap that warrants continued monitoring and potential threshold calibration or post-processing if group parity targets are strict.