发现打车软件和传统出租车的打赏行为完全不同，不能用一个模型通吃。
One Size Fits None: Modeling NYC Taxi Trips
- 对比线性回归与深度网络，发现两类打车方式打赏模式差异巨大
- 传统出租车打赏预测准确率高（R²≈0.72），App打赏几乎无法预测（R²≈0.17）
- 揭示合并建模可能掩盖分类偏差，需为不同车型定制模型
基于打车应用的兴起彻底改变了纽约市的小费文化。我们分析了2024年2.8亿次乘车记录，试图预测传统出租车与高流量网约车的小费。通过测试从线性回归到深度神经网络的方法，发现两种情况截然不同：传统出租车因车内支付屏幕，小费高度可预测（R² ≈ 0.72）；而基于应用的小费则近乎随机，难以建模（R² ≈ 0.17）。结论表明，构建统一模型是错误的，由于辛普森悖论，整体模型看似准确，却无法有效预测具体类别，必须采用专用模型。
原文摘要 · Abstract (English)
The rise of app-based ride-sharing has fundamentally changed tipping culture in New York City. We analyzed 280 million trips from 2024 to see if we could predict tips for traditional taxis versus high-volume for-hire services. By testing methods from linear regression to deep neural networks, we found two very different outcomes. Traditional taxis are highly predictable ($R^2 \approx 0.72$) due to the in-car payment screen. In contrast, app-based tipping is random and hard to model ($R^2 \approx 0.17$). In conclusion, we show that building one universal model is a mistake and, due to Simpson's paradox, a combined model looks accurate on average but fails to predict tips for individual taxi categories requiring specialized models.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。