跳转至

面向超大规模电商的时序感知复购预测：多场景生鲜推荐的生存分析模型

文章背景与核心概要

传统的电商复购推荐系统通常采用二分类任务来构建（即预测“客户是否会在 \(W\) 天内购买商品 \(X\)”），这种做法导致针对每一个感兴趣的时间视界（time horizon）都需要训练一个独立模型，带来了巨大的模型维护负担。本文提出用生存分析模型（survival analysis models）直接预测复购时间（time-to-repurchase），以此取代繁琐的模型堆栈。研究团队依托某大型生鲜电商平台的数百万名客户数据，在 30 种以上的消融配置下进行了评估，证明了单一的加速失效时间（AFT）模型不仅能够成功替代多个二分类器，还能在准确性、特征效率和校准性能上实现全面提升。

该研究的核心贡献包括：通过实证风险分析发现生鲜复购的边际风险略有下降（\(k \approx 0.9\)），打破了“距离上次购买时间越久复购概率越高”的传统直觉；验证了单个 AFT 模型在显著降低树模型总数量（约减少 3 倍）的同时，特征重要性也发生了合理的转变（渠道节奏和近因信号上升，总量统计下降）；并提出了一种四参数参数化校准方法，实现了零跨视界单调性违背。研究揭指出了同一 AFT 家族内部在校准与排序之间的原则性权衡（如选择指数 AFT 用于概率敏感场景，Log-Normal 用于纯排序场景），为工业界大规模推荐系统提供了重要的实践指导。

📌 执行摘要 (Executive Summary)

Traditional e-commerce repurchase recommenders rely on a binary formulation ("Will the customer buy item \(X\) within \(W\) days?"), which requires training a separate model for every time horizon of interest. This paper proposes replacing this burdensome model stack with survival analysis models that predict time-to-repurchase directly. Evaluated on millions of customers from a major grocery e-commerce platform across over 30 ablation configurations, the study demonstrates that a single Accelerated Failure Time (AFT) model can successfully replace multiple binary classifiers while improving accuracy, feature efficiency, and calibration.

传统电商的复购推荐系统依赖于二分类构架（“客户是否会在 \(W\) 天内购买商品 \(X\)？”），这种构架要求为每一个感兴趣的时间视界单独训练一个模型。本文提出使用直接预测复购时间的生存分析模型（survival analysis models）来替代这种沉重的模型堆栈。通过在某大型生鲜电商平台的数百万客户数据上对 30 多种消融配置进行评估，本研究证明了单个加速失效时间（AFT）模型可以成功替代多个二分类器，同时在准确性、特征效率和校准度上均有所提升。

🔍 核心贡献与发现 (Key Contributions & Findings)

  1. Empirical Hazard Analysis & Model Fits
  2. Surprising Hazard Trend: The study reveals a slightly decreasing marginal hazard (\(k \approx 0.9\)), contradicting the common intuition that grocery items become more likely to be repurchased the longer it has been since the last purchase (\(k > 1\)).
  3. Fit Discrepancy: While the Log-Normal model achieves the best marginal fit (\(R^2 = 0.998\)) and superior ranking capabilities, the Weibull model provides the best conditional residual fit.
  1. 实证风险分析与模型拟合
  2. 令人意外的风险趋势： 研究揭示了一个略微下降的边际风险（\(k \approx 0.9\)），这与“距离上次购买时间越长，生鲜商品越容易被再次购买（递增风险，\(k > 1\)）”的常见直觉相悖。
  3. 拟合差异： 尽管对数正态（Log-Normal）模型实现了最佳的边际拟合（\(R^2 = 0.998\)）和卓越的排序能力，但威布尔（Weibull）模型提供了最佳的条件残差拟合。
  1. Model Efficiency & Feature Importance
  2. Reduced Complexity: A single AFT model replaces three per-horizon binary classifiers, matching or exceeding their performance while utilizing roughly \(3\times\) fewer total trees.
  3. Feature Shift: Under the survival objective, feature importance reshuffles: channel-cadence and recency signals rise in prominence, whereas aggregate frequency counts decline.
  1. 模型效率与特征重要性
  2. 降低复杂度： 单个 AFT 模型取代了三个针对不同视界的二分类器，在匹配或超越其各自视界性能的同时，总树木数量减少了约 \(3\) 倍。
  3. 特征转变： 在生存分析目标下，特征重要性发生了重组：渠道节奏（channel-cadence）和近因（recency）信号的重要性上升，而总量统计计数（aggregate frequency counts）则有所下降。
  1. Parametric Calibration & The Ranking-Calibration Trade-off
  2. Zero Violations: A 4-parameter parametric calibration maps raw survival Cumulative Distribution Functions (CDFs) to per-horizon probabilities with zero cross-horizon monotonicity violations.
  3. Calibration vs. Ranking: Calibration quality varies significantly across the AFT family. The Exponential AFT model (Weibull with \(k=1\)) achieves an Expected Calibration Error (ECE) of \(\sim 10^{-4}\) (roughly \(10\times\) lower than Log-Normal), whereas their ranking metrics remain within 0.3% of each other.
  4. Practical Deployment: The authors adopt Exponential AFT for probability-consuming surfaces and Log-Normal for pure ranking applications, highlighting a principled trade-off within a single AFT family.
  1. 参数化校准与排序-校准权衡
  2. 零违背： 四参数参数化校准将原始生存累积分布函数（CDFs）映射为各视界的概率，实现了零跨视界单调性违背。
  3. 校准与排序： 校准质量在不同的 AFT 模型家族中存在显著差异。指数 AFT 模型（\(k=1\) 的威布尔模型）实现的期望校准误差（ECE）约为 \(\sim 10^{-4}\)（比对数正态模型低大约 \(10\) 倍），而它们的排序指标彼此相差在 0.3% 以内。
  4. 实际部署： 作者在消耗概率的场景中采用指数 AFT，而在纯排序应用中采用对数正态模型，凸显了在单一 AFT 家族内部进行有原则的校准-排序权衡。

📄 摘要 (Abstract)

Repurchase recommenders in e-commerce are commonly framed as a binary question asking "will this customer buy this item within \(W\) days", a formulation that requires a separately trained model for every horizon of interest. We replace this stack with survival models that predict time-to-repurchase directly, and evaluate them on millions of customers from a major grocery e-commerce platform across more than thirty ablation configurations. Our study makes three contributions.

First, an empirical hazard analysis reveals a slightly decreasing marginal hazard (\(k \sim 0.9\)), differing from the common intuition that grocery items become more likely to be repurchased the longer since the last purchase (increasing hazard, \(k > 1\)). Log-Normal achieves the best marginal fit (\(R^2 = 0.998\)) and the best ranking, despite Weibull providing the best conditional residual fit, revealing an apparent discrepancy we analyze in detail.

Second, a single Accelerated Failure Time (AFT) model replaces three per-horizon binary classifiers, matching or exceeding each at its own horizon while using roughly 3x fewer total trees. Feature importance reshuffles under the survival objective: channel-cadence and recency signals rise while aggregate frequency counts fall.

Third, a 4-parameter parametric calibration maps raw survival CDFs to per-horizon probabilities with zero cross-horizon monotonicity violations. Calibration quality varies by an order of magnitude across the AFT family: Exponential AFT (Weibull \(k=1\)) achieves expected calibration error (ECE) \(\sim 1\text{e-4}\), roughly 10x lower than Log-Normal, while ranking metrics agree within 0.3% relative. We adopt Exponential AFT for probability-consuming surfaces and Log-Normal for pure ranking, exposing a principled calibration-ranking trade-off within a single AFT family.

电商领域的复购推荐系统通常被构架为一个二分类问题，即询问“该客户是否会在 \(W\) 天内购买此商品”，这种设定要求针对每一个感兴趣的时间视界单独训练一个模型。我们用直接预测复购时间的生存模型替代了这一模型堆栈，并在某大型生鲜电商平台的数百万客户数据上，通过 30 多种消融配置对其进行了评估。我们的研究做出了三项贡献。

首先，实证风险分析揭示了一个略微下降的边际风险（\(k \sim 0.9\)），这与“距离上次购买时间越久，生鲜商品越容易被再次购买（递增风险，\(k > 1\)）”的常见直觉不同。尽管威布尔模型提供了最佳的条件残差拟合，但对数正态（Log-Normal）模型实现了最佳的边际拟合（\(R^2 = 0.998\)）和最佳的排序能力，我们对此明显的差异进行了详细分析。

其次，单个加速失效时间（AFT）模型替代了三个分视界的二分类器，在各自的视界上匹配或超越了它们，同时总树木数量减少了约 3 倍。在生存目标下，特征重要性发生了重组：渠道节奏和近因信号上升，而总量统计计数下降。

第三，四参数参数化校准将原始生存 CDF 映射为各视界的概率，且实现了零跨视界单调性违背。在 AFT 家族中，校准质量相差一个数量级：指数 AFT（威布尔 \(k=1\)）实现的期望校准误差（ECE）约为 \(\sim 1\text{e-4}\)，比对数正态模型低约 10 倍，而排序指标的相对误差在 0.3% 以内。我们在概率敏感的场景中采用指数 AFT，在纯排序中采用对数正态，这展现了在单一 AFT 家族内部进行有原则的校准-排序权衡。

license icon