Skip to main content
QUICK REVIEW

[论文解读] Length-Adaptive Interest Network for Balancing Long and Short Sequence Modeling in CTR Prediction

Zhicheng Zhang, Zhaocheng Du|arXiv (Cornell University)|Jan 27, 2026
Recommender Systems and Techniques被引用 0
一句话总结

LAIN 引入了面向长度的条件化来 CTR 模型,使用光谱长度编码器、长度条件提示和长度调制注意力,在整体 CTR 和短序列用户表现上都实现了提升,开销极小。

ABSTRACT

User behavior sequences in modern recommendation systems exhibit significant length heterogeneity, ranging from sparse short-term interactions to rich long-term histories. While longer sequences provide more context, we observe that increasing the maximum input sequence length in existing CTR models paradoxically degrades performance for short-sequence users due to attention polarization and length imbalance in training data. To address this, we propose LAIN(Length-Adaptive Interest Network), a plug-and-play framework that explicitly incorporates sequence length as a conditioning signal to balance long- and short-sequence modeling. LAIN consists of three lightweight components: a Spectral Length Encoder that maps length into continuous representations, Length-Conditioned Prompting that injects global contextual cues into both long- and short-term behavior branches, and Length-Modulated Attention that adaptively adjusts attention sharpness based on sequence length. Extensive experiments on three real-world benchmarks across five strong CTR backbones show that LAIN consistently improves overall performance, achieving up to 1.15% AUC gain and 2.25% log loss reduction. Notably, our method significantly improves accuracy for short-sequence users without sacrificing longsequence effectiveness. Our work offers a general, efficient, and deployable solution to mitigate length-induced bias in sequential recommendation.

研究动机与目标

  • 识别当前 CTR 模型中长序列用户与短序列用户之间的性能不平衡。
  • 提出一个长度自适应框架,在不同序列长度之间平衡兴趣建模。
  • 证明长度条件化在不牺牲长序列性能的情况下提升短序列精度。
  • 展示 LAIN 在多种骨干网络和数据集上的鲁棒性与可部署性。

提出的方法

  • 引入 Spectral Length Encoder,将原始序列长度映射到可训练傅里叶基的连续嵌入。
  • 开发 Length-Conditioned Prompting,在短期和长期行为序列前置长度感知提示。
  • 实现 Length-Modulated Attention,基于长度先验对查询/键进行条件化并调整 softmax 温度。
  • 通过提示解耦共享参数与长度特定参数,建立长度相关的归纳偏置。
  • 提供端到端训练,参数/计算开销微小。
  • 在三个真实世界数据集与五个强 CTR 骨干网络上评估 LAIN。

实验结果

研究问题

  • RQ1显式对序列长度进行条件化是否能够缓解 CTR 模型中的注意力偏极化和长度信号不足?
  • RQ2LAIN 是否在不同骨干与数据集上提升短序列用户 CTR,同时不损害长序列性能?
  • RQ3每个 LAIN 组件(SLE、LCP、LMA)及其相互作用的经验影响如何?
  • RQ4对于超参数选择和不同序列长度分布,LAIN 的鲁棒性如何?

主要发现

ModelEBNeRD-small GAUCEBNeRD-small AUCEBNeRD-small loglossKuaiVideo GAUCKuaiVideo AUCKuaiVideo loglossMicroVideo1.7M GAUCMicroVideo1.7M AUCMicroVideo1.7M logloss
DIN0.70530.70800.26950.67160.69790.44980.70230.70930.4196
+LAIN0.70960.71560.26780.67420.70430.44570.70190.71010.4179
Rel. Gain0.61%1.07%-0.65%0.40%0.93%-0.92%-0.05%0.12%-0.40%
DIEN0.70900.70990.27370.66780.69770.45090.71430.71850.4171
+LAIN0.70780.71180.26760.66880.69910.44760.71460.72190.4161
Rel. Gain-0.17%0.27%-2.25%0.14%0.20%-0.71%0.04%0.47%-0.23%
SIM0.69600.69920.27200.66720.68750.45740.70170.70950.4168
+LAIN0.69900.70370.27060.66780.68960.45660.70700.71310.4151
Rel. Gain0.43%0.65%-0.50%0.08%0.31%-0.16%0.75%0.50%-0.41%
SDIM0.70390.70990.27190.67290.69240.45360.69840.71610.4129
+LAIN0.70990.71230.27040.67320.69490.45250.69930.71630.4121
Rel. Gain0.85%0.34%-0.56%0.05%0.35%-0.23%0.13%0.03%-0.18%
TWIN0.69300.69930.27180.67290.69180.45300.70600.71580.4164
+LAIN0.70120.70740.26980.67480.69760.45080.70930.72330.4097
Rel. Gain1.19%1.15%-0.73%0.29%0.84%-0.48%0.47%1.05%-1.63%
  • LAIN 在不同骨干和数据集上持续提升整体 CTR 性能,GAUC 最高提升至 +1.2%,logloss 最多降低至 -1.6%。
  • LAIN 显著提升短序列用户(少于 100)的准确性,在 MicroVideo1.7M 的 TWIN 上 AUC 提升至 +1.08%,logloss 降至 -2.17%。
  • 注意力极化被 LAIN 缓解,在短、中、长序列下的 Gini 系数显著低于基线。
  • 消融研究显示所有组件均有贡献;移除 LMA 会导致最大的性能下降。
  • 超参数敏感性分析显示在不同配置下仍能获得鲁棒提升(选定设置:傅里叶维度 64,隐藏层 512,提示数量 4)。

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。