中文

面向点击率预测的高效 LLM 训练范式

信息检索 2025-03-18 v3

摘要

大型语言模型(LLM)已显示出作为下一代基于排序的推荐系统的巨大潜力。许多最近的工作表明,LLM 显著优于传统的点击率(CTR)预测方法。尽管如此,当前训练范式中固有的计算效率低下,使得在大数据集上进行基于排序的推荐任务训练 LLM 具有挑战性。为了进行 CTR 预测,现有大多数研究采用流行的“滑动窗口”范式。给定长度为 mm 的用户交互序列,为每个交互构建唯一的训练提示,通过将其指定为预测目标以及其前 nn 个交互作为上下文。在此基础上,滑动窗口范式导致整体复杂度为 O(mn2)O(mn^2),随用户交互长度线性增长。因此,直接采用此策略训练 LLM 可能会因交互长度增长而导致训练成本高昂。为了缓解计算效率低下,我们提出一种新颖的训练范式,即动态目标隔离(Dynamic Target Isolation, DTI),该范式结构化地并行训练 kk(其中 k>>1k >> 1)个目标交互。 Furthermore, we identify two major bottlenecks - hidden-state leakage and positional bias overfitting - that limit DTI to only scale up to a small value of kk (e.g., 5) then propose a computationally light solution to effectively tackle each. Through extensive experiments on three widely adopted public CTR datasets, we empirically show that DTI reduces training time by an average of \textbf{92%} (e.g., from 70.570.5 hrs to 5.315.31 hrs), without compromising CTR prediction performance.

关键词

引用

@article{arxiv.2503.01001,
  title  = {Towards An Efficient LLM Training Paradigm for CTR Prediction},
  author = {Allen Lin and Renqin Cai and Yun He and Hanchao Yu and Jing Qian and Rui Li and Qifan Wang and James Caverlee},
  journal= {arXiv preprint arXiv:2503.01001},
  year   = {2025}
}