AutoLossGen: Automatic Loss Function Generation for Recommender Systems
AutoLossGen: Automatic Loss Function Generation for Recommender Systems
复制标题
DOI：
10.1145/3477495.3531941
复制
发表时间：
2022-04
期刊：
影响因子：
--
通讯作者：
Zelong Li;Jianchao Ji;Yingqiang Ge;Yongfeng Zhang
中科院分区：
文献类型：
--
作者：
Zelong Li;Jianchao Ji;Yingqiang Ge;Yongfeng Zhang
在推荐系统中，损失函数的选择是至关重要的，因为良好的损失可以显着提高模型的性能。然而，由于问题的复杂性，手动设计良好的损失是一个很大的挑战。之前的大部分工作都集中在手工制作的损失函数上，这需要大量的专业知识和人力。在本文中，受到自动机器学习的最新发展的启发，我们提出了一个自动损失函数生成框架AutoLossGen，它能够直接从基本的数学运算符中生成损失函数，而无需损失结构的先验知识。更具体地说，我们开发了一个由强化学习驱动的控制器模型来生成损失函数，并开发了迭代和交替优化计划来更新控制器模型和推荐器模型的参数。推荐系统中自动损失生成的一个挑战是推荐数据集的极端稀疏性，这导致了损失生成和搜索的稀疏奖励问题。为了解决这个问题，我们进一步开发了一个奖励过滤机制，以实现高效和有效的损失生成。实验结果表明，我们的框架管理创建量身定制的损失函数，为不同的推荐模型和数据集，生成的损失提供了更好的推荐性能比常用的基线损失。此外，所产生的大部分损失是可转移的，即，基于一个模型和数据集生成的损失对于另一个模型或数据集也很有效。该工作的源代码可在https://github.com/rutgerswiselab/AutoLossGen上获得。
In recommendation systems, the choice of loss function is critical since a good loss may significantly improve the model performance. However, manually designing a good loss is a big challenge due to the complexity of the problem. A large fraction of previous work focuses on handcrafted loss functions, which needs significant expertise and human effort. In this paper, inspired by the recent development of automated machine learning, we propose an automatic loss function generation framework, AutoLossGen, which is able to generate loss functions directly constructed from basic mathematical operators without prior knowledge on loss structure. More specifically, we develop a controller model driven by reinforcement learning to generate loss functions, and develop iterative and alternating optimization schedule to update the parameters of both the controller model and the recommender model. One challenge for automatic loss generation in recommender systems is the extreme sparsity of recommendation datasets, which leads to the sparse reward problem for loss generation and search. To solve the problem, we further develop a reward filtering mechanism for efficient and effective loss generation. Experimental results show that our framework manages to create tailored loss functions for different recommendation models and datasets, and the generated loss gives better recommendation performance than commonly used baseline losses. Besides, most of the generated losses are transferable, i.e., the loss generated based on one model and dataset also works well for another model or dataset. Source code of the work is available at https://github.com/rutgerswiselab/AutoLossGen.