一种基于Meta-learning改进的特征交互算法  

Improved Feature Interaction Algorithm Based on Meta-learning

作　　者：白静 耿新宇[1] 易流 穆禹锟 陈琴 宋杰 BAI Jing;GENG Xinyu;YI Liu;MU Yukun;CHEN Qin;SONG Jie(School of Computer Science,Southwest Petroleum University,Chengdu 610000,China)

机构地区：[1]西南石油大学计算机科学学院,成都610000

出　　处：《计算机科学》2023年第S02期606-613,共8页Computer Science

基　　金：四川省科技计划项目(2022NSFSC0555)。

摘　　要：特征交互在推荐系统领域的广告点击率(Click-Through Rate,CTR)预测任务中至关重要,当前业界做的特征交互往往是基于内积、外积等矩阵变换,这些操作没有引入额外的信息,可以作为衡量两个向量相似性的手段,但作为特征交互的表示不一定是可靠的,许多特征交互无法有效提高点击率预测性能。首先从改善特征交互方式的角度入手引入额外的参数来学习一个映射,假设这个映射能够将两个向量的表征映射成交互的表征。学习映射的过程能够通过元学习(Meta-learning)来实现,故构建一个学习器以函数的方式表征特征交互。另外,不同的特征对不一定采取相同的方式交互,不能通过同一种交互方式得到所有特征对,因此设计一组元学习器(meta-learner)来学习映射函数,引入门控网络(GateNet)学习模型中元学习器的分布,那么不同的特征嵌入可以由一组元学习器得到表征。基于以上两点提出了一种融合多个元学习器并结合门控网络(Multiple meta-learners combined with GateNet,gate-MML)的特征交互算法,通过学习不同特征的联系和差异提高每个特征交互的质量。为了验证所提算法的性能,在xDeepFM模型上采用gate-MML做进一步的特征交互,采用2个真实广告点击率预测的数据集进行实验,并使用Logloss作为损失函数,AUC作为评价指标。实验结果表明与传统的CTR预测模型相比,改进算法提升了广告点击率预测任务的预测性能。Feature interaction is crucial in the field of advertising click-through rate(CTR)prediction in recommendation systems.However,current industry practices for feature interaction often rely on matrix transformations such as inner and outer products,which do not introduce additional information and can only serve as a means of measuring the similarity between two vectors.Therefore,such methods may not reliably represent feature interaction and may not effectively improve the performance of CTR prediction.To address this issue,this paper first introduces additional parameters to learn a mapping from the perspective of improving the feature interaction,assuming that this mapping can map the representation of two vectors to the representation of interaction.The process of learning mapping can be achieved through meta-learning,which constructs a learner to represent feature interactions in a functional manner.Additionally,different features may not adopt the same interaction method,and it is impossible to obtain all feature pairs through a single interaction method.Therefore,a set of meta-lear-ners is designed to learn the mapping function,and a GateNet is introduced to learn the distribution of meta-learners in the model,so that a set of meta-learners can represent different feature embeddings.Based on these two points,a feature interaction algorithm is proposed that combines multiple meta-learners with GateNet(gate-MML),which improves the quality of each feature interaction by learning the connections and differences between different features.To verify the performance of the proposed algorithm,gate-MML is used for further feature interaction in the xDeepFM model,and experiments are conducted on two real advertising click-through rate prediction datasets using Logloss as the loss function and AUC as the evaluation metric.Experimental results show that compared with traditional CTR prediction models,the improved algorithm enhances the prediction performance of advertising click-through rate prediction tasks.

关 键 词：特征交互 广告点击率预测 元学习 门控网络 推荐系统 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]