摘要
点击率预测模型是主流推荐系统中十分重要的部分.根据点击率预测的打分来调整商品的展示策略,对提高业务的转化率、改进用户体验等有着重要的意义.传统的点击率预测模型是利用用户特征和商品特征,对点击率进行预测.然而,用户行为序列的结构特征,如周期性规律、趋势等也能一定程度地体现用户行为的倾向.针对部分信息利用上的空缺,使用时间序列分析单元,将提取用户行为序列的特征作为用户特征的扩展,结合因子分解机结构将其与用户、商品特征进行交叉,能够有效提高特征质量,优化点击率预测模型的性能.实验表明,结合用户行为序列特征进行交叉优化的方法能够对点击率预测模型的表现带来很大提升,提高点击率预测的精度.
The click-through rate(CTR)prediction model is an important component of mainstream recommendation systems.The model assigns a score to recommended items according to the predicted CTR and generates an optimized scoring function which in-turn influences an item’s display strategy;this helps generate improved business conversion rates and a better user experience.Generally,CTR prediction models utilize both user and item features to predict CTR.However,structural characteristics of user behavior,such as frequency and trends,can also reflect behavioral tendencies.Given the absence of this information,this paper analyzes user behavior sequences as a time series and extracts latent features.Factorization machines are then used to learn from user/item features combined with sequence features to improve the quality of prediction.Experiments show that the sequence feature-based methods improve the performance of CTR prediction models and make CTR prediction more accurate.
作者
朱思涵
浦剑
ZHU Sihan;PU Jian(School of Computer Science and Technology,East China Normal University,Shanghai 200062,China)
出处
《华东师范大学学报（自然科学版）》
CAS
CSCD
北大核心
2020年第4期134-146,共13页
Journal of East China Normal University(Natural Science)
基金
国家自然科学基金(61702186)。
关键词
点击率预测
推荐系统
自回归滑动平均模型
因子分解机
click-through rate(CTR)prediction
recommendation system
auto-regressive moving average model(ARMA)
factorization machine
