一个基于低通图滤波器的推荐模型

A Recommendation Model Based on Low-pass Graph Filters

推荐系统已经成为大数据时代帮助用户挖掘其偏好的有力工具,也创造了大量的经济价值.在实际推荐场景下,尽管用户的偏好或者项目(也称“商品”)的特性都是相对稳定的,可以通过用户与商品的历史交互来捕获.但是,用户对商品存在误触的点击行为,这实际上是噪音信号.如何对用户-商品进行降噪并学习精确的用户偏好是推荐系统的基本需求.我们将用户-商品交互建模为二部图,并从图信号处理的角度设计了一种低通图滤波器,其可以抑制和过滤高频噪音并筛选出低频的用户偏好.最后,2个真实数据集上的大量实验验证了所设计算法的有效性.

Recommendation systems have become useful tools in the era of big data to mine user preferences with economic value. In recommendation applications, user preferences or characteristics of commodity are relatively stable and can be gathered through historical interactions between user and commodity. Users may also make false clicks and create noisy signals. A basic requirement of recommendation system is to reduce noises in user-commodity interactions and mine user preferences accurately. In this paper, the user-commodity interaction is modeled as a bipartite graph. A low-pass graph filter is designed to filter noises in high-frequency and retain user preferences in low-frequency.Experiments verify the effectiveness of the algorithm in applying to open data sets.