yuehuda koren 是yahoo推荐系统方面的大牛。Advances inCollaborative Filtering是他2011年的文章。
下面简单介绍一下:
大量的人对推荐系统感兴趣,研究机构、企业等大量投入推荐系统的研究和应用。
推荐系统的输入:
explicitfeedback:打分;点击向上/向下的大拇指;
implicitfeedback:隐式的反馈;购买、浏览、搜索模式,甚至鼠标的移动。
推荐系统需要联系两种实体。items和users。
CF有两种方法:
the neighborhoodapproachandlatent factor models
the neighborhood approach:关键是找相关的item,相关的user。
Netflix的数据情况:
Netflix customers between Nov11, 1999 and Dec 31, 2005 [5]. Ratingsare integers ranging between 1 and 5. The data spans
17,770 movies rated by over480,000 users. Thus, on average, a movie receives 5600 ratings, while a userrates 208 movies, with substantial variation around each
of these averages.
SVD:
最小化的两种解法:Minimization is typically performed by either stochastic gradient descent or alternating least squares。SGD和ALS方法