基于矩阵分解的单类协同过滤推荐算法

新闻网页和书签的推荐被认为是单类协调过滤问题。通常这类数据是相当稀疏的,仅仅一小部分数据是正例,在非正例数据中负例和没有标记的正例是混合在一起的,难以区分开来,因此,就如何解释非正例数据出现了歧义。为了解决该问题,提出了一种加权的带正则化的基于迭代最小二乘法的单类协同过滤算法。即通过对正例赋予权值1,负例赋予一个较小的正实数权值来反映数据的正负置信度。在两个真实的实验数据集上验证了该算法在性能上均优于几个经典的单类协同过滤推荐算法。

One-class collaborative filtering based on matrix factorization

News item recommendation and bookmarks recommendation are most naturally thought of as OOCF problems. Usually this kind of data are extremely sparse, just a small fraction are positive examples. Negative examples and unlabeled positive examples are mixed together and are typically unable to distinguish them, therefore ambiguity arises in the interpretation of the non-positive example. This paper proposed a CF algorithm-weighted alternating least squares(wALS). That was, by using weighting scheme assigning "1" to observed examples and low positive real number weights to unobserved examples to reflect the confidence of positive examples and negative examples. The experimental evaluation using two real-world datasets shows that wALS achieves better results in comparison with several classical one-class collaborative filtering recommendation algorithms.