鲁棒的单类协同排序算法

单类协同过滤(One-class collaborative filtering, OCCF)问题是当前的一大研究热点.之前的研究所提出的算法对噪声数据很敏感,因为训练数据中的噪声数据将给训练过程带来巨大影响,从而导致算法的不准确性.文中引入了Sigmoid成对损失函数和Fidelity成对损失函数,这两个函数具有很好的灵活性,能够和当前最流行的基于矩阵分解(Matrix factorization, MF)的协同过滤算法和基于最近邻(K-nearest neighbor, KNN)的协同过滤算法很好地融合在一起,进而提出了两个鲁棒的单类协同排序算法,解决了之前此类算法对噪声数据的敏感性问题.基于Bootstrap抽样的随机梯度下降法用于优化学习过程.在包含有大量噪声数据点的实际数据集上实验验证,本文提出的算法在各个评价指标下均优于当前最新的单类协同排序算法.

Robust Ranking Algorithms for One-class Collaborative Filtering

The problem of ranking for one-class collaborative filtering (OCCF) is a research focus. One drawback of the existing ranking algorithms for OCCF is noise sensitivity, because the noisy data of training data might bring big influences to the training process and lead to inaccuracy of the algorithm. In this paper, in order to solve the noise sensitivity problem of the ranking algorithms, we propose two robust ranking algorithms for OCCF by using the pairwise sigmoid/fidelity loss functions that are flexible and can be easily adopted by the popular matrix factorization (MF) model and the K-nearest-neighbor (KNN) model. We use stochastic gradient descent with bootstrap sampling to optimize the two robust ranking algorithms. Experimental results on three practical datasets containing a large number of noisy data show that our proposed algorithms outperform several state-of-the-art ranking algorithms for OCCF in terms of different evaluation metrics.