基于级联过滤的多模型融合的推荐方法

Recommendation Method of Multi-model Combination Based on the Cascaded Filtering

针对当前主流推荐算法无法甄别离群样本和弱贡献率样本,且单模型算法泛化能力较弱等问题,提出一种基于级联过滤的多模型融合的推荐方法.该方法先采用级联回归模型过滤掉离群样本和弱贡献率样本;然后,把推荐问题抽象成二分类问题和回归问题,分别采用基于Bagging的随机森林和基于Boosting的梯度提升回归树两种树型算法、线性的逻辑回归算法来拟合用户兴趣;最后,将这三种算法分别训练若干模型进行线性融合,取Top-N推荐.实验表明,该方法不仅有效提高了推荐精度,还增强了模型的泛化能力,具有较强的实用价值.

The current recommendation algorithms are unable to identify outlier and weak-contribution-rate samples, and single modelalgorithms show poor generalization ability, which seriously affect the quality of recommendation. To solve these problems, an im-proved method named Recommendation method of Multi-Model Combination based on the Cascaded Filtering is presented. This meth-od firstly uses cascade regression model to filter out outlier and weak-contribution-rate samples. Then the recommendation problem isabstracted into two-category and regression problems, respectively, the random forest algorithm based on bagging and gradient boostingregression trees based on boosting, logistic regression algorithm are used to fit the user＇s interest. Finally, several models of each algo-rithm are trained for linear fusion to do Top-N recommendation. The experiments demonstrate this method not only improves the accu-racy of recommendation, but also enhances the model generalization ability, and has a strong practical value.