Recommender system extracts useful information from massive data to learn user preferences. The current mainstream recommended algorithms are collaborative filtering algorithms which based on ratings data and social data.Collaborative filtering method contains traditional user or commodity collaborative filtering algorithms and matrix decomposition methods.User or commodity collaborative filtering algorithmis interpretable,but is slow in training;Matrix decomposition method is fast but lack of concrete interpretation.Moreover,rating data and social data are more difficult to obtain than attribute information.Aiming at these problems,this paper proposes a recommend method based on attributes preference self-learning.Based on rating data and attribute information,the algorithm firstly uses the user's initial preference for attribute and attribute rating for commodity to construct predict rating model,then utilizes normalization coefficient and the square difference of predict rating and actual rating to construct loss function,uses gradient descent method to train attribute preference of predict rating model in loss function.Finally,the trained attribute preference and attribute rating are used to predict user's ratings.The experimental results show that the model training is faster and is superior to the traditional collaborative filtering methods and basic matrix decomposition model in the case of ratings data sparse.%推荐系统从海量数据中挖掘出有用信息来学习用户偏好.目前主流研究的推荐算法是在考虑评分数据和社交数据的基础上执行协同过滤算法.协同过滤方法包括传统的用户、商品协同过滤算法和经典的矩阵分解方法.用户、商品协同过滤算法具有可解释性但训练速度慢,矩阵分解模型虽然训练速度快但缺乏解释性.此外,评分数据和社交数据在获取难度上高于属性信息.针对这些问题,提出一种基于属性偏好的自学习算法.该算法在评分数据和属性信息的基础上,首先利用用户对属性的初始偏好程度与属性对商品的评分构建预测评分模型,通过预测评分和实际评分的平方差及正规化项构造损失函数,使用梯度递减方法对损失函数中预测评分模型的属性偏好程度进行迭代训练,最后使用训练后的属性偏好程度和属性值评分来预测用户评分.实验在两个经典数据集上证明了该模型运行时间较快,且在评分数据稀疏的情况优于传统的协同过滤方法和基本矩阵分解模型.
展开▼
