论文部分内容阅读
为了吸引用户,最新的推荐算法注重于所推荐物品的新颖性和推荐列表的多样性.而传统的基于协同过滤的推荐算法只专注于提高准确性使得推荐的物品列表种类单一,因此在保持准确性的同时寻找新颖多样的物品列表成为研究热点.大多数现有研究提出的模型分为两阶段:先优化准确率后优化多样性.由于优化目标(多样性和准确性)的冲突,两阶段优化模型只能在牺牲准确性的情况下生成多样化的推荐列表.因此,文章提出了一个新的矩阵分解模型,该模型可以同时优化新颖性、多样性和准确性三个目标.此外,还设计了两个新的约束项:第一个约束项使目标用户的隐因子向量接近那些对长尾物品评过分的用户的平均隐因子向量,从而提高了推荐的新颖性;另一个约束项使每个物品的隐因子向量接近所有物品隐因子向量的均值,从而使推荐列表多样化.为了验证所提模型的有效性,我们在Movielens100K,Epinions和Rym数据集上进行了综合实验.实验结果表明,在准确性、系统多样性、个体多样性和新颖性方面,所提模型均具有卓越的性能.“,”Modern recommendation algorithms focus on novel items and diverse recommendation list for at-tracting users.Because a collaborative filtering based recommendation algorithm usually generates similar items for accuracy,it is a challenge to find novel and diverse items while keeping accuracy.Most of the ex-isting studies developed two-step recommendation models that optimize accuracy first and then diversity,and the two-step optimized model generated diverse items at the sacrifices of accuracy due to the conflict of the optimization goals(diversity and accuracy).We propose a new matrix factorization model,that simul-taneously optimizes novelty,diversity and accuracy.The new constraint that makes the latent vector of the target user close to the average latent factors of the users who have rated long-tail items was developed for novel recommendations.And the other new constraint that makes each item latent close to the mean of all i-tem latent,was designed for diversity recommendation lists.The comprehensive experiments were conduc-ted on the Movielens100K,Epinions and Rym dataset.Experimental results demonstrated the superior performance in terms of accuracy,aggregate diversity,individual diversity and novelty to the state of the art models.