基于改进权重计算的协同过滤算法研究

发布时间：2018-08-29 19:27

【摘要】：自进入21世纪以来,互联网技术发展迅速。随着互联网的普及,电子商务逐渐兴起,网络信息传播等亦得到了长足发展。随之而来的是,形形色色的商品或者网络信息铺天盖地地呈现在我们的面前,面对种类和数量如此繁多的商品或信息,用户很难快速选出最想要的内容。个性化推荐技术应运而生,它是解决上述问题的有效途径之一,其功能主要是通过收集和分析用户的历史浏览信息,为用户推荐出其可能感兴趣的产品。一个功能完善的推荐系统通常是由三部分组成:记录板块,主要作用是记录用户的历史行为;分析建模板块;算法板块。推荐算法作为整个推荐系统的核心,近年来其相关研究内容成为推荐系统研究的热点方向,因为推荐结果的好坏与推荐算法的性能密切相关。由于不同用户的推荐需求各异,推荐工作就要具有个体化差异。实现推荐需求个性化所用的技术和算法复杂多样,就目前而言,大概有以下4种,基于内容的推荐算法、协同过滤推荐算法、混合推荐算法以及网络推荐算法。协同过滤推荐技术是个性化推荐技术中应用领域比较广的成功的技术之一,它是利用集体智慧,在海量的用户中发掘出一小部分兴趣爱好比较类似的用户即“邻居”,根据分析记录这些“邻居”喜欢的其他内容,生成一个具有排序的目录,我们称之为推荐结果,把推荐结果推送给这群用户,从一定程度上减少了用户“挑选”过程的工作量。传统的协同过滤算法在计算相似用户或者相似项目时,未将用户行为时间或者项目之间具有相同的标签等因素考虑在相似性计算中,而是直接使用用户评分进行相似性计算,这就暴露出如新用户或新产品加入后的冷启动问题、稀疏矩阵问题、推荐的可扩展性问题等等,导致推荐结果精度不高,难以满足用户的实际需求。本研究是在基于项目的协同过滤算法基础之上,将用户行为时间以及项目本身所具有的标签属性等信息纳入相似性计算之中,改善新用户或者新产品进入后的冷启动问题,进而提高推荐结果的质量,尽可能满足不同用户的实际需求,实现个性化推荐服务。具体研究如下:首先,在实验数据准备过程中,将用户对资讯新闻、短视频等信息产生的行为时间引入到数据集中,并在计算项目之间相似性时,将时间因子经过时间衰减函数的预处理,整合到资源热度的热度评分上,然后再进行基于项目的协同过滤推荐。其次,针对新入资源库的项目得不到合适的推荐权重,利用短视频项目本身具有标签这一特性,将短视频标签引入到相似性计算的过程中。因为短视频在发布之前会被预先定义标签,所以在提取短视频标签后,对标签进行空间余弦相似度(Cosine Similarity)计算,对新入库项目进行预测资源热度评分,然后再将该资源的发布者的权威性进行权重调节,然后使用基于项目的协同过滤进行推荐。最后,在基于用户对资讯APP的真实行为日志上,设计实验方案,对比经典协同过滤算法与改进后的协同过滤的推荐结果,验证改进后算法的性能。通过实验结果,我们不难发现,改进后的协同过滤算法确实改善了新用户或新产品进入后的冷启动问题,推荐精度在一定程度上确实得到了提高。
[Abstract]:Since the beginning of the 21st century, Internet technology has developed rapidly. With the popularity of the Internet, e-commerce has gradually risen, and network information dissemination has also been greatly developed. Personalized recommendation technology emerges as the times require. It is one of the effective ways to solve the above problems. Its function is to recommend products that users may be interested in by collecting and analyzing users'historical browsing information. As the core of the whole recommendation system, recommendation algorithm has become a hot research direction in recent years, because the recommendation results are closely related to the performance of the recommendation algorithm. The techniques and algorithms used to personalize the recommendation requirements are complex and varied. At present, there are about four kinds of recommendation algorithms: content-based recommendation algorithm, collaborative filtering recommendation algorithm, hybrid recommendation algorithm and network recommendation algorithm. One of the more successful technologies in the field of application is to use collective wisdom to discover a small number of users with similar interests and hobbies, namely "neighbors". According to the analysis and recording of other content that these "neighbors" like, a catalog with ranking is generated, which we call recommendation results, and recommendation results are obtained. Pushing to this group of users reduces the workload of the user's "selection" process to a certain extent. The traditional collaborative filtering algorithm does not consider such factors as user's behavior time or the same label between items in the similarity calculation, but directly uses the user's score for similarity. Sexual computing, which exposes such as cold start-up problems, sparse matrix problems, recommendation scalability problems, and so on, leads to the recommendation results are not accurate enough to meet the actual needs of users. This study is based on the Project-based Collaborative Filtering algorithm, the user behavior time and the project itself. Information such as tag attributes is included in similarity calculation to improve the cold start problem of new users or new products, and then improve the quality of recommendation results, satisfy the actual needs of different users as much as possible, and realize personalized recommendation service. The behavior time generated by news, short video and other information is introduced into the data set. When calculating the similarity between items, the time factor is integrated into the heat score of resource heat through the pretreatment of time attenuation function, and then the item-based collaborative filtering recommendation is carried out. Secondly, the new items are not suitable for the new ones. Recommended weights are used to introduce short video labels into similarity computation by using the tagging feature of short video items. Since short video labels are pre-defined before publishing, the spatial cosine similarity (Cosine Similarity) of the labels is calculated after extracting the short video labels. Finally, an experimental scheme is designed based on the user's log of the actual implementation of the information APP. The proposed scheme is validated by comparing the recommendation results of the classical collaborative filtering algorithm with the improved collaborative filtering algorithm. Experimental results show that the improved collaborative filtering algorithm improves the cold start problem of new users or new products, and the recommendation accuracy is improved to a certain extent.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】