【作者】 刘炯；

【导师】 姚天昉； 吕栋；

【作者基本信息】 上海交通大学 ， 计算机技术（专业学位）， 2018， 硕士

【摘要】 随着互联网技术的飞速发展,我们逐步迈入大数据时代。近年来,不少社区、论坛等互联网社交、问答的平台迅速崛起。由于互联网上数据量飞速增长,使得用户无从快速地选择所需要的信息。因此,帮助用户快速地搜索到符合用户喜好的信息变得尤为重要。信息推荐技术是实现这一需求的重要手段。通过挖掘、分析用户的历史行为数据、以及社区里的评论文本数据,推荐系统可以快速定位到用户所需要的信息,为用户提供可信的推荐结果。这样一来,用户检索信息的时间就可以大大减少了,而且用户的产品体验也改善了。因此,推荐功能也就起到了改善用户产品体验满意程度的作用,也为用户带来了更方便、更快捷的生活。协同过滤推荐方法是最早的推荐算法之一。其算法原理是:通过参考用户的历史行为和文章的不同属性来预测用户可能偏好的文章,选取预测得分较高的文章推荐给用户。但是,传统的推荐算法往往存在一些问题。其主要方面归纳如下:(1)数据稀疏性问题;(2)冷启动(Cold Start)问题;(3)同质化问题。为了解决传统的推荐算法所存在的问题,本文的研究工作提出了一种基于传统协同过滤的改进方法。首先,使用了皮尔森相似度对用户打分进行了归一化,剔除了零得分的文章和由于用户使用习惯差异引入的噪声。。另外,基于隐语义模型的推荐算法解决了在推荐上的稀疏性问题。除此之外,用户历史行为数据只能捕捉到一部分用户的行为习惯,对于没有游览、点击行为的用户便遗漏了信息。因此,本文还提出了使用组合推荐策略的方法。结合协同过滤算法和LDA主题模型相融合的推荐算法,一方面从用户评论数据中挖掘出用户的兴趣偏好,根据LDA主题模型中用户在不同主题上的概率用来计算用户的相似度。另一方面,改进的协同过滤算法用来计算文章的相似度。结合这两方面的信息,从不同维度上搜集更全面的用户信息用于推荐服务。这样的组合策略大大减轻了推荐中冷启动问题和同质化问题的影响。通过对实验结果分析得知,基于组合模型的推荐算法在准确率、召回率和多样性方面的优于传统推荐算法。为了实验验证组合模型的推荐效果,我们采集某企业的真实数据作为数据集,通过对比多种算法,并在多种评估指标上进行了实验。实验结果表明:基于融合模型的推荐算法在上述数据集上取得了良好的成绩。其准确性、查全率和广泛性都获得了提高。因此,该算法明显优于其他传统的推荐算法,从而证明了融合组合模型的推荐算法具有可行性、合理性和有效性。更多还原

【Abstract】 With the rapid development of Internet technology,we are gradually entering the era of big data.In recent years,many communities,BBS,Internet social,question and answer platform etc.are rising rapidly.Because of the rapid growth of the amount of data on the Internet,which makes users fail to select the information needed quickly.Therefore,It becomes quite urgent to help users to search the information which meets users’ preferences quickly.Information recommendation technology is the important means to realize this requirement.By mining,analysis of user behavior data,as well as the comment text data in the community,recommendation system can quickly locate the information which user needs,and provide users with reliable results.As a result,not only the time to retrieve information can be greatly reduced,but also the user experience was of the product was improved.Therefore,the recommendations have played an important role in improving user satisfaction with product experience,also provided users more convenient and faster life.Collaborative filtering recommendation technology is one of the most common recommendation methods.By reference to the history of user behavior and articles’ different attributes to predict users’ favorite articles,we can select highest scores articles to recommend to the user.The research work of this thesis is to put forward to improve and optimize the traditional collaborative filtering technology.Although traditional collaborative filtering technology is widely used,it has following disadvantages:(1)sparsity of the rating matrix.In the large-scale recommendation system,both the number of user and number of item are very large.But the number of items in fact users expressed preferences is quite few.That is to say,the known data is very sparse.Sparsity leads to the deviation in the process of calculation,such as calculating the similarity between users or items when making recommendations,but sparsity makes the calculating of similarity between users or items inaccuracy.(2)Cold Start.New user with few rating record,so it is difficult to analyze his preference and it is impossible to make effectively recommendation for him.(3)Homogenization.Users’ interests are diverse,the explicit or implicit expressed interest is very limited,the recommendation method based on content filter recommends items which matches with current interest,therefore,this recommendation results homogenization,it is hard for items which users didn’t express preference but actually interested to get recommendations.In view of the above problems of the traditional collaborative filtering,this thesis puts forward the corresponding improved algorithm:Latent Factor Model(LFM)connects user interests and articles by implicit features.LFM splits the two-dimensional users-articles matrix into user-hidden theme matrix and hidden theme-article matrix,which avoids the occurrence of zero that makes similarity inaccuracy.Use provided data such as age,gender in user registration to do coarse granularity personalized recommendation.Asking the user to feedback of some items when logging in,collecting user’s interest in these items,then recommends the similar items to the user.In feature extraction and similarity calculation,the use of TFIDF,Word2 Vec and some other methods will be applied in the research.Through the experimental analysis,it is found that introducing category features can improve the accuracy of the article recommendation.To verify the recommend effect of combined model,we collected some real data from enterprise as datasets,by comparing various algorithms,and experiment on a variety of performance metrics,recommendation algorithm based on combined model achieved great score in the above datasets.The accuracy,recall and diversity have been improved.As a result,the algorithm was superior to other traditional recommendation algorithm was proposed,which proved the effectiveness of the recommendation algorithm based on combined model.更多还原