实时新闻推荐系统的设计与实现

发布时间：2018-04-08 19:15

本文选题：新闻推荐　切入点：推荐系统　出处：《北京交通大学》2017年硕士论文

【摘要】：信息的指数爆炸带来了信息过载问题,从而产生了分类目录技术和搜索引擎技术,然而分类目录只能覆盖热门分类,搜索引擎只能由用户主动输入关键词检索信息,于是个性化新闻推荐系统应运而生。单一的算法很难从多个角度为用户进行推荐,易导致推荐结果多样性欠缺。为提高推荐的准确率和多样性,本文就目前已有的推荐算法展开研究,结合传统的推荐技术设计了混合加权的新闻推荐策略,将基于内容的推荐算法和基于用户的协同过滤算法按不同权重进行加权混合,使之达到取长补短的目的,提高了推荐结果的准确性,更好的为用户进行个性化的新闻推荐。本文将新闻内容建模、用户兴趣建模和混合算法建模三部分作为推荐系统的核心内容。对于新闻内容建模,首先介绍了新闻文本预处理的相关理论,针对新闻内容的特点,采取线性加权的方式进行新闻关键词的提取,并使用支持向量机实现了对新闻的分类;对于用户兴趣建模,通过对用户行为日志的收集,分析用户的新闻浏览偏好,进而完成对用户兴趣模型的建立与更新;对于混合算法建模,基于内容的推荐算法主要通过计算新闻内容向量和用户兴趣向量的夹角余弦相似度确定新闻推荐列表,基于用户的协同过滤算法通过建立用户相似度矩阵来推荐相似用户喜欢的新闻,然后将两者召回的结果按不同权值进行加权混合,并通过多次训练得出加权效果最好的权值比,确保推荐系统的准确性。另外还设置了新闻时间阀值,对推荐返回的结果进行适当过滤,在一定程度上保障了推荐结果的时效性。论文首先通过介绍系统的背景意义及国内外研究现状确立了基本工作内容,然后就典型推荐算法进行详细描述,分析了系统需求。针对新闻推荐系统数据规模大、用户兴趣时效性高等需求,构建了本系统,勾勒出实时新闻推荐系统的框架,然后叙述了推荐系统的总体设计,并对系统的架构及关键模块的实现过程进行了详细分析,为用户提供了更加个性化、实时化的新闻推荐。
[Abstract]:The exponential explosion of information brings about the problem of information overload, which leads to the classification catalogue technology and search engine technology. However, the classified directory can only cover the popular classification, and the search engine can only input the keyword information actively by the user.So personalized news recommendation system came into being.It is difficult for a single algorithm to recommend users from multiple angles, which leads to the lack of diversity of recommendation results.In order to improve the accuracy and diversity of recommendation, this paper studies the existing recommendation algorithms and designs a mixed weighted news recommendation strategy combined with traditional recommendation technology.The content-based recommendation algorithm and the user-based collaborative filtering algorithm are weighted according to different weights to achieve the purpose of complementing each other, improving the accuracy of the recommendation results and making personalized news recommendations for usersIn this paper, news content modeling, user interest modeling and hybrid algorithm modeling are taken as the core contents of recommendation system.For news content modeling, this paper first introduces the relevant theory of news text preprocessing, according to the characteristics of news content, adopts the method of linear weighting to extract news keywords, and uses support vector machine to realize the classification of news.For user interest modeling, through collecting user behavior log, analyzing user's news browsing preference, and then completing the establishment and updating of user interest model; for hybrid algorithm modeling,The content-based recommendation algorithm determines the news recommendation list by calculating the angle cosine similarity between the news content vector and the user's interest vector.Based on the user collaborative filtering algorithm, the user similarity matrix is established to recommend the news that similar users like, and then the recall results are weighted and mixed according to different weights, and the weight value ratio with the best weighting effect is obtained through multiple training.Ensure the accuracy of the recommendation system.In addition, the news time threshold is set and the recommended return results are filtered properly to ensure the timeliness of the recommended results to a certain extent.This paper first introduces the background significance of the system and the current research situation at home and abroad to establish the basic work content, then describes the typical recommendation algorithm in detail, and analyzes the system requirements.Aiming at the large scale of news recommendation system and the high demand of user interest, this paper constructs the system, outlines the frame of the real-time news recommendation system, and then describes the overall design of the recommendation system.The architecture of the system and the implementation process of the key modules are analyzed in detail, which provides users with more personalized and real-time news recommendation.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】