个性化新闻推荐系统的研究与设计

发布时间：2018-06-12 15:38

本文选题：协同过滤 + 新闻特点　；参考：《重庆理工大学》2017年硕士论文

【摘要】：个性化新闻推荐系统是根据每个登录过推荐系统的用户的历史行为,使用推荐算法为每个用户推荐其感兴趣的新闻。基于协同过滤算法的个性化新闻推荐算法是根据用户的历史行为计算新闻的相似度,并完成相似新闻的推荐。这种相似度的计算方法没有挖掘新闻本身的特点,存在数据稀疏的问题。同时,协同过滤算法没有考虑用户的兴趣随时间发生动态变化的问题。针对推荐算法新闻相似度计算存在数据稀疏问题,本文着重研究了国内外文本相似度的计算方法,提出了适合新闻特点的混合相似度计算方法。改进的相似度计算方法是在现有的相似度计算方法的基础上,考虑了新闻文本中不同词性的词语重要性不同、新闻标题中的词语重要高于新闻正文中的词语这两个特点,并融合了基于用户行为的相似度计算方式,最后将改进的新闻相似度计算方式用于新闻推荐算法中。针对协同过滤算法没有考虑用户兴趣变化的问题,本文着重研究了国内外现有个性化新闻推荐算法,提出了适应用户兴趣变化的个性化新闻推荐算法。一般来说,用户近期浏览的新闻对用户的兴趣模型贡献较大,但用户兴趣具有反复性的特点,即早期的兴趣也有可能对用户有影响。因此,在协同过滤算法的基础上,建立了用户的近期兴趣模型和基于行为反复的兴趣模型,融合得到用户稳定的兴趣模型,并用于推荐算法中。论文中的数据集采用的是DataCastle的财新网阅读记录,评测指标是F-measure值和平均绝对误差值。适合新闻特点的混合相似度计算方法与现有的相似度计算方法都用于推荐算法进行对比,推荐结果显示,改进后的相似度计算方法的推荐结果的Fmeasure值比其他的算法最大高出10.5%,这说明了改进后的算法能更精确地计算新闻相似度值,有效避免了数据稀疏问题;适应用户兴趣变化的个性化新闻推荐算法的F-measure值与传统的协同过滤算法、现有的推荐算法最大高出11.5%,平均绝对误差值最高下降了8%,这说明了改进后的算法能更好地反映用户的兴趣。论文最后完成了个性化新闻推荐系统的设计与实现。通过对个性化新闻推荐系统进行总体分析和需求设计,并将改进的推荐算法应用于系统设计中,最终完成了整个新闻推荐系统。
[Abstract]:The personalized news recommendation system is based on the historical behavior of each user who has logged into the recommendation system and uses the recommendation algorithm to recommend the news of interest to each user. The personalized news recommendation algorithm based on collaborative filtering algorithm calculates the similarity of news according to the user's historical behavior and completes the recommendation of similar news. The similarity calculation method does not mine the characteristics of news itself and has the problem of sparse data. At the same time, the collaborative filtering algorithm does not consider the dynamic change of user's interest over time. In order to solve the problem of data sparsity in news similarity calculation of recommendation algorithm, this paper focuses on the calculation methods of text similarity at home and abroad, and puts forward a hybrid similarity calculation method suitable for the characteristics of news. The improved similarity calculation method is based on the existing similarity calculation methods, considering the two characteristics of different parts of speech words in news texts, the importance of words in news headlines is higher than the words in news text. Finally, the improved news similarity calculation method is used in news recommendation algorithm. Aiming at the problem that the collaborative filtering algorithm does not consider the change of user interest, this paper focuses on the existing personalized news recommendation algorithm at home and abroad, and proposes a personalized news recommendation algorithm to adapt to the change of user interest. Generally speaking, the news that the user browses recently contributes a lot to the user's interest model, but the user's interest has the characteristic of repetition, that is, the early interest may also have the influence on the user. Therefore, based on the collaborative filtering algorithm, the user's short-term interest model and the behavioral repeated interest model are established, and the user's stable interest model is fused and used in the recommendation algorithm. The data set in this paper uses the data Castle's Caixin net reading record, and the evaluation index is F-measure value and average absolute error value. The hybrid similarity calculation method, which is suitable for news features, is compared with the existing similarity calculation methods, and the recommended results are shown. The recommended Fmeasure value of the improved similarity calculation method is 10.5% higher than that of other algorithms, which shows that the improved algorithm can calculate the news similarity value more accurately and effectively avoid the problem of data sparsity. The F-measure value of personalized news recommendation algorithm which adapts to the change of user's interest and the traditional collaborative filtering algorithm, The existing recommendation algorithm is 11.5 higher than the maximum, and the average absolute error decreases by 8%, which shows that the improved algorithm can better reflect the interest of the user. Finally, the design and implementation of personalized news recommendation system are completed. Through the overall analysis and requirement design of the personalized news recommendation system, the improved recommendation algorithm is applied to the system design, and the whole news recommendation system is finally completed.
【学位授予单位】：重庆理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】