【作者】 刘洋；

【导师】 冯勇；

【作者基本信息】 辽宁大学 ， 计算机应用技术， 2021， 硕士

【摘要】 随着移动互联和大数据的飞速发展,互联网即将从Web2.0时代迈入Web3.0时代,这意味着互联网每日产生的数据量将不断暴涨,“信息过载”问题愈发严重,用户如何在海量数据中得到想要的信息成为当下亟待解决的难题。推荐技术的发展与普及很大程度上缓解了该难题,通过分析用户历史数据获取其喜好从而进行推荐。籍此,该技术迅速在学术界和工业界盛行并应用于诸多领域。目前,越来越多研究者致力于研发各种优良的推荐算法,但现有成果仍然存在一些不足。一是数据稀疏性问题仍然没有得到很好的解决,虽然已经有大量研究证明采用评论缓解评分的稀疏性切实有效,却忽略了评论本身的稀疏性。二是缺少对推荐多样性的深入思考,即使有些方法考虑了多样性也是采用单一的策略,推荐效果欠佳。本文针对以上问题对提升推荐的准确性和多样性开展研究,提出了面向用户多样性需求的个性化商品推荐方法,该方法包括如下两个部分:一、针对评论的稀疏问题,提出了融合近邻评论的GRU商品推荐模型。首先,引入近邻评论的思想,用近邻评论挖掘用户的潜在偏好;然后将矩阵分解得到的隐向量送入多层感知机,获取评分数据的深层非线性特征表示;最后,将基于评分数据和评论文本得到的用户特征和商品特征加以拼接,进行评分预测,将评分高的商品推荐给用户。二、针对采用单一多样性推荐方式的弊端,提出了基于用户需求的生成对抗推荐算法。首先进行多样性用户划分,拥有不同多样性需求的用户采用不同的推荐方法。该算法的生成模型由多生成器组成,每个生成器关注一个项目属性,从而生成多样性项目表示,判别模型评估生成项目并反馈给生成模型使其不断完善,最终推荐与生成项目相似度高的项目。本文在Amazon数据集上对所提出的推荐方法进行对比实验,验证了近邻评论的思想能够有效缓解评论稀疏,并在准确性指标上优于其他模型,而且多生成器网络结构在多样性指标上的也取得了不错的效果。通过对用户的划分,实现为不同的用户提供最适合的推荐方式,满足用户的多样性需求。另外,本文提出的推荐方法使用的数据有评分、评论、商品属性等,这些数据在商品交易网站容易获取,所以本文研究拥有广阔的应用领域和良好的发展前景。更多还原

【Abstract】 With the rapid development of mobile internet and big data,the Internet is about to enter the Web3.0 era from the Web2.0 era,which means that the daily amount of data generated by the Internet will continue to skyrocket,and the problem of information overload will become more and more serious.Obtaining the desired information from the massive data has become an urgent problem to be solved.The development and popularization of recommendation technology have largely alleviated this problem.The recommendation is made by analyzing users’ historical data to obtain their preferences.As a result,this technology has quickly become popular in academia and industry and has been applied in many fields.At present,more and more researchers are devoted to the development of various excellent recommendation algorithms,but the existing results still have some shortcomings.First,the problem of data sparsity has not been solved well.Although there have been a large number of studies that have proved that using comments to alleviate the sparsity of ratings is effective,the sparsity of comments itself has been ignored.The second is the lack of depth thinking about the diversity of recommendations.Even if some methods consider diversity,a single strategy is adopted,and the recommendation effect is not good.This article focuses on the above problems to improve the accuracy and diversity of recommendations,and proposes a personalized product recommendation method oriented to the diverse needs of users.The method includes the following two parts:1.Aiming at the sparseness of reviews,a GRU product recommendation model incorporating neighbor reviews is proposed.First,introduce the idea of neighbor comments,and use neighbor comments to mine the user’s potential preferences.Then send the hidden vector obtained by matrix decomposition to the multi-layer perceptron to obtain the deep nonlinear feature representation of the rating data.Finally,the user characteristics and product characteristics obtained based on the rating data and the review text are spliced together to make a rating prediction,and the products with high ratings are recommended to the user.2.Aiming at the disadvantages of using a single diverse recommendation method,generative adversarial recommendation algorithm based on user requirements is proposed.Diversified users are divided first,and users with different diversified needs adopt different recommendation methods.The generative model of the algorithm is composed of multiple generators.Each generator focuses on a project attribute to generate diverse project representations.The discriminant model evaluates the generated project and feeds it back to the generated model for continuous improvement,and the final recommendation is highly similar to the generated project s project.This paper conducts comparative experiments on the proposed recommendation methods on the Amazon data set,and verifies that the idea of neighbor comments can effectively alleviate comment sparsity,and is better than other models in accuracy indicators,and the multi-generator network structure is in diversity indicators.The also achieved good results.Through the division of users,the most suitable recommendation method is provided for different users to meet the diverse needs of users.In addition,the recommended method proposed in this paper uses data such as ratings,reviews,product attributes,etc.These data are easily available on commodity trading websites,so the research in this paper has broad application fields and good development prospects.更多还原