【作者】 孙静；

【导师】 孙静宇；

【作者基本信息】 太原理工大学 ， 软件工程（专业学位）， 2020， 硕士

【摘要】 互联网技术的飞速发展使“信息过载”现象日趋明显,用户从庞大的信息量中筛选出有用信息也变得愈发困难。因此,如何精准挖掘用户信息并提供个性化推荐成为研究的热点。传统推荐算法大多都是基于用户的显性反馈行为挖掘用户与项目的关联关系,依赖于用户评分信息,而隐性行为较显性反馈而言包含着更多反映用户喜好的深层信息,但却常常被忽略。用户上网日志数据作为最具代表性的隐性行为,有助于提高推荐系统性能。通过深入分析推荐系统在用户画像和深度学习两个方向上的研究,针对深度神经网络参数众多和计算繁重的限制,引入深度森林模型对用户画像的兴趣标签进行分类预测,该模型将随机森林作为训练基本单元,既利用了集成学习方法中子模型差异性和多样性对学习效果的提升,又吸收了深度神经网络逐层处理、模型内特征转化和模型高复杂度的优势。本文从建立用户画像的角度入手获取用户信息,通过深度学习模型对用户行为信息和项目相关性进行深入挖掘,实现协同过滤推荐。为实现这一目的,本文从以下方面做出研究:(1)在研究了用户标签体系的建立和用户画像建模方法基础上,确定了本课题使用多维度建模方式刻画用户画像,并且利用关键意见领袖(Key Opinion Leader,KOL)对用户的粘着性和对信息的过滤作用,提出KOL用户画像的建模方法。(2)针对传统的直译式表征方法把项目转化为0和1组成的一串数字序列,不完全适合针对用户上网行为的兴趣偏好表征这一问题,提出了用户兴趣向量表征方法,将多示例学习中包(Bag)与示例(Instance)的概念引入到对用户上网日志的表征,用相似项目聚成的簇代表用户的兴趣簇。(3)针对深度森林模型在特征选择过程中对所有特征赋予相同权重这一问题,提出利用兴趣簇重要性作为特征选择权重对模型进行改进。并提出引入注意力机制深化时间上下文信息,将其与gc Forest模型融合,使用基于K-means聚类的用户协同过滤算法产生推荐。本文在运营商DPI用户日志上分析和建立了用户画像,并利用统计方法得到KOL用户设备和兴趣标签分布。并且,在课题自采数据集、YOOCHOOSE数据集和Iris数据集上进行预测模型和推荐模型实验,验证了本文所提的用户兴趣向量表征、深度森林模型及其改进、融合时间注意力机制对协同过滤推荐的有效性。更多还原

【Abstract】 With the rapid development of Internet technology,the phenomenon of "information overload" becomes more obvious,and it becomes more difficult for users to select useful information from the huge amount of information.Therefore,how to accurately mine user information and provide personalized recommendation has become a research hotspot.Most of the traditional recommendation algorithms are based on the explicit feedback behavior of users to mine the relationship between users and items,which depends on user rating information,while the implicit behavior contains more deep information reflecting user preferences than the explicit feedback,but it is often ignored.As the most representative implicit behavior,user log data is helpful to improve the performance.Through in-depth analysis of the research on user profile and deep learning in the recommendation system,the deep forest model is introduced to classify and predict the interest labels of user profile.The deep forest model takes random forest as the basic unit of training,which not only absorbs the advantages of deep neural network layer-by-layer processing,in-model feature conversion and high complexity,but also uses the influence of the difference and diversity of sub-models in the ensemble learning method on the learning effect,and overcomes the limitations of the numerous parameters and heavy calculation of the depth neural network.In this paper,user information is obtained from the perspective of user profile,which aims to mine user behavior information and project relevance through deep learning model,and realize collaborative filtering and recommendation.In order to achieve this goal,this paper studies from the following aspects:(1)Based on the study of the establishment of user label system and the modeling method of user profile,this paper determines to use the multi-dimensional modeling method to depict the user profile,and uses the key opinion leader(KOL)to the user adhesiveness and information filtering function,and proposes the modeling method of KOL user profile.(2)In order to solve the problem that the traditional literal representation method transforms the item into a series of numbers composed of 0 and 1,which is obviously not suitable for user’s interest preference representation,the user interest vector representation algorithm is proposed.The algorithm introduces the concepts of bag and instance into the representation of user online log,and represents the user interest cluster by clustering similar items.(3)In order to solve the problem that all features are given the same weight in the process of feature selection,the importance of interest cluster is used as feature selection weight to improve the model.The attention mechanism is proposed to deepen the time context information and is integrated with the gc Forest model.And the recommendation results are generated by the user collaborative filtering algorithm based on K-means clustering.In this paper,the user profile is analyzed and established on the DPI user log used in the project,and the KOL user devices and interest label distribution are obtained by using statistical methods.In addition,the prediction model and recommendation model experiments are carried out on the real data set,YOOCHOOSE data set and Iris data set to verify the user interest vector representation,the gc Forest model and the improvement and integration of time attention mechanism is effective for collaborative filtering recommendation.更多还原