推荐系统相关技术的研究与应用

发布时间：2018-04-26 09:02

本文选题：推荐系统 + 用户特征　；参考：《江西理工大学》2016年硕士论文

【摘要】：随着4G时代的到来,推荐系统作为大数据时代的产物在电子商务、信息检索以及移动应用、社交网络等众多领域都取得了不错的发展。伴随着这些应用的发展用户和产品数目不断增长,数以亿计的资源信息被产生,数据的规模不断壮大,致使推荐系统面临着前所未有的挑战。主要包括数据稀疏、可扩展性和冷启动等问题,这些问题俨然已成为了影响推荐质量的主要因素,研究者们在学习过程中提出了多种方法来克服这些问题。但是针对形态各异的问题不同的推荐算法存在各自的优劣点。为此本文综合分析了国内外的研究现状,针对推荐系统中的数据稀疏、冷启动、用户特征随时间动态变化及不同用户对同一特征依赖程度不同等问题,分别从改进算法及构造新的数学模型等方面对推荐系统开展了细致地研究,主要工作成果如下:1.提出结合用户特征分类和动态时间的协同过滤推荐。本文主要通过加入用户特征信息及特征动态变化因子来改善协同过滤推荐存在的不足,首先构建特征分类树、并根据用户间相同特征数目引入一种自适应的权重模型改进了计算特征相似度的方法。同时还分析了用户特征动态变化这一因素对推荐结果的影响,通过引入一个包含时间跨度的衰减函数来改善用户特征动态变化问题对推荐精度的限制。最后针对冷启动问题利用sigmoid函数将上述基于用户特征模型与传统协同过滤推荐模型相融合,通过平滑地过渡来灵活地计算用户间相似度。为了突出本文算法的有效性,从推荐结果的准确率、精度及召回率等方面进行了验证。实验结果表征结合用户特征分类和动态时间的协同过滤推荐不但能够提升推荐精度、准确率和召回率,同时对推荐系统面临的冷启动问题也起到了一定的缓解作用。2.提出基于降噪自编码的推荐算法。自编码神经网络隶属于模型的方法,它能够提取数据从低层到高层的特征,发现隐藏在样本间潜在的相关性,为了提高推荐系统的推荐质量将该模型应用于推荐系统。针对原始数据之间的冗余度问题,本文采用ZCA白化技术对原始数据进行转化、去相关。同时为了增强自编码神经网络的鲁棒性,本文还对自编码模型进行了改进,提出了基于降噪自编码的推荐算法,主要是在利用自编码模型进行训练时加入了随机噪声来克服自编码神经网络将输入样本直接复制输出的问题。实验结果表明引入ZAC白化和降噪后的自编码模型能够有效地提高推荐的精度。
[Abstract]:With the arrival of 4G era, recommendation system as a product of big data era has made good progress in many fields such as e-commerce, information retrieval, mobile applications, social networks and so on. With the development of these applications, the number of users and products is increasing, hundreds of millions of resource information has been generated, and the scale of data is growing, which makes the recommendation system face unprecedented challenges. The main problems include data sparsity scalability and cold start. These problems have become the main factors affecting the quality of recommendations. Researchers have put forward a variety of methods to overcome these problems in the learning process. However, different recommendation algorithms have their own advantages and disadvantages. For this reason, this paper synthetically analyzes the domestic and foreign research situation, aiming at the problems of sparse data, cold start, dynamic change of user characteristics with time and different degree of dependence of different users on the same feature in recommendation system. The recommendation system is studied in detail from the aspects of improving the algorithm and constructing a new mathematical model. The main results are as follows: 1. A collaborative filtering recommendation based on user feature classification and dynamic time is proposed. In this paper, the shortcomings of collaborative filtering recommendation are improved by adding user feature information and feature dynamic change factor. Firstly, the feature classification tree is constructed. According to the same number of features among users, an adaptive weight model is introduced to improve the method of computing feature similarity. At the same time, the influence of the dynamic change of user features on the recommendation results is analyzed, and a attenuation function including time span is introduced to improve the limitation of recommendation accuracy caused by the dynamic change of user features. Finally, the sigmoid function is used to fuse the user feature model and the traditional collaborative filtering recommendation model for the cold start problem, and the similarity between users is calculated flexibly through a smooth transition. In order to highlight the effectiveness of the proposed algorithm, the accuracy, accuracy and recall rate of the proposed algorithm are verified. The experimental results show that the collaborative filtering recommendation combined with user feature classification and dynamic time can not only improve the recommendation accuracy, accuracy and recall rate, but also play a role in alleviating the cold start problem faced by the recommendation system. A recommendation algorithm based on noise reduction and self-coding is proposed. The self-coding neural network belongs to the model. It can extract the features from the lower level to the higher level and find the potential correlation between the samples. In order to improve the recommendation quality of the recommendation system, the model is applied to the recommendation system. Aiming at the problem of redundancy between raw data, this paper uses ZCA whitening technology to transform and decorrelate the original data. At the same time, in order to enhance the robustness of the self-coding neural network, the self-coding model is improved, and the recommendation algorithm based on de-noising self-coding is proposed. In order to overcome the problem of direct replication and output of input samples by self-coding neural network, random noise is added to train the self-coding model. The experimental results show that the self-coding model with ZAC whitening and noise reduction can effectively improve the accuracy of the recommendation.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】