基于矩阵分解的个性化推荐系统

发布时间：2018-05-17 02:13

本文选题：矩阵分解 + 局部信息　；参考：《华东师范大学》2017年博士论文

【摘要】：随着信息技术和互联网的发展,网民用户和网络产品数量成爆炸式增长,用户从信息匮乏时代进入信息过载时代。个性化推荐系统对用户行为和企业商品特性数据建模,为用户提供满足他们兴趣和需求的信息,同时为企业推广提供目标客户。现代互联网服务提供商,例如淘宝等在线购物网站、爱奇艺等在线视频网站、大众点评等生活信息服务网站,提供大量商品给用户消费,让用户给商品评分以及使用标签描述商品。针对以上用户行为数据,本文以矩阵分解相关理论为基础,建立模型以解决推荐系统中的三个典型任务:(1)在隐式反馈数据上构建商品推荐模型,推荐用户感兴趣的商品;(2)利用显式评分数据建立评分预测模型,预测用户对商品的喜好值;(3)利用显式标签数据建立标签推荐模型,方便用户输入来描述商品属性,帮助推荐系统良性循环。本文的研究问题和技术贡献总结如下:1.基于加权局部矩阵分解的商品推荐:现有的针对隐式反馈数据的矩阵分解模型往往只从数据的全局信息出发,忽略了数据之中的局部信息。为了利用隐式反馈数据的局部信息,本文提出了一种加权局部矩阵分解模型进行商品推荐,并为该模型设计了高效的子矩阵选择算法和改进的交替最小二乘参数优化算法,对用户和商品的局部特性建模,同时缓解了数据稀疏性问题。真实数据上的实验结果表明该模型有较优的推荐效果,并验证了考虑隐式反馈数据的局部信息有助于商品推荐。2.基于多主题矩阵分解的评分预测:为克服现有工作中针对显式评分数据局部信息建模的不可解释性和目标函数的不一致性,本文提出了多主题矩阵分解模型。它结合主题模型和概率矩阵分解模型,利用主题模型建模数据局部信息和矩阵分解来刻画用户和商品的局部内在特征。本文使用贝叶斯方法建模主题矩阵分解模型,使得模型只需少量的经验设置参数以得到更高的推荐准确率。实验结果说明该模型在评分预测中优于其他局部矩阵分解模型,并对局部建模信息具有一定的可解释性。3.时间感知的标签推荐:为了利用标签数据中用户标注标签的时间信息,本文提出了时间感知的张量分解模型。该模型利用Hawkes时间点过程对用户使用标签的时间信息建模,并利用指数函数将Hawkes过程中的叠加形式转化为递归形式,使得计算用户当前时间对标签的喜好值只跟上次使用标签的时间有关,减少了大量的计算时间,并将其以权重的方式加入到逐对排序张量分解模型。实验结果表明该模型能够有效地利用时间信息,提高标签准确度,同时在冷启动问题上也有较好的表现,并且具有可接受的推荐新颖性。
[Abstract]:With the development of information technology and Internet, the number of Internet users and network products has increased explosively, and users have entered the era of information overload from the era of information scarcity. The personalized recommendation system models user behavior and enterprise commodity characteristics data, provides users with information to meet their interests and needs, and provides target customers for enterprise promotion. Modern Internet service providers, such as online shopping sites such as Taobao, online video sites such as iqiyi, and life information services such as Dianping, provide large quantities of goods for consumers to consume. Ask the user to rate the product and use labels to describe the product. Aiming at the above user behavior data, based on the matrix decomposition theory, this paper establishes a model to solve the problem of three typical tasks in the recommendation system: 1) construct the commodity recommendation model on the implicit feedback data. Using explicit scoring data to set up a rating prediction model, and predict the user's preference for the product / value / 3) using explicit label data to establish a label recommendation model to facilitate the user's input to describe the attributes of the product. Help recommendation system virtuous cycle. The research problems and technical contributions of this paper are summarized as follows: 1. Commodity recommendation based on weighted Local Matrix decomposition: the existing matrix decomposition model for implicit feedback data is usually based on the global information of the data and neglects the local information in the data. In order to make use of the local information of implicit feedback data, a weighted local matrix factorization model is proposed for commodity recommendation, and an efficient submatrix selection algorithm and an improved alternating least square parameter optimization algorithm are designed for the model. The local characteristics of users and commodities are modeled and the problem of data sparsity is alleviated. The experimental results on real data show that the model has a better recommendation effect, and it also verifies that the local information considering implicit feedback data is helpful to commodity recommendation. Score prediction based on multi-topic matrix factorization: in order to overcome the inexplicable modeling of local information of explicit scoring data and the inconsistency of objective function, a multi-topic matrix factorization model is proposed in this paper. It combines the topic model with the probability matrix decomposition model and uses the topic model to model the local information and matrix decomposition of the data to describe the local intrinsic characteristics of users and commodities. In this paper, Bayesian method is used to model the topic matrix decomposition model, so that the model needs only a small amount of experience to set the parameters to get a higher recommendation accuracy. The experimental results show that the model is superior to other local matrix decomposition models in score prediction, and has a certain interpretability for local modeling information. Time aware label recommendation: in order to use the time information of the label, this paper proposes a Zhang Liang decomposition model of time perception. The model uses the Hawkes time point process to model the time information of the user using the label, and uses the exponential function to transform the superposition form of the Hawkes process into the recursive form. It makes the preference value of the user's current time to the label only related to the last time of using the label, which reduces a lot of computing time, and adds it to the sorting Zhang Liang decomposition model by weight. The experimental results show that the model can effectively utilize the time information and improve the label accuracy, at the same time, it has a good performance on cold start problem, and it has the acceptable novelty of recommendation.
【学位授予单位】：华东师范大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】