基于特征表示的推荐算法实践与研究

发布时间：2018-03-20 14:09

本文选题：推荐系统　切入点：特征表示　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：推荐系统作为一种信息过滤工具诞生至今已有20余年,推荐算法的应用场景亦早已不限于电商领域,转而在诸多关联人和信息的领域发挥作用。早期的推荐算法多采用单一的用户物品交互数据,基于上下文信息的推荐算法致力于通过额外场景信息的引入来对传统推荐算法进行改良。受神经网络和矩阵分解算法普及的影响,特征表示的算法日新月异。本文致力于通过基于上下文的特征表示,在隐式反馈数据上引入上下文信息以提升推荐精度,具体包括三方面工作:·归纳总结并实现常见的基于特征表示的推荐模型。具体地,针对引入了上下文的隐式数据推荐场景,改进并实现矩阵分解和因子机算法;因推荐中神经网络模型应用较少,且一般只使用单一评分数据,设计了一种可引入多源特征数据的DeepRec算法。·在隐因子模型基础上,提出一种基于上下文信息的矩阵分解方法(Context Aware Matrix Factorization.,CAMF)。建模非线性的特征交互关系,同时又缓解数据稀疏性。引入按对排序学习框架,更加贴近基于隐式反馈数据的Top-K推荐场景,提升相关推荐算法性能。·针对基于上下文推荐的场景,基于Stacking模型融合框架,定制了一种不同深浅层次的多模型融合框架(Joint-training)。相比传统模型集成方法,在不降低预测准确率的前提下,更加的轻量快速。上述三方面的工作始于对APP推荐问题的研究,但不限于APP领域。为此,本文在IJCAI SocInf 2016推荐比赛数据上进行实验,以进一步验证本文相关工作在基于上下文的推荐场景中,具有一定的通用性。
[Abstract]:As a kind of information filtering tool, recommendation system has been born for more than 20 years, and the application scene of recommendation algorithm is not limited to the field of e-commerce. In turn, they play a role in a wide range of related people and information fields. Early recommendation algorithms used a single user object to interact with data. Context-based recommendation algorithms are designed to improve the traditional recommendation algorithms through the introduction of additional scenario information, which is influenced by the popularity of neural networks and matrix decomposition algorithms. The algorithm of feature representation is changing with each passing day. In this paper, context information is introduced into implicit feedback data to improve recommendation accuracy through context-based feature representation. It includes three aspects: 路generalizing and implementing the common recommendation model based on feature representation. Specifically, the matrix decomposition and factoring algorithm are improved and implemented in the context of the implicit data recommendation scene. Because the neural network model is seldom used in recommendation, and only a single score data is generally used, a DeepRec algorithm which can introduce multi-source feature data is designed. A method of matrix decomposition based on context information is proposed to model nonlinear feature interaction and reduce the sparsity of data. This method is more close to the Top-K recommendation scenario based on implicit feedback data. Improve the performance of related recommendation algorithms. 路for context-based recommendation scenarios, based on the Stacking model fusion framework, a multi-model fusion framework with different depth and shallow level is customized. Compared with the traditional model integration method, a multi-model fusion framework is proposed. Under the premise of not reducing the accuracy of prediction, it is more light and fast. The work of the above three aspects begins with the research of APP recommendation problem, but is not limited to the field of APP. Therefore, this paper carries on the experiment on the IJCAI SocInf 2016 recommendation contest data. In order to further verify that the related work in the context of the recommendation scenario, a certain degree of versatility.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】