基于图论的个性化视频推荐算法研究

发布时间：2018-03-13 22:43

本文选题：个性化推荐　切入点：视频推荐　出处：《华南理工大学》2013年硕士论文　论文类型：学位论文

【摘要】：得益于数码设备的发展以及几乎无处不在的宽带互联网连接，普通用户也可以很容易的通过手机、单反等设备记录下各种视频，并把它们上传到互联网。互联网上的视频内容正以爆炸式的速度增长。这些视频能够为用户提供丰富的可能感兴趣的内容。用户可以通过观看视频进行免费的学习、娱乐，同时可以通过各大社交平台将喜欢的视频分享给朋友，增进朋友之间的感情。庞大的视频数量使得寻找感兴趣视频的过程变成一项艰巨的任务。搜索是解决信息过载问题的常用方法。但是，，只有当用户十分明确自己想要的内容并能很好的用一些搜索关键词来表达的时候，视频搜索才能得到很好的效果。在实践中，普通用户很难通过一些文本来准确的表达搜索意图。此外，传统的搜索对所有用户会返回同样的搜索结果，无法根据用户的喜好为用户提供优化过的个性化的服务。为解决这一信息过载问题，本文提出了一种个性化视频推荐算法。不同于传统方法只利用用户点击数据构成用户与视频之间的二分图，并基于该图进行推荐，我们首先结合用户点击和搜索信息构建一个三分图。在三分图中，搜索关键词顶点扮演桥梁的角色，连接用户顶点和视频顶点。之后，为了丰富用户顶点与视频顶点之间的连接，三个同类型顶点之间的子图被添加到三分图中。这三个子图是基于内容的信息（视频标签和文本搜索）构建的。我们提出一种迭代式消息传播算法，利用该图计算每个用户的喜好信息。并且基于计算所得的用户喜好信息为每位用户提供个性化的视频推荐。实验数据来自一个商用的视频搜索引擎，并包含1369个用户，8765个搜索关键词以及17712个视频。实验结果显示，在给出相同推荐视频个数的情况下，我们的算法比起TP[91]，Co-view[6]，Adsorption[2]等算法有着更高的查准率和查全率（Precision andRecall）。实验结果证明了我们提出的方法，通过结合用户行为信息和基于内容的信息，能够获得更好的推荐效果。
[Abstract]:Thanks to the development of digital devices and the almost ubiquitous broadband Internet connection, ordinary users can easily record all kinds of videos via mobile phones, SLR and other devices. And upload them to the Internet. Video content on the Internet is exploding. These videos provide users with a wealth of potentially interesting content. Users can learn and entertain free of charge by watching videos. At the same time, you can share your favorite videos with your friends on various social platforms. The sheer number of videos makes the search for interesting videos a daunting task. Searching is a common way to solve the problem of information overload. However, Video search can only get good results when the user is very clear about what he wants and can express it well with some search keywords. It is difficult for ordinary users to express their search intention accurately through some text. In addition, traditional search will return the same search results to all users, and can not provide users with optimized personalized services according to their preferences. In order to solve the problem of information overload, a personalized video recommendation algorithm is proposed in this paper. We first build a three-point graph with user click and search information. In the three-point graph, the search keyword vertex acts as a bridge, connecting the user vertex with the video vertex. To enrich the connection between user vertices and video vertices, Three subgraphs between vertices of the same type are added to the three-point graph. The three subgraphs are constructed based on content-based information (video tag and text search). We propose an iterative message propagation algorithm. The graph is used to calculate the preferences of each user, and based on the calculated user preference information, a personalized video recommendation is provided for each user. The experimental data comes from a commercial video search engine and contains 1369 users, 8765 search keywords and 17,712 videos. Our algorithm has higher precision and recall ratio than TP [91] Co-view [6] Adsorption [2]. The experimental results show that the proposed method can obtain better recommendation effect by combining user behavior information with content-based information.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.3;O157.5

【共引文献】