基于用户行为分析的个性化推荐系统设计与实现

发布时间：2018-04-18 04:06

本文选题：协同过滤 + 用户行为序列　；参考：《南京大学》2012年硕士论文

【摘要】：随着Internet迅速普及,如何从浩如烟海的互联网数据中迅速找到相关信息,是互联网用户面临的重要问题,也是互联网技术研究的重点之一。目前,搜索引擎和信息过滤是解决该问题最常用到的两种主要技术。个性化推荐是一种新兴的信息过滤技术。它从用户的历史行为数据中发现用户的兴趣偏好,采用“推送”的方式,将用户感兴趣的信息从大量数据中过滤出来,并根据用户对信息“感兴趣”的程度,按一定的方式将相关信息呈现在用户面前。对于电子商务平台而言,使用个性化推荐技术,有助于提升平台的“长尾”优势,增加利益攸关方的收益。本文将个性化推荐相关技术引入“搜房网”垂直搜索引擎升级的设计中,分析历史用户的操作行为,提取其的兴趣模型,使用基于用户协同的过滤方式,发现当前用户兴趣,在项目库中找出当前用户可能感兴趣的信息并将之推荐给当前用户,缓解垂直搜索引擎面临的“过度筛选”问题。本文主要工作如下：概述了个性化推荐领域的经典算法、理论、研究热点及相关技术,比较了基于规则发现、基于内容过滤和基于协同过滤等相关算法和理论的优缺点,并阐述了它们各自的应用场景。同时还简要介绍了隐马尔可夫模型的相关理论。基于“搜房网”搜索引擎的用户行为特点,分析了搜索引擎系统的用户搜索行为日志,从而给出了用户行为、用户行为序列的相关定义。设计了一个序列融合算法,提取日志中的用户行为序列,同时,提出了一种计算用户行为序列相似度的方法。根据用户行为序列对用户进行了建模,并基于隐马尔可夫模型理论,设计了预测用户行为序列的模型及模型参数的估计方法。进而设计了一套基于用户行为序列分析,综合考虑了用户协同、用户行为序列相似性、项目时效性等因素的项目推荐算法。此外,还制定了相关的“冷启动”策略。最后,结合“搜房网”的实际需求,设计并实现了一个房屋信息个性化推荐系统。设计相关实验,在真实的数据集上,验证了系统的用户行为预测效果,结合隐马尔可夫模型特点,分析了系统关于用户行为预测设计上的一些局限性。并结合系统特性,讨论了评价推荐项目相关性和推荐列表排序正确性的相关指标。设计实验,评估系统在推荐列表排序、推荐项目相关性等方面的实际效果,并在此基础上分析了系统设计的不足,对系统的下一步工作进行了展望。
[Abstract]:With the rapid popularization of Internet, how to quickly find the relevant information from the vast amount of Internet data is an important problem facing Internet users, and also one of the key points in the research of Internet technology.At present, search engine and information filtering are the two most commonly used technologies to solve this problem.Personalized recommendation is a new information filtering technology.It finds the user's interest preference from the user's historical behavior data, filters out the information of the user's interest from a large amount of data by "pushing" the information, and according to the degree of the user's "interest" in the information,The relevant information is presented to the user in a certain way.For e-commerce platform, the use of personalized recommendation technology will help to enhance the platform's "long tail" advantage and increase the benefits of stakeholders.This paper introduces the personalized recommendation technology into the design of vertical search engine upgrade, analyzes the operation behavior of historical users, extracts its interest model, and finds out the current user's interest by using the filtering method based on user cooperation.Find out the information that the current user may be interested in in the project library and recommend it to the current user to alleviate the problem of "excessive filtering" faced by the vertical search engine.The main work of this paper is as follows:This paper summarizes the classical algorithms, theories, research hotspots and related technologies in the field of personalized recommendation, and compares the advantages and disadvantages of the algorithms and theories related to rule-based discovery, content-based filtering and collaborative filtering.Their respective application scenarios are described.At the same time, the theory of hidden Markov model is briefly introduced.Based on the characteristics of user behavior in search engine, the user search behavior log of search engine system is analyzed, and the relevant definitions of user behavior and user behavior sequence are given.A sequence fusion algorithm is designed to extract user behavior sequences from logs and a method to calculate the similarity of user behavior sequences is proposed.The user is modeled according to the user behavior sequence, and based on the hidden Markov model theory, the model of predicting user behavior sequence and the estimation method of model parameters are designed.Then, a set of project recommendation algorithm based on user behavior sequence analysis is designed, which considers the factors of user collaboration, user behavior sequence similarity, project timeliness and so on.In addition, the related "cold start" strategy has been developed.Finally, a personalized recommendation system for building information is designed and implemented in combination with the actual demand of Sou Fang net.Experiments are designed to verify the effectiveness of user behavior prediction in real data sets. Combined with the characteristics of hidden Markov model, some limitations of user behavior prediction design are analyzed.Combined with the characteristics of the system, this paper discusses the related indexes of evaluating the relevance of recommendation items and the correctness of recommendation list ranking.Design experiments, evaluate the system in the recommended list ranking, recommendation item correlation and other aspects of the actual results, and on the basis of the analysis of the system design deficiencies, the next work of the system is prospected.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.3

【引证文献】