基于用户行为的移动电子商务推荐算法研究

发布时间：2018-05-09 10:28

本文选题：移动电子商务推荐 + 概率隐因子模型　；参考：《江苏大学》2017年硕士论文

【摘要】：随着移动互联网、智能终端技术的快速发展,具有移动性、便捷性的移动电子商务成为电子商务发展的新方向。受移动端显示器大小的限制,移动电子商务面临严重的信息过载现象,迫切需要为用户提供个性化推荐服务。现有的推荐技术主要利用用户评分或者隐式反馈建立兴趣模型,根据设定的时间间隔或者数据累积量定期更新该模型,为用户提供个性化推荐。然而移动用户的需求会随着时间和位置的变化而发生改变,导致即时兴趣变化迅速,现有算法的兴趣模型主要反映用户的长期兴趣,需要对模型定期更新,不能满足移动电商环境下为用户提供实时推荐的要求。此外,由于新用户没有行为数据或历史数据非常稀疏,现有算法无法为新用户提供可靠的推荐结果,影响推荐质量。针对上述问题,本文提出了基于用户购买倾向和兴趣度的个性化推荐算法和基于多源信息融合的协同过滤推荐算法。本文的主要研究工作如下:(1)针对移动电子商务中用户的即时兴趣变化迅速,而现有推荐算法不能实时响应用户需求的问题,提出了一种基于用户购买倾向和兴趣度的个性化推荐算法(Purchase Intention and Interest Degree,PIID)。该算法从长期兴趣和即时兴趣两方面对用户兴趣建模,用购买倾向来量化用户对最近交互过却没有购买的商品的即时兴趣。在离线状态下,提取用户的行为特征,利用逻辑回归模型训练影响购买的行为特征所对应的回归系数,建立购买倾向预测模型;为了更精确地定位用户可能购买的商品,将购买倾向与兴趣度线性结合作为购买概率,针对购买记录,利用概率隐因子模型通过最大化购买概率来学习用户兴趣度。在线推荐时,将用户的实时行为特征带入预测模型中,得到用户的购买倾向,再与兴趣度结合,将购买概率较大的商品列表推荐给用户。在真实数据集上的实验表明,PIID算法与现有算法相比,有较高的准确率和F1指标,而且能够提供高效的实时推荐。(2)针对现有推荐算法是在大量评分或隐式反馈的基础上建立兴趣模型,无法为缺乏行为数据的新用户提供可靠推荐的问题,提出了一种基于多源信息融合的协同过滤推荐算法。算法首先利用位置信息,建立位置消费图,计算位置远近和不同用户的兴趣偏好对目标用户的影响,找到最近邻居集,可以根据近邻的偏好为新用户进行推荐;接着利用PIID算法计算用户的个人兴趣;最后利用购买概率将近邻偏好和个人兴趣对购买的影响建模,为用户推荐购买概率最大的商品列表。在真实数据集上的实验表明,该算法有较高的准确率和F1指标,而且能够为新用户提供可靠的推荐列表。(3)为了验证本文所提出算法的可行性,采用面向对象和模块化的设计思想,使用java编程语言设计并实现了一个基于用户行为的移动电子商务原型推荐系统。
[Abstract]:With the rapid development of mobile Internet and intelligent terminal technology, mobile electronic commerce with mobility and convenience has become a new direction of electronic commerce. Limited by the size of mobile display, mobile e-commerce is faced with serious information overload phenomenon, so it is urgent to provide personalized recommendation services for users. The existing recommendation techniques mainly use user rating or implicit feedback to build interest model, update the model periodically according to the time interval or data accumulation, and provide personalized recommendation for users. However, the needs of mobile users will change with the change of time and location, which leads to the rapid change of immediate interest. The interest models of existing algorithms mainly reflect the long-term interests of users, and need to be updated regularly. Can not meet the mobile e-commerce environment for users to provide real-time recommendation requirements. In addition, due to the lack of behavioral data or sparse historical data for new users, the existing algorithms cannot provide reliable recommendation results for new users, which affects the quality of recommendation. To solve the above problems, this paper proposes a personalized recommendation algorithm based on user purchase tendency and interest degree and a collaborative filtering recommendation algorithm based on multi-source information fusion. The main research work of this paper is as follows: (1) aiming at the problem that the instant interest of users in mobile e-commerce changes rapidly, but the existing recommendation algorithms can not respond to the needs of users in real time. This paper presents a personalized recommendation algorithm based on user purchase tendency and interest degree. The algorithm models user interest in terms of both long-term interest and immediate interest, and quantifies users' immediate interest in recently interacted but not purchased items with purchasing tendency. In the off-line state, the user's behavior characteristics are extracted, and the regression coefficients corresponding to the behavior characteristics that affect the purchase are trained by using the logical regression model, and the prediction model of purchase tendency is established, in order to locate the goods that the user may buy more accurately. The linear combination of purchase propensity and interest degree is taken as the purchase probability. According to the purchase record, the probability implicit factor model is used to learn the user interest degree by maximizing the purchase probability. When online recommendation, the real-time behavior features of the user are brought into the prediction model, and then the user's purchase tendency is obtained, and then combined with the interest degree, the list of items with a high purchase probability is recommended to the user. Experiments on real data sets show that the PIID algorithm has higher accuracy and F1 index than the existing algorithms. Moreover, it can provide efficient real-time recommendation. (2) aiming at the problem that the existing recommendation algorithm is based on a large number of ratings or implicit feedback, it can not provide reliable recommendation for new users who lack behavioral data. A collaborative filtering recommendation algorithm based on multi-source information fusion is proposed. Firstly, the location consumption graph is built by using location information, and the influence of location distance and different user's interest preference on the target user is calculated, and the nearest neighbor set is found, and the new user can be recommended according to the nearest neighbor's preference. Then the PIID algorithm is used to calculate the personal interest of the user, and the influence of the nearest neighbor preference and personal interest on the purchase is modeled by using the purchase probability, and the list of items with the highest purchase probability is recommended for the user. Experiments on real data sets show that the proposed algorithm has high accuracy and F1 index, and can provide a reliable recommendation list for new users in order to verify the feasibility of the proposed algorithm. A prototype recommendation system for mobile e-commerce based on user behavior is designed and implemented by using object-oriented and modular design ideas and java programming language.
【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】