基于用户动态偏好的异构隐式反馈推荐算法研究

发布时间：2018-05-17 23:07

本文选题：异构隐式反馈 + 推荐系统　；参考：《浙江大学》2017年硕士论文

【摘要】：互联网时代的大潮带来了数据的爆炸式增长,大数据的概念也在过去的几年内持续升温,用户从海量数据中获得有用信息的代价也越来越高。推荐系统为解决这一问题带来了曙光。推荐系统通过用户画像、用户的历史行为数据及物品的相关数据等对用户的偏好进行建模,从而帮助用户快速发现其真正感兴趣的信息。目前非时间敏感的推荐算法通常认为用户的偏好不随时间变化,在此假设下进行相关算法的研究。然而,在真实的世界中,用户的偏好随着时间的推移不断变化。时间敏感的推荐算法存在没有考虑用户长期、有重复性的偏好,计算效率不高等问题。因此,研究用户的动态偏好,对于提升个性化推荐算法的准确度、召回率等具有十分重要的意义。本文首先对电商环境下用户的动态偏好进行了详细分析,然后围绕基于用户动态偏好的异构隐式反馈推荐算法展开研究,主要工作包括:1)提出一种基于用户偏好置信度时间衰减的推荐算法时间敏感的推荐算法通常会采用时间衰减的方式,通过降低用户较远时间前的评分值来预测用户未来购买兴趣。此外,现有的研究主要是将时间衰减用在相对简单的基于用户的协同过滤算法中,尚没有应用于基于模型的推荐算法中。针对上述两个问题,本文提出一种基于置信度时间衰减的用户偏好度量方法。我们认为,用户的评分是确定的,不随时间变化,随时间变化的是用户对这个评分的置信程度,以此表征用户的短期偏好随时间变化。通过在基于模型推荐算法上的大量实验,结果表明该方法可以更好地表征用户的偏好,从而提升推荐了算法的准确度、召回率等指标。2)提出了一种基于隐马尔可夫模型的用户动态偏好推荐算法对用户的长期、有重复性的偏好进行建模对于提升个性化推荐系统的准确度有重要作用。基于时间衰减的偏好模型可以识别出用户的短期偏好,但不足以识别出用户的长期、有重复性的偏好。针对上述问题,本文提出了利用隐马尔可夫模型来预测用户未来偏好的方法。该方法利用历史行为数据,为每一个用户建立隐马尔克夫模型,通过该模型来预测该用户的长期、有重复性的偏好。通过在基于模型推荐算法上的大量实验,结果表明该方法可以识别出用户的长期、有重复性的偏好,从而提升了推荐算法的准确度、召回率等指标。
[Abstract]:The tide of the Internet era has brought about the explosive growth of data, the concept of large data has also been rising in the past few years, and the price of users getting useful information from mass data is also getting higher and higher. The recommended system has brought dawn to solve this problem. The user preferences are modeled to help the user to quickly discover the information that they are really interested in. Currently, the non time sensitive recommendation algorithms usually think that the user's preference does not change with time, and the related algorithms are studied under this assumption. However, in the real world, the user's preference is not in the process of time. The time sensitive recommendation algorithm has no consideration of long term user, repetitive preference and low computational efficiency. Therefore, it is of great significance to study user's dynamic preference for improving the accuracy and recall of personalized recommendation algorithm. Detailed analysis, and then around the heterogeneous implicit feedback recommendation algorithm based on user dynamic preference, the main work includes: 1) a recommendation algorithm based on the time attenuation of user preferences is time sensitive, which usually uses a time attenuation method by reducing the value of the user before the longer time. In addition, the current research is mainly to use the time attenuation in a relatively simple user based collaborative filtering algorithm, and it is not yet applied to the model based recommendation algorithm. In this paper, a user preference measurement method based on the confidence time decline is proposed for the two problems. The score is determined and does not vary with time, and the user's confidence in the score is changed over time to represent the user's short-term preference over time. By a lot of experiments on the model based recommendation algorithm, the results show that the method can improve the user's preference, thus improving the accuracy of the proposed algorithm. Recall rate and other indicators.2) proposed a user dynamic preference recommendation algorithm based on Hidden Markov model for long-term users. Modeling with repetitive preference plays an important role in improving the accuracy of the personalized recommendation system. The time attenuated preference model can recognize the user's short-term preference, but not enough to identify the user's short-term preference. In this paper, a hidden Markov model is proposed to predict the user's future preference. This method uses historical behavior data to establish a hidden Markov model for each user to predict the user's long-term, repetitive preference through the model. A large number of experiments on the model recommendation algorithm show that the method can identify the long-term and repeatable preference of the user, thus improving the accuracy of the recommendation algorithm and the recall rate and so on.
【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】