基于神经网络集成和用户偏好模型的协同过滤推荐算法研究

发布时间：2019-06-24 11:15

【摘要】：随着互联网技术的发展和电子商务的蓬勃生长,网络数据信息以指数级别的趋势增长,用户不得不耗费大量的时间去搜索自己想要的信息和商品,人们进入了一个“信息超载”的时代。推荐系统应运而生,它的主要任务是主动从海量的资源中为用户推送其可能需要的资源,缓解信息检索的压力。在目前的应用中,协同过滤算法取得的成就无疑是最大的,但是在发展中同样不可避免的遇到很多障碍,数据稀疏问题就是阻碍其发展的一个重要难题。针对数据稀疏性问题,本文从用户兴趣的角度出发,利用当前的用户数据,构建用户偏好模型预测未评分项目的分值,填充数据到用户评分矩阵中。然而,用户偏好存在描述上的模糊性和不确定性,给用户偏好建模带来了一定的困难,需要引入机器学习方法来构建用户偏好模型。神经网络集成算法具有很好的泛化能力,是目前机器学习领域的一个研究热点,可以用来模拟用户的偏好。但是面对用户偏好的复杂性,神经网络集成算法也会存在这样或那样的不足。针对这样的情况,本文首先对传统的神经网络集成算法提出自己的改进思想,提出了基于差分进化的负相关神经网络集成算法,提高了神经网络集成算法的泛化能力；其次,利用改进的算法,结合现有的用户数据,构建用户偏好模型；最后利用构建好的偏好模型预测未评分项目的分值,填充用户评分矩阵,并针对可能产生的填充过度问题,对相似度的计算也做了一个改进。基于差分进化的负相关神经网络集成算法的基本思想是：在集成个体的生成中,为了增加集成个体的差异性,引入负相关学习方法并行训练成员网络；在结论的生成中,利用差分进化算法的良好的寻优能力,对成员网络的权重系数进行优化。通过实验仿真,并将它与其他算法进行对比,结果表明该算法无论是在泛化性能方面还是在鲁棒性方面都表现得更好。基于差分进化神经网络集成的用户偏好模型的基本思想是：充分利用项目特征属性,构建项目特征向量,通过项目特征向量和用户偏好的映射,构建用户偏好模型,并采用提出的差分进化负相关神经网络集成算法模拟用户的兴趣爱好。通过实验证明,提出的差分进化负相关神经网络集成算法能够很好的模拟用户的喜好,预测对未评分项目的分值。基于用户偏好模型的协同过滤推荐算法的基本思想是：使用构建好的偏好模型预测未评分项目的分值,填充用户评分矩阵,形成了一个伪用户评分矩阵。在利用伪用户评分矩阵计算用户的相似度时,针对可能产生的填充过度问题,只选择部分的项目进行计算。通过MovieLens数据集测试,该算法比传统的协同过滤推荐算法有更好的性能。
[Abstract]:With the development of Internet technology and the vigorous growth of electronic commerce, the network data information is growing at the exponential level, users have to spend a lot of time searching for the information and goods they want, and people have entered an era of "information overload". Recommendation system emerges as the times require, its main task is to actively push the resources it may need for users from a large number of resources, and alleviate the pressure of information retrieval. In the current application, the cooperative filtering algorithm has made the greatest achievement, but it is also inevitable to encounter many obstacles in the development. The problem of data sparse is an important problem that hinders its development. In order to solve the problem of data sparsity, from the point of view of user interest, this paper constructs a user preference model to predict the score of ungraded items from the point of view of user interest, and fills the data into the user scoring matrix. However, there are fuzziness and uncertainty in the description of user preference, which brings some difficulties to user preference modeling. It is necessary to introduce machine learning method to construct user preference model. Neural network integration algorithm has good generalization ability, and it is a hot research topic in the field of machine learning, which can be used to simulate the preferences of users. However, in the face of the complexity of user preferences, neural network integration algorithms will also have some shortcomings. In view of this situation, this paper first puts forward its own improved idea for the traditional neural network integration algorithm, and proposes a negative correlation neural network integration algorithm based on differential evolution, which improves the generalization ability of the neural network integration algorithm. Secondly, using the improved algorithm, combined with the existing user data, the user preference model is constructed. Finally, the constructed preference model is used to predict the score of ungraded items, fill the user scoring matrix, and improve the calculation of similarity in order to solve the problem of overfilling. The basic idea of negative correlation neural network integration algorithm based on differential evolution is: in order to increase the difference of integrated individuals, negative correlation learning method is introduced to train member networks in parallel, and in the generation of conclusions, the weight coefficients of member networks are optimized by using the good optimization ability of differential evolution algorithm. Through the experimental simulation and comparing it with other algorithms, the results show that the algorithm performs better both in terms of generalization performance and robustness. The basic idea of user preference model based on differential evolution neural network integration is: make full use of project feature attributes, construct project feature vector, construct user preference model through the mapping of project feature vector and user preference, and use the proposed differential evolution negative correlation neural network integration algorithm to simulate users' interests and hobbies. The experimental results show that the proposed differential evolution negative correlation neural network integration algorithm can simulate the preferences of users and predict the score of ungraded items. The basic idea of collaborative filtering recommendation algorithm based on user preference model is to use the constructed preference model to predict the score of ungraded items, fill in the user scoring matrix, and form a pseudo-user scoring matrix. When the pseudo-user rating matrix is used to calculate the similarity of users, only part of the items are selected to calculate the possible overfilling problem. Through the test of MovieLens dataset, this algorithm has better performance than the traditional collaborative filtering recommendation algorithm.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】