基于机器学习的众筹平台个性化推荐算法研究

发布时间：2019-06-06 07:49

【摘要】：随着互联网技术的不断进步,众筹平台成为了一种新的网络融资途径。在众筹平台所产生的数据不断增加的同时,其数据效益并没有成正比增长,因此产生了所谓的“信息超载”现象。个性化推荐系统从大量数据中挖掘用户的兴趣偏好,可以很好的解决这个问题,其在电子商务、社交媒体、广告系统、搜索引擎等领域都取得了一定的成功。但是,在迅速发展的网络众筹领域,目前还未有众筹网站为用户提供专业的个性化推荐服务。本论文对众筹平台的整体情况进行了分析,并对常用个性化推荐算法进行了研究和比较,完成了对推荐系统输入、输出、以及推荐算法的选取与设计。同时,本文应用机器学习算法建立基于协同过滤的推荐系统,并对其中存在的问题设计了相应的改进方案。一方面,本文针对数据稀疏性问题,设计了基于隐语义模型的协同过滤算法,利用统计学习方法解决模型最优化问题。算法通过学习用户评分数据的特征,训练预测模型,得出预测评分后填充至原始评分矩阵,再以填充后的评分矩阵为数据源,基于协同过滤算法得到预测评分。另一方面,本文针对数据源单一带来的冷启动问题,结合众筹平台的用户评分与项目属性特征,对协同过滤算法进行了改进。本文通过网络通信技术,获取众筹平台用户评分与项目属性数据,对推荐算法进行了可行性验证,对特征学习模型中的参数进行了调节,并比较了改进前后算法的平均绝对误差与准确度。经过实验验证,本文所设计的众筹平台个性化推荐算法能够提供精准、快速的个性化推荐服务,为用户提供便利的同时,也有利于众筹平台的发展。本论文提出的改进算法,在一定程度上解决了数据稀疏性与冷启动问题,相较于传统的推荐算法,预测的准确度有了明显的提升。该方案还可以根据用户偏好的变化不断学习修正,能够取得不错的实时推荐效果。另外,本文采用实际运行的众筹平台数据来完成算法的性能验证,具有更好的实用性。
[Abstract]:With the continuous progress of Internet technology, crowdfunding platform has become a new way of network financing. With the increasing data generated by crowdfunding platform, the data efficiency of crowdfunding platform is not proportional to the increase, so the so-called "information overload" phenomenon has emerged. Personalized recommendation system can solve this problem by mining users' interest preferences from a large number of data. It has achieved some success in e-commerce, social media, advertising system, search engine and other fields. However, in the rapidly developing field of network crowdfunding, there is no crowdfunding website to provide users with professional personalized recommendation services. In this paper, the overall situation of crowdfunding platform is analyzed, and the commonly used personalized recommendation algorithms are studied and compared, and the input and output of the recommendation system, as well as the selection and design of the recommendation algorithm are completed. At the same time, this paper uses machine learning algorithm to establish a recommendation system based on collaborative filtering, and designs the corresponding improvement scheme for the existing problems. On the one hand, aiming at the problem of data sparsity, a collaborative filtering algorithm based on implicit semantic model is designed, and the statistical learning method is used to solve the model optimization problem. By learning the characteristics of user scoring data and training the prediction model, the algorithm obtains the prediction score after filling into the original score matrix, and then takes the filled score matrix as the data source, and then obtains the prediction score based on collaborative filtering algorithm. On the other hand, aiming at the cold start problem caused by a single data source, combined with the user score and project attribute characteristics of crowdfunding platform, the collaborative filtering algorithm is improved. In this paper, the user score and project attribute data of crowdfunding platform are obtained by network communication technology, the feasibility of the recommendation algorithm is verified, and the parameters in the feature learning model are adjusted. The average absolute error and accuracy of the improved algorithm are compared. The experimental results show that the personalized recommendation algorithm of crowdfunding platform designed in this paper can provide accurate and fast personalized recommendation service, provide convenience for users, and is also conducive to the development of crowdfunding platform. The improved algorithm proposed in this paper solves the problem of data sparsity and cold start to a certain extent. Compared with the traditional recommendation algorithm, the accuracy of prediction is obviously improved. The scheme can also be revised according to the changes of user preferences, and can achieve a good real-time recommendation effect. In addition, this paper uses the actual crowdfunding platform data to verify the performance of the algorithm, which has better practicability.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3;TP181

【参考文献】