基于网络演化的推荐算法分析与网络压缩重建算法设计

发布时间：2018-08-14 13:40

【摘要】：随着互联网技术的蓬勃发展和电子商务的规模不断扩大,个性化推荐技术给人们的生活带来了巨大便利。然而,传统的推荐算法大多局限于静态数据和单次推荐场景,忽略了推荐应用场景随时间演化特征和推荐算法有效性的伴随变化。本文结合网络科学基本理论,采用二分网络刻画推荐问题,结合推荐场景动态演化的特征,建立用户在线选择模型,研究在线推荐算法的有效性和在线系统的协同演化问题,并针对大规模网络压缩提出了新的办法。主要内容有:1.研究了推荐算法的性能在系统中的长期演化特征。本文设计了一种用户选择模型模拟在线系统与推荐算法协同演化的过程,系统地检测了几种经典推荐算法的推荐性能处于在线系统演化下的长期变化情况。研究发现,在系统演化完全依赖推荐算法的情况下,推荐算法的单步推荐性能会逐渐变差。有趣的是,研究还发现了用户的随机选择会改善推荐算法的长期性能。当系统采用混合推荐算法时,研究发现算法的最优参数值向着是推荐多样性改善的方向移动,这表明推荐多样性的改善对保持长期推荐准确性很重要。最后在实证中验证了模型的结果。本研究为设计长期有效的推荐算法提供了理论支撑。2.提出了一种层次化的动态网络压缩算法。本文针对大规模网络压缩算法存在的问题,提出了一种新的层次化动态网络压缩算法-HDSLN(Hierarchical Dynamic Summarization of Large Networks),通过网络分割,边的重连和迭代压缩的方法,将一个大规模网络层次化地压缩成小规模网络,同时尽可能地保留网络的原有结构。此外,本文还提出了一种新的基于Super-Net的网络重建算法,使得我们可以根据Super-Net尽可能相似地还原出原网络。同时,为了验证算法的性能,我们采用人工和真实数据集对HDSLN算法进行了实验和分析。
[Abstract]:With the rapid development of Internet technology and the expansion of e-commerce, personalized recommendation technology has brought great convenience to people's life. However most of the traditional recommendation algorithms are limited to static data and single recommendation scenarios ignoring the evolution characteristics of recommendation scenarios over time and the validity of recommendation algorithms. Combined with the basic theory of network science, the bipartite network is used to describe the recommendation problem, and the dynamic evolution of recommendation scene is combined to establish the online selection model of users. The effectiveness of online recommendation algorithm and the co-evolution of online system are studied. A new method for large-scale network compression is proposed. The main content is: 1. The long-term evolution characteristics of the performance of the recommendation algorithm in the system are studied. In this paper, we design a user selection model to simulate the collaborative evolution of online systems and recommendation algorithms, and systematically detect the long-term variation of the recommendation performance of several classical recommendation algorithms under the evolution of online systems. It is found that the single-step recommendation performance of the recommendation algorithm will deteriorate gradually when the system evolution is completely dependent on the recommendation algorithm. Interestingly, the study also found that random selection of users improves the long-term performance of recommendation algorithms. When the hybrid recommendation algorithm is used in the system, it is found that the optimal parameter value of the algorithm moves towards the direction of the improvement of the recommendation diversity, which indicates that the improvement of the recommendation diversity is very important to maintain the accuracy of the long-term recommendation. Finally, the results of the model are verified in the empirical analysis. This study provides theoretical support for the design of long-term effective recommendation algorithm. 2. A hierarchical dynamic network compression algorithm is proposed. In this paper, a new hierarchical dynamic network compression algorithm, HDSLN (Hierarchical Dynamic Summarization of Large Networks), is proposed to solve the problems of large scale network compression algorithm, which is based on network segmentation, edge reconnection and iterative compression. A large scale network is hierarchically compressed into a small scale network while preserving the original network structure as much as possible. In addition, a new network reconstruction algorithm based on Super-Net is proposed, which enables us to restore the original network as similar as possible according to Super-Net. At the same time, in order to verify the performance of the algorithm, we use artificial and real data sets to test and analyze the HDSLN algorithm.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】