The Study Of Ranking And Recommendation Algorithms Based On Heterogeneous Graph

As the Internet has become more popular and especially the mobile Internet has explosively developed recently, resources on the Internet is becoming richer: 1). exponential growth of the amount of data; 2). increasingly diverse types of data resulting from different sources. How to discover a variety of valuable information from amounts of data becomes more important and has attracted strong interest from academia and industry. With the continuous development of machine learning techniques, using machine learning methods forthe massive data mining has become a very popular and effective method.Ranking and recommendation algorithms are two common types of algorithms forthe massive data mining. Ranking algorithms are to solve the problem, that is, how to sort the results according to the relevance to the query entered by the users in the search engine. Recommendation algorithms seek to predict the ratingthat user would give to an item according to the user preference. They can provide the information the users wants to get to them and help them find what they seek, therefore, they are very important for the industry and have also attracted a great deal of academic attention.This paper studies the ranking and recommendation algorithms based on heterogeneous graph.Such algorithms can not only use machine learning methods to mine valuable information from the data, but also make use of heterogeneous graph to improve the performance with richer informationfrom different sources. In this paper, two aspects were studied: 1). A heterogenous automatic feedback semi-supervised method for image re-ranking algorithm. Visual graph-based re-ranking methods due to the excellent performance have attracted a lot of attentions. In these methods, graph is firstly constructed, where the nodes of the graph are visual features and weights of the edges are visual similarity between the nodes, and then some sort algorithm is run to get the score of each node. Such graph-based methods are based on ranking score consistency assumption: the adjacent nodes have the similar ranking scores. But for multimedia retrieval, such as image search, the use of only a single-mode data (visual features) is not very good. Based on the assumption that the semantic meanings of visual features and textualfeatures are correlated and they are the representations of the image from differentviews, the proposed method considers visual andtextual features simultaneously. Specifically, a heterogenous graph is firstly constructed in which each node representingan image includes visualfeature and textual feature. Then, a graph-based semi-supervised learningmethod with an automatic feedbackis proposed to propagate theranking scores onthe graph, which can update the weights of heterogeneous graph automatically.Finally, the result is used to re-rank the images. 2). A hybrid recommendation method with implicit social information based on heterogeneous graph. Recently more and more work began to study how to use social information to improve the performance of collaborative filtering (especially matrix factorization) algorithm. Some authors have begun to study the use of implicit social information in the matrix factorization algorithm. However, only traditional Pearson correlation coefficient and cosine similarity function are used to tap the implicit social information. We firstly construct the heterogeneous graph, and then use the random walk approach to mining implicit social information. Therefore, we propose a hybrid recommendation method with implicit social information based on heterogeneous graph to improve the performance of collaborative filtering method with memory based recommendation method.Compared with the classical algorithms, the above-mentioned two algorithms have achieved good results in published datasets. The proposed ranking and recommendation algorithms havea good potential for applicationin the Internet industry.