Learning To Rank Relational Objects Based On The Listwise Approach

In order to improve the precision of users'retrieval and returning much more relational and important pages to users, how to improve the ranking quality of search engines has become an essential problem. When searching in the WWW, not only the importance degree of a webpage but also the relevance is a significant feature. But it is inadequate to evaluate a webpage only by the importance and relevance, so it is necessary to consider some other facts, such as the relationships among webpages. Up to now with the development of network technique, many factors which affect the ranking quality have been found. Using all these facts as features, and then considering them together with some methods to obtain a more reasonable webpages sorting list is just the problem which learning to rank must solve.This thesis is based on the webpages, and because webpages are the formatted texts with special tags, our research concentrates on the research of documents ranking. This thesis introduces traditional 8 feature extraction methods, and propose a new feature extraction method which is an improved TFIDF with the position weight, and combine the entropy of the word for a query. The new method can excellent represent an article and distinguish one document and others for umost. Formost, this method can represent the importance and relevance of a term in one document. Meanwhile, we propose a method to construct a ground truth sorting list for research of listwise learning to rank.Most existing learning to rank approaches only considered the content of the objects when building a ranking model, the relationship between objects are ignored. To address this issue, in this thesis a listwise approach using both the content information of objects and the relations between objects is proposed to build a ranking model. The relationships between objects are represented as an affinity graph. The ranking function for an object depends not only on the content of the object but also on the relations between these objects.The surrogate loss function is defined as the likelihood loss and cross entropy loss respectively. A stochastic gradient descent search is employed to find the minimum value of the surrogate loss function. Experimental results show that the proposed method outperforms the baseline listwise methods ListMLE, ListNet.