登录
Microsoft.com
 
Microsoft
 
 
 
Microsoft >
> 
> 
FAQ: How are results ranked?

Sticky FAQ: How are results ranked?

  • 2010年7月12日 12:26
    所有者：
     
     
    登录进行投票
    1

    The Ranking Mechanism in Microsoft Academic Search

    Microsoft Academic Search extracts and integrates information about academic objects (i.e. entities), including scientific publications, authors, conferences, journals, and organizations. To help users locate desired information quickly, these objects are ranked differently in different user scenarios based on their popularity and relevance.

    This document explains the ranking mechanism for the following user scenarios:

    Retrieve Objects through Keyword Search

    In this scenario, users issue a keyword query and expect that the most relevant objects are ranked at the top of the search result page. The objects in the search results are sorted based on two factors: their relevance to the query and their popularity. The relevance score of an object is computed based on its text information integrated from all sources, and the popularity score of an object is calculated using the object relationship graph of all objects in Microsoft Academic Search. For detailed information about our object ranking mechanism for keyword search, please refer to the following two papers:

    Browse Objects within a Specific Academic Domain

    In this scenario, users just want to browse the objects within a specific academic domain. For example, a student may just want to browse the researchers in the “data mining” domain.

    Microsoft Academic Search has different ranking mechanism for different types of objects. Specifically,

    · Papers are purely ranked by the number of citations;

    · Authors are ranked by the total number of citations of their papers published within the domain.

    · Conferences and journals are ranked mainly according to the number of publications and citations.More specifically, Microsoft Academic Search considers several factors: total citations, total number of papers, the starting year of the conference, and the PopRank (please read our WWW2005 paper: Object-Level Ranking: Bringing Order to Web Objects) of a conference/journal. For the young conferences/journals, their citation numbers will be much less than those of the established ones. The PopRank of these young conferences/journals becomes a better indicator than citations.

    · Organizations are ranked based on citations of all papers from its current and previous affiliated authors.

    For example, paper A is written by Author B and Author C, when this paper was published in year 2000, Author B is affiliated with Organization 1, and Author C was affiliated with Organization 2, provided such information was presented in the paper full text or meta data. Now Author B is affiliated with Organization 3, and Author C is with Organization 4. Our extraction and matching algorithm constructs the following relationship:

    Paper A is related to Organization 1, 2, 3, and 4.

    Therefore, all the citations to Paper A will be contributed to all 4 organizations.

    We also provide a time range feature for users to better browse objects arising in recent years. We currently have “All Years”, “Last 10 Years”, and “Last 5 Years”. For each range, we only consider the papers and citations within the corresponding timeframe.

    Note that the relative position of an object is designed to help users locate desired information easier, and it is by no means an authoritative indicator of the overall academic impact of the object. How to measure the impact of a scientific work is a very interesting and difficult research problem. We encourage researchers in the related fields to conduct experiments leveragingour API, and share with us their research findings.