Facilitating Entity Navigation over Semantic Web

Presentation on theme: "Facilitating Entity Navigation over Semantic Web"— Presentation transcript:

1 Facilitating Entity Navigation over Semantic Web
Liang Zheng 8.31

2 Navigation is a process whereby people determine where they are, where everything else is, and how to get to particular objects or places. Psychology of "real world" navigation stresses the importance of landmarks, routes and survey for orientation.

3 Related Work Path (&Link)
Human Wayfinding in Information Networks. WWW12 The link-prediction problem for social networks.  Kleinberg07 Mining Missing Hyperlinks from Human Navigation Traces. WWW15 Influencing web-browsing behavior with intriguing and informative hyperlink wording. JIS05 Improved Search Engines and Navigation Preference in Personal Information Management. TOIS06 Modeling human navigation ScentTrails: Integrating Browsing and searching on the web. CHI03

4 Our Task (Related) Entity Navigation Link (&property)
Entity (entity type)

5 Related Work Hearst M A. Clustering versus faceted categories for information exploration[J]. Communications of the ACM, 2006, 49(4): Scatter/Gather: A cluster-based approach to browsing large document collections. SIGIR92 Facet Navigation: FacetMap, Longwell, BrowseRDF, Flamenco, gFacet, tFacet, Humboldt, Exhibit, /facet, and Visor. These tools focus on: how the facets are specified; how the facets are presented to users; and which operations can be applied to filtering items.

6 现有大多数的方法只是单向的，即要么是聚类entity set，要么是通过facet过滤entity set，没有考虑到facet和entity (entity type)之间存在的自然关系。
我们考虑对facet与entity双向聚类生成，即在自动生成facet时结合entity type的信息，反之亦然。

7

8 Our Method Co-clustering Link set Type set . .
最终得到Link和Type的k个聚类，{L1, L2,…, Lk}和{T1, T2,…, Tk}. 若li ∈Lm ,则表明li经常指向Tm 中的type;同样,若tj ∈Tm ,则表明tj 经常被Lm 中的link指向。 li tj . . lm tn Hierarchical co-clustering

9 Hierarchical co-clustering 相关算法
主要思路：1：将二部图转化成二维矩阵。2：对矩阵的行列同时进行聚类，行聚类和列聚类交替进行，通过行列之间自然存在的二元关系实现行聚类过程和列聚类过程的彼此约束。 Direct clustering( J.A. Hartigan, 1972)，该方法结合局部贪心分裂过程将文档-词协同矩阵划分成许多块，以聚类后的各个块的方差之和为目标函数，目标函数越小说明聚类效果越好。 Dhillon将协同聚类转化为二部图的划分问题. Hierarchical co-clustering 相关算法 每一层节点的分裂都是一次 Co-clustering聚类，簇的数目随着层次的加深而递增，直到节点满足合适的停止标准，树的分裂停止。

10 Q & A