Improving Web Search Results Using Affinity Graph
description
Transcript of Improving Web Search Results Using Affinity Graph
SIGIR 1Intelligent Database Systems Lab
國立雲林科技大學National Yunlin University of Science and Technology
Improving Web Search Results Using Affinity Graph
Advisor : Dr. HsuPresenter : Jia-Hao YangAuthor :Benyu Zhang , Hua Li , Yi Liu , Wensi Xi , Weiguo Fan
SIGIR2Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Outline Motivation Objective Definition Methods (Affinity Ranking) Experiments Conclusion Opinion
SIGIR3Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Motivation situation
─ Many of the queries are ambiguous. ─ the user’s information needs are unknown.
Ex : “ 足球” , 是只想要足球還是要找足球賽 In traditional, precision and recall are two metr
ics, but these didn’t consider the content of documents.
Hyperlink
SIGIR4Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Objective Two metrics, diversity and information
richness, have been proposed to improve this problem.
Re-ranking the top search results to satisfy the user’s information needs.
SIGIR5Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Definition Diversity measures the variety of topics in a gr
oup of documents. Information richness measures how many dif
ferent topics a single document contains.
SIGIR6Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Methods AG : According to vector space model, each
document can be represented ,
If we consider documents as nodes, the document collection can be modeled
as a graph by generating the link between
documents.
d1
d5d6
d4
d3d2
SIGIR7Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Methods(cont.) Information richness : 1st
2nd
SIGIR8Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Methods(cont.) Diversity penalty : 1st :
2nd
3rd ,
4th
5th 2nd
Re-ranking :─ The score-combination scheme uses a linear combination of two parts:
─ The rank-combination scheme of re-ranking uses a linear combination of the ranks based on full-text search and Affinity Ranking :
SIGIR9Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (In Yahoo & ODP) Affinity Ranking vs. K-Means Clustering
SIGIR10Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (cont.)
SIGIR11Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (cont.)
SIGIR12Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (In Newsgroup)
Improve in Top 10 Search Results : As the top 10 search results always receive the most attention of end-users,
we show how Affinity Ranking affects the top 10 search results from the newsgroup data set.
SIGIR13Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (cont.) Improve within Top 50 Search Results
SIGIR14Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (cont.)
SIGIR15Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Experiments (α & β)
SIGIR16Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.A Case Study Outlook print error :
SIGIR17Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Conclusion This paper proposed two new metrics, diversity and
information richness, and a novel ranking scheme, Affinity Ranking, to measure the search performance.
By presenting wider topic coverage and more highly informative results in each topic in the top results, this method can effectively improve the search performance.
SIGIR18Intelligent Database Systems Lab
N.Y.U.S.T.
I. M.Opinion Future work : scaling the AR computation, to
the Web scale.