1 / 31

Harikrishnan Karunakaran Sulabha Balan

Keyword Searching and Browsing in Databases using BANKS Gaurav Bhalotia, Arvind Hulgeri, Charuta Nakhe , Soumen Chakrabarti , S. Sudarshan. Harikrishnan Karunakaran Sulabha Balan. CSE 6339 . Outline. Introduction Database and Query Model Informal Model

renata
Download Presentation

Harikrishnan Karunakaran Sulabha Balan

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Keyword Searching and Browsing in Databases using BANKSGaurav Bhalotia, Arvind Hulgeri, CharutaNakhe, SoumenChakrabarti, S. Sudarshan Harikrishnan Karunakaran Sulabha Balan • CSE 6339

  2. Outline • Introduction • Database and Query Model • Informal Model • Formal Model • Query and Answer Model • Searching for the Best Answers • Backward Expanding Search Algorithm • Browsing through BANKS • Experience and Performance • Related Work • Conclusion

  3. Introduction • With the onset of the web, number of users needing to access online databases have increased • Search engines have popularized use of unstructured querying which just needs the user to type in the keyword and follow links • Same methodology cannot be used in querying databases as knowledge of schema and querying language like SQL is needed • Keyword searching will not also work on datatbases because the data is usually spread across tables/tuples due to normalization

  4. BANKS • Browsing ANd Keyword Searching • Enables Keyword-Based search on Relational Databases along with Data and Schema Browsing • User interacts with data through typing keywords, following hyperlinks and using controls made available. • Absolutely no knowledge of querying or programming languages required of the user

  5. BANKS contd…

  6. BANKS contd… • Makes Joins Implicit and Transparent • Incorporates Notions of Proximity and Prestige • Methods to publish relational data which would otherwise remain invisible on the web are provided • Creates hierarchical and graphical views of data with hyperlinks to navigate through them

  7. Database and Query Model

  8. Schema

  9. Database Fragment Snapshot

  10. Informal Model • Answer to query should be sub-graph connecting nodes matching the keywords • Central node that connects all keyword nodes is Information Nodeand tree is Connection Tree. Foreign Key Dependencies constitute edges in the graph • Edges (References) are given weights according to its type • Weight of the tree is proportional to total of its edge weights and relevance is inversely proportional to its weight

  11. Informal Model contd… • To obtain model with edges directed away from information node and preserve directionality we make use of backward edges • Backward Edge assigned weight proportional to in-degree • Information Node is selected from certain sets of nodes in the graph • Backward Edges ensure that tree is rooted at the Information Node • To avoid problem of “hubs”, edges connecting popular nodes are given a higher weight thus lowering proximity

  12. Informal Model contd… • Concept of Node weights introduced to include Prestige Rankings • Nodes having more pointers given higher prestige • In BANKS node prestige is assigned based on in-degree of the node • Node weights and Tree Weights combined to obtain relevance score

  13. Formal Database Model • Each tuple Τ has a corresponding node uτ • Each node u has a node weight N(u) depending upon the prestige of the node • Between each pair of related tuples T1 & T2 , graph contains edge between uτ1 to uτ2 and back edge from uτ2to uτ1 • Similarity between two relations R1 and R2 depends upon the type of link from R1 to R2 and is set to infinity if R1 does not refer to R2

  14. Formal Database Model Edge weights • Depending upon importance of the link we set a value to the edge. Default value is 1 • The weight of the directed edge(u, v) depends on factors: • If (u, v) exists but (v, u) does not, assign the weight s(R(u),R(v)) to (u, v) • If (u, v) does not exist and (v, u) does, assign the weight INv (u) s(R(v), R(u)) to (u, v) where INv is the indegree of u contributed by the tuples belonging to relation R(v) • If both (u, v) and (v, u) exist in the graph , assign the weight as the minimum of two values min{s(R(u),R(v)) , INv (u) s(R(v), R(u))

  15. Query and Answer Model • Query • A set of keywords e.g.{k1,k2,…kn} • A set of nodes Si= {S1,S2,…Sn} • Locate nodes matching search terms t1,t2,…tn • Answer Model • A rooted directed tree connecting keyword nodes (at least one node from Si). • Note: Tree may also contain nodes not in any Si, Steiner Tree • Relevance score of an answer tree • Combination of its nodes and its edge weight presented in decreasing order

  16. Query and Answer Model • Calculating Relevance Score involves adjustment of both node weights and edge weights along with a factor to control individual weight variations • Node weights • Scaled to Nmax and depressed using log • Nscore(v) = N(v)/ Nmax or log(1+N(v)/Nmax • Overall Nscore taken to be average of node scores • Edge Weights • Normalized Escore(e) obtained by diving edge weight by minimum edge weight • Escore(e) = log(1+w(e)/wmin) • Overall Edge Score = 1/(1 + Σe Escore(e)) • Combination of Overall Edge Score and Node Score • Additive : (1-λ)Escore + λNscore • Multiplicative : Escore * Nscoreλ

  17. Searching for the Best Answers • We have to use not just the tree with the highest relevance score but also those with high scores • Answers have to be generated incrementally so that the user are provided with the ‘best’ answers at the beginning • Resultant Graph is assumed to fit in memory since only Row IDs and index to map RowIDs to nodes in the graph need to be stored by us.

  18. Backward Expanding Search • Incrementally computes search results • Start at leaf nodes each containing a query keyword • Run concurrent single source shortest path algorithm from each such node • Traverses the graph edges backwards • Confluence of backward paths identify answer tree roots • Output a node whenever it is on the intersection of the sets of nodes reached from each keyword • Answer trees may not be generated in relevance order • Insert answers to a small buffer (heap) • Output highest ranked answer from buffer to user when buffer is full

  19. Backward Expanding Search

  20. Backward Expanding Search

  21. Searching for Best Answers • Model (Query : Roy Sudarshan)

  22. Backward Expanding Search • Due to the graphs being Steiner Trees lot of time is spent doing wasteful exploration of the graph • As keyword nodes increase, the feasibility of the algorithm decreases • Connection Trees are only approximately sorted in their increasing order of weights • Node weights are not considered, hence trees may not be produced in exact decreasing order of relevance

  23. Browsing through BANKS • BANKS system provides • A rich interface to browse data stored in a relational database • Automatically generates browsable views of database relations and query results • Schema browsing and data browsing • A hyperlink to the referenced tuple

  24. Browsing through BANKS contd… • Functionalities • Columns can be projected away • Selections can be imposed on columns • Joins can be performed with foreign key columns by joining them with referencing tables • Results can be grouped by on columns which returns only distinct values in column being displayed • Sorting can be done on columns

  25. Sample Browsing

  26. Templates • Cross Tabs • Group By Template to view Data hierarchially • Folder Views modeled after the folder view supported by Windows Explorer etc. • In the form of bar chart, line chart or pie chart with HTML image maps to embed hyperlinks in the graphics

  27. Performance Evaluation • Datasets of varying sizes have been tested • No agreed upon benchmarks for Ranking Algorithms in this domain • System was found to return the most intuitive answers

  28. Evaluating Error Score • Ideal answers were obtained using different queries • Compute absolute value of rank difference of the ideal answers with rank in the answers for given parameter setting • Sum of rank differences gives the raw error score for that parameter • We map error scores against λ and log-scaling of edge weights

  29. Error Scores

  30. Observations • Setting λ = 0.2 produced best results while λ = 1 produced worst with error scores of around 15 • Log scaling of edge weight is important as otherwise back-edges from popular nodes would result in correct answers getting low relevance scores • Additive or Multiplicative combination has no effect on ranking • Node weights were abandoned as log-scaling and no log scaling produced same ranking

  31. Conclusions • Effective when using queries matching non-metadata keywords • Brings to light data that might not be readily available on the web to the non technical user • Higher the no. of keywords, the less useful backward-expanding search algorithm becomes • Over reliance on Java can at times cause slow down of application

More Related