1 / 33

Lecture 13

Lecture 13. Housekeeping Term Projects Evaluations

addo
Download Presentation

Lecture 13

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. Lecture 13 • Housekeeping • Term Projects • Evaluations • Morse, E., Lewis, M., and Olsen, K. (2002) Testing Visual Information Retrieval Methodologies Case Study: Comparative Analysis of Textual, Icon Graphical and 'Spring' DisplaysJournal of the American Society for Information Scienceand Technology (JASIST) PDF • Reiterer H., Mußler G., Mann T.: Visual Information Retrieval for the WWW, in: Smith M.J. et al. (eds.), Usability Evaluation and Interface Design, Lawrence Erlbaum, 2001 PDF • searchCrystal Studies

  2. Prototype Project • Motivate domain choice. • Perform task and need analysis. • Describe design approach and information visualization principles used. • Develop prototype. • Have an "domain expert" use the prototype and provide feedback. • Class PresentationYou have 15 min. to describe task analysis and your design approach.Demonstrate your prototype.Report on the "domain expert" feedback. • Create Report20 to 25 pages, written as a standard paper  10pt, double-spacedProvide screenshots of prototype and explain design approach.Include URL of prototype. • Hand-inHardcopy of report.Post report online and send instructor an email with the URL.

  3. Text Retrieval Visualizations – Evaluations : Morse et al. • Many Tools Proposed • Few Tested and Often Inconclusive / Fare Poorly • Simplify Evaluation  Focus on Method (instead of implementation)  Only Static Aspects • POI = Point of Interest Visualizations • Position Coding • Glyph = Graphical Entity • Conveys data values via attributes such as shape, size, color

  4. Glyph = Graphical Entity

  5. Evaluation – Morse et al.

  6. Evaluation – Morse et al. : Two-Term Boolean Test

  7. Evaluation – Morse et al. : Two-Term Boolean Test

  8. Evaluation – Morse et al. : Three-Term Boolean Test

  9. Evaluation – Morse et al. : Vector Studies – Text List

  10. Evaluation – Morse et al. : Vector Studies – Table

  11. Evaluation – Morse et al. : Vector Studies – Icons

  12. Evaluation – Morse et al. : Vector Studies – VIBE

  13. Evaluation – Morse et al. : Vector Studies Time

  14. Evaluation – Reiterer et al.

  15. Evaluation – Reiterer et al.

  16. Evaluation – Reiterer et al.

  17. Evaluation – Reiterer et al.

  18. Evaluation – Reiterer et al.

  19. searchCrystal – Studies • Validate Design Approach • How does Overlap between Results Actually Correlate with Relevance? • User Study

  20. Overlap between Search Results Correlated with Relevance? • Method • Use Ad-hoc track data for TREC 3, 6, 7, 8 • Systems search the SAME Database • Automatic Short Runs • 50 Topics and 1,000 Documents per topic 50,000 documents • Retrieval systems can submit multiple runs  Select Best Run based Mean Average Precision TREC 319systems928,709documents found TREC 624systems1,192,557documents found TREC 728systems1,327,166documents found TREC 835systems1,723,929documents found • Compute Average by summing over all 50 topics and divide by 50

  21. How does Overlap Correlate with Relevance? Percentage of Documents that are Relevant Systems  Authority Effect

  22. TREC 8 – Impact of Average Rank Position? • Compute overlap structure between top 50 search results of 35 random groupings of 5 retrieval systems for 50 topics. Percentage of Documents that are Relevant Systems  Ranking Effect

  23. searchCrystal – Studies • How does Overlap between Search Results Correlate with Relevance? • Authority Effect– the more systems that find a document, the greater the probability that it is relevant • Ranking Effect– the higher up a document in a ranked list and the more systems that find it, the greater the probability of its relevance • Validates searchCrystal’s Design Approach • searchCrystal Visualizes Authority & Ranking Effects • searchCrystal can Guide User’s Exploration Toward Relevant Documents

  24. searchCrystal – Studies • Validate Design Approach • How does Overlap between Results Actually Correlate with Relevance? • User Study • http://www.scils.rutgers.edu/~aspoerri/study/UserStudy.swf

  25. User Study – Cluster Bulls-Eye

  26. User Study – RankSpiral

  27. User Study – Compare Cluster Bull’s Eye and RankSpiral • Nine undergraduates. • Short Introduction and No Training. • Randomized presentation order of data sets and display type. • Subject selects ten document; • Visual feedback about correct top 10 • http://www.scils.rutgers.edu/~aspoerri/study/UserStudy.swf • Test for Cluster Bull’s Eye and RankSpiral displays: • 1) How well can novices use visual cues to find the documents that are most likely to be relevant? • 2) Performance difference in terms of effectiveness and/or efficiency? • 3) How much document’s distance from the display center will interfere with the size coding used to encode its probability of being relevant

  28. User Study – Results • Hypothesis 1: “Novices can perform the task.” • Error is minimal for the top 7 documents and increases rapidly after the top 7 documents for both displays. • Novice users can use the Cluster Bulls-Eye and RankSpiral displays to select highly relevant documents, especially the top 7 documents. • Hypothesis 2: “RankSpiral outperforms Cluster Bulls-Eye.” • 8 of the 9 subjects performed the task faster using the RankSpiral. • Average time difference was 7.89 seconds. • The one-sided T-test value is 0.033, which is significant at the 0.05 level. • 7 out of 9 subjects performed the task more effectively using the RankSpiral. • Average “relevance score” difference is 0.034. • The one-sided T-test value is 0.037, which is significant at the 0.05 level. • Hypothesis 3: “Distance from center dominant cue.”

  29. Discussion • Relax searchCrystal’s design principles? • Mapping documents found by the same number of engines into the same concentric ring. • Option: Distance and Size encode likelihood that a document is relevant. • Internet search results: • Concentric rings are of value, because it is much harder to estimate a document’s probability of being relevant.

  30. Cluster Bulls-Eye Size = Distance from Center

  31. Cluster Bulls-Eye Size = Distance from Center

  32. searchCrystal - Studies • Authority & Ranking Effects Comparing Results of All Retrieval Systems at once Comparing Results of Random Subsets of Five Systems  Validating searchCrystal’s Design Principles • User Study Identify Top 10 Docs in Cluster Bull’s Eye and RankSpiral • Novice Users can use the two searchCrystal displays • Statistical Difference between two displays • Distance from center is dominant visual feature

  33. What is Popular on Wikipedia?Why? • Please read the two papers published by me in First Monday:http://www.firstmonday.org/ISSUES/issue12_4/ • Approach • 1 Visualize Popular Wikipedia Pages • Overlap between 100 Most Visited Pages on Wikipedia for September 2006 to January 2007 • Information Visualization helps to gain quick insights • 2 Categorize Popular Wikipedia Pages • 3 Examine Popular Search Queries • 4 Determine Search Result Position of Popular Wikipedia pages • 5 Implications

More Related