Professor John Canny Fall 2001 Nov 29, 2001 CS 160: Lecture 24 Professor John Canny Fall 2001 Nov 29, 2001 9/19/2018
Review What is the difference between late and early fusion in MM interfaces? Give an example of each. List some input modes and their advantages and weaknesses. 9/19/2018
Multimodal Interfaces The OAI model (lecture 20) was our starting point for information organization. 9/19/2018
Information Tasks Specific Fact-finding: Extended Fact-finding: Find the phone number of Bill Clinton Extended Fact-finding: What kinds of music is Sony publishing? Open-ended browsing: Is there new work on voice recognition in Japan? Exploration of availability: What genealogy information is at the National Archives? 9/19/2018
Database queries Query languages like SQL are widely used, but are hard to learn and easy to make mistakes with. SELECT DOCUMENT# FROM JOURNAL-DB WHERE (DATE >= 1994 AND DATE <= 1997) AND (LANGUAGE = ENGLISH OR FRENCH) AND (PUBLISHER = ASIS OR HFES OR ACM) 9/19/2018
Visual Query Builders 9/19/2018
QBE: Query By Example User chooses a record (Database) or document (search engine) and specifies “more like this”. User can also pick a segment of text, even a paragraph, from a good document and use it as a search query (search engines only). 9/19/2018
Visualizing Search Results 9/19/2018
Multidimensional Scaling 9/19/2018
Multidimensional Scaling Multi-Dimensional Scaling (MDS) is a general technique for displaying n-dimensional data in 2D. It preserves the notion of “nearness”, and therefore clusters of items in n-dimensions still look like clusters on a plot. 9/19/2018
Multidimensional Scaling MDS applied to hand-classified discussion topics. 9/19/2018
Multidimensional Scaling Clustering of the MDS datapoints (discussion topics) 9/19/2018
Multidimensional Scaling MDS can be applied to search engine results easily because they automatically have a high-dimensional representation (used internally by the search engine). The MDS plot helps organize the data into meaningful clusters. You can search either near your desired result, or scan for an overview. 9/19/2018
Tasks for a visualization system Overview: Get an overview of the collection Zoom: Zoom in on items of interest Filter: Remove uninteresting items Details on demand: Select items and get details Relate: View relationships between items History: Keep a history of actions for undo, replay, refinement Extract: Make subcollections 9/19/2018
Visualization principles To support tasks 1 & 2, a general design pattern called “focus+context” is often used. Idea is to have a focal area at high resolution, but keep all of the collection at low resolution. Mimics the human retina. 9/19/2018
Distortion Several visualization systems use distortion to allow a focus+context view. “Fisheye lenses” are an example of strongly enlarging the focus while keeping a lot of context (sometimes the entire dataset). Many of these were developed at Xerox PARC. 9/19/2018
Focus+Context: Document lens 9/19/2018
Focus+Context: Webbook lens 9/19/2018
Focus+Context: Table lens 9/19/2018
Navigation: Hyperbolic trees 9/19/2018
Navigation: Hyperbolic trees 9/19/2018
Navigation: Hyperbolic trees 9/19/2018
Navigation: Hyperbolic trees 9/19/2018
Navigation: Hyperbolic trees 9/19/2018
Navigation: Animation 9/19/2018
Using 3D People perceive a 3D world from 2D views, so it seems like we could use 3D structure to advantage. Several systems (also Xerox PARC) have tried this. Use 3D spatial memory and organization to speed up navigation. 9/19/2018
WebBook 9/19/2018
Web Forager 9/19/2018
Representing Hierarchies 9/19/2018
Summary High-dimensional data can be visualized by 2D via MultiDimensional Scaling. Focus+Context is a design pattern for showing the area of interest and its relationship to the entire dataset. 3D techniques can leverage the spatial capabilities of the human visual system. 9/19/2018