660 likes | 804 Views
Lecture 20 : Time Series. April 23, 2013 COMP 150-2 Visualization. Definition. Data set containing a temporal (chronological) component
E N D
Lecture 20:Time Series April 23, 2013 COMP 150-2Visualization
Definition • Data set containing a temporal (chronological) component • Random sample of 4000 graphics from 15 of the world’s newspapers and magazines from 1974-1980 found that 75% of graphics published were time series (Tufte)
Problem • Problem is that time is abstract and not easily “seen”. • In everyday life, we understand temporal relations mostly by relying on memory. We cannot compare two temporal “objects” by placing them side-by-side. • Human memory is shoddy.
Datasets • Each data point in a time-series is likely an event, or an observation in time. • Each variable therefore has a timestamp. • Numerous examples: patient health information, stock prices, sports stats, news, photos taken, etc.
Example • Visualize these datasets: • Dataset 1: • Game 1, Paul Pierce scored 25 points • Game 2, Paul Pierce scored 12 points • Game 3, Paul Pierce scored 20 points • : • Dataset 2: • Game 1, Paul Pierce, 25 pts, 2 rebounds, 6 turnovers • Game 2, Paul Pierce, 12 pts, 8rebounds, 0 turnovers • Game 3, Paul Pierce, 20 pts, 4 rebounds, 2turnovers
Example • Visualize these datasets: • Dataset 3: • 1994-1997: Remco went to Johns Hopkins Univ in MD • 1997-2000: Remco went to Brown Univ in RI • 2000-2001: Remco worked for a startup in Boston • 2001-2002: Remco travelled the world • 2002-2003: Remco worked for Boeing in DC • 2003-2010: Remco went to UNC Charlotte in NC
Example • Visualize these datasets: • Dataset 4: • Day 1: 5 articles about Obama, 2 about Iraq, 5 about national debt • Day 2: 3 articles about Obama, 4 about Iraq, 6 about national debt • Day 3: 1 articles about Obama, 6 about Iraq, 7 about national debt
Meta-Level Perspective • Once you know of a temporal pattern, there are methods to find them. • E.g. Did Paul Pierce have a bad game today compared to yesterday? • However, the data is often not single dimensional… • E.g. considering Rebounds and Turnovers, the question is less clear
Meta-Level Perspective • Often, identifying the question is the hard part. • Visualizations can help with displaying the temporal data so that a person can begin asking questions.
Task Analysis • What do people usually ask? • When something was at its peak / bottom (stock prices) • Is there a pattern? • Are two series similar? • Does a series contain (or match) a known pattern?
Task Analysis • More tasks… • What happened at time t? • When does an event occur? • How often does an event occur? • How long does an event last? • How quickly do events change? • Do two events happen next or near each other? • In what order do certain events take place?
Taxonomy • Discrete points vs. interval points • (specific time vs. a span of time) • Linear time vs. cyclic time • (1-2pm vs. Monday-Saturday) • Ordinal time vs. continuous time • (game 1, game 2, etc. vs. 1-2pm) • Ordered time vs. branching time vs. multiple perspectives of time • (game 1 is before game 2, vs. if I go to work instead of staying at home, vs. what I measure in those two paths)
Design Decision • Animation or not? • Static • Shows history, multiple perspectives, allows comparison. • Dynamic (animation) • Gives feel for process and changes over time. Can be concise in visualization.
Standard Presentation • Present time data as a 2D line graph with time on x-axis and other variable(s) on the y-axis.
LifeLines Project • Visualize personal history in some domain Plaisant et al. CHI 96
Pros and Cons • Pros: • Reduce chances of missing information • Facilitate spotting trends or anomalies • Streamline access to details • Remain simple and generalizable to many domains • Cons: • Scalability • Multiple records
LifeLines 2 • Allow querying • Allow align -> rank –> filter • Find temporal coincidence of two events • First pneumonia and asthma attack • Reduce panning and zooming
LifeLines 2 • http://www.cs.umd.edu/hcil/lifelines2/
TimeSearcher Create rectangles that function as matching regions Dark gray = query matches
TimeSearcher http://www.cs.umd.edu/hcil/timesearcher/
Challenges with Querying Time • How to determine if two time series are the same (or similar)?
Stacked Graph http://www.nytimes.com/interactive/2008/02/23/movies/20080223_REVENUE_GRAPHIC.html
Structured Time • Most line-based visualizations do not take into account how we perceive time – days, weeks (weekdays, weekends), months, years, which don’t always fall on regular intervals. • Can we create visualizations based on these structured times?
For Example Number of flu cases over 3 years
Archimedes Spiral • Polar coordinates (r, θ) • r = a θ • Where a controls the tightness of the spiral
Pros and Cons • Pros • Scales well to large datasets • Finds periodic structures in the data • Comparison between cycles • Cons • The cyclic pattern might not be known • Inner circles receive fewer pixels
Different colors represent different time series • Is the period pattern still visible? • Can one compare between the different series?
Calendar Visualization • Calendar time is irregular but hierarchical in structure (event view, day view, week view, month view, year view, task view, etc.) • Visualization task: • See commonly available times for groups of people • Show both details and broader context
Spiral Calendar Mackinlay et al. UIST 94
Calendar View + Cluster Analysis • Task • Find the similar days in your calendar and merge them into a composite • Repeat process until no new patterns are found • What would you find? • Repeated days? • Outliers?
Calendar View + Cluster Analysis van Wijk, InfoVis 99
Characteristics • Unique types of days get their own color • Contextually placed in a calendar and a line graph • Clustering stops when a threshold is met or a pre-determined number of clusters is met