I590 Interactive Visual Analytics Week 13 | Nov 16, 2016 Filtering - - PowerPoint PPT Presentation

Khairi Reda | redak@iu.edu School of Informa5cs & Compu5ng, IUPUI

Week 13 | Nov 16, 2016 Filtering and Aggregation Models in Visual Analytics

I590 Interactive Visual Analytics

http://www.michelecoscia.com/wp-content/uploads/2012/08/demon2.png

Filtering & Aggrega1on

• Too much data can overwhelm the

visualiza5on

• Some5mes we need to show less data

points

• Filter: eliminate irrelevant items
• Aggregate: group similar items

Filter

• Any func5on that par55ons

the data into two set based

• n aGributes
• Larger / smaller than X
• Within a specified

geographic extents

• Noisy / significant readings
• Filtering can also be apply to

aGributes, as opposed to the data point themselves

Based on a slide by Alex Lex

Filtering with Dynamic Queries

Schneiderman

Filtering with menus

Scented Widgets

Willett 2007, Via Alex Lex

• Provide cues (scent) to the users to aid in

filtering and explora5on

• Usually come in the form of small visual

representa5ons that bind to interface elements

Interac1ve Legends

• Provides filtering controls from the

legend

Riche 2010, Via Alex Lex

Aggrega1on

Histogram

• Aggregate items into bins
• Display the number of items (i.e., frequency)

in each bin

Histogram

Number of bins can affect the shape of the histogram

Distribution of passengers by Age 10 Bins 20 Bins

Based on a slide by Alex Lex

http://web.stanford.edu/~mwaskom/software/seaborn/tutorial/plotting_distributions.html

Density plots

Box plots (aka Box-and- Whisker) plots

• First quar5le: splits the lowest

25% of the data

• Median: splits data into half
• Third quar5le: splits the

highest 25% of the data

http://image.mathcaptain.com/cms/images/106/box-plot.png

Box plots (aka Box-and- Whisker) plots

Wikipedia

• An alterna5ve

representa5on to the min/max is to scale the whiskers by the Interquar5le Range (Q3-Q1)

One box plot, four distribu1ons

http://stat.mq.edu.au/wp-content/uploads/2014/05/Can_the_Box_Plot_be_Improved.pdf

Distribu1on, errors bars, and box plots

Streit & Gehlenborg, PoV, Nature Methods, 2014 Via Alex Lex

Violin plots

http://web.stanford.edu/~mwaskom/software/seaborn/tutorial/plotting_distributions.html

Heatmaps

• Aggregate 2D points into 2D bins

Heatmaps (for scaNerplots)

Spa1al Aggrega1on

Changing the boundaries / structure of the aggrega5on bins yields different results

Based on a slide by Alex Lex

Spa1al Aggrega1on

Gerrymandering

Based on a slide by Alex Lex

Clustering

• Classifica5on of items into “similar” bins
• Typically based on a similarity measure
• Euclidean distance, Pearson correla5on, etc…
• Many different clustering algorithms, with

weaknesses and strengths

• K-Means
• Hierarchical clustering

K-Means

• Pick K star5ng points as centroids.

Those eventually will comprise the clusters

• Calculate distance of every point to

centroid, assigning the point to the closest centroid

• Update the centroid to the average
• f the cluster’s members
• Repeat

K-Means

• Have to pick K
• Assump5ons about the data:

roughly “circular” clusters of equal size

http://stats.stackexchange.com/questions/133656/how-to- understand-the-drawbacks-of-k-means

Limita5ons

K-Means

Limita5ons

http://stats.stackexchange.com/questions/133656/how-to-understand-the-drawbacks-of-k-means

Dimensionality Reduc1on

Item ANr 1 ANr 2 ANr 3 ANr 4 ANr 5 ANr 6 ANr 7 ANr 8 ANr 9 ANr 10 ANr 11 …

A B C …

• High-dimensional data: large number of

aGributes

• Dimensionality reduc5on: Reduce number of

dimensions (aGributes) while keeping as much varia5on as possible

Dimensionality Reduc1on

Item ANr 1 ANr 2 ANr 3 ANr 4 ANr 5 ANr 6 ANr 7 ANr 8 ANr 9 ANr 10 ANr 11 …

A B C …

• Principle component analysis
• Mul5dimensional scaling
• And other techniques…

Principle Component Analysis (PCA)

• Find a new set of

dimensions (axes) that explains the majority of the variance in the data

• Order the new dimensions

by variance

• The first principle

component accounts for most variance

Principle Component Analysis (PCA)

http://setosa.io/ev/principal-component-analysis/

Mul1dimensional scaling (MDS)

• Project the high-dimensional

space onto a much lower space (e.g, 2D)

• Relies on similarity between

points (usually have to compute pairwise similarity between every pair of points)

• Non-linear transforma5on:

More difficult to interpret than PCA, but can maintain structures beGer in some cases

Models in Visual Analy1cs

Adapted from: http://slideplayer.com/slide/4659134/ and from Remo Chang, 2010

Models in Visual Analy1cs

• Abstrac5ons of how visualiza5on works:
• Provide a way of talking about how

humans interact with visualiza5ons

• Language for describing different parts of

the visual analy5c process

• Every model is (overly) simplified:

beware!

Terminology / Assump1ons

• Sense making: The act of processing incomplete

informa5on in order to improve one’s understanding

• f a situa5on and/or to make decisions
• A person’s decision making is bound by [1]
• incomplete informa5on
• the amount of 5me they have to decide
• the finite processing power of their brain
• Mental model: An abstracted versions of the real-

world that are more tractable

[1] H. Simon 1957. “A Behavioral Model of Rational Choice”

Models in Visual Analy1cs

Informa1on Visualiza1on Reference Model

Card, Mackinlay, and Scneiderman. Readings in Information Visualization: Using Vision to Think, Morgan Kaufmann, 1999, pp. 17

Van Wijk’s Model

Van Wijk, J. “The value of visualization”, 2005

D=Data V=Visualiza5on S=Specifica5on I=Image P=Percep5on K=Knowledge E=Explora5on

Keim’s Visual Analy1cs Model

Keim, D et al. “Visual Analytics: Definition, process, and challenges”, 2008

Pirolli and Card Sensemaking model

Pirolli, P and Card, S. “The sense making process and leverage points for analyst technology as identified through cognitive task analysis”, 2005

Next week

Class canceled — Happy Thanksgiving!

Week 15 - Nov 30

Time series and temporal data Inference and uncertainty in visualiza5on