2016-02-11 09:37:51 -05:00
<!doctype html>
< html lang = "en" >
< head >
< meta charset = "utf-8" >
< title > Embracing Uncertainty< / title >
< meta name = "description" content = "Mimir, an awesome system for embracing uncertainty" >
< meta name = "author" content = "Oliver Kennedy" >
< meta name = "apple-mobile-web-app-capable" content = "yes" / >
< meta name = "apple-mobile-web-app-status-bar-style" content = "black-translucent" / >
< meta name = "viewport" content = "width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no, minimal-ui" >
< link rel = "stylesheet" href = "../reveal.js-3.1.0/css/reveal.css" >
< link rel = "stylesheet" href = "ubodin.css" id = "theme" >
<!-- Code syntax highlighting -->
< link rel = "stylesheet" href = "../reveal.js-3.1.0/lib/css/zenburn.css" >
<!-- Printing and PDF exports -->
< script >
var link = document.createElement( 'link' );
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = window.location.search.match( /print-pdf/gi ) ? '../reveal.js-3.1.0/css/print/pdf.css' : '../reveal.js-3.1.0/css/print/paper.css';
document.getElementsByTagName( 'head' )[0].appendChild( link );
< / script >
<!-- [if lt IE 9]>
< script src = "../reveal.js-3.1.0/lib/js/html5shiv.js" > < / script >
<![endif]-->
< / head >
< body >
< div class = "reveal" >
<!-- Any section element inside of this container is displayed as a slide -->
< div class = "header" >
<!-- Any Talk - Specific Header Content Goes Here -->
Embracing Uncertainty
< / div >
< div class = "footer" >
<!-- Any Talk - Specific Footer Content Goes Here -->
< div style = "float: left; margin-top: 15px; " >
Exploring < u > < b > O< / b > < / u > nline < u > < b > D< / b > < / u > ata < u > < b > In< / b > < / u > teractions
< / div >
< img src = "graphics/FullText-white.png" height = "40" style = "float: right;" / >
< / div >
< div class = "slides" >
< section >
< h2 > Embracing Uncertainty< / h2 >
< h4 > Oliver Kennedy< / h4 >
< / section >
< section >
< h2 > Embracing Uncertainty< / h2 >
< h4 > Oliver Kennedy< / h4 >
< h4 style = "color: blue" > Ying Yang, Niccolo Meneghetti, < br / > Arindam Nandi, Vinayak Karuppasamy< br / > (UB)< / h3 >
< h4 style = "color: red" class = "fragment" > Ronny Fehling, Zhen-Hua Liu, Dieter Gawlick< br / > (Oracle)< / h3 >
< / section >
< section >
< section >
< h3 > A Big Data Fairy Tale< / h3 >
< / section >
< section >
< img src = "graphics/dagobert83-female-user-icon-800px.png" height = "300" / >
< h4 > Meet Alice< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< section >
< img src = "graphics/dagobert83-female-user-icon-800px.png" height = "300" / >
< img src = "graphics/littlestorefront-800px.png" height = "300" / >
< h4 > Alice has a Store< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< section >
< img src = "graphics/littlestorefront-800px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > →< / span >
< img src = "graphics/matt-icons_text-x-log-300px.png" height = "300" style = " vertical-align: middle;" / >
< h4 > Alice's store collects sales data< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< section >
< img src = "graphics/dagobert83-female-user-icon-800px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > +< / span >
< img src = "graphics/matt-icons_text-x-log-300px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > =< / span >
< img src = "graphics/saco-800px.png" height = "300" style = " vertical-align: middle;" / >
< h4 > Alice wants to use her sales data to run a promotion< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< section >
< img src = "graphics/matt-icons_text-x-log-300px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > →< / span >
< img src = "graphics/database-server-800px.png" height = "300" style = " vertical-align: middle;" / >
< h4 > So Alice loads up her sales data in her trusty database/hadoop/spark/etc... server.< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< section >
< img src = "graphics/database-server-800px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > + ?< / span >
< h4 > ... asks her question ...< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< section >
< img src = "graphics/database-server-800px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > + ? →< / span >
< img src = "graphics/crystalball-800px.png" height = "300" style = " vertical-align: middle;" / >
< h4 > ... and basks in the limitless possibilities of big data.< / h4 >
< attribution > (OpenClipArt.org)< / attribution >
< / section >
< / section >
< section >
< section >
< h2 > Why is this a fairy tale?< / h2 >
< / section >
< section >
< img src = "graphics/matt-icons_text-x-log-300px.png" height = "300" style = " vertical-align: middle;" / >
< span style = "font-size: 3em; vertical-align: middle;" > →< / span >
< img src = "graphics/database-server-800px.png" height = "300" style = " vertical-align: middle;" / >
< h4 > It's never this easy...< / h4 >
< / section >
< section >
< h2 > Loading Data< h2 >
< small >
< ul >
< li class = "fragment" > Validating and Fixing Outliers< / li >
< li class = "fragment" > Handling Missing Data< / li >
< li class = "fragment" > Matching Schemas< / li >
< li class = "fragment" > Fixing Schemas< / li >
< li class = "fragment" > Managing Stale Data< / li >
< li class = "fragment" > Deduplicating Records< / li >
< li class = "fragment" > ... and lots more< / li >
< / ul >
< / small >
< / section >
< / section >
< section >
< section >
< h2 > Data Cleaning is Hard!< / h2 >
< / section >
< section >
< h3 > State of the Art< / h3 >
< img src = "graphics/BI-Analyst.jpg" height = "400" / >
< attribution > (skilledup.com)< / attribution >
< p > Alice spends weeks cleaning her data before using it.< / p >
< / section >
< section >
< h3 > Newer State of the Art< / h3 >
< img src = "graphics/azure-data-lake.png" height = 500 / >
< attribution > (azure.microsoft.com)< / attribution >
< / section >
< section >
< img src = "graphics/data-lake-to-data-swamp.jpg" height = 500 / >
< attribution > (timoelliott.com)< / attribution >
< / section >
< / section >
< section >
< section >
< h2 > Making Cleaning Easier< / h2 >
< svg width = 500 height = 300 >
< polygon
points="60,50 60,60 40,50 60,40 60,50 440,50 440,40 460,50 440,60 440,50"
style="
stroke: black;
fill: black;
stroke-width: 2;
"
/>
< text x = 0 y = 30 style = "font-size: 0.75em" > Scalability< / text >
< text x = 370 y = 30 style = "font-size: 0.75em" > Reliability< / text >
< text class = "fragment" x = -220 y = 400 style = "font-size: 0.75em" transform = "rotate(-90 20,20)" > Expert Analysis< / text >
< text class = "fragment" x = -220 y = 250 style = "font-size: 0.75em" transform = "rotate(-90 20,20)" > Crowdsourcing< / text >
< text class = "fragment" x = -180 y = 100 style = "font-size: 0.75em" transform = "rotate(-90 20,20)" > Automation< / text >
< / svg >
< p class = "fragment" > Can we start with automation and work our way up?< / p >
< / section >
< section >
< h3 >
In the name of Codd,< br / > < span class = "fragment grow highlight-current-blue" > thou shalt not give the user a wrong answer.< / span >
< / h3 >
< h4 class = "fragment" >
... but what if we did?
< / h4 >
< h4 class = "fragment" >
What would it take for that to be ok?
< / h4 >
< / section >
< / section >
< section >
< ul >
< li > Automate educated guesses for fast cleaning< ul >
< li > < b class = "fragment highlight-blue" data-fragment-index = "5" > Lenses< / b > : A family of simple data-cleaning operators< / li >
< li class = "fragment" data-fragment-index = "1" > ... but what if the guesses are wrong?< / li >
< / ul > < / li >
< li class = "fragment" data-fragment-index = "2" > Annotate 'best guess' relations with the guesses< ul >
< li class = "fragment shrink fade-out" data-fragment-index = "5" > < b > Virtual C-Tables< / b > : A lineage model based on views, labeled nulls, and lazy evaluation.< / li >
< li class = "fragment" data-fragment-index = "3" > ... so now the user needs to interpret your guesses?< / li >
< / ul > < / li >
< li class = "fragment" data-fragment-index = "4" > Rank guesses by their impact on result uncertainty< ul >
< li > < b class = "fragment highlight-blue" data-fragment-index = "5" > CPI< / b > : A greedy heuristic for ranking sources of uncertainty.< / li >
< / ul > < / li >
< / ul >
< / section >
< section >
< section >
< h3 > Lenses< / h3 >
< p class = "fragment" > Here's a problem with my data. < span class = "fragment" > Fix it.< / span > < / p >
< ul >
< li class = "fragment" > What type is this column? (majority vote)< / li >
< li class = "fragment" > How do the columns of these relations line up? (pick your favorite schema matching paper)< / li >
< li class = "fragment" > How do I query heterogeneous JSON objects? (see above)< / li >
< li class = "fragment" > What should these missing values be? (learning-based interpolation)< / li >
< ul >
< / section >
<!--
## Points to get across ##
* Lenses have minimal configuration
* Lens = VG-RA Query + Model
-->
< section >
< h3 > VG-Relational Algebra< br / > < small > (Variable-Generating Relational Algebra)< / small > < / h3 >
< ul >
< li > Relational Algebra< / li >
< li > Labeled Nulls< / li >
< li > Lazy Evaluation< / li >
< / ul >
< / section >
< section >
< h3 > Labeled Nulls< / h3 >
< p > $Var(\ldots)$ constructs new variables< / p >
< ul >
< li class = "fragment" > $Var('X')$ constructs a new variable $X$< / li >
< li class = "fragment" > $Var('X', 1)$ constructs a new variable $X_{1}$< / li >
< li class = "fragment" > $Var('X', ROWID)$ evaluates $ROWID$ and then constructs a new variable $X_{ROWID}$< / li >
< / ul >
< / section >
< section >
< h3 > Lazy Evaluation< / h3 >
< p > Variables can't be evaluated until they are bound.< br / > So, we allow arbitrary expressions to be values.< / p >
< ul >
< li class = "fragment" > $X$ is a legitimate data value.< / li >
< li class = "fragment" > $X+1$ is a legitimate data value.< / li >
< li class = "fragment" > $1+1$ is a legitimate data value< span class = "fragment" > , but can be reduced to $2$.< / span > < / li >
< / ul >
< p class = "fragment" > A lazy value without variables is < b > deterministic< / b > < / p >
< / section >
< section >
< p > The $Var()$ operator can be inlined into SQL< / p >
< pre > < code >
SELECT A, VAR('X', B)+2 AS C FROM R;
< / code > < / pre >
< center > < div style = "width: 600px" class = "fragment" >
< table style = "float: left" >
< thead >
< tr > < th > A< / th > < th > B< / th > < / tr >
< / thead > < tbody >
< tr > < td > 1< / td > < td > 2< / th > < / tr >
< tr > < td > 3< / td > < td > 4< / th > < / tr >
< tr > < td > 5< / td > < td > 6< / th > < / tr >
< / tbody >
< / table >
< table style = "float: right" class = "fragment" >
< tr > < th > A< / th > < th > C< / th > < / tr >
< tr > < td > 1< / td > < td > $X_2+2$< / th > < / tr >
< tr > < td > 3< / td > < td > $X_4+2$< / th > < / tr >
< tr > < td > 5< / td > < td > $X_6+2$< / th > < / tr >
< / table >
< / div > < / center >
< div style = "clear: both;" > < / div >
< / section >
< section >
< p > Selects on $Var()$ need to be deferred too...< / p >
< pre > < code >
SELECT A FROM R WHERE VAR('X', B) > 2;
< / code > < / pre >
< center > < div style = "width: 600px" >
< table style = "float: left" >
< thead >
< tr > < th > A< / th > < th > B< / th > < / tr >
< / thead > < tbody >
< tr > < td > 1< / td > < td > 2< / th > < / tr >
< tr > < td > 3< / td > < td > 4< / th > < / tr >
< tr > < td > 5< / td > < td > 6< / th > < / tr >
< / tbody >
< / table >
< table style = "float: right" class = "fragment" >
< tr > < th > A< / th > < th > $\phi$< / th > < / tr >
< tr > < td > 1< / td > < td > $X_2>2$< / th > < / tr >
< tr > < td > 3< / td > < td > $X_4>2$< / th > < / tr >
< tr > < td > 5< / td > < td > $X_6>2$< / th > < / tr >
< / table >
< / div > < / center >
< div style = "clear: both;" > < / div >
< p class = "fragment" > When evaluating the table, rows where $\phi = \bot$ are dropped.< / p >
< / section >
< / section >
< section >
< section >
< h3 > C-Tables< / h3 >
< ul >
< li > Original Formulation < small > [Imielinski, Lipski 1981]< / small > < / li >
< li class = "fragment" > PC-Tables < small > [Green, Tannen 2006]< / small > < / li >
< li class = "fragment" > Systems< ul >
< li > Orchestra < small > [Green, Karvounarakis, Taylor, Biton, Ives, Tannen 2007]< / small > < / li >
< li > MayBMS < small > [Huang, Antova, Koch, Olteanu 2009]< / small > < / li >
< li > Pip < small > [Kennedy, Koch 2009]< / small >
< li > Sprout < small > [Fink, Hogue, Olteanu, Rath 2011]< / small > < / li >
< / ul > < / li >
< li class = "fragment" > Generalized PC-Tables < small > [Kennedy, Koch 2009]< / small > < / li >
< / ul >
< / section >
< section >
< h3 > Lenses< / h3 >
< ul >
< li class = "fragment" > A VG-RA Expression< / li >
< li class = "fragment" > A 'Model' < span class = "fragment" > that defines for each variable...< / span > < ul >
< li class = "fragment" > A sampling process< / li >
< li class = "fragment" > A best guess estimator< / li >
< li class = "fragment" > A human-readable description< / li >
< / ul > < / li >
< / ul >
< p class = "fragment" > < b > Lenses implement PC-Tables< / b > < / p >
< / section >
< section >
< pre > < code >
CREATE LENS PRODUCTS
AS SELECT * FROM PRODUCTS_RAW
USING DOMAIN_REPAIR(DEPARTMENT NOT NULL);
< / code > < / pre >
< ul >
< li > < code > AS< / code > clause defines source data.< / li >
< li > < code > USING< / code > clause requests repairs.< / li >
< / ul >
< / section >
< section >
< pre > < code >
CREATE LENS PRODUCTS
AS SELECT * FROM PRODUCTS_RAW
USING DOMAIN_REPAIR(DEPARTMENT NOT NULL);
< / code > < / pre >
< div >
< h4 > The Query< / h4 >
< pre > < code >
CREATE VIEW PRODUCTS
AS SELECT ID, NAME, ...,
CASE WHEN DEPARTMENT IS NOT NULL THEN DEPARTMENT
ELSE VAR('PRODUCTS.DEPARTMENT', ROWID)
END AS DEPARTMENT
FROM PRODUCTS_RAW;
< / code > < / pre >
< / div >
< small class = "fragment" >
< table >
< tr > < th > ID< / th > < th > Name< / th > < th > ...< / th > < th > Department< / th > < / tr >
< tr > < td > 123< / td > < td > Apple 6s, White< / td > < td > ...< / td > < td > Phone< / td > < / tr >
< tr > < td > 34234< / td > < td > Dell, Intel 4 core< / td > < td > ...< / td > < td > Computer< / td > < / tr >
< tr > < td > 34235< / td > < td > HP, AMD 2 core< / td > < td > ...< / td > < td class = "fragment" > $Prod.Dept_3$< / td > < / tr >
< tr > < td > ...< / td > < td > ...< / td > < td > ...< / td > < td > ...< / td > < / tr >
< / table >
< / small >
< / section >
< section >
< pre > < code >
CREATE LENS PRODUCTS
AS SELECT * FROM PRODUCTS_RAW
USING DOMAIN_REPAIR(DEPARTMENT NOT NULL);
< / code > < / pre >
< div >
< h4 > The Model< / h4 >
< pre > < code >
SELECT * FROM PRODUCTS_RAW;
< / code > < / pre >
< / div >
< div class = "fragment" >
< div style = "font-size: 1em; vertical-align: middle;" > ↓< / div >
< div >
< img src = "graphics/weka.png" / >
< / div >
< / div >
< div class = "fragment" >
< div style = "font-size: 1em; vertical-align: middle;" > ↓< / div >
< div > < p > An estimator for < small style = "vertical-align: baseline;" > $PRODUCTS.DEPARTMENT_{ROWID}$< / small > < p > < / div >
< / div >
< / section >
< / section >
< section >
< section >
< h2 > The User's View< / h2 >
< pre > < code >
SELECT NAME, DEPARTMENT FROM PRODUCTS;
< / code > < / pre >
< table class = "fragment" data-fragment-index = "1" >
< tr > < th > Name< / th > < th > Department< / th > < / tr >
< tr > < td > Apple 6s, White< / td > < td > Phone< / td > < / tr >
< tr > < td > Dell, Intel 4 core< / td > < td > Computer< / td > < / tr >
< tr > < td > HP, AMD 2 core< / td > < td class = "fragment highlight-red" data-fragment-index = "2" > Computer< / td > < / tr >
< tr > < td > ...< / td > < td > ...< / td > < / tr >
< / table >
< p class = "fragment" data-fragment-index = "2" > < b > Simple UI:< / b > Highlight values that are based on guesses.< / p >
< / section >
< section >
< pre > < code >
SELECT NAME, DEPARTMENT FROM PRODUCTS;
< / code > < / pre >
< small >
< table >
< tr > < th > Name< / th > < th > Department< / th > < / tr >
< tr > < td > Apple 6s, White< / td > < td > Phone< / td > < / tr >
< tr > < td > Dell, Intel 4 core< / td > < td > Computer< / td > < / tr >
< tr > < td > HP, AMD 2 core< / td > < td style = "color: red;" > Computer< / td > < / tr >
< tr > < td > ...< / td > < td > ...< / td > < / tr >
< / table >
< / small >
< svg xmlns = "http://www.w3.org/2000/svg" xmlns:xl = "http://www.w3.org/1999/xlink" version = "1.1" viewBox = "241 277 265 125" width = "265pt" height = "125pt" xmlns:dc = "http://purl.org/dc/elements/1.1/" class = "fragment" data-fragment-index = "1" >
< metadata > Produced by OmniGraffle 6.2.5 < dc:date > 2015-09-20 14:45:55 +0000< / dc:date > < / metadata >
< defs > < font-face font-family = "Helvetica Neue" font-size = "16" panose-1 = "2 0 8 3 0 0 0 9 0 4" units-per-em = "1000" underline-position = "-100" underline-thickness = "50" slope = "0" x-height = "517" cap-height = "714" ascent = "975.0061" descent = "-216.99524" font-weight = "bold" > < font-face-src > < font-face-name name = "HelveticaNeue-Bold" / > < / font-face-src > < / font-face > < font-face font-family = "Helvetica Neue" font-size = "16" panose-1 = "2 0 5 3 0 0 0 2 0 4" units-per-em = "1000" underline-position = "-100" underline-thickness = "50" slope = "0" x-height = "517" cap-height = "714" ascent = "951.99585" descent = "-212.99744" font-weight = "500" > < font-face-src > < font-face-name name = "HelveticaNeue" / > < / font-face-src > < / font-face > < / defs >
< g stroke = "none" stroke-opacity = "1" stroke-dasharray = "none" fill = "none" fill-opacity = "1" >
< title > Canvas 1< / title >
< g >
< title > Layer 1< / title >
< path d = "M 279 351 L 243 369 L 279 387 L 279 389 C 279 394.52285 283.47715 399 289 399 L 494 399 C 499.52285 399 504 394.52285 504 389 L 504 289 C 504 283.47715 499.52285 279 494 279 L 289 279 C 283.47715 279 279 283.47715 279 289 Z" fill = "white" / >
< path d = "M 279 351 L 243 369 L 279 387 L 279 389 C 279 394.52285 283.47715 399 289 399 L 494 399 C 499.52285 399 504 394.52285 504 389 L 504 289 C 504 283.47715 499.52285 279 494 279 L 289 279 C 283.47715 279 279 283.47715 279 289 Z" stroke = "black" stroke-linecap = "round" stroke-linejoin = "round" stroke-width = "1" / >
< text transform = "translate(293 293)" fill = "black" > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "bold" x = "0" y = "16" textLength = "16.896" class = "fragment" data-fragment-index = "2" > Pr< / tspan > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "bold" x = "16.608" y = "16" textLength = "69.28" class = "fragment" data-fragment-index = "2" > obability:< / tspan > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "500" x = "85.888" y = "16" textLength = "38.24" class = "fragment" data-fragment-index = "2" > 95%< / tspan > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "bold" x = "0" y = "53" textLength = "62.224" class = "fragment" data-fragment-index = "3" > Reason:< / tspan > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "500" x = "62.224" y = "53" textLength = "144.912" class = "fragment" data-fragment-index = "3" > Because I guessed < / tspan > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "500" x = "0" y = "71" textLength = "206.592" class = "fragment" data-fragment-index = "3" > ‘ Computer’ for ‘ Department’ < / tspan > < tspan font-family = "Helvetica Neue" font-size = "16" font-weight = "500" x = "0" y = "89" textLength = "196.16" class = "fragment" data-fragment-index = "3" > on Row ‘ 3’ of ‘ PRODUCTS’ < / tspan > < / text >
< / g >
< / g >
< / svg >
< p class = "fragment" data-fragment-index = "1" > Allow users to < code > EXPLAIN< / code > uncertain outputs< / p >
< p class = "fragment" data-fragment-index = "3" > Explanations include reasons given in English< / p >
< / section >
< section >
< div style = "padding: 30px;" >
< p > $PRODUCTS.DEPARTMENT_{3}$< / p >
< div style = "font-size: 2em" > ⬍< / div >
< p > "I guessed 'Computer' for 'Department' on Row '3'"< / p >
< / div >
< p class = "fragment" > < b >
(Generalized) C-Tables are a form of lineage.
< / b > < / p >
< / section >
< / section >
< section >
< section >
< h2 > Selection (Filtering)< / h2 >
< pre > < code >
SELECT NAME FROM PRODUCTS
WHERE DEPARTMENT='PHONE'
AND ( VENDOR='APPLE'
OR PLATFORM='ANDROID' )
< / code > < / pre >
< p class = "fragment" > Recall, row-level uncertainty is a boolean formula $\phi$.< / p >
< p class = "fragment" >
For this query, $\phi$ can be as complex as:
< small > $$DEPT_{ROWID}='P\ldots' \wedge \left( VEND_{ROWID}='Ap\ldots' \vee PLAT_{ROWID} = 'An\ldots' \right)$$< / small > < / p >
< p class = "fragment" > < b > Too many variables! Which is the most important?< / b > < / p >
< / section >
< section >
< h2 > What is important?< / h2 >
< p class = "fragment" > Data Cleaning< / p >
< h2 class = "fragment" > Which variables are important?< / h2 >
< p class = "fragment" > The ones that keep us from knowing everything< / p >
< / section >
< section >
< p > < small > $$D_{ROWID}='P' \wedge \left( V_{ROWID}='Ap' \vee PLAT_{ROWID} = 'An' \right)$$< / small > < / p >
< div style = "font-size: 2em" > ⬍< / div >
< p > $$A \wedge (B \vee C)$$< / p >
< / section >
< section >
< h3 > Naive Approach< / h3 >
< p > Consider a game between a database and an impartial oracle.< / p >
< ul >
< li > The DB picks a variable $v$ in $\phi$ and pays a cost $c_v$.< / li >
< li > The Oracle reveals the truth value of $v$.< / li >
< li > The DB updates $\phi$ accordingly and repeats until $\phi$ is deterministic.< / li >
< / ul >
< p class = "fragment" > < b > Naive Algorithm: < / b > Pick all variables!< / p >
< p class = "fragment" > < b > Less Naive Algorithm: < / b > Minimize $E\left[\sum c_v\right]$.< / p >
< / section >
< section >
< h2 > Exponential Time Bad!< / h2 >
< / section >
< / section >
< section >
< section >
< h3 > The Value of What We Don't Know< / h3 >
< p > $$\phi = A \wedge (B \vee C)$$< / p >
< ol >
< li class = "fragment" data-fragment-index = "1" > Generate Samples for $A$, $B$, $C$< / li >
< li class = "fragment" data-fragment-index = "2" > Estimate $p(\phi)$< / li >
< li class = "fragment" data-fragment-index = "3" > Compute $H[\phi] = -\log\left(p(\phi) \cdot (1-p(\phi))\right)$< / li >
< / ol >
< p class = "fragment" data-fragment-index = "4" > < b > Entropy is intuitive: < / b > < br / > $H = 1$ means we know nothing, < br / > $H = 0$ means we know everything.< / p >
< / section >
< section >
< h3 > Information Gain< / h3 >
< p > $$\mathcal I_{A \leftarrow \top} (\phi) = H\left[\phi\right] - H\left[\phi(A \leftarrow \top)\right]$$< / p >
< p > < b > Information gain of< / b > $v$: The reduction in entropy from knowing the truth value of a variable $v$.< / p >
< / section >
< section >
< h3 > Expected Information Gain< / h3 >
< p > $$\mathcal I_{A} (\phi) = \left(p(A)\cdot \mathcal I_{A\leftarrow \top}(\phi)\right) + \left(p(\neg A)\cdot \mathcal I_{A\leftarrow \bot}(\phi)\right)$$< / p >
< p > < b > Expected information gain of< / b > $v$: The probability-weighted average of the information gain for $v$ and $\neg v$.< / p >
< / section >
< section >
< h3 > The Cost of Perfect Information< / h3 >
< p > Combine Information Gain and Cost< / p >
< p > $$f(\mathcal I_{A}(\phi), c_A)$$< / p >
< p class = "fragment" > < b > For example: < / b > $EG2(\mathcal I_{A}(\phi), c_A) = \frac{2^{\mathcal I_{A}(\phi)} - 1}{c_A}$< / p >
< p class = "fragment" > < b > Greedy Algorithm: < / b > Minimize $f(\mathcal I_{A}(\phi), c_A)$ at each step< / p >
< / section >
< section >
< h3 > Experimental Data< / h3 >
< ul >
< li > Start with a large dataset.< / li >
< li > Delete random fields (~50%).< / li >
< / ul >
< / section >
< section >
< h3 > Experimental Queries< / h3 >
< p > Simulate an analyst trying to manually explore correlations.< / p >
< ul >
< li > Train a tree-classifier on the base data.< / li >
< li > Convert the decision tree to a query for all rows where the tree predicts a specific value.< / li >
< / ul >
< / section >
< section >
< h3 > Cost vs Entropy: Credit Data< / h3 >
< img src = "graphics/credit_entropy.png" height = 400 / >
< p > < small >
< b > EG2:< / b > Greedy Cost/Value Ordering< br / >
< b > NMETC:< / b > Naive Minimal Expected Total Cost< br / >
< b > Random:< / b > Completely Random Order
< / small > < / p >
< / section >
< section >
< h3 > Cost vs Entropy: Product Data< / h3 >
< img src = "graphics/product_entropy.png" height = 400 / >
< p > < small >
< b > EG2:< / b > Greedy Cost/Value Ordering< br / >
< b > NMETC:< / b > Naive Minimal Expected Total Cost< br / >
< b > Random:< / b > Completely Random Order
< / small > < / p >
< / section >
< / section >
< section >
< section >
< h2 > Demo (Mimir)< / h2 >
2017-08-31 17:18:47 -04:00
< p > < a href = "http://demo.odin.cse.buffalo.edu" > < img src = "https://odin.cse.buffalo.edu/wp-content/uploads/2015/08/Mimir_Screenshot.png" height = "400" / > < / a > < / p >
2016-02-11 09:37:51 -05:00
< / section >
< section >
< h2 > Intuitive Uncertainty< / h2 >
< p > < b > UB< / b > : Ying Yang, Niccolo Meneghetti, < br / > Arindam Nandi, Vinayak Karuppasamy< / p >
< p > < b > Oracle< / b > : Ronny Fehling, Zhen-Hua Liu, Dieter Gawlick< / p >
< h4 > Thanks to Oracle for multiple gifts that made this research possible< / h4 >
< / section >
< / section >
< / div > < / div >
< script src = "../reveal.js-3.1.0/lib/js/head.min.js" > < / script >
< script src = "../reveal.js-3.1.0/js/reveal.js" > < / script >
< script >
// Full list of configuration options available at:
// https://github.com/hakimel/../reveal.js#configuration
Reveal.initialize({
controls: false,
progress: true,
history: true,
center: true,
slideNumber: true,
transition: 'fade', // none/fade/slide/convex/concave/zoom
// Optional ../reveal.js plugins
dependencies: [
{ src: '../reveal.js-3.1.0/lib/js/classList.js', condition: function() { return !document.body.classList; } },
{ src: '../reveal.js-3.1.0/plugin/math/math.js',
condition: function() { return true; },
mathjax: '../reveal.js-3.1.0/js/MathJax.js'
},
{ src: '../reveal.js-3.1.0/plugin/markdown/marked.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: '../reveal.js-3.1.0/plugin/markdown/markdown.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: '../reveal.js-3.1.0/plugin/highlight/highlight.js', async: true, condition: function() { return !!document.querySelector( 'pre code' ); }, callback: function() { hljs.initHighlightingOnLoad(); } },
{ src: '../reveal.js-3.1.0/plugin/zoom-js/zoom.js', async: true },
{ src: '../reveal.js-3.1.0/plugin/notes/notes.js', async: true }
]
});
< / script >
< / body >
< / html >