People, Computers, and the Hot Mess of Real Data

Keynote, KDD 2016

Transcript

1. People, Computers, and the 
 Hot Mess of Real Data Joe Hellerstein
2. WHO AM I 2 ?
3. THE MISSING THIRD INGREDIENT: PEOPLE 3 Research imperative: 
 Dramatically simplify labor-intensive tasks … in the analytic lifecycle. 2010 Computing is free. Storage is free. Data is abundant. The remaining bottlenecks lie with people.
4. A SIDE PROJECT 4 dp = datapeople http://deepresearch.org
5. dp (c. 2012) 5 Jeff Heer
 Stanford Tapan Parikh Berkeley Maneesh Agrawala Berkeley Joe Hellerstein Berkeley Sean Diana Ravi Kandel MacLean Parikh Kuang Nicholas Wesley
 Chen Kong Willett
6. THE ANALYTIC LIFECYCLE 6 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION
7. THE ANALYTIC LIFECYCLE 7 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION KDD, SIGMOD, SOSP, NIPS, etc.
8. THE ANALYTIC LIFECYCLE 8 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION
9. THE ANALYTIC LIFECYCLE 9 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION Shreddr Wrangler MADlib d3 [Chen et al., DEV12] [Kandel, et al. CHI 11] [Hellerstein, et al. VLDB 12] [Bostock et al. Infovis 11] CommentSpace [Willett et al. CHI 11]
10. THE ANALYTIC LIFECYCLE 10 Shreddr Wrangler MADlib d3 CommentSpace ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION
11. THREE CHAPTERS ➔ Data Acquisition. (Shreddr —> Captricity) ➔ Data Wrangling (Potter’s Wheel —> Wrangler —> Trifacta) ➔ Data Context (Ground) 11
12. THE ANALYTIC LIFECYCLE 12 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION
13. Data in the First Mile
14. 14 Extracting value from data without waiting for infrastructure
15. 15 Shreddr
16. 16 Shreddr: Columnar Data Entry & Confirmation
17. Select the values are not: Michael 17 Shreddr: Columnar Data Entry & Confirmation
19. ANALYTICS ENABLEMENT
 Extracting Data from 1M+ Death Claims 19 CHALLENGE… No easy access to “cause of death” data 100’s of templates to identify, sort and capture UNLOCKED Improve fraud detection by leveraging patterns found in historical customer data
21. SOME LESSONS ➔(Problems from the field) × (Ideas from the lab) ➔Apply systems ideas to remove UX bottlenecks ➔Column compression ➔Batch processing & instruction locality ➔Filter pipelines ➔Crowdsourcing: first hints of Human/Machine collaboration ➔Humans as algorithmic agents ➔Challenge: optimize the human work
22. THE ANALYTIC LIFECYCLE ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION COLLABORATION ACQUISITION
23. DATA WRANGLING: A USER-CENTRIC TASK 23 Designing For Humans Not Designing for SciFi
24. Talk to Humans
25. WHERE DOES THE TIME GO IN ANALYTICS? PROCESSING ANALYTICS 80% of the work in any data project is preparing the data. Patil, Data Jujitsu, 2012. Kandel et al. “Enterprise Data Analysis and Visualization: An Interview Study”, IEEE VAST, 2012.
26. Interview study of 35 analysts: 25 companies Healthcare Retail, Marketing Social networking Media Finance, Insurance Various titles Data analyst Data scientist Software engineer Consultant Chief technical officer [Kandel et al., VAST12] KANDEL SURVEY 26
27. “I spend more than half of my time integrating, cleansing and transforming data without doing any actual analysis. Most of the time I’m lucky if I get to do any ‘analysis’ at all.” Friction “Most of the time once you transform the data ... the insights can be scarily obvious.” Lost potential
28. “It’s easy to just think you know what you are doing and not look at data at every intermediary step. An analysis has 30 different steps. It’s tempting to just do this then that and then this. You have no idea in which ways you are wrong and what data is wrong.” Interactivity and Visualization
30. A PROGRAMMING PROBLEM THE DATA TRANSFORMATION PROBLEM 30 DATA TRANSFORMATION Business System Data Machine Generated Data Log Data Data Visualization Fraud Detection Recommendations DATA SOURCE Complexity DATA PRODUCT Simplicity … …
31. TRANSFORMATION PROGRAMMING Languages: Python, Bash, Ruby, Perl… DSLs: DataStep, AJAX, Pandas, dplyr, Wrangle, Ibis… 31 Domain Specific Language (DSL) Data Output write code, compile, run
32. POTTER’S WHEEL (2001): ENTER THE VISUAL ➔ Visual DSL ➔ Immediate feedback ➔ Ongoing discrepancy detection ➔ Data lineage, redo/undo 32 [Raman & Hellerstein, VLDB11]
33. Lifting from DSL to Visual Language 33 Domain Specific Language (DSL) Data Output write code, compile, run Visualization and Interaction View Result visualize interact Lift Ground compile Problem: Remaining burden of specification for users.
34. My software doesn’t understand what I’m trying to do.
35. I don’t (yet) know what I’m trying to do.
36. HINTS OF INTELLIGENT INTERFACES Type-ahead uses context and data to predict search terms and preview results.
37. SEARCH QUERY AUTO-COMPLETE 37 Search Engine Query Textbox Query Response Suggestions pick type GUIDE DECIDE predict What about more complex input/output relations? The input and output domains are the same: text.
38. WRANGLER (2011): ADD INTELLIGENCE 38 [Kandel, et al. CHI 11] [Guo, et al. UIST11] ➔ Automatic inference of transforms ➔ Predictive preview of results ➔ Interactive history ➔ User Studies http://vis.stanford.edu/wrangler
39. TRADITIONAL DATA TRANSFORMATION 39 Visualization and Interaction Data Transformation Code User authors a draft transformation script User tests the script on a small amount of data User inspects output data to assess effects 1. 2. 3.
40. Trifacta. Confidential & Proprietary. PREDICTIVE INTERACTION 40 Visualization and Interaction Data Transformation Code User highlights visual features of the data Data previews allow user to choose, adjust and confirm Algorithms predict a ranked list of scalable transforms 1. 3. 2. GUIDE DECIDE
41. PREDICTIVE INTERACTION 41 Domain Specific Language (DSL) Visualization and Interaction Data Output write code, compile, run View Result visualize compile Response Preview pick interact predict GUIDE DECIDE codegen present Lift Ground [Heer, Hellerstein, Kandel, CIDR15]
42. Empowering businesses to innovate with data.
43. Wrangling Web Chat Log Data 43 Business Challenge: Understanding web chat interactions to personalize the customer experience Data Challenge: Only 0.01% of web chat logs analyzed due to complexity • Large volumes of unstructured, difficult to prep, web chat data being created • Only 200 chats manually extracted per month and analyzed for quality assurance • Valuable frontline time taken up by manual processing • Limited insight into what their customers are speaking to them about • In retail banking, web-based self- service has surpassed both in person and call center usage • At RBS, 250,000 customer chats per month launched for multiple banking needs • Analyzing web chat data can provide valuable information about customer needs and pain points Trifacta: Providing a self-service solution to wrangle 100% of logs • 100% of web chat logs now prepped and analyzed • Went from processing 200 logs to 250,000 logs…and now automated, not manual! • Have new insight into customer needs
44. © 2016 Royal Bank of Scotland Group. All rights Reserved The classification of this document is PUBLIC. “The dashboard is transforming the way I run my business. It is improving the customer-centric approach in our chats and it is showing in the output that we now see” Akshay Vats - Head of Web Chat Operation (India) Empowering RBS’s frontline staff
45. SOME LESSONS ➔Predictive Interaction: Guide and Decide ➔A UX model for AI-assisted, human-driven tasks ➔DSLs at the center ➔A formal “narrow waist” ➔Targetable to multiple runtimes ➔Provides a modest, factored search space for learning & prediction ➔Interactive Profiling ➔Continuous data vis feedback during transformation ➔Data profile qua data interface
46. 46 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION
47. 47 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY ACQUISITION CONTEXT
48. 48 ACQUISITION TRANSFORMATION ANALYSIS VISUALIZATION DECIDE/DEPLOY CONTEXT
49. A broader context for big data ground
50. ground A broader context for big data
51. WHAT CHANGED WITH BIG DATA? Shift in technology
 Data representations Shift in behavior
 Data-driven organizations
52. Shift in behavior
 Data-driven organizations
53. By 2017: 
 marketing spends more on tech than IT. Data escapes IT GARTNER GROUP
54. By 2017: 
 marketing spends more on tech than IT. Data escapes IT GARTNER GROUP By 2020: 
 90% of IT budget controlled outside of IT.
55. MANY USE CASES MANY CONSTITUENCIES MANY INCENTIVES MANY CONTEXTS
56. Shift in technology
 Data representations
57. What does it
 mean? It depends on
 the context. Raw data in the data lake
 Simplifies capture Encourages exploration
58. MANY SCRIPTS MANY MODELS MANY APPLICATIONS MANY CONTEXTS
59. It’s time to establish a bigger context for big data. Historical context
 Because
 things change Behavioral context
 Because behavior determines meaning Application context Because truth
 is subjective THE MEANING AND VALUE OF DATA DEPENDS ON CONTEXT
60. APPLICATION CONTEXT Metadata Models for interpreting
 the data for use • Data structures • Semantic structures • Statistical structures Theme: services must provide an unopinionated model of context
61. HISTORICAL CONTEXT Versions Web logs Code to extract user/ movie rentals Recommender for movie licensing Point in time
 A promising new
 movie is similar to older hot movies at time of release! Trends over time
 How does a movie
 with these features
 fare over time?
62. BEHAVIORAL CONTEXT Why Dora?! Lineage & Usage
63. 2 4 8 7 9 BEHAVIORAL CONTEXT Lineage & Usage Data Science Recommenders “You should compare with book sales from last year.” Curation Tips “Logistics staff checks weather data the 1st Monday of every month.” Proactive
 Impact Analysis “The Twitter analysis script changed. You should check the boss’ dashboard!”
64. 7 7 9 9 THE BIG CONTEXT A NEW WORLD NEEDS NEW SERVICES
65. ABOVEGROUND API TO APPLICATIONS UNDERGROUND API TO SERVICES CONTEXT MODEL COMMON GROUND Parsing &
 Featurization Catalog &
 Discovery Wrangling Analytics &
 Vis Reference
 Data Data
 Quality Reproducibility Model
 Serving Scavenging
 and Ingestion Search &
 Query Scheduling &
 Workflow Versioned
 Storage ID & Auth
66. COMMON GROUND Version-Model-Lineage (VML) Graphs Model Graphs Version Graphs Usage Graphs: Lineage
67. ABOVEGROUND API TO APPLICATIONS UNDERGROUND API TO SERVICES CONTEXT MODEL COMMON GROUND Parsing &
 Featurization Catalog &
 Discovery Wrangling Analytics &
 Vis Reference
 Data Data
 Quality Reproducibility Model
 Serving Scavenging
 and Ingestion Search &
 Query Scheduling &
 Workflow Versioned
 Storage ID & Auth ABOVEGROUND API TO APPLICATIONS UNDERGROUND API TO SERVICES RESEARCH OPPORTUNITIES ACROSS THE STACK
68. IN SUM: PEOPLE + DATA + COMPUTATION ➔Dealing with Data: involves much more than algorithms ➔Human Component: a huge opportunity for tech innovation ➔Context is Key: for grounding analysis 68
69. @joe_hellerstein hellerstein@berkeley.edu