Quantifying the intangible
from Data.
Sagar Joglekar, Ph.D. Candidate, Computer Science
@SagarJoglekar
1
What is the intangible?
● Drives human cognitive processes. (aesthetics)
● Subject to a deeper perception (Safety, beauty)
● Grounded in context, culture and community. (Support)
2
What “Intangible” do you “Quantify”?
● Online:
○ How do we share and seek help online ?
● Offline
○ Can we quantify how citizens perceive spaces around them?
○ Can we use this knowledge to create better cities ?
3
Our online lives
Part I
4
Online life is stream of gushing information
5
6
Internet
Anxiety
Depression
Support
Comfort 7
Why online support?
8
Data:
• Online support communities for Mental
distresses and respiratory distress
• More than 2 million posts over a period of 5
years
• More than 100,000 active users between the
two.
9
Key Question:
• Can we operationalize offline theories of Social support,
online?
• Can we quantify these processes?
10
Technical
challenges
Find established theories to
draw parallels from
Find effective
operationalization of theories
Drawling Causal links*
11
Technical
challenges
Find established theories to
draw parallels from
Find effective
operationalization of theories
Drawling Causal links*
User Graphs:
• Node => Users
• Edges => Who replies to
whom
• Edge weights => Topical
alignment between user
pairs
12
Topic Modelling
• Modeling corpus of text
as a mixture of topics
• Each topic is a mixture
of words.
• Represent text as a
vector of topics.
• Easier understanding of
salient phrases
Taking offline
support
theories,
online
Social Capital* -> Embeddedness
of help seekers (Topology of
community)
Critical consciousness *->
Topical relevance of responses
(LDA topical coherence of posts)
Community capacity* ->
Urgency, Mobilization, Involvement
(temporal aspects of activity)
*Community organizing and community building for health– Meredith Minkler et.a 1997
15
How embedded are the support seekers ?
Life sucks, I want to kill
myself
16
How embedded
are the support
seekers ?
Support seekers are Very Central
17
Topical Alignment:
Life sucks, I want to kill
myself
Don’t say that, you can
make your life better.
Life is all about creating
experiences.
Where:
18
How topically
aligned are the
conversations?
Responses are Topically aligned
19
Propensity to Answer: Z-score
Questions = 1/3 Answers = 2/3
20
Supportive agents develop over
time
Triads and cohesion.
21
Supportive agents add cohesion to the conversations
22
Key Takeaways
• Online support communities shows
characteristics of offline support communities
• Online supportive conversations show
characteristics of offline supportive
conversations
• People evolve and take over the mantle as they
stay on
23
Why is this important ?
Finding supportive agents in an online community
Learn language traits in online support. Fill gaps in the
care provider’s lexicon.
24
Our offline Lives
Part II
25
Spaces we occupy, influence our mental well being.
26
So what if A.I. could help us design them ?
Why Urban design?
Well Being City Planning
Profit!!
Beauty, Safety, Security
27
The challenge
A.I. decisions are often ``Black
Box’’ in nature
Decisions need to be explained
in the language of users
(Architects, designers, planners)
Framework needs to be quickly
adaptable and scalable
28
The Data
29
How do we do it?
http://facelift.datadissonance.org/
30
Transformation Example:
Original Template Transform Beautified
31
Do humans find results actually beautified?
Original - Transformed Mturk comparison
At least 3 votes per comparison.
Overall agreement 78%
32
The Explainability Metrics from
Urban planning:
•Walkability
•Openness
•Complexity
•Green Spaces
33
34
35
Expert
insights and
feedback
36
37
Still not
perfect
Generative models are very
tricky to train
No control over how much to
maximize to produce realistic
image
Now way as of now to
incorporate zoning laws
38
Why is this relevant ?
39
Milestones:
• Make human conditions and
human processes part of the
learning function
• Domain specific transparency
in A.I. Systems
40
ML
NLP
REC SYSCURATION
OTHERS
…………….
INTERNET
“Learning” From Humans
41
ML
NLP
REC SYSCURATION
OTHERS
INTERNET
“Learning” For Humans
42
43
Nishanth Sastry (Primary Advisor)
Miram Redi, Wikimedia research Daniele Quercia, Bell Labs UK Luca Aiello, Bell Labs UK Anna Simoni, Blizzard
Institute, QMUL
Pietro Panzarasa, QMUL
44
Thank You!!
https://sagarjoglekar.github.io

Center For telecomm research workshop 2018

Editor's Notes

  • #4 So through my research, I tried to answer such questions, driven by data and methods from inference science and complex networks. In the next part of the talk, I will give examples from each use case mentioned above Reduce Text here.
  • #5 We live in a constant state compulsive access to the flux of information and we are getting addicted to it.
  • #6 Our goal is to study how people interact Rise of Innovative creative methods Short videos are the new status update.
  • #7 What you see may not what you want to see in your state of mind.
  • #8 Reality -> anxiety, Depression But we want to look at the positive processes like support and comfort and draw parallels to real lives.
  • #9 And there is a reason for that.
  • #10 Compress the text.
  • #11 Operationalize
  • #18 Add formulas for weighted Centrality Label graphs with symbols from formulae
  • #19 Explain formula in slide
  • #20 Write a different slide with this Jaccard equation.
  • #21 Explain Zscore Explain symbols with some text
  • #23 Spell out what we are measuring
  • #24 Too much text. Online equal to offline in terms of of conversations and community stricture People mature in their roles
  • #27 Urban Mind,
  • #28 Creating places that look beautiful or safe, based on people’s perception. Unique combination of using Crowd perception to help design create real world places can be taken as inputs to automated design methods which can not only improve mental health but also would be economically beneficial
  • #31 Show slide selectively to unveil each block. Explain each block in relation to what we propose. Explain the formula terms
  • #32 Retrieval is done using some deep network We try overlaying the template on the original to give it an additive feel, but did not succeed without corrupting the original
  • #34 Not just transform but also help designers improve a place by pointing out changes based on popular urban design metrics (Explainable AI)
  • #37 We recruited around 30 experts from the fields of Urban planning, architecture and Urban informatics, to use the app and fulfill certain tasks that required them to get data drivien insights. And then we asked them how useful was such a toll for those tasks.
  • #38 Break the slides in to two. Walk through one and then show two side by side
  • #41 SMEs are deploying ML for every little task Make human behavior as part of the learning process by either discounting or accounting for it Finally pragmatic transparency, which means explain A.I. decisions in the language of the users
  • #42 Every interaction passes through layers of algorithm and a highly curated stream of content comes back. Egotistical inferences from the crowd causes us to consume content that is popular, but not what suits our state of mind. This has given rise to not only phenomena like Filter bubbles and echo chambers, which are huge problems for society in general. So the systems are taking a lot from the users, but not giving any
  • #43 Keep human psyche and condition at the center of the learning process. Have a symbiotic relationship.