Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
SOCIAL	MEDIA	&	POLICING:	Computational	
Approaches	to
Enhancing	Collaborative	Action
between	Residents	and	Law	Enforcement...
Who	am	I?
– Ph.D.	student
– Senior	Research	Scientist	@Philips	Research,	India	
– TCS	Research	Scholar
– Done	work	in	comp...
India	is	Biggest	Police	Department
3
238	Police	Officers	per	100,000
129	Police	Officers	per	100,000
327	Police	Officers	p...
4
Most	Overworked	– Need	Help!
Collaborating	with	Residents
One	– way	communication	
5
Two	– way	communication
Asynchronous,	Remote	and	
Public	platform	...
How	about	Interacting	with	Police	on	OSM?
6
– How	many	of	you	are	on	Facebook	/	Twitter?
– How	many	of	you	know	about	soci...
New	Media	to	Stay	Connected
7
Three	Dimensions	for	Successful	Collaboration	
High
More	accurate	analytical and	
modeling	tools
Low
High
High
More	people...
Challenges:	Successful	Collaboration	
– Identifying	how	social	media	can	support	day-to-day	
interaction	between	police	an...
Challenges:	Successful	Collaboration	
– Identifying	how	social	media	can	support	day-to-
day	interaction	between	police	an...
Challenges:	Successful	Collaboration	
– Identifying	how	social	media	can	support	day-to-
day	interaction	between	police	an...
Challenges:	Successful	Collaboration	
– Identifying	how	social	media	can	support	day-to-day	
interaction	between	police	an...
Core	Thesis	Question	
How	can	social	media	platforms	be	utilized	to	support,	analyze,	
and	enhance day-to-day	collaborativ...
Contributions
– Identify	need	for	social	media	support in	collaborative	policing
– Quantify	and	mine	unstructured	data	to	...
Contributions
– Identify	need	for	social	media	support in	collaborative	policing
– Quantify	and	mine	unstructured	data	to	...
Research	
Problem
Initial	
Coding
Advanced	
Memo
Theoretical	
Sampling	new	
data
Integrating	for	
dimensions
Need	for	Supp...
Why Which For	Whom Challenges
17
EDUCATED	YOUNG
ANYONE
Need	for	Support	:	Requirement	Elicitation
Police	Officers Resident	Communities
Collaborative	Platform
Collaborative	Ecosystem	and	Actors
Interaction	Layer
Meaningfu...
Contributions
– Identify	need	for	social	media	support in	collaborative	
policing
– Quantify	and	mine	unstructured	data	to...
Quantifying	Interaction
– Exploring	the	feasibility	of	social	media	in	quantifying	attributes	of	
communication
– Identify...
Mixed	Method	Approach:	Data	Collection
21
85	Public	and	official	Police	Department
Average	age	3	years	(from	2010	– April	...
Mixed	Method	Approach:	Data	Collection
22
85	Public	and	official	Police	
Department
Average	age	3	years	(from	2010	–
April...
Quantifying	Interaction	for	Meaningful	Information
– Content	Cluster	Identification
– Nature	of	content	and	topics	
– Emot...
Mixed	Method	Approach:	Methods
24
Topics
• Unigram (N) Gram	Analysis
• K-means	Clusters with	K-means++	seeding
Emotional	
...
Mixed	Method	Approach:	Methods
25
Topics
• Unigram (N) Gram	Analysis
• K-means	Clusters with	K-means++	seeding
Emotional	
...
26
Unigram Freq. Unigram Freq.
rules 0.015 safety 0.012
safety 0.014 following 0.011
violations 0.014 notice 0.010
challan...
K-Means++	Seeding:	Clusters	of	
Topics
– Police	initiated	discussions	are	more	focused	than	citizen	initiated.
27
Awarenes...
Quantifying	Interaction
– Topic	Identification
– Nature	of	content	and	topics	
– Emotional	Exchange Quantification
– Natur...
– Negative	sentiment	higher	in	resident	initiated	threads	than	police
29
CP&C CC
Avg Std.	dev Avg Std.	dev
NA 0.021 0.03 0...
– Discussion	threads	involving	just	the	citizens	are	highly	self-attention	
focused
30
Likely	citizens	mostly	express	thei...
Self-focused
31
My	Vehicle	KA-02-HW-3183	white	color	Honda	Dio
was	stolen	from	Kadamba Hotel(Near	Modi
Hospital),	RajajiNa...
Accountability:	Meaningful	Information
32
Word	Tree	visualizations	of	posts	in	which	residents	
questioned	 police	using	t...
Quantifying	Interaction
– Topic	Identification
– Nature	of	content	and	topics	
– Emotional	Exchange Quantification
– Natur...
Engagement	Quantification
– Content	Generation
34
Police	+	Citizens 55,028 1,79,176 17,124 12,630
Citizens	Only 54,982 1,7...
Engagement	(Response)	Type
35
Ignored	(#83)
Acknowledged (21.3%)
Reply
Follow	Up	(10%)
Dear	X,	We	will	take	all	possible	l...
Lessons
36
Responsiveness
Accountability
Quantify
And	
Extract
Need	for	
Enhancing	Police	
Responsiveness
Contributions
– Identify	need	for	social	media	support in	collaborative	policing
– Quantify	and	mine	unstructured	data	to	...
Serviceable	Requests
“A message	that	solicits	a	response	in	
a	form	of	an	action	or	information	
from	the	police”
38Tucson...
Serviceable	v/s	Non-Serviceable	Posts
39
Need	more	
Information
Forward
Give	Solution
Cops	driving	wrong	side	[of	road]	ne...
Research	Questions:	Serviceability
– RQ	1:	What	attributes	differentiates
– Serviceable	posts	from	non-serviceable	request...
Dataset
85	Public	and	official	Police	Department
22,213	wall	posts
1000	Posts	annotated	by	Police
41
Post	Type #Posts Like...
Attributes
Emotion
(Alchemy	&	
LIWC)
States: Anger,	disgust,	fear,	joy,	sadness
Valence:	Positive,	negative,	Anxiety	
Cogn...
Attributes
– Top	Topics	Serviceable	and	Non-Serviceable	Posts	using	LDA
43
LDA	topic	 Vocabulary
Traffic	congestion	 Traff...
Attributes
– Non-negative	Matrix	Factorization	(NMF)	for	detecting	closely	
connected	topics	in	sub-types
44
NMF	topic	 Vo...
RQ1:	Attributes	Defining	Serviceability
– Serviceable	requests	show	significantly	higher	value	of	negative	
emotional	stat...
RQ1:	Attributes	Defining	Serviceability
– Serviceable	requests	show	significantly	higher	use	of	1st	person	
singular	prono...
RQ1:	Attributes	Defining	Serviceability
– Serviceable	posts	showed	higher	Objectivity	
47
Serv. Non-Serv. Frwd Give Thanks...
Research	Questions:	Serviceability
– RQ	1:	What	attributes	differentiates
– Serviceable	posts	from	non-serviceable	request...
RQ2:	Police	Response	Time
– Survival	Time
– Time	until	the	event	of	interest	occurs	
– Censoring	Event	
– Posts	which	did	...
RQ2:	Police	Response	Time
50
Mean	Est. Sd.	Error Median	
Est.
Sd.	Error
Frwd 1062.52 82.22 21.33 2.05
Give 1064.37 138.08 ...
Research	Questions:	Serviceability
– RQ	1:	What	attributes	differentiates
– Serviceable	posts	from	non-serviceable	request...
Formulation
52
RQ3:	Automatic	Classification	Performance:	Cost	
weights
– Ten-fold	Cross	Validation	Performance	of	different	algorithms	t...
– +Model	1
– Explains	15.6%	of	the	variance	
– Reduces	deviance	significantly	to	1,127.58	(178.14	less).	
– Better	predict...
– +Model	3
– Explains	26.3%	of	the	variance	
– Reduces	deviance	significantly	to	984.92.	
– Not	helpful:	Tenses	(present	a...
Serviceability	Plugin	
56
Contributions
– Identify	need	for	social	media	support in	collaborative	policing
– Quantify	and	mine	unstructured	data	to	...
Challenges	in	similar	images	retrieval
58
Image	
processing
1.	Scaling	
2.	Cropping
3.	Stitching	
4.	Multiple
59
Data	Collection
Sample	images Event		 Total	
images
Similar	
images
Dissimilar	
images
Charlie	Hebdo 568 118 450
Kulkar...
Image	features	for	similarity
1. Hand-Crafted	Features
a. 3D-colour	histogram
b. Daisy	features
c. ORB	(Oriented	FAST	Rota...
Original	
Image	
Modified	
Image
Improved	
ORB	
(%	accuracy)
CNN
(%	accuracy)
Modification
93.7 98.3 Scaled,	stitched	
ima...
PicHunt
62
Contributions
– Identify	need	for	social	media	support in	collaborative	policing
– Quantify	and	mine	unstructured	data	to	...
Why	it	matters?
– Propose	a	Data-Driven	Technique	to	complement	Overworked	departments
– Detecting	posts	that	should	elici...
Technological	Advancement
– Designing	early	warning	systems	that	indicate:
– Need	for	emotional	&	social	support	needs	to	...
Limitations
– Cultural	Limitation
– Strong	history	of	community	policing	may	be	helpful
– Restricted	Modality
– Only	text	...
67
Acknowledgement
– Google	for	travel	support
– Tata	Consultancy	Services	for	funding	the	thesis	work
– All	participants	and...
Acknowledgement
– Dr.	Aaditeshwar Seth
– Dr.	Carlos	Castillo
– Dr.	Maura	Conway
– Dr.	Ponnurangam Kumaraguru
69
Publications
– Peer	Reviewed	Conferences
– Sachdeva,	N.	and	Kumaraguru,	P.	Online	Social	Media	- New	face	of	policing?	A	S...
Publications
– Peer-reviewed	Conference	Papers	
– Lamba,	 H.,	Bharadhwaj,	V.,	Vachher,	M.,	Agarwal,	D.,	Arora,	M.,	Sachdev...
Thank	you!	
niharikas@iiitd.ac.in
http://precog.iiitd.edu.in/
Upcoming SlideShare
Loading in …5
×

Social Media and Policing: Computational Approaches to Enhancing Collaborative Action between Residents and Law Enforcement

590 views

Published on

Social Media and Policing: Computational Approaches to Enhancing Collaborative Action between Residents and Law Enforcement

Law and order concerns are one of the major disquiets of urban societies in day-to-day life. Var- ious crime prevention theories show the importance of collaboration between residents and police for maintaining law and order and addressing concerns. Collaborative action across public orga- nizations such as police shows different challenges like enabling collective action, problem-solving, accountability, and responsiveness of the organizational actors towards residents. To enable col- lective action and problem solving with the help of residents, modern police departments explore innovative mechanisms to overcome the barrier of reachability and communication. Using these mechanisms, residents can convey their concerns and enquire/provide information useful for police contributing towards the collaborative process. With growing reach of web 2.0, social media has emerged as an effective platform to enable collaboration between police and resident. Social media use for communication between police and resident introduces various challenges for organizations.

Published in: Engineering
  • Be the first to comment

Social Media and Policing: Computational Approaches to Enhancing Collaborative Action between Residents and Law Enforcement

  1. 1. SOCIAL MEDIA & POLICING: Computational Approaches to Enhancing Collaborative Action between Residents and Law Enforcement Niharika Sachdeva PhD Thesis Defense TCS Research Scholar niharikas@iiitd.ac.in
  2. 2. Who am I? – Ph.D. student – Senior Research Scientist @Philips Research, India – TCS Research Scholar – Done work in computer mediated communication and usable security (HCI) – Research interests – Collaboration and communication – Machine Learning – Human computer interaction – Usable security and privacy 2
  3. 3. India is Biggest Police Department 3 238 Police Officers per 100,000 129 Police Officers per 100,000 327 Police Officers per 100,000 Which is India? South Africa? USA?
  4. 4. 4 Most Overworked – Need Help!
  5. 5. Collaborating with Residents One – way communication 5 Two – way communication Asynchronous, Remote and Public platform for Interaction Need for Improved Collective Action and Accountability
  6. 6. How about Interacting with Police on OSM? 6 – How many of you are on Facebook / Twitter? – How many of you know about social media police pages / accounts or use them to interact with police?
  7. 7. New Media to Stay Connected 7
  8. 8. Three Dimensions for Successful Collaboration High More accurate analytical and modeling tools Low High High More people involved More data available Police Resident Charalabidis, Yannis, and Sotirios Koussouris, eds. Empowering open and collaborative governance: Technologies and methods for online citizen engagement in public policy making. Springer Science & Business Media, 2012. 8
  9. 9. Challenges: Successful Collaboration – Identifying how social media can support day-to-day interaction between police and residents – Analyzing and Extracting meaningful and actionable information from enormous data – Unstructured and unconstrained – Inferring actionable information – Quantifying behavior (emotions and linguistic attributes) – Maintaining responsiveness to residents – Promptness and timely action by police on social media – Engaging with people 9 High More accurate analytical and modeling methods Low High High More people involved More data available
  10. 10. Challenges: Successful Collaboration – Identifying how social media can support day-to- day interaction between police and residents – Analyzing and Extracting meaningful and actionable information from enormous data – Unstructured and unconstrained – Inferring actionable information – Quantifying behavior (emotions and linguistic attributes) – Maintaining responsiveness to residents – Promptness and timely action by police on social media – Engaging with people 10 High Low High More people involved More data available More accurate analytical and modeling methods High
  11. 11. Challenges: Successful Collaboration – Identifying how social media can support day-to- day interaction between police and residents – Analyzing and Extracting meaningful and actionable information from enormous data – Unstructured and unconstrained – Inferring actionable information – Quantifying behavior (emotions and linguistic attributes) – Maintaining responsiveness to residents – Promptness and timely action by police on social media – Keep engaging with people 11 High Low High More people involved More data available More accurate analytical and modeling methods High
  12. 12. Challenges: Successful Collaboration – Identifying how social media can support day-to-day interaction between police and residents – Analyzing and Extracting meaningful and actionable information from enormous data – Unstructured and unconstrained – Inferring actionable information – Quantifying behavior (emotions and linguistic attributes) – Maintaining responsiveness to residents – Promptness and timely action by police on social media – Engaging with people 12 High More accurate analytical and modeling methods Low High High More people involved More data available
  13. 13. Core Thesis Question How can social media platforms be utilized to support, analyze, and enhance day-to-day collaborative interaction between police and residents using computational methods? 13
  14. 14. Contributions – Identify need for social media support in collaborative policing – Quantify and mine unstructured data to analyzecommunication attributes and actionable information – Propose a method to enhancecollaboration – Police responsiveness usingrequest–response detection framework and police response quantification to residents – Images retrieval architecture for improved collective action using user generated content 14
  15. 15. Contributions – Identify need for social media support in collaborative policing – Quantify and mine unstructured data to analyzecommunication attributes and actionable information – Propose a method to enhance – Police responsiveness usingrequest–response detection framework and police response quantification to residents – Images retrieval architecture for improved collective action using user generated content 15
  16. 16. Research Problem Initial Coding Advanced Memo Theoretical Sampling new data Integrating for dimensions Need for Support : Requirement Elicitation – What opportunitiessocial media offers for supporting collaboration? – What challenges police and residents can face while adopting social media for collaboration? 16 • 17 Interviews • 200 Surveys • 20 Interviews • 402 Surveys Multi-stakeholder & Mixed Method Limited Grounded Theory Approach
  17. 17. Why Which For Whom Challenges 17 EDUCATED YOUNG ANYONE Need for Support : Requirement Elicitation
  18. 18. Police Officers Resident Communities Collaborative Platform Collaborative Ecosystem and Actors Interaction Layer Meaningful Information Acknowledgement / Response System Verification and Credibility Assessment Lessons 18
  19. 19. Contributions – Identify need for social media support in collaborative policing – Quantify and mine unstructured data to analyze communication attributes and actionable information – Propose a method to enhance – Police responsiveness usingrequest–response detection framework and police response quantification to residents – Images retrieval architecture for improved collective action using user generated content 19 High Low High More people involved More data available More accurate analytical and modeling methods High
  20. 20. Quantifying Interaction – Exploring the feasibility of social media in quantifying attributes of communication – Identifying behavioral attributes like affective expression, engagement and social and cognitive response processes 20 Resident to Resident Resident to Police Police to Resident Police to Police
  21. 21. Mixed Method Approach: Data Collection 21 85 Public and official Police Department Average age 3 years (from 2010 – April 2015) 47,474 wall posts and 85,408 status updates DT w/ ≥ 1 Comment P&C C Total DT 85,408 47,474 46,845 24,984 5,519 17,196 41,326 7,788 PP&C RP&C PC CC Quantitative Data Qualitative Analysis
  22. 22. Mixed Method Approach: Data Collection 22 85 Public and official Police Department Average age 3 years (from 2010 – April 2015) 47,474 wall posts and 85,408 status updates DT w/ ≥ 1 Comment P&C C Total DT 85,408 47,474 46,845 24,984 5,519 17,196 41,326 7,788 PP&C CP&C PC CC Quantitative Data 1600 comments on 255 posts Posts & Comments Collected public posts, 21 July - 21 Aug 2014 Qualitative Analysis
  23. 23. Quantifying Interaction for Meaningful Information – Content Cluster Identification – Nature of content and topics – Emotional Exchange Quantification – Nature of emotions and affective expression – Cognitive and Social Orientation Quantification – Type of linguistic attributes that characterize cognitive and social orientation – Engagement (Response) Quantification – Quantity and nature of engagement 23
  24. 24. Mixed Method Approach: Methods 24 Topics • Unigram (N) Gram Analysis • K-means Clusters with K-means++ seeding Emotional • Valence • Arousal Social and cognitive • Interpersonal Focus • Social Orientation • Cognition Engagement • No. of police and resident who comment in DTs • Distinct citizens who comment in DTs • Shannon’s Wiener Diversity index • Average no. of likes and comments LIWC and Anew Dictionary LIWC Dictionary Quantitative Data Thematic Inductive Analysis
  25. 25. Mixed Method Approach: Methods 25 Topics • Unigram (N) Gram Analysis • K-means Clusters with K-means++ seeding Emotional • Valence • Arousal Social and cognitive • Interpersonal Focus • Social Orientation • Cognition Engagement • No. of police and resident who comment in DTs • Distinct citizens who comment in DTs • Shannon’s Wiener Diversity index • Average no. of likes and comments LIWC and Anew Dictionary LIWC Dictionary Quantitative Data Thematic Inductive Analysis Validate and Characterize Type of sub-topics in Residents Posts Direct / Indirect Concerns + Style of communication Type of Engagement
  26. 26. 26 Unigram Freq. Unigram Freq. rules 0.015 safety 0.012 safety 0.014 following 0.011 violations 0.014 notice 0.010 challans 0.011 prosecuted 0.009 please 0.011 movement 0.008 citizens 0.01 complaint 0.008 Focus on advisories, the status of different cases being investigated (Mann Whitney U test, p < .05, z = −3.57) Most posts tend to request police to take action on their complaints Unigram Freq. Unigram Freq. please 0.026 people 0.022 take 0.021 please 0.02 action 0.019 one 0.019 people 0.019 take 0.016 one 0.019 action 0.015 time 0.017 time 0.015 Higher Reference to “people” Quantifying Interaction for Meaningful Information
  27. 27. K-Means++ Seeding: Clusters of Topics – Police initiated discussions are more focused than citizen initiated. 27 Awareness drive / safety campaigns Prosecuted / action taken reports Advisories on situations Newspaper articles Citizen tips and complaints Neighbourhood problems Missing people Appreciation Quantifying Interaction for Meaningful Information
  28. 28. Quantifying Interaction – Topic Identification – Nature of content and topics – Emotional Exchange Quantification – Nature of emotions and affective expression – Cognitive and Social Orientation Quantification – Type of linguistic attributes that characterize cognitive and social orientation – Engagement Quantification – Quantity and nature of engagement 28
  29. 29. – Negative sentiment higher in resident initiated threads than police 29 CP&C CC Avg Std. dev Avg Std. dev NA 0.021 0.03 0.018 0.04 Anx 0.001 0.01 0.003 0.02 Anger 0.006 0.02 0.005 0.02 Arousal 4.4 1.74 3.9 2.16 16.67% higher in CP&C 12.82% higher in CP&C Higher arousal and negative affect to be markers of sensitisationbecause of crime! Cp&c Cc Avg Std. dev Avg Std. dev NA 0.021 0.03 0.018 0.04 Anx 0.001 0.01 0.003 0.02 Anger 0.006 0.02 0.005 0.02 Arousal 4.4 1.74 3.9 2.16 200% higher in Cc (Mann-Whitney U, p < .01) (Mann-Whitney U p < .01) Meaningful Information Quantification: Emotions
  30. 30. – Discussion threads involving just the citizens are highly self-attention focused 30 Likely citizens mostly express their own concerns that they face with others CP&C CC ppron 0.062 0.059 0.045 0.056 i 0.008 0.017 0.014 0.033 shehe 0.002 0.01 0.003 0.003 they 0.005 0.013 0.008 0.008 75% More I have lived in the UK and all the time I have never heard anyone honking. …. if I see anyone who don't comply ? (U Test p < .01, z = −16.02) Meaningful Information Quantification: Social and Cognitive Orient.
  31. 31. Self-focused 31 My Vehicle KA-02-HW-3183 white color Honda Dio was stolen from Kadamba Hotel(Near Modi Hospital), RajajiNagar on Friday(25th July) evening between 6:30-7:45PM. Please help in tracing my vehicle. Dear BCP, though I stay at JP Nagar, but being part of KSFC Layout RWA (BanaswadiPolice station) , I got to know that there are frequent problem at KSFC Layout near BBMP Hall . . . ….
  32. 32. Accountability: Meaningful Information 32 Word Tree visualizations of posts in which residents questioned police using the word why.
  33. 33. Quantifying Interaction – Topic Identification – Nature of content and topics – Emotional Exchange Quantification – Nature of emotions and affective expression – Cognitive and Social Orientation Quantification – Type of linguistic attributes that characterize cognitive and social orientation – Engagement Quantification – Quantity and nature of engagement 33
  34. 34. Engagement Quantification – Content Generation 34 Police + Citizens 55,028 1,79,176 17,124 12,630 Citizens Only 54,982 1,79,176 17,081 12,630 Entropy 4.39 4.96 3.23 3.6 Police Resident 26% lower 10.28% lower Lower entropy: large number of comments are posted by a small number of citizens and police
  35. 35. Engagement (Response) Type 35 Ignored (#83) Acknowledged (21.3%) Reply Follow Up (10%) Dear X, We will take all possible legal measures in this regard. Thank you. Dear X, Please provide the police station details. Thank you. [Received no reply] Dear X, This post has been forwarded to appropriate Police Station…. Dear X, Please lodge a complaint at your nearest police station with the details ….. 44.3% 22% 172 posts
  36. 36. Lessons 36 Responsiveness Accountability Quantify And Extract Need for Enhancing Police Responsiveness
  37. 37. Contributions – Identify need for social media support in collaborative policing – Quantify and mine unstructured data to analyzecommunication attributes and actionable information – Propose a method to enhance – Police responsiveness usingrequest–response detection framework and police response quantification to residents – Images retrieval architecture for improved collective action using user generated content 37
  38. 38. Serviceable Requests “A message that solicits a response in a form of an action or information from the police” 38Tucson Police. 2016. Calls for Service. https://www.tucsonaz.gov/police/terms. (May 2016). Low High High High More people involved More data available
  39. 39. Serviceable v/s Non-Serviceable Posts 39 Need more Information Forward Give Solution Cops driving wrong side [of road] near XXX hotel .. what action will be taken against them Date : 4/11/2015 (Wednesday), Time : 10:17 pm, Number : [withheld], Location : [withheld], Violations : Crossing line by way too much obstructing the vehicles which were coming from [withheld] entrance later he jumped the signal . Admin !! Can U Explain to me rules and regulations for transferring vehicle from Chennai to Bangalore? Ignored XXX shared NowThis Future's video. 21 February at 10:07 · BENGALURU CITY POLICE Interesting piece of handgun bullet-proof shield in development. Acknowledge Chennai City Traffic Police: a humble salute from a fellow Chennaiite for the commendable job in such rains!!
  40. 40. Research Questions: Serviceability – RQ 1: What attributes differentiates – Serviceable posts from non-serviceable requests and – Sub-types of serviceable requests w.r.t content characteristics such as linguistic and emotional attributes? – RQ 2: How does police response time vary between serviceable and non-serviceable posts made on social media? – RQ 3: Can machine learning techniques be used to automatically identify serviceable requests – Can we further classify them into different sub-types using post characteristics (content and metadata)? 40
  41. 41. Dataset 85 Public and official Police Department 22,213 wall posts 1000 Posts annotated by Police 41 Post Type #Posts Likes Comments Serviceable Posts Forward 286 1383 661 Give Solution 88 183 121 Thanks 72 1288 63 Need More Info. 104 1245 258 Total 550 4099 1103 Non-Serviceable Total 113 316 32 0.77 agreement using Fleiss Agreement
  42. 42. Attributes Emotion (Alchemy & LIWC) States: Anger, disgust, fear, joy, sadness Valence: Positive, negative, Anxiety Cognitive (LIWC) Cognitive Mechanism:Tentativeness and Discrepancy Inter- Personal (LIWC) 1st person singular & plural, 2nd person, 3rd person singular & plural, and impersonal pronouns Linguistic (LIWC) Objectivity, Tenses, Lexical Density & Parts- Of-Speech. Question Asked (Heuristics) who, how, why, what, where, whom and containing a “?” Entities (Alchemy) people, companies, organizations, cities, geographic features, facility, date andtime 42
  43. 43. Attributes – Top Topics Serviceable and Non-Serviceable Posts using LDA 43 LDA topic Vocabulary Traffic congestion Traffic, road, signal, bus, people,turn, jam Shared photos websites com, www, facebook, https, videos, traffic, http, type, old, photos, job Appreciation signal, great, good, taking, act, action Question posed asked, rules, vehicle, sir, said, car, know, what Places Telangana, state, hyderabad, city, nagar, Fines issued Challan [fine charged], violation, documents Cyber crime Police, city, cyber, crime nampally, complaint, better, safe
  44. 44. Attributes – Non-negative Matrix Factorization (NMF) for detecting closely connected topics in sub-types 44 NMF topic Vocabulary Police incorrect decision Police, asked, said, constable, taken, public wrong, driving, pay, vehicle, come, way Awareness Don’t, mobile, rules,need, people, let, share, helmet, circle, Dangerous driving complains wrong, dangerous, action, driving, turn, going junction Fines issued Vehicle, challan [fine charged], number, violation fine, documents, driving, guys, stopped, pay Parking issues Parking, people, bus, stop, parked, time, action Used Frobenius Norm
  45. 45. RQ1: Attributes Defining Serviceability – Serviceable requests show significantly higher value of negative emotional states 45 Serviceable Non-Serviceable Avg Std. dev Avg Std. dev Man. Anger 0.15 0.13 0.13 0.17 -3.43** Disgust 0.34 0.25 0.23 0.27 -3.88** Fear 0.24 0.21 0.15 0.18 -6.09** Sad 0.11 0.10 0.10 0.14 -5.45** +15.38% +60% Presumably, emotional states are experienced due to distress caused because of encounters with law and order situation.
  46. 46. RQ1: Attributes Defining Serviceability – Serviceable requests show significantly higher use of 1st person singular pronouns – highly self-attention 46 Serv. Non-Serv. Frwd Give Thanks Need 1st person Singular ** Avg. 1.68 1.54 1.61 2.56 0.70 1.80 Sd. 2.96 9.50 2.45 3.54 2.36 3.77 I am just worried if Hyderabad Traffic Police [HTP] makes things worse like always
  47. 47. RQ1: Attributes Defining Serviceability – Serviceable posts showed higher Objectivity 47 Serv. Non-Serv. Frwd Give Thanks Need Objectivity** Avg 2.86 2.04 +40% 3.47 2.64 2.07 1.9 Sd 2.84 2.63 3.16 2.8 2.6 1.29 Serv. Non-Serv. Frwd Give Thanks Need Past Tense Avg 1.75 0.81 1.88 1.68 0.78 2.14 Sd 2.99 2.87 2.86 3.55 2.13 3.23 Serviceable posts contain factual information on which the police can act upon. – Serviceable posts showed use of Past tense Cops were driving on the wrong side near [withheld] hotel .. what action was taken against them?
  48. 48. Research Questions: Serviceability – RQ 1: What attributes differentiates – Serviceable posts from non-serviceable requests and – Sub-types of serviceable requests w.r.t content characteristics such as linguistic and emotional attributes? – RQ 2: How does police response time vary between serviceable and non-serviceable posts made on social media? – RQ 3: Can machine learning techniques be used to automatically identify serviceable requests – Can we further classify them into different sub-types using post characteristics (content and metadata)? 48
  49. 49. RQ2: Police Response Time – Survival Time – Time until the event of interest occurs – Censoring Event – Posts which did not receive a reply during our observation period – Survival Probability – Probability that a post survives longer than some specific time (t) given by survival function S(t) i.e. it does not receive a reply 49 Total N N of Events Censored %Censored Frwd 286 182 104 34.60 Give 88 53 35 39.80 Thanks 72 5 67 93.10 Need 104 60 44 42.30 Serv. 550 300 250 45.50 Police responses are maximum for Forward Sub-type posts
  50. 50. RQ2: Police Response Time 50 Mean Est. Sd. Error Median Est. Sd. Error Frwd 1062.52 82.22 21.33 2.05 Give 1064.37 138.08 20.43 8.45 Thank 2693.14 86.42 -- -- Need 1136.31 127.48 28.26 10.45 Serv. 1326.94 61.23 33.33 -- statistically significant difference between all four sub-types Log Rank (Mantel-Cox) test (χ2=57.03, df=3, p<0.005). Police reply to posts that can be given solution immediately followed by Forward – Kaplan Meier Estimator
  51. 51. Research Questions: Serviceability – RQ 1: What attributes differentiates – Serviceable posts from non-serviceable requests and – Sub-types of serviceable requests w.r.t content characteristics such as linguistic and emotional attributes? – RQ 2: How does police response time vary between serviceable and non-serviceable posts made on social media? – RQ 3: Can machine learning techniques be used to automatically identify serviceable requests – Can we further classify them into different sub-types using post characteristics (content and metadata)? 51
  52. 52. Formulation 52
  53. 53. RQ3: Automatic Classification Performance: Cost weights – Ten-fold Cross Validation Performance of different algorithms to correctly identify serviceable posts. – Content attributes such as emotions and linguistic attributes are highly predictive of serviceable posts in addition to bag-of-words model 53 Algorithm Recall F1 Accuracy RF 0.97 0.85 0.87 LR 0.82 0.77 0.76 ADT 0.96 0.80 0.86 DT 0.84 0.78 0.77 GBC 0.94 0.83 0.84 Attributes R2 Deviance Emotion 0.23 437.88 Linguistic 0.19 401.83 Bag-of-words 0.53 260.07
  54. 54. – +Model 1 – Explains 15.6% of the variance – Reduces deviance significantly to 1,127.58 (178.14 less). – Better predictors are sadness, fear, and joy. – +Model 2 – Explains 20% of the variance – 1st person singular pronouns have statistically significant 54 RQ3: Automatic Classification Performance
  55. 55. – +Model 3 – Explains 26.3% of the variance – Reduces deviance significantly to 984.92. – Not helpful: Tenses (present and future) and lexical terms (verbs and adverbs) – +Model 4 – Explains 32.4% of the variance & deviance is 902.79 i.e. 82.13 less – Reliable predictors: question, date, time, and entity count – Topic does not help much 55 RQ3: Automatic Classification Performance
  56. 56. Serviceability Plugin 56
  57. 57. Contributions – Identify need for social media support in collaborative policing – Quantify and mine unstructured data to analyzecommunication attributes and actionable information – Propose a method to enhance – Police responsiveness usingrequest–response detection framework and police response quantification to residents – Images retrieval architecture for improved collective action using user generated content 57
  58. 58. Challenges in similar images retrieval 58 Image processing 1. Scaling 2. Cropping 3. Stitching 4. Multiple
  59. 59. 59 Data Collection Sample images Event Total images Similar images Dissimilar images Charlie Hebdo 568 118 450 Kulkarni Ink 1,905 354 1,551 Insults Hanuman 664 277 387 ShaniShingnapur 180 70 110 RamRahim 408 97 311
  60. 60. Image features for similarity 1. Hand-Crafted Features a. 3D-colour histogram b. Daisy features c. ORB (Oriented FAST Rotated BRIEF) features d. Improved ORB (ORB + RANSAC) 60 2. Trainable Features a. Deep CNN (Convolution Neural Network) Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks.
  61. 61. Original Image Modified Image Improved ORB (% accuracy) CNN (% accuracy) Modification 93.7 98.3 Scaled, stitched image, added text, cropped 77.4 93.8 Cropped, stitched, text added 84.1 99.4 Scaled (7.4 ✕ 5.2) 61 Competing on modified images
  62. 62. PicHunt 62
  63. 63. Contributions – Identify need for social media support in collaborative policing – Quantify and mine unstructured data to analyzecommunication attributes and actionable information – Propose a method to enhancecollaboration – Police responsiveness usingrequest–response detection framework and police response quantification to residents – Images retrieval architecture for improved collective action using user generated content 63
  64. 64. Why it matters? – Propose a Data-Driven Technique to complement Overworked departments – Detecting posts that should elicit police response make social media streams more listenable for resident’s concerns – Help police improve policing and responsiveness – Taking cognizance of prominent constituents’ concern and unsafe regions can help police plan their resources better to provide improved safety – Measure resident’s reactions in a fine-grained manner – Information (e.g., emotions and interpersonal attributes) improve the understanding from factual information to a more nuanced understanding of psychological. 64
  65. 65. Technological Advancement – Designing early warning systems that indicate: – Need for emotional & social support needs to enhance police response to residents experiencing safety issues. – A feedback system on social media platforms that complements lack of physical signals of communication – informs about the likely time duration to respond to service request – Sense and record the reactions of citizens and share these records with decision makers – Take timely measures and gain better insights 65
  66. 66. Limitations – Cultural Limitation – Strong history of community policing may be helpful – Restricted Modality – Only text based serviceability detection – Consider other modalities such as videos, images etc. – Urban and Sub-urban Resident Communities – Rural areas may have different needs – Causality – Analysis is based on correlations 66
  67. 67. 67
  68. 68. Acknowledgement – Google for travel support – Tata Consultancy Services for funding the thesis work – All participants and police officers who helped us in various stages of the thesis – My collaborators – special thanks to Dr. Nitesh Saxena (UAB), Dr. Munmun De Choudhury (GaTech), Dr. Iulia Ion (Google) – Monitoring Committee Dr. Rahul Purandare, Dr. Sambuddho Chakravarty, Dr. Amarjeet Singh – Dr. Aditi Gupta, Dr. Paridhi Jain, Siddhartha Asthana, PrateekDewan, Anupama Aggarwal, Srishti Gupta, Rishabh Kaushal, Anuradha Gupta – Shrey Bagroy, Sonal Gupta, Divam Gupta, Megha Arora, Indira Sen, Neha Jawalkar, Bhavana Nagpal, Tushar Gupta, Vedant Swain – Members of Cybersecurity Education and Research Centre (CERC) and Precog who have given us continued support throughout the project – My Family 68
  69. 69. Acknowledgement – Dr. Aaditeshwar Seth – Dr. Carlos Castillo – Dr. Maura Conway – Dr. Ponnurangam Kumaraguru 69
  70. 70. Publications – Peer Reviewed Conferences – Sachdeva, N. and Kumaraguru, P. Online Social Media - New face of policing? A Sur- vey Exploring Perceptions, Behavior, Challenges for Police Field Officers and Residents. Accepted at18th International Conference on Human-Computer Interaction (HCII), 2016. – Sachdeva, N. and Kumaraguru, P. Deriving requirements for social media based com- munity policing: insights from police. Accepted at ACM 16th International Digital Government Research Conference (dg.o 2015), 2015. – Sachdeva, N. and Kumaraguru, P. Online Social Networks and Police in India - Under- standing the Perceptions, Behavior, Challenges. Accepted at the European Conference on Computer-Supported Cooperative Work (ECSCW), 2015. – Sachdeva, N. and Kumaraguru, P. Characterising Behavior and Emotions on Social Media for Safety: Exploring Online Communication between Police and Citizens. Accepted at 30th British Human Computer Interaction Conference (HCI) 2016. – Goel, S., Sachdeva, N., Kumaraguru, P., Subramanyam, A., and Gupta, D. PicHunt: Social Media Image Retrieval for Improved Law Enforcement. Accepted at 8th International Conference on Social Informatics. 2016. – Sachdeva, N., and Kumaraguru, P. Social Networks for Police and Residents in India: Exploring Online Communication for Crime Prevention. Accepted at the ACM16th Annual International Conference on Digital Government Research (dg.o), 2015. [Best paper award]. – Sachdeva, N., and Kumaraguru, P. Call for Service: Characterizing and Modeling Police Response to Serviceable Requests on Facebook. Accepted at the ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW), 2017. 70
  71. 71. Publications – Peer-reviewed Conference Papers – Lamba, H., Bharadhwaj, V., Vachher, M., Agarwal, D., Arora, M., Sachdeva, N., Ku- maraguru, P. From Camera to Deathbed: Understanding Dangerous Selfies on Social Media. 11th International Conference on Web and Social Media (ICWSM), 2017 – Mohamed, M., Sachdeva, N., Georgescu, M., Gao, S., Saxena, N., Zhang, C., Kumaraguru, P., Van Oorschot, P., and Chen, W. A Three-Way Investigation of a Game-CAPTCHA: Automated Attacks, Relay Attacks and Usability. Accepted at 9th ACM Symposium on Information, Computer and Communications Security (ASIACCS), 2014. – Sachdeva, N., Saxena, N., and Kumaraguru, P. On the Viability of CAPTCHAs for Use in Telephony Systems: A Usability Field Study. 16th Information Security Conference November 2013 in Dallas, Texas (ISC), 2013. – Sachdeva, N., Saxena, N., and Kumaraguru, P. On the Viability of CAPTCHAs for Use in Telephony Systems: A Usability Field Study [Poster]. (APCHI) 2013 – Ion, I., Sachdeva, N., Kumaraguru, P., and Capkun, S. Home is safer than the cloud! privacy concerns for consumer cloud storage. In Symposium on Usable Privacy and Security (SOUPS) (2011). – Journal Papers – Manar Mohamed, Song G ao, Niharika Sachdeva, Nitesh Saxena, Chengcui Zhang, Pon- nurangam Kumaraguru and Paul van Oorschot. On the Security and Usability of Dynamic Cognitive Game CAPTCHAs. In Journal of Computer Security (JCS), 2017. 71
  72. 72. Thank you! niharikas@iiitd.ac.in http://precog.iiitd.edu.in/

×