Yuhui Wang is defending his Ph.D. thesis on "Fusing Physical and Social Sensors for Situation Awareness" at the National University of Singapore. The document discusses using both physical sensors like cameras and social sensors like Twitter to gain a more complete understanding of situations and events. It presents a framework that detects low-level concepts from camera images, applies filtering and analysis, and then fuses information from social media to detect real-world events with both visual and social data. Examples are given detecting parades and protests by analyzing camera feeds and geotagged tweets from New York City.
A SiGe BiCMOS E-Band Power Amplifier with 22% PAE at 18dBm OP1dB and 8.5% at ...aiclab
Huawei and the University of Pavia present a SiGe power amplifier at 80GHz. A common-base output stage causes the DC current to track the signal current and improve efficiency at back-off power. Realized prototype shows OP1dB of 18dBm with Psat of 19dBm. The efficiency at OP1dB and at 6dB are 22% and 8.5%,
respectively.
This slide file was presented in ISSCC 2015. This paper shows a padless chip concept which enables CMOS chip to be tested in bare chip state.
This padless concept can be applied to sensors, memories, etc to reduce the extra system cost.
Analysis and optimization of wireless power transfer linkAjay Kumar Sah
In this presentation, a high efficiency Gallium nitride (GaN), HEMT (High Electron Mobility Transistor) class-E power amplifier for the wireless power transfer link is designed and simulated on PSpice. A four-coil wireless power transfer link is modeled for maximum power transfer efficiency on ADS (Advanced Design System) and frequency splitting phenomenon is demonstrated, explained and analyzed. Two resonant coupling structures, series & mixed, are presented and compared. The efficiency performance of the link is studied using spiral and helical antennas of different wire make. In addition, techniques for improving efficiency of the wireless power transfer systems with changing coupling coefficient viz. frequency splitting phenomenon of the coils are proposed.
A SiGe BiCMOS E-Band Power Amplifier with 22% PAE at 18dBm OP1dB and 8.5% at ...aiclab
Huawei and the University of Pavia present a SiGe power amplifier at 80GHz. A common-base output stage causes the DC current to track the signal current and improve efficiency at back-off power. Realized prototype shows OP1dB of 18dBm with Psat of 19dBm. The efficiency at OP1dB and at 6dB are 22% and 8.5%,
respectively.
This slide file was presented in ISSCC 2015. This paper shows a padless chip concept which enables CMOS chip to be tested in bare chip state.
This padless concept can be applied to sensors, memories, etc to reduce the extra system cost.
Analysis and optimization of wireless power transfer linkAjay Kumar Sah
In this presentation, a high efficiency Gallium nitride (GaN), HEMT (High Electron Mobility Transistor) class-E power amplifier for the wireless power transfer link is designed and simulated on PSpice. A four-coil wireless power transfer link is modeled for maximum power transfer efficiency on ADS (Advanced Design System) and frequency splitting phenomenon is demonstrated, explained and analyzed. Two resonant coupling structures, series & mixed, are presented and compared. The efficiency performance of the link is studied using spiral and helical antennas of different wire make. In addition, techniques for improving efficiency of the wireless power transfer systems with changing coupling coefficient viz. frequency splitting phenomenon of the coils are proposed.
Conference talk: Understanding Vulnerabilities of Location Privacy Mechanisms...Zohaib Riaz
Slides for our work presented at MobiQuitous 2017 Conference (http://mobiquitous.org/).
Full paper text: ftp://ftp.informatik.uni-stuttgart.de/pub/library/ncstrl.ustuttgart_fi/INPROC-2017-46/INPROC-2017-46.pdf
This paper focused on revealing weaknesses of existing location obfuscation approaches when an attacker possesses accurate or obfuscated location history information.
This is the presentation file for the Smart Citizen Workshop held in Banff, Canada, as part of the Cybera Summit 2016. 15 smart citizens participated the workshop, assembled their own air quality sensor, connected to the OGC SensorThings API cloud (powered by SensorUp), and analyzed the data with SensorThings API SDK libraries.
Snap4City November 2019 Course: Smart City IOT Data AnalyticsPaolo Nesi
• Data Analytics: Examples from Snap4City
o Smart parking: Predictions
o User Behavior Analysis, via Wi-Fi, OD, Trajectories
o Recognition of Used Transportation means
o Traffic Flow Reconstruction, from Traffic Sensors Data
o Quality of Public Transport Service
o Origin Destination Matrices from: Wi-Fi, Mobile Apps, etc.
o Demand of Mobility vs Offer of Transportation
o Modal and Multimodal Routing for Navigation and Travel Planning
o Environmental Data Analysis and Predictions, early Warning
o Prediction of Air Quality Conditions
o Anomaly Detection
o What-IF Analysis
• Data Analytics: Enforcing and Exploiting
o Real Time Data Analytics: using R Studio Exploitation in IOT Applications
• Decision Support Systems, Smart DS and Resilience DS
• Twitter Vigilance: Social Media Analysis: Early Warning, Predictions
Twitter Vigilance: a Multi-User platform for Cross-Domain Twitter Data Analyt...Paolo Nesi
Citizens as sensors to
Assess sentiment on services, events, …
Response of consumers wrt…
Early detection of critical conditions
Information channel
Opinion leaders
Communities
Formation
Predicting volume of visitors for tuning the services
Collecting Tweets
on the basis of several criterial, searches
Multiple users may have multiple searches and multiple purposes (views on those searches) minimization of searches
With high reliable model exploiting Twitter Search and/or Stream API
Performing NLP and Sentiment Analysis
Real time or daily
Multiple languages
Compute Metrics:
different kinds, etc.: volume, users, etc..
Low and High Level,
Work on daily, hourly and real time
Downloading data: raw data and derived metrics
Providing support for drilling down on data
to perform inspection and analysis, faceted search for instance
perform data analytics
exploiting metrics on Machine Learning and predictive tools
Providing API for querying, dashboard
The IEEE Smart World Congress originated from the 2005 Workshop on Ubiquitous Smart Worlds (USW, Taipei) and the 2005 Symposium on Ubiquitous Intelligence and Smart World (UISW, Nagasaki). SmartWorld 2017 in San Francisco is the next edition after the successful SmartWorld 2016 in Toulouse France and SmartWorld 2015 in Beijing China. SmartWorld 2017 is to provide a high-profile, leading-edge platform for researchers and engineers to exchange and explore state-of-art advances and innovations in graceful integrations of Cyber, Physical, Social, and Thinking Worlds for the theme
http://ieee-smartworld.org/2017/smartworld/
CQA services are collaborative platforms where users ask and answer questions. We investigate the influence of national culture on people’s online questioning and answering behavior. For this, we analyzed a sample of 200 thousand users in Yahoo Answers from 67 countries. We use a number of cultural factors extracted from Geert Hofstede’s cultural dimensions and Robert Levine’s Pace of Life and show that behavioral cultural differences exist in community question answering platforms. We find that national cultures differ in Yahoo Answers along a number of dimensions such as temporal predictability of activities, contribution related behavioral patterns, privacy concerns, and power inequality.
Cultures in Community Question Answering. Imrul Kayes, Nicolas Kourtellis, Daniele Quercia, Adriana Iamnitchi, and Francesco Bonchi. ACM 26th Conference on Hypertext and Social Media (HT'15), 2015
Slides for my Associate Professor (oavlönad docent) lecture.
The lecture is about Data Streaming (its evolution and basic concepts) and also contains an overview of my research.
Multimedia Information Retrieval: Bytes and pixels meet the challenges of hum...maranlar
Within computer science, "Multimedia" is a field of research that investigates how computers can support people in communication, information finding, and knowledge/opinion building. Multimedia content is defined broadly. It includes not only video, but also images accompanied by text and other information (for example, a geo-location). It can be professionally produced, or generated by users for online sharing. Computer scientists historically have a “love-hate” relationship with multimedia. They “love” it because of the richness of the data sources and the wealth of available data, which leads to interesting problems to tackle with machine learning. They “hate” it because multimedia is a diffuse and moving target: the interpretation of multimedia differs from person to person, and changes over time in the course of its use as a communication medium. This talk gives a view onto ongoing research in the area of multimedia information retrieval algorithms, which help people find multimedia. We look at a series of topics that reveal how pattern recognition, text processing, and crowdsourcing tools are used in multimedia research, and discuss both their limitations and their potential.
Digital scholarship and identifiers - Geoffrey Bilder, CrossReff
Share update – Elliott Shore, Association of Research Libraries
Jisc Monitor update – Neil Jacobs, Jisc
Infrastructure and services to track research activity – Daniel Hook, Digital Science
Jisc and CNI conference, 6 July 2016
Data Science: History repeated? – The heritage of the Free and Open Source GI...Peter Löwe
Data Science is described as the process of knowledge extraction from large data sets by means of scientific
methods. The discipline draws heavily from techniques and theories from many fields, which are jointly used to
furthermore develop information retrieval on structured or unstructured very large datasets. While the term Data
Science was already coined in 1960, the current perception of this field places is still in the first section of the hype cycle according to Gartner, being well en route from the technology trigger stage to the peak of inflated
expectations.
In our view the future development of Data Science could benefit from the analysis of experiences from
related evolutionary processes. One predecessor is the area of Geographic Information Systems (GIS). The
intrinsic scope of GIS is the integration and storage of spatial information from often heterogeneous sources, data
analysis, sharing of reconstructed or aggregated results in visual form or via data transfer. GIS is successfully
applied to process and analyse spatially referenced content in a wide and still expanding range of science
areas, spanning from human and social sciences like archeology, politics and architecture to environmental and
geoscientific applications, even including planetology.
This paper presents proven patterns for innovation and organisation derived from the evolution of GIS,
which can be ported to Data Science. Within the GIS landscape, three strategic interacting tiers can be denoted: i) Standardisation, ii) applications based on closed-source software, without the option of access to and analysis of the implemented algorithms, and iii) Free and Open Source Software (FOSS) based on freely accessible program code enabling analysis, education and ,improvement by everyone. This paper focuses on patterns gained from the synthesis of three decades of FOSS development. We identified best-practices which evolved from long term FOSS projects, describe the role of community-driven global umbrella organisations such as OSGeo, as well as the standardization of innovative services. The main driver is the acknowledgement of a meritocratic attitude.
These patterns follow evolutionary processes of establishing and maintaining a web-based democratic culture
spawning new kinds of communication and projects. This culture transcends the established compartmentation and
stratification of science by creating mutual benefits for the participants, irrespective of their respective research
interest and standing. Adopting these best practices will enable
Conference talk: Understanding Vulnerabilities of Location Privacy Mechanisms...Zohaib Riaz
Slides for our work presented at MobiQuitous 2017 Conference (http://mobiquitous.org/).
Full paper text: ftp://ftp.informatik.uni-stuttgart.de/pub/library/ncstrl.ustuttgart_fi/INPROC-2017-46/INPROC-2017-46.pdf
This paper focused on revealing weaknesses of existing location obfuscation approaches when an attacker possesses accurate or obfuscated location history information.
This is the presentation file for the Smart Citizen Workshop held in Banff, Canada, as part of the Cybera Summit 2016. 15 smart citizens participated the workshop, assembled their own air quality sensor, connected to the OGC SensorThings API cloud (powered by SensorUp), and analyzed the data with SensorThings API SDK libraries.
Snap4City November 2019 Course: Smart City IOT Data AnalyticsPaolo Nesi
• Data Analytics: Examples from Snap4City
o Smart parking: Predictions
o User Behavior Analysis, via Wi-Fi, OD, Trajectories
o Recognition of Used Transportation means
o Traffic Flow Reconstruction, from Traffic Sensors Data
o Quality of Public Transport Service
o Origin Destination Matrices from: Wi-Fi, Mobile Apps, etc.
o Demand of Mobility vs Offer of Transportation
o Modal and Multimodal Routing for Navigation and Travel Planning
o Environmental Data Analysis and Predictions, early Warning
o Prediction of Air Quality Conditions
o Anomaly Detection
o What-IF Analysis
• Data Analytics: Enforcing and Exploiting
o Real Time Data Analytics: using R Studio Exploitation in IOT Applications
• Decision Support Systems, Smart DS and Resilience DS
• Twitter Vigilance: Social Media Analysis: Early Warning, Predictions
Twitter Vigilance: a Multi-User platform for Cross-Domain Twitter Data Analyt...Paolo Nesi
Citizens as sensors to
Assess sentiment on services, events, …
Response of consumers wrt…
Early detection of critical conditions
Information channel
Opinion leaders
Communities
Formation
Predicting volume of visitors for tuning the services
Collecting Tweets
on the basis of several criterial, searches
Multiple users may have multiple searches and multiple purposes (views on those searches) minimization of searches
With high reliable model exploiting Twitter Search and/or Stream API
Performing NLP and Sentiment Analysis
Real time or daily
Multiple languages
Compute Metrics:
different kinds, etc.: volume, users, etc..
Low and High Level,
Work on daily, hourly and real time
Downloading data: raw data and derived metrics
Providing support for drilling down on data
to perform inspection and analysis, faceted search for instance
perform data analytics
exploiting metrics on Machine Learning and predictive tools
Providing API for querying, dashboard
The IEEE Smart World Congress originated from the 2005 Workshop on Ubiquitous Smart Worlds (USW, Taipei) and the 2005 Symposium on Ubiquitous Intelligence and Smart World (UISW, Nagasaki). SmartWorld 2017 in San Francisco is the next edition after the successful SmartWorld 2016 in Toulouse France and SmartWorld 2015 in Beijing China. SmartWorld 2017 is to provide a high-profile, leading-edge platform for researchers and engineers to exchange and explore state-of-art advances and innovations in graceful integrations of Cyber, Physical, Social, and Thinking Worlds for the theme
http://ieee-smartworld.org/2017/smartworld/
CQA services are collaborative platforms where users ask and answer questions. We investigate the influence of national culture on people’s online questioning and answering behavior. For this, we analyzed a sample of 200 thousand users in Yahoo Answers from 67 countries. We use a number of cultural factors extracted from Geert Hofstede’s cultural dimensions and Robert Levine’s Pace of Life and show that behavioral cultural differences exist in community question answering platforms. We find that national cultures differ in Yahoo Answers along a number of dimensions such as temporal predictability of activities, contribution related behavioral patterns, privacy concerns, and power inequality.
Cultures in Community Question Answering. Imrul Kayes, Nicolas Kourtellis, Daniele Quercia, Adriana Iamnitchi, and Francesco Bonchi. ACM 26th Conference on Hypertext and Social Media (HT'15), 2015
Slides for my Associate Professor (oavlönad docent) lecture.
The lecture is about Data Streaming (its evolution and basic concepts) and also contains an overview of my research.
Multimedia Information Retrieval: Bytes and pixels meet the challenges of hum...maranlar
Within computer science, "Multimedia" is a field of research that investigates how computers can support people in communication, information finding, and knowledge/opinion building. Multimedia content is defined broadly. It includes not only video, but also images accompanied by text and other information (for example, a geo-location). It can be professionally produced, or generated by users for online sharing. Computer scientists historically have a “love-hate” relationship with multimedia. They “love” it because of the richness of the data sources and the wealth of available data, which leads to interesting problems to tackle with machine learning. They “hate” it because multimedia is a diffuse and moving target: the interpretation of multimedia differs from person to person, and changes over time in the course of its use as a communication medium. This talk gives a view onto ongoing research in the area of multimedia information retrieval algorithms, which help people find multimedia. We look at a series of topics that reveal how pattern recognition, text processing, and crowdsourcing tools are used in multimedia research, and discuss both their limitations and their potential.
Digital scholarship and identifiers - Geoffrey Bilder, CrossReff
Share update – Elliott Shore, Association of Research Libraries
Jisc Monitor update – Neil Jacobs, Jisc
Infrastructure and services to track research activity – Daniel Hook, Digital Science
Jisc and CNI conference, 6 July 2016
Data Science: History repeated? – The heritage of the Free and Open Source GI...Peter Löwe
Data Science is described as the process of knowledge extraction from large data sets by means of scientific
methods. The discipline draws heavily from techniques and theories from many fields, which are jointly used to
furthermore develop information retrieval on structured or unstructured very large datasets. While the term Data
Science was already coined in 1960, the current perception of this field places is still in the first section of the hype cycle according to Gartner, being well en route from the technology trigger stage to the peak of inflated
expectations.
In our view the future development of Data Science could benefit from the analysis of experiences from
related evolutionary processes. One predecessor is the area of Geographic Information Systems (GIS). The
intrinsic scope of GIS is the integration and storage of spatial information from often heterogeneous sources, data
analysis, sharing of reconstructed or aggregated results in visual form or via data transfer. GIS is successfully
applied to process and analyse spatially referenced content in a wide and still expanding range of science
areas, spanning from human and social sciences like archeology, politics and architecture to environmental and
geoscientific applications, even including planetology.
This paper presents proven patterns for innovation and organisation derived from the evolution of GIS,
which can be ported to Data Science. Within the GIS landscape, three strategic interacting tiers can be denoted: i) Standardisation, ii) applications based on closed-source software, without the option of access to and analysis of the implemented algorithms, and iii) Free and Open Source Software (FOSS) based on freely accessible program code enabling analysis, education and ,improvement by everyone. This paper focuses on patterns gained from the synthesis of three decades of FOSS development. We identified best-practices which evolved from long term FOSS projects, describe the role of community-driven global umbrella organisations such as OSGeo, as well as the standardization of innovative services. The main driver is the acknowledgement of a meritocratic attitude.
These patterns follow evolutionary processes of establishing and maintaining a web-based democratic culture
spawning new kinds of communication and projects. This culture transcends the established compartmentation and
stratification of science by creating mutual benefits for the participants, irrespective of their respective research
interest and standing. Adopting these best practices will enable
1. National University of Singapore
Yuhui Wang
Advisor: Prof. Mohan Kankanhalli
NUS Graduate School for Integrative Sciences & Engineering
National University of Singapore
16 November 2016
Fusing Physical and Social
Sensors for Situation Awareness
Ph.D. Thesis Defense
1
2. National University of Singapore
24th International World Wide Web Conference
Big Sensor Data
Physical Sensors
2
3. National University of Singapore
24th International World Wide Web Conference
Big Sensor Data
Social Sensors
3
4. National University of Singapore
24th International World Wide Web Conference
Big Sensor Data
Physical Sensors
• Camera
Social Sensors
• Twitter
82 % access Twitter via mobile devices
-- https://www.statista.com/statistics/282087/number-of-monthly-active-twitter-users/
-- T. Huang, “Surveillance Video: The
Biggest Big Data,” Computing Now, vol. 7,
no. 2, Feb. 2014, IEEE Computer Society
65% global big data come
from surveillance video by
2015
4
5. National University of Singapore
24th International World Wide Web Conference
Live WebCams
5
6. National University of Singapore
24th International World Wide Web Conference
Live WebCams
6
7. National University of Singapore
24th International World Wide Web Conference
Motivation
Physical
Sensors
Social
Sensors
Physical and social sensors are observing same
situation from different perspectives
7
8. National University of Singapore
24th International World Wide Web Conference
Literature Review
TwitterStand [2009]
Sakaki et al. [2010]
Weng & Lee [2011]
Walther & Kaisser [2013]
Twevent [2012]
Yang et la. [2016]
...
Kulkarni et al. [2005]
Atrey et al. [2007]
Jacobs et al. [2009]
Babari et al. [2012]
Jing et al. [2016]
Lin et al. [2016]
…
Mediamill101 [2006]
VIREO-374 [2010]
Snoek et al. [2006]
Karpathy et al. [2014]
Vinyals et al. [2014]
Markatopoulou et al. [2016]
Eventshop [2012]
Vivek et al. [2010]
Pan et al. [2013]
Wu et al. [2015]
Semantic Understanding &
Image/Video Concept
Detection
Situation Understanding
using Social Sensors
Event Detection
using Physical Sensors
This work
8
9. National University of Singapore
24th International World Wide Web Conference
Problem & Challenges
Work independently
Incomplete information
Different Modalities
Numeric (physical) vs Symbolic (social)
Different Spatio-Temporal Density
Spatial: physical < social
Temporal: physical > social
Uncertain Data (Noise)
No restrictions on content
Failure of sensors
Maintenance of devices
9
10. National University of Singapore
24th International World Wide Web Conference
Work I:
Tweeting Cameras for Event Detection
10
11. National University of Singapore
24th International World Wide Web Conference
Motivation
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101
010101
010110
110101010101
010110
110101010101
010110
110101Big
Visual Data
?
!? …
Traditional Camera System Tweeting Camera System
11
12. National University of Singapore
24th International World Wide Web Conference
Multi-Layer Tweeting Cameras Framework
PST = < Probability, Space, Time, Label>
Isensor
Low-level Concept Detection
Crowd
Action
Parade
Face
Car
Mid-level Concept Filtering
Filtering & Analytic
Operators
High-level Social Sensor Fusion
Social Media
Social Information
Extraction
Cross Media Analysis
Database
(Building
Blocks)
Event Signal Detection
Feature Extraction
…
Traffic
User
Cmage
Physical
Sensors
Social
Sensors
Filtering &
Analytic
Operators
Event
Detection
Social
Information
Crawler
Sensor Data
Collector
Parade
Is going
12
13. National University of Singapore
24th International World Wide Web Conference
Multi-Layer Tweeting Cameras Framework
PST = < Probability, Space, Time, Label>
Low-level Concept Detection
Crowd
Action
Parade
Face
Car
Database
(Building
Blocks)
Feature Extraction
…
Traffic
Sensor Data
Collector
13
14. National University of Singapore
24th International World Wide Web Conference
Multi-Layer Tweeting Cameras Framework
PST = < Probability, Space, Time, Label>
Isensor
Low-level Concept Detection
Crowd
Action
Parade
Face
Car
Mid-level Concept Filtering
Filtering & Analytic
Operators Database
(Building
Blocks)
Event Signal Detection
Feature Extraction
…
Traffic
Sensor Data
Collector
14
15. National University of Singapore
24th International World Wide Web Conference
Multi-Layer Tweeting Cameras Framework
PST = < Probability, Space, Time, Label>
Isensor
Low-level Concept Detection
Crowd
Action
Parade
Face
Car
Mid-level Concept Filtering
Filtering & Analytic
Operators
High-level Social Sensor Fusion
Social Media
Social Information
Extraction
Cross Media Analysis
Database
(Building
Blocks)
Event Signal Detection
Feature Extraction
…
Traffic
Social
Information
Crawler
Sensor Data
Collector
15
16. National University of Singapore
24th International World Wide Web Conference
Multi-Layer Tweeting Cameras FrameworkLow-level Concept Detection
• Concept Detectors
Columbia 374
VIERO-374
Mediamill (101)
VIREO-WEB81
CU-VIREO 374
…
16
17. National University of Singapore
24th International World Wide Web Conference
Columbia 374
VIERO-374
Mediamill (101)
VIREO-WEB81
CU-VIREO 374
…
Low-level Concept Detection
• Concept Detectors
Concept Label Confidence
Crowd 0.9
Parade 0.8
Car 0.1
Outdoor 0.5
… …
6 Avenue @ 23 Street 15:10 13th Dec, 2014
Location Time
6 Ave@ 23 St 15:10, Dec13th
6 Ave@ 23 St 15:10, Dec13th
6 Ave@ 23 St 15:10, Dec13th
6 Ave@ 23 St 15:10, Dec13th
… … 17
18. National University of Singapore
24th International World Wide Web Conference
Low-level Concept Detection
I’ve seen crowd here
now, 90% sure
Low Level Camera Tweet
18
19. National University of Singapore
24th International World Wide Web Conference
Low-level Concept Detection
server storage
I’ve seen crowd
here now, 90%
sure
I’ve seen crowd
here now, 90%
sure
I’ve seen crowd
here now, 90%
sure
I’ve seen crowd
here now, 90%
sure
I’ve seen crowd
here now, 90%
sure
19
20. National University of Singapore
24th International World Wide Web Conference
Mid-level Concept Filtering
• Filtering & Analytic Operators
o Query Operators:
E.g. Show the March 17th data for the concept of “parade” at 5th
Avenue with a confidence higher than 0.8:
𝑄𝑢𝑒𝑟𝑦: 𝜃 𝑃_𝑃𝑅𝑂𝐵⋀𝑃_𝐿𝐴𝐵𝐸𝐿⋀𝑃_𝐿𝑂𝐶⋀𝑃_𝑇𝐼𝑀𝐸 𝑆
Where 𝑃_𝑃𝑅𝑂𝑃 = 𝑃_𝑝𝑟𝑜𝑏(0.8 ≤ 𝑝), 𝑃𝐿𝐴𝐵𝐸𝐿 = 𝑃𝑙𝑎𝑏𝑒𝑙 𝑙𝑎𝑏𝑒𝑙=𝑝𝑎𝑟𝑎𝑑𝑒 ,
𝑃_𝐿𝑂𝐶 = 𝑃_𝑙𝑜𝑐(𝐶𝐴𝑀𝑖 = 5 𝑡ℎ
𝐴𝑣𝑒𝑛𝑢𝑒), 𝑃_𝑇𝐼𝑀𝐸 = 𝑃_𝑡𝑖𝑚𝑒(𝑡 = 𝑀𝑎𝑟𝑡ℎ 17 𝑡ℎ
)
o Statistical Functions
o mean, max, min, sum
o Processing Operators
o Extremes
o Smooth
o Trend
o Outlier (Anomaly)
20
21. National University of Singapore
24th International World Wide Web Conference
Operators
St Patrick’s Day Parade Event (“parade” concept)
Probability
Hour
PROJECTION(SELECTION(EXTRE
ME(SMOOTH(MAP
𝑡1, 𝑙𝑜𝑐1, raw_image1
, 0.2
𝑡2, 𝑙𝑜𝑐2, raw_image2
, 0.3
…
𝑡 𝑛, 𝑙𝑜𝑐 𝑛, raw_image 𝑛
, 0.7)))) = (t 𝑥, 𝑙𝑜𝑐1,
parade, 0.7)
21
22. National University of Singapore
24th International World Wide Web Conference
What about Social Tweets ?
Tokenizer
Normalizer
Preprocessing
English Words
Slang Words
Dictionary
2015-01-26 12:05:32
I'm ready for you snowpocalypse! #madisonsqpark !!!! @madisonsqpark zzzzzzz
#snowpocalypse @ Madison Square ParkQ& http://t.co/KoRJ4FYOkZ
40.7421 -73.988283
2015-01-26 12:05:32
I'm ready for you snowpocalypse #madisonsqpark #snowpocalypse Madison Square Park
40.7421 -73.988283
22
23. National University of Singapore
24th International World Wide Web Conference
What about Social Tweets ?
Representative Term Mining
𝐷𝑎𝑦 𝑒𝑣𝑒𝑛𝑡
(𝑡 𝑠−𝑡 𝑒)
𝐷𝑎𝑦 𝑝𝑟𝑣1
(𝑡 𝑠−𝑡 𝑒)
… … ……
𝐷𝑎𝑦 𝑝𝑟𝑣2
(𝑡 𝑠−𝑡 𝑒)
23
24. National University of Singapore
24th International World Wide Web Conference
Representative Term Mining
What about Social Tweets ?
Loc 1 Loc 2 Loc 3 Loc N
𝑇𝐶: tweets posted during events
𝑇 𝐻: tweets posted before events
𝑡𝑓: term frequency
i𝑑𝑓: inverse document frequency
𝑤𝑡𝑒𝑟𝑚 = 𝑡𝑓 𝑡𝑒𝑟𝑚, 𝑇𝐶 × 𝑖𝑑𝑓 𝑡𝑒𝑟𝑚, 𝑇 𝐻
𝑠. 𝑡.
𝑡𝑓 𝑡𝑒𝑟𝑚, 𝑇𝐶 =
𝑓(𝑡𝑒𝑟𝑚, 𝑇𝐶)
max{𝑓 𝜔, 𝑇𝐶 , ∀𝜔 ∈ 𝑇𝐶}
𝑖𝑑𝑓 𝑡𝑒𝑟𝑚, 𝑇 𝐻 = 𝑙𝑜𝑔
𝑇 𝐻
{𝑡ℎ ∈ 𝑇 𝐻: 𝑡𝑒𝑟𝑚 ∈ 𝑡ℎ}
…
…
24
25. National University of Singapore
24th International World Wide Web Conference
Representative Term Mining
What about Social Tweets ?
Loc 1 Loc 2 Loc 3 Loc N
𝑇𝐶: tweets posted during events
𝑇 𝐻: tweets posted before events
𝑡𝑓: term frequency
i𝑑𝑓: inverse document frequency
𝑤𝑡𝑒𝑟𝑚 = 𝑡𝑓 𝑡𝑒𝑟𝑚, 𝑇𝐶 × 𝑖𝑑𝑓 𝑡𝑒𝑟𝑚, 𝑇 𝐻
𝑠. 𝑡.
𝑡𝑓 𝑡𝑒𝑟𝑚, 𝑇𝐶 =
𝑓(𝑡𝑒𝑟𝑚, 𝑇𝐶)
max{𝑓 𝜔, 𝑇𝐶 , ∀𝜔 ∈ 𝑇𝐶}
𝑖𝑑𝑓 𝑡𝑒𝑟𝑚, 𝑇 𝐻 = 𝑙𝑜𝑔
𝑇 𝐻
{𝑡ℎ ∈ 𝑇 𝐻: 𝑡𝑒𝑟𝑚 ∈ 𝑡ℎ}
…
…
25
26. National University of Singapore
24th International World Wide Web Conference
Data Analysis & Real World Events
26
27. National University of Singapore
24th International World Wide Web Conference
DataSet
• NYC Traffic CCTV Camera
149 cameras all over Manhattan
Sampling rate: ~10s/f
Period: 2014 ~ May 2016
• Twitter Data (Geo-tagged)
o Region: Manhattan
o Oct 4th 2014 ~ Sep 2016
o Attributes: text, time, geo-coordinates, etc.
o Size: ~40,000/day
Event Date Time Location
CBGB Music Festival 12 Oct 10am-7pm Broadway 51st
Columbus Day Parade 13 Oct 11am-5pm 5th Avenue
Hispanic Parade 12 Oct 12pm-5pm 5th Avenue
Million March NYC Protest 13 Dec 2pm-5pm Washington Square
Park, 5th Avenue,
Foley Square 27
28. National University of Singapore
24th International World Wide Web Conference
Real-world Events
28
29. National University of Singapore
24th International World Wide Web Conference
Real-world Events
Hispanic ParadeCBGB Musical Festival
Columbus Day Parade
twtw
tw tw
tw tw
twtw
tw tw
tw tw
twtw
tw tw
tw tw
twtw
tw tw
tw tw
Historic TweetsRecent Tweets
Event time and location
Retrieve recent tweets Retrieve historic tweets
29
30. National University of Singapore
24th International World Wide Web Conference
“Million March NYC Protest” Event
30
31. National University of Singapore
24th International World Wide Web Conference
Twitter Images
“people marching” : 0.5
“parade” : 0.4, “crowd” : 0.9
31
32. National University of Singapore
Demo : Tweeting Camera
A New Paradigm of Event-based Smart
Sensing Device
32
33. National University of Singapore
24th International World Wide Web Conference
From 1925:
35mm Leica A
From 1942: CCTV Camera Network
“Looks like a fire ball here ?
”
Fire Event Parade Event Meeting Event Jogging Event
NOW:New Tweeting Cameras Paradigm
Motivation
33
34. National University of Singapore
24th International World Wide Web Conference
Tweeting Camera (Group Meeting Event)
34
https://www.youtube.com/watch?v=eXn89Z_MZwI
35. National University of Singapore
24th International World Wide Web Conference
Summary
• Aggregation of physical sensors and social sensors
• Multi-layer tweeting camera framework
• Probabilistic Spatio-temporal Data (Camera Tweet)
• Analytic functions & operators
• Concept Based Image (Cmage)
• Feasibility via Real-world Events Data
35
36. National University of Singapore
24th International World Wide Web Conference
Work II:
Cmage Based Hybrid Fusion of Multimodal
Event Signals
36
37. National University of Singapore
24th International World Wide Web Conference
Motivation
Geo-tagged Multisensor Data
Noise and Sparsity
PST = <Loc, Time, Label, Prob>
Isensor
Low-level Concept Detection
Crowd
Actio
nPara
de Face
Car
Mid-level Concept Filtering
Filtering & Analytic
Operators
High-level Social Sensor Fusion
Social
Media
Social Information
Extraction
Cross Media Analysis
Database
(Building
Blocks)
Event Signal Detection
Feature Extraction
…
Traffic
User
Cmage
Physical
Sensors
Social
Sensors
Filtering &
Analytic
Operators
Event
Detection
Social
Informati
onCrawle
r
Sensor
Data
Collector
Parade
Is going
Event Locating
Better Visualization (Where happens what)
Goal:
37
38. National University of Singapore
24th International World Wide Web Conference
Event Signals to Event Cmage
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
“crowdedness”
38
39. National University of Singapore
24th International World Wide Web Conference
Cmage-based Fusion Pipeline
Gaussian
Process
Prediction
Bayesian
Decision
Fusion
Spatial
Fusion
Sensor Cmage
Social Cmage Event CmageFused CmageExtract Social
Signals
Extract Sensor
Signals
Sensor Concepts
(Crowd, People marching, Car, Traffic, Building)
Social Terms
(“MillionMarchNYC”, “HappyNewYear”, “SGHaze”)
… …
39
40. National University of Singapore
24th International World Wide Web Conference
Sensor Cmage Pixel Value Estimation
Gaussian Process
- using noisy and sparse observations
Observed pixel values
Predicted pixel values
Gaussian Process
based Prediction
40
42. National University of Singapore
24th International World Wide Web Conference
Evaluation Metrics
Saliency Metric S
Low S => more salient & concentrated region
Mean Square Error compared with Ground Truth
Experiments
42
43. National University of Singapore
24th International World Wide Web Conference
Experiments
Evaluation Metrics
Saliency Metric S
• Noise Removal & Saliency Enhancement
S=122.86 S=53.49 S=21.18
“MillionMarchNYC”“Marching” Fused 43
44. National University of Singapore
24th International World Wide Web Conference
Experiments
• Saliency Enhancement
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Saliency Metric S
sensor image social image fused image
44
45. National University of Singapore
24th International World Wide Web Conference
Experiments
• More Events
Events Sensor
Cmage
Social
Cmage
Fused Sensor
Concept
Social
Term
Columbus Day
Parade
124.6 0.43 0.34 Crowd ColumbusDay
MillionMarchNYC
protest
124.5 0.47 0.40 People_marching MillionMarchNYC
StPatricks Day
Parade
1.49 0.61 0.53 Crowd StPatriks
45
46. National University of Singapore
24th International World Wide Web Conference
Experiments
• MSE compared with Ground Truth
(MillionMmarchNYC Events)
0
20
40
60
80
100
120
MSE
MSE (Fused Cmage – Ground Truth)
MSE(Sensor Cmage)
MSE(Social Cmage)
MSE(Fused Cmage)
46
47. National University of Singapore
24th International World Wide Web Conference
Experiments
• Effectiveness of Gaussian process
0
0.1
0.2
0.3
0.4
0.5
0.6
Fused with GP Fused without GP
S
47
48. National University of Singapore
24th International World Wide Web Conference
Summary
• Leveraged multimodal information for better
situation understanding
• Proposed an image-based hybrid fusion method
featuring sensor decision and spatial information
• Reduced noise in sensor data for better event
detection
• Limitation
Concepts to be fused are predefined
48
49. National University of Singapore
24th International World Wide Web Conference
Work III:
A Matrix Factorization Based Framework for
Fusion of Physical and Social Sensors
49
50. National University of Singapore
24th International World Wide Web Conference
Motivation
Physical sensors generate sparse
or inaccurate readings
Social Information implicitly
explains readings
• Help us discovery different dimensions or
aspects of events
• Useful for inferring & predicting ongoing
situations
• More than a single reading; tells why
Same events have similar social
topics and physical sensor readings
Goal: utilize physical & social
correlation to predict events
Crowdedness Prob
Hour
50
51. National University of Singapore
24th International World Wide Web Conference
Spatio-Temporal-Semantic Representation
Time
Stamps
Locations
1 40 80 120 150
3-4pm
1-2pm
2-3pm
12-1pm
4-5pm
5-6pm
“People_marching” Situation
51
52. National University of Singapore
24th International World Wide Web Conference
Approach: Matrix Factorization Based
Fusion Framework
≈ ×
52
53. National University of Singapore
24th International World Wide Web Conference
Formalization & Notations
26 ? 20 56 ?
? 102 80 ? 90
89 35 ? 21 35
14 ? 16 ? 109
Word 1, Word 2,
Word 3, word 4
…
Times
Physical
Readings 𝑆𝑖𝑗
𝛿
Locations
Locations
Time
Stamps
3-4pm
1-2pm
2-3pm
12-
1pm
4-5pm
5-6pm
1 40 80 120 150
N locations: 𝑗 = 1, … , 𝑁
M time stamps: i = 1, … , 𝑀
Temporal window: 𝛿
Situation matrix: 𝑆 𝛿
⊆ ℝ 𝑀×𝑁
Physical readings: 𝑆𝑖𝑗
𝛿
∈ 𝑆 𝛿
Word 1, Word 2,
Word 3, word 4
…
Word 1, Word 2,
Word 3, word 4
…
Word 1, Word 2,
Word 3, word 4
…
Word 1, Word 2,
Word 3, word 4
…
Location Document: 𝐿𝐷𝑗
𝛿,𝑟
= {𝑝1, … , 𝑝 𝑘}
Social post : 𝑝 𝑘 = 𝜔1, … , 𝜔 𝑅
Word: 𝜔 𝑞 ∈ 𝐷
Social Information: 𝕊 𝛿
= {𝐿𝐷1
𝛿,𝑟
, … , 𝐿𝐷 𝑁
𝛿,𝑟
}
Sensor Signal S :
Social Signal 𝕊 :
𝑝𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒 𝔽 𝑆, 𝕊 𝛿
> 𝑝𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒 𝑆
Fusion Function 𝔽 :
Matrix Factorization (MF) Model:
MF on Physical Signals: Basic Model
MF incorporating Social Signals: Latent Topics
𝐿𝐷 𝛿,𝑟
53
55. National University of Singapore
24th International World Wide Web Conference
MF Incorporating Social: Latent Topics
Object Function
min 𝑓 𝑆 Γ = 𝑆 𝑝ℎ𝑦=
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗
2
+ Ω𝑖,𝑗 Γ
55
56. National University of Singapore
24th International World Wide Web Conference
MF Incorporating Social: Latent Topics
Object Function
min 𝑓 𝑆 Γ = 𝑆 𝑝ℎ𝑦=
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗
2
+ Ω𝑖,𝑗 Γ
56
57. National University of Singapore
24th International World Wide Web Conference
MF Incorporating Social: Latent Topics
Object Function
min 𝑓 𝑆 Γ = 𝑆 𝑝ℎ𝑦=
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗
2
+ Ω𝑖,𝑗 Γ
min 𝑓 𝑆, 𝕊 Γ , 𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟(𝕊) = 𝑆 𝑝ℎ𝑦 + 𝑆𝑠𝑜𝑐
57
58. National University of Singapore
24th International World Wide Web Conference
MF Incorporating Social: Latent Topics
Object Function
min 𝑓 𝑆 Γ = 𝑆 𝑝ℎ𝑦=
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗
2
+ Ω𝑖,𝑗 Γ
Location document: 𝐿𝐷𝑗 = {𝑝1, … , 𝑝 𝑘}
Social post : 𝑝 𝑘 = 𝜔1, … , 𝜔 𝑅
Word: 𝜔 𝑞 ∈ 𝐷
Social Information: 𝕊 = {𝐿𝐷1, … , 𝐿𝐷1}
Social Signal 𝕊 :
min 𝑓 𝑆, 𝕊 Γ , 𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟(𝕊) = 𝑆 𝑝ℎ𝑦 + 𝑆𝑠𝑜𝑐
Locations
Time
Stamps3-
4pm
1-
2pm
2-
3pm
12-
1pm
4-
5pm
5-
6pm
1 40 80 120
150
58
59. National University of Singapore
24th International World Wide Web Conference
MF Incorporating Social: Latent Topics
LDA Model
𝑆𝑠𝑜𝑐 = − 𝑝 𝕊|𝜃, 𝜙, 𝑧 =
−
𝑗
𝑁
𝑞=1
𝑁 𝐿𝐷 𝑗
𝜃𝑧 𝐿𝐷 𝑗,𝑞
𝜙 𝑧 𝐿𝐷 𝑗,𝑞,𝜔 𝐿𝐷 𝑗,𝑞
min( 𝑓 𝑆, 𝕊 Γ , 𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟 𝕊 = 𝑆 𝑝ℎ𝑦 + 𝑆𝑠𝑜𝑐 )
t
l
s
θ
w
ф
N
M
|LD|
min 𝑓 𝑆, 𝕊 Γ, Θ, 𝑘, 𝑧 = 𝑆 𝑝ℎ𝑦 − 𝜆 𝑠𝑜𝑐𝑖𝑎𝑙 𝑝 𝕊|𝜃, 𝜙, 𝑘, 𝑧
=
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗
2
+ 𝜆 𝑟𝑒𝑔 𝑡𝑖
2
+ 𝑙𝑗
2
+ 𝛽𝑖
2
+ 𝛽𝑗
2
− 𝜆 𝑠𝑜𝑐𝑖𝑎𝑙
𝑗
𝑁
𝑞=1
𝑁 𝐿𝐷 𝑗
𝜃𝑧 𝐿𝐷 𝑗,𝑞
𝜙 𝑧 𝐿𝐷 𝑗,𝑞,𝜔 𝐿𝐷 𝑗,𝑞
spatial latent -> social topic
𝜃𝑗,𝑓 =
𝑒
𝑘𝑙 𝑗,𝑓
𝑓 𝑒
𝑘𝑙 𝑗,𝑓
Parameters Learning: Gradient Descent
59
60. National University of Singapore
24th International World Wide Web Conference
Situation Prediction for Missing Readings
“Cold-start” problem
𝑆 𝑝ℎ𝑦 =
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗 − 𝛽𝑖
2
+ 𝜆 𝑟𝑒𝑔 𝑡𝑖
2 + 𝑙𝑗
2
+ 𝛽𝑖
min(𝑓) =
𝑖,𝑗 ∈𝜅
𝑆𝑖𝑗 − 𝑠𝑖𝑗 − 𝛽𝑖
2
+ 𝜆 𝑟𝑒𝑔 𝑡𝑖
2 + 𝑙𝑗
2
+ 𝛽𝑖 − 𝑝 𝕊|𝜃, 𝜙, 𝑘, 𝑧
26 ? ? 56 ?
? 102 ? ? 90
89 35 ? 21 35
14 ? ? ? 109
Times
Locations
60
61. National University of Singapore
24th International World Wide Web Conference
Tweets,
FB,
Flickr
…
Matrix Factorization Based Fusion
26 ? 20 56 ?
? 102 80 ? 90
89 35 ? 21 35
14 ? 16 ? 109
1 12 2
3.6 2 2
1 -3 1
2 2 -3
-0.9 -0.4 5 3 1
10 1 7 -0.3 9
3 2 -9 6 12
= ×Times
Goal: minimize the error of predicted values & maximize likelihood of social observations
Tweets,
Youtube,
Flickr …
Tweets,
words,
Instagram
…
WeChat
News
Media
Flickr …
Tweets,
FB,
Flickr
…
Temporal Latent Factors Social Embedded
Latent Factors
Physical
Readings
𝑡𝑝1 𝑡𝑝2 𝑡𝑝3
𝑡,𝑙
(𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 𝑙𝑎𝑡𝑒𝑛𝑡
𝑝𝑎𝑟𝑎𝑚𝑠
𝑡, 𝑙 − 𝑃ℎ𝑦𝑠𝑖𝑐𝑎𝑙 𝑅𝑒𝑎𝑑𝑖𝑛𝑔𝑠𝑡,𝑙)2
− 𝜆 ∗ 𝑝𝑟𝑜𝑏(𝑆𝑜𝑐𝑖𝑎𝑙|𝑠𝑜𝑐𝑖𝑎𝑙 𝑝𝑎𝑟𝑎𝑚𝑠)
Physical Readings Error Social Observations Likelihood
𝑙𝑎𝑡𝑒𝑛𝑡, 𝑠𝑜𝑐𝑖𝑎𝑙
𝑝𝑎𝑟𝑎𝑚𝑠
min
Locations
61
62. National University of Singapore
24th International World Wide Web Conference
Situation Awareness – Singapore Haze
Noise Filtering – NYC Large Scale Events
Experiments :
62
63. National University of Singapore
24th International World Wide Web Conference
DataSet
• Historical PSI Readings
5 stations
3 weeks: 1st-7th, 12-19th August,
22nd-29th September
• Geo-tagged Tweets
Attributes: text, time, geo-
coordinates, etc
• 149 cameras all over Manhattan
Sampling rate: ~10s/f
Period: 2014 ~ Now
Size: > 2 TB
• Geo-tagged Tweets
Attributes: text, time, geo-
coordinates, etc.
Period: 2014 ~ Now
Size: ~60,000/day
SG Haze Data NYC Traffic Data
1
2
3
4
5
6
7
#tweets #words #words per location
SGHaze 19073 178825 8515
NYCTraffic 10005 90381 669 63
64. National University of Singapore
24th International World Wide Web Conference
Situation Awareness – Singapore Haze
Physical & Social Sensors Correlation
1
2
3
4
5
6
7
=
×
×
without
tweets
with
tweets
PSI Situation Matrix
1
14
28
42
56
1 2 3 4 5 6 7 8 9
1 2 3 4 5 6 7 8 9
Spearman’s rank correlation 𝜌𝑖𝑗
Location pairs: 1-2, 1-3, 2-1, …, 8-9
Ground truth: original PSI readings 𝜌𝑖𝑗
𝑝𝑠𝑖
With tweets 𝜌𝑖𝑗
𝑙+𝑡𝑤
vs without tweets 𝜌𝑖𝑗
𝑙
Evaluate: 𝑑𝑖𝑠𝑡 𝜌𝑖𝑗
𝑙+𝑡𝑤
, 𝜌𝑖𝑗
𝑝𝑠𝑖
, 𝑑𝑖𝑠𝑡 𝜌𝑖𝑗
𝑙
, 𝜌𝑖𝑗
𝑝𝑠𝑖
64
66. National University of Singapore
24th International World Wide Web Conference
Situation Awareness – Singapore Haze
Spatio-Temporal Situation Prediction
1
2
3
4
5
6
7
PSI Situation Matrix
1
1
14
28
42
56
52 3 4 76 1 52 3 4 76 1 52 3 4 76
Week 1 Week 2 Week 3
66
67. National University of Singapore
24th International World Wide Web Conference
Situation Awareness – Singapore Haze
Spatio-Temporal Situation Prediction
1
2
3
4
5
6
7
PSI Situation Matrix
1
1
14
28
42
56
52 3 4 76 1 52 3 4 76 1 52 3 4 76
Physical Only
LDA Topics:
1. hazy, hari, haze, gardens, uffc
2. Internationalcosplayday(icds), icds, sghaze, psi
3. iphone, airport, changi, terminal
Physical + Social
Week 1 Week 2 Week 3
67
68. National University of Singapore
24th International World Wide Web Conference
Situation Awareness – Singapore Haze
Spatio-Temporal Situation Prediction
1
2
3
4
5
6
7
PSI Situation Matrix
1
1
14
28
42
56
52 3 4 76 1 52 3 4 76 1 52 3 4 76
Physical Only
Physical + Social
Week 1 Week 2 Week 3
LOC1 LOC2 LOC3
51.38 49.45 57.73
42.48 26.80 22.86
Cross Validation Measured by MSE
No Tweets
With Tweets
68
69. National University of Singapore
24th International World Wide Web Conference
Event Classification
Events:
1.MillionMarchNYC Protest (M)
2.St Patrick’s Day Parade (S)
3.Columbus Day Parade (C)
Evaluation:
Root Mean Squared Error (RMSE)
69
70. National University of Singapore
24th International World Wide Web Conference
Event Classification Performance
Performance: Precision, Recall, 𝐹1-score
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 0.5 1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.5 1
0 0.5 1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Precision
𝐹1 Score
Recall 70
71. National University of Singapore
24th International World Wide Web Conference
Summary
• Social signals tells why for physical sensor readings
• Matrix factorization is used to fuse physical and
social information with spatial and temporal
aspects
• MF solves “cold-start” problem of predicting
missing readings
• Correlation exist between physical and social
signals that reflect events
• Fusing two sources has better performance in real-
world situations understanding than using only one
71
72. National University of Singapore
24th International World Wide Web Conference
Thesis Contributions
Proposed multilayer tweeting camera framework
can bridge the gap between physical and social
sensors, makes data analysis efficient, enable the
sematic details of occurring events
Cmage based fusion method removes noise
effectively, locates event accurately, and results a
better visualization of situations
MF based fusion results higher performance in event
classification and situation prediction with the help
of correlation between physical and social sensors
72
73. National University of Singapore
24th International World Wide Web Conference
Future work
Create an interactive framework extending
multilayer tweeting camera framework
Investigate semantic relatedness among concept
(utilizing ontology from various lexical databases,
e.g. WordNet), build up event ontology
Predict how event will evolve in temporal aspect
Apply fusion methods in different scenarios
(semantic, sentiment, stock analysis)
73
74. National University of Singapore
24th International World Wide Web Conference
Publication List
Journal paper
Yuhui Wang, Francesco Gelli, Christian von der Weth and Mohan
Kankanhalli, “A Matrix Factorization Based Framework for Fusion of
Physical and Social Sensors", in revision, IEEE Transactions on Multimedia,
2016. (In Peer Review)
Conference papers
1) Yuhui Wang and Mohan Kankanhalli, “Tweeting Cameras for Event
Detection”, 24th International Conference on World Wide Web (WWW’15),
pp. 1231-1241, Florence, Italy, May 2015.
2) Yuhui Wang, “Socializing Multimodal Sensors for Information Fusion”, 23rd
ACM international conference on Multimedia (MM’15),Doctoral Symposium.
653-656, Brisbane, Australia, October 2015. (Best Paper Award).
3) Yuhui Wang, Christian von der Weth, Thomas Winkler and Mohan
Kankanhalli, “Demo: Tweeting Camera - A New Paradigm of Eventbased
Smart Sensing Device", pp. 210-211, 10th International Conference on
Distributed Smart Cameras (ICDSC’16), Paris, France, September 2016.
4) Yuhui Wang, Yehong Zhang, Christian von der Weth, Kian Hsiang
Low, Vivek Singh and Mohan Kankanhalli, “Concept Based Fusion of
Multimodal Event Signals", IEEE International Symposium on Multimedia
(ISM’16), San Jose, USA, December 2016.
74
75. National University of Singapore
24th International World Wide Web Conference
Acknowledgement
Prof Mohan Kankanhalli
Prof Roger Zimmermann
Prof Qi Zhao
Prof Terence Sim
Prof Ramesh Jain
Prof Vivek Singh
Dr. Christian von der Weth
Dr. Prabhu Natarajan
Dr. Tian Gan
Dr. Yongkang Wong
Dr. Thomas Winkler
Lab mates in SeSaMe
75
Vintage camera early 19s , digital CCTV camera mi19
Gradual evolution digital eyes distributed, passively
Traditional cctv camera system , infront of the wall , look into footage , what’s happening, Event camera capture, human interpret
Why not let camera actively send information,
You know human , good at tweet, social media, smart , what worth to tweet
Iot era, smart nation, understand situation, not only be human tweets, but also camera tweets.
New tweeting camera paradigm, socially connected, configurable applications, for example infer meeting going by cheking lighting
Know people by dectecting face
we assume the sensor readings over an area to be realized from, Gaussian process incorporates noise model and allows the spatial correlation of sensor readings (sensor pixels) to beformally characterized in terms of their locations
<- why matrix factorization is good ->
We can imagine
In one hand , Continuously coming
In the other hand , Social Streaming to
------------------
Fuse them togeter
This can be cast as matrix completion on apartially observed matrix of users’ ratings
This can be cast as matrix completion on apartially observed matrix of users’ ratings
Maximaize the Likelyhood of observing the words give this parameters
Optimize theta by so that opitmized l_j,f
The colums where the haze happens are more similar if social information is corporated
nonparametric measure of statistical dependence between two variables.
False alarm LDA just check occurrence
False alarm LDA just check occurrence
with tweets: Precision goes up faster to 1
Recall goes down slower to 0