lowering user burden in mobile crowdsourcing through compressive sensing
TRANSCRIPT
More with Less:Lowering User Burden in Mobile Crowdsourcing through Compressive Sensing
Presenter: Harshitha
Chidanan
da
Liwen Xu , Xiaohong Hao , Nicholas D. Lane , Xin
Liu , Thomas Moscibroda – UbiComp 2015
CS290F: Smartphone centric systems and applications
o Introductiono Main problemo Definitiono Main pointso Contributionso Other Solutions
o Challengeso Key Technical pointso Evaluationo Strengths and
weaknesseso Open Issueso Thoughts and summary
Table Of Contents
Analyzing is criticalAnalysis has opened up new domains
Main problem
Making analysis requires large amount of data High burden placed on the user
Users won’t join the system Users stop participating eventually
Definition
Mobile crowdsourcing is a promising way to collect largescale data about ourselves and the urban areas we live in
Sensor data from environment(noise) User provided information
Main points and contributions
Present compressive crowdsensing (CCS) – a framework that enables compressive sensing techniques to be applied to mobile crowdsourcing scenarios.
Reduced amounts of manually collected data Acceptable levels of overall accuracy
First time CS has been applied to mobile crowdsourcing
CS has the ability to utilize inherent structure that may not be obvious /natural ways of considering the data
Other solutions
Already known approaches : General-purpose statistical methods
Sub-sampling Interpolation
Domain-specific techniques Population surveying Geospatial
All of these methods presuppose, and then leverage, certain relationships within the collected data.
CS has the ability to utilize inherent structure that may not be obvious and may not correspond to more “natural” ways of considering the data
Conventional CS Deal with 1-D vectors Easily vectorized data
Crowdsourcing datasets
Multiple columns Multi-dimensional
Unknown as to how to apply CS to crowdsourced dataNovel processing steps need to be introduced: Complex data correlation preserve the important ones base training handle missing data
Challenges
Key Technical Points
Demonstrate the feasibility of applying CS to large-scale question-based user surveys
Propose a technique to use the data that do not have obvious representations
Evaluate compressive crowdsourcing by applying real-world datasets
COMPRESSIVE SENSING PRIMERCompressive Sensing (CS), an efficient technique of sampling data with an underlying sparse structure
Example
Traffic speeds Speeds at intersections Reduces sampling rate
-Sparse Structure-Random Sampling-Data Reconstruction-Base Learning-Stages of CS
Sparse structure
Signal of interest x ->coefficient vector
Names kx is called k-sparse k non-zero entriesy is called compressible small
< - - 5-sparse
Sampling is the reduction of a continuous signal to a discrete signal
Most popular CS sampling methods
Random Sampling
Random Sampling
To Capture signal y• m samples• n entries• m<<n
25 random readings ~ 100 readings
Data reconstruction
Linear interpolation
Linear interpolation
Compressive sensing
Base Learning
Base Ψ plays a critical role in transforming the signal of interest y to a sparse signal x
Sparsifying Base can be:
Standard : fourier base Discrete cosine transform Good base: trained using historical data
Stages of CS
Count rat sightings in different areas of a city within a particular period of time. • Use a set of
historical data {y1, y2, · · · , yN}
• Select a small random sample of areas
• Recover the data
Stages
COMPRESSIVE CROWDSENSINGA framework that enables compressive sensing techniques to be applied to mobile crowdsourcing scenarios.
Crowdsourced data
A global view of a phenomenon
Data through sensors
Manually entered data
Step 1: Data structure conversion
Step 2: Base training
Uncovers the inherent correlation in the data structure
K-SVD algorithm is used
Given data Y, finds base:1)Represents each yi=Ψxi
2)Minimizes total error
Step 3: Sampling
Passively Users provide data
when they wish Contributed data is
grouped Sampling group Training group
Only data from users who have information is used during reconstruction
Pro-actively Selects randomly within
the range of sampling values
Users with characteristics are directly asked to provide data
If such characteristics about the user are not known then all users are asked.
Step 4: Reconstruction
Arranging a matrix representation according to:• Training• Sampling
Matrix is projected into the trained base to recover a sparse representation of the target
Missing target values is recovered by multiplying base with the recovered sparse representation
Evaluation
Methodology: diverse group of real life datasets perform the same random sampling
Evaluation Metric: Data in vector format Format
Comparison Baselines Conventional CS Linear Interpolation Spline Interpolation Kriging Interpolation Sampling Only
Conventional CS
Conventional CS-exact same CS stages as CCS
Linear Interpolation
• A method of curve fitting using linear polynomials.
Spline interpolation
Curve fitting is performed but this time using piecewise cubic splines.
Kriging Interpolation
A well-known method for geographical interpolation that is popular in the GIS community
Sensing data can be represented more accurately with more coefficients and their corresponding bases
Strengths and Weaknesses
Strengths
First to use compressive sensing for crowdsourcing
Succeeds Good results Diversity of datasets
Temporal Spatial Demographic
Weaknesses
If original data shows no correlation, CS would not apply.
Overall vector space may be very large.
Within broader CS research, how to predict dataset performance is unknown
Open issues/ Directions for future scope
Analysis across different datasets with different crowd-based scenarios
Analysis out of limited correlated data. Target a variety of application domains and monitor:
Traffic conditions Place categories Noise pollution Wifi conditions Happiness
Thoughts and summary
Thoughts Very very hard to
understand Required lot of
background reading Needed better term
explanation Less graphical
representation Good results
Summary Recovers large-scale
urban information Two fold
Demonstrates novel crowd-based applications of compressive sensing
Develops key new techniques that allow CS to be generically applied to many scenarios
Thank You!QUESTIONS?