3rd Ph.D.
Review
Tracking trends on the web using novel Machine Learning
methods
Vasileios Lampos
advised by
Professor Nello Cristianini
Intelligent Systems Laboratory, University of Bristol
May 25, 2010
Vasileios Lampos 3rd Ph.D. Review
Vasileios Lampos 3rd Ph.D. Review
Aims
The general aims of our research project can be summarised in
the following points:
1 Track trends on the Web by applying Machine Learning
methods (track expresses the notions of infer or predict as
well)
2 Extend current or invent new methodologies (where and if
needed) for accomplishing our primary aim
3 Build tools that apply the experimental/theoretical results in
real and large-scale applications (featured research)
Vasileios Lampos 3rd Ph.D. Review
Being more specific
1 Trends about what? Examples?
Predict flu rates (epidemics)
Infer vote intensions (politics)
Infer traffic/weather conditions (toy problems)
Vasileios Lampos 3rd Ph.D. Review
Being more specific
1 Trends about what? Examples?
Predict flu rates (epidemics)
Infer vote intensions (politics)
Infer traffic/weather conditions (toy problems)
2 Methodologies?
Feature extraction/selection
Exploit probabilistic relationships (PGMs)
Regression/classification/ranking scenarios
Active learning
Vasileios Lampos 3rd Ph.D. Review
Being more specific
1 Trends about what? Examples?
Predict flu rates (epidemics)
Infer vote intensions (politics)
Infer traffic/weather conditions (toy problems)
2 Methodologies?
Feature extraction/selection
Exploit probabilistic relationships (PGMs)
Regression/classification/ranking scenarios
Active learning
3 Applications?
Back-end infrastructure for data collection/retrieval/mining
Real time online tools for making and displaying predictions
(like the Flu detector)
Vasileios Lampos 3rd Ph.D. Review
P1 - Summary (1 of 3)
Title: Tracking the flu pandemic by monitoring the Social Web
Authors: V. Lampos and N. Cristianini
Submitted to: IAPR Cognitive Information Processing 2010 (accepted)
Twitter and Health Protection Agency data for weeks 26-49,
2009 (on average 160,000 tweets collected per day geolocated
in 54 urban centres in the UK)
Frequency of 41 flu related words (markers) in Twitter
corpus had a correlation of >80% with the HPA flu rates in
all UK regions
Learn a better list of weighted markers automatically:
Generate a list of candidate markers (1560 words taken from
flu related web pages)
Use LASSO for feature selection
Vasileios Lampos 3rd Ph.D. Review
P1 - Summary (2 of 3)
Validation schemes:
1 Train on one region, validate regularisation parameter on
another, test on the remaining regions (for all possible
combinations)
Train/Validate (regions) A B C D E
A - 0.9594 0.9375 0.9348 0.9297
B 0.9455 - 0.9476 0.9267 0.9003
C 0.9154 0.9513 - 0.8188 0.908
D 0.9463 0.9459 0.9424 - 0.9337
E 0.8798 0.9506 0.9455 0.8935 -
Total Avg. 0.9256
97 selected words : lung, unwel, temperatur, like, headach, season, unusu, chronic, child,
dai, appetit, stai, symptom, spread, diarrhoea, start, muscl, weaken, immun, feel, liver, plenti, antivir,
follow, sore, peopl, nation, small, pandem, pregnant, thermomet, bed, loss, heart, mention, condit, ...
2 Aggregate data from all regions, test on weeks 28 and 41
(2009) and train using the rest of the data set
Vasileios Lampos 3rd Ph.D. Review
P1 - Summary (3 of 3)
1 Inferred vs Official flu rate in North England
HPA
100 Inferred
Flu rate
50
0
180 200 220 240 260 280 300 320 340
Day Number (2009)
2 Inferred vs Official rates in all regions (aggregated data set)
150
HPA
Inferred
100
Flu rate
A B C D E
50
0
0 10 20 30 40 50 60 70 80 90
Days
Vasileios Lampos 3rd Ph.D. Review
P2 - Summary
Title: Flu detector - Tracking epidemics on Twitter
Authors: V. Lampos, T. De Bie, and N. Cristianini
Submitted to: ECML PKDD 2010 Demos (under review)
Extending and making more robust the methodology of P1
Larger data sets (bigger time series) and more (2675)
candidate features
Select a list of features (markers) using BoLASSO (bootstrap
version of LASSO)
Then learn weights of those markers via linear least squares
regression
Stricter evaluation of the methodology - Available online
Put all this into practice and come up with the Flu detector
Vasileios Lampos 3rd Ph.D. Review
Other activities
Studied/implemented the necessary statistical tools and
algorithms (in MATLAB or Java)
Extended further the infrastructure for conducting large scale
experiments and data retrieval on demand
TA for Intro to AI, Data Analysis and Pattern Analysis &
Statistical Learning
Attended some of the ISL meetings and seminars
Vasileios Lampos 3rd Ph.D. Review
General goals & activities
1 ... (content omitted)
2 ... (content omitted)
3 ... (content omitted)
4 ... (content omitted)
Vasileios Lampos 3rd Ph.D. Review
A more time specific tentative plan
In June: ... (content omitted)
In July: ... (content omitted)
In August: ... (content omitted)
In September - November: ... (content omitted)
Vasileios Lampos 3rd Ph.D. Review
This is the last slide.
Vasileios Lampos 3rd Ph.D. Review
This is the last slide.
Any questions?
Vasileios Lampos 3rd Ph.D. Review
