The document analyzes the effectiveness of the 1% Twitter data sample for spatio-temporal analysis tasks compared to a 10% sample, revealing that while both samples share similar properties, the larger sample yields more reliable data for less popular retweets. Experiments conducted on geo-location coverage, sentiment analysis, popular topic detection, and graph evolution show varying degrees of reliability based on the sample size. The findings indicate that the default 1% sample may not be sufficient for all analytical tasks, particularly in detailed explorations of less prevalent topics.

1. Mining Twitter Data with Resource Constraints
Geoge Valkanas, Ioannis Katakis,
Dimitrios Gunopulos, Anthony Stefanidis
August 12, 2015
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 1 / 18

2. Research Question
Is the 1% sample provided by the Twitter API sucient for
spatio-temporal analysis tasks? ... which tasks?
! We compare with the 10% sample (Garden Hose)
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 2 / 18

3. Outline
1 Problem and Motivation
2 Data Collection
3 Experiments in Various Tasks
Geo-location Coverage
Sentiment Analysis
Popular Topic Detection
Graph Evolution
4 Conclusions
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 3 / 18

4. Introduction
Twitter Samples
Two ways to access the stream
Public Stream: 1% Sample
Garden Hose: 10% Sample
... in both cases, we don't know details about the sampling method.
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 4 / 18

5. Introduction
Constraints
Financial cost
Licences of larger samples, are costly and dicult to obtain.
Computational cost
7 Giga Bytes per minute
O the shelf approaches are unable to operate in such settings
In practice: those who engage in social media analytical tasks have
practically no choice but to resort to the downsized information. However,
being only a small fraction of the entire stream, it is unclear how reliable
this information is for each type of application.
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 5 / 18

6. Introduction
A more concrete example
The INSIGHT Project: Improve understanding, prediction and warning of
emergencies through real-time processing of data streams including social
data.
(a) Floods in Germany (2013) (b) Control Center in Dublin CC
How much data are ecient for our task?
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 6 / 18

7. Introduction
Tasks we look into...
Sentiment Analysis
Geo-located information
Popular tweets
Social Graph Evolution
Linguistic Analysis
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 7 / 18

8. Data
The data
10M
1M
100K
Default Gardenhose
0 20 40 60 80 100
Tweet Count
Hours
(c) All tweets
100K
10K
1K
Default Gardenhose
0 20 40 60 80 100
GPS Tweet Count
Hours
(d) GPS-tagged tweets
Figure : Comparing default and gardenhose samples for volume over time
4 day period - November 2013
The two samples dier by an order of magnitude
Exhibit the same temporal pattern
Geotagged tweets are between 1-2% of their respective sampled data
Geotagged are more
attened out
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 8 / 18

9. Experiments
Geo-location coverage - Experiment 1
Bounding Box
Twitter also allows its users to ask for geotagged information.
The user provides a bounding box, by specifying 4 coordinates in the
form [(latmin; lonmin)(latmax ; lonmax )], and Twitter returns tweets that
fall within this region.
25
0
−25
−50
60 90 120 150
lon
lat
. In this particular case, where geotagged tweets are asked for instead of a
general sample, the volume of the returned results is the same for the two
samples!.
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 9 / 18

10. Experiments
Geo-location coverage - Experiment 2
4 dierent crawls in London area
Loc1 Loc2 Loc3 Loc4
1400
1200
1000
800
600
400
200
0
0 5 10 15 20 25 30 35 40 45
Count
Half-Hour Interval
. As the overlap increases between the bounding boxes, so does the
similarity between two dierent crawls.
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 10 / 18

11. Experiments
Sentiment Analysis
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
Sample 1% Sample10%
0 20 40 60 80 100
Ratio
Hour
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
Sample 1% Sample10%
0 20 40 60 80 100
Ratio
Hour
Positive and Negative Sentiment Ratio
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0 20 40 60 80 100
Ratio
Hours
Pos 1%
Neg 1%
Pos 10%
Neg 10%
- Dictionary based
sentiment analysis
- Ratio of tweets is
the same in both
samples
- Ratios in geo-tagged
tweets are lower,
meaning that
geottagged tweets
oer less
sentiment-oriented
information
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 11 / 18

12. Experiments
Popular Topic Detection - Experiment
1 Extract the top-k most retweeted posts, that appear in our data
(both samples).
2 Compare the two lists (Kendall Correlation)
3 Compare the two lists with the ground truth (= actual retweet count
information included in the tweet)
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 12 / 18

13. Experiments
Popular Topic Detection - Results
1
0.98
0.96
0.94
0.92
0.9
0.88
0.86
10 100 1000 10000
Kendall Correl.
List Items
S1-S10
S1-S10P1
S1-S10P2
S10P1-S10P2
S1-S1P1
(a) Kendall
1
0.99
0.98
0.97
0.96
0.95
0.94
10 100 1000 10000
Common Items (%)
List Items
S1-S10
S1-S10P1
S1-S10P2
S10P1-S10P2
S1-S1P1
(b) Common Items
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Sample 1% Sample 10%
1
5
10
100
500
1000
2500
5000
7500
10000
Kendall Correl.
Iteration
(c) Vs the ground truth
Figure : Comparing the top-N most retweeted items
Conclusions
For up to 10 items, 1% is adequate. That is not however the case for
list with more than 1000 items.
Comparison with Ground Truth: 10% has higher correlation.
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 13 / 18

14. Experiments
Graph Evolution Study - Experiment
Study the re-tweet graph (directed)
Edges are weighted (more re-tweets ! larger weight) and decay over
time
Edges are removed when their weight drops below a certain threshold
Method 1: Iter At each time interval extract a new graph
Method 2: Glb At each time interval aggregate the new nodes to the
current graph
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 14 / 18

15. Experiments
Results
300000
250000
200000
150000
100000
50000
0
0 200 400 600 800 1000 1200
Value
Iteration
Iter 1%
Glb 1%
Iter 10%
Glb 10%
(a) Size
100
90
80
70
60
50
40
30
20
10
0
Glb 1% Glb 10%
0 200 400 600 800 1000 1200
Value
Iteration
(b) Lar. Con. Comp. Size
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0 200 400 600 800 1000 1200
Value
Iteration
Iter 1%
Glb 1%
Iter 10%
Glb 10%
(c) Clustering Coecient
Figure : Statistical properties of the extracted retweet graph, over time
Conclusions
No signi

16. cant dierences between the two samples
LCC does not follow the 24-hour pattern
Clustering coecient of 10% similar 100%
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 15 / 18

17. Experiments
More on the paper...
Retweet Burstiness
The rate at which users retweet information plays an important role
in capturing trending topics
We investigate wether there is a dierence between the rates of
receiving retweets in both samples
Linguistic Analysis
Is there a correlation between the spoken languages in Twitter, and
the ground truth obtained from studies in the physical world?
What are the dierences between the two samples in this context?
We use language detection tools and ground truth information from
Wikipedia.
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 16 / 18

18. Summary and Conclusions
Conclusions
Research question: Is the default sample sucient? For which tasks?
Focused on spatio-temporal tasks
We compared 1% with 10% sample
The samples have quite similar properties
However when you get into the details (less popular re-tweets) the
bigger sample is better
Valkanas et al (UoA and GMU) Mining Twitter with Resource Constraints August 12, 2015 17 / 18