This document describes the EventSense framework for analyzing large-scale events by mining social media streams. The framework extracts entities of interest from messages using TF-IDF vectors and cosine similarity. It detects topics in event messages using clustering and sentiment analysis using naive Bayes classifiers. The framework was applied to a film festival, extracting 834 topics from tweets and analyzing sentiment with 75% accuracy for English. Aggregation provided insights like more popular films having higher ratings and more tweets/bookmarks. Future work includes applying the framework to more events and data sources.
Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
EventSense: Capturing the Pulse of Large-scale Events by Mining Social Media Streams
1. #1
EventSense: Capturing the Pulse of Large-scale
Events by Mining Social Media Streams
Case study: Thessaloniki International Film Festival
E. Schinas, S. Papadopoulos, S. Diplaris, Y. Kompatsiaris,
Y. Mass, J. Herzig, L. Boudakidis
2. #2
Capturing & mining large-scale events
• Large-scale events attended by
thousands of people captured by
mobile devices in the form of status
updates, photos, ratings, etc.
• SXSW Music, Film and Interactive
Conferences and Festivals
– 30000+ attendees
– ~300,000 tweets between Mar 3 and 7
– 40,247 tweets even the last month
• Sundance film festival
– 200 films, 10 days, 50,000+ attendees
– 200,000+ tweets during the festival
– 20,438 tweets even the last month
A search for #tiff53 in
twitter returns an
unstructured list of tweets
3. #3
Capturing & mining large-scale events
• The online representation of an event as a sequential list of
posts and status updates is ineffective
• A more effective means of event representation would
employ facets, such as entities, sub-events and sentiment.
• Challenge:
– Organize information around
– entities of interest
– Extract meaningful
insights, obtain informative
summaries
• EventSense framework
4. #4
Entity Detection (1/3)
• Entities are defined as lists of properties:
– a film consists of a title, description, names of
director(s)/actors
• Matching status updates (tweets) to entities relies on
representing both as tf * idf vectors
m: message (tweet), f: feature (term), M: set of all event messages
boost(f): boosting factor when f is a named entity
6. #6
Entity Detection (3/3)
Entity of interest
1. Select a combination of properties
e.g.title, director and actors
2. Aggregate selected properties to a
single string «Μούχλα Αλί Αιντίν»
3. Calculate tf * idf vector of n-grams
using the same vocabulary with tweets
4. Calculate cosine similarity between an
incoming message and the set of all
entities of interest.
5. Assign message to the entities that
similarity exceeds a predefined
threshold.
7. #7
Topic analysis
• 1 NN clustering algorithm to create clusters/topics
Assign an incoming message to the nearest topic, if cosine similarity
exceeds a predefined threshold. Else create a new topic.
• Similarity threshold sensitivity analysis similar to entity extraction
• LSH approximation to scale up (Petrovich at al., NAACL 2010)
hash the input items so that similar items are mapped to the same
buckets with high probability. Reduce search only to this bucket.
• Title Extraction per Topic
For the set of the items of a topic we find the largest sequence of words
with the highest frequency.
8. #8
Sentiment Analysis
• Training using tweets with emoticons. E.g. positive, negative
(A. Go, R. Bhayani, and L. Huang)
• For each message we extract two types of features. The first is n-grams.
The second includes the existence of user mentions and URLs,
punctuation, repeated letters
• Naive Bayes (NB) classifier for positive and negative data. Assuming a
uniform prior for all classes, independence between features, and using
the Bayes rule we get:
9. #9
Aggregation & summarization
• For each entity we retrieve the set of associated messages
and calculate the mean value of sentiment, Polarity and
Subjectivity
• Calculate the same sentiment measures per topic and per user
• Several other statistics: top shared messages, URLs and images,
top active & influential users
10. #10
Dataset: 53rd
Thessaloniki International Film Festival
Three sources of data
1. A detailed set of the 168 films included in the official festival
program of tiff53
2. 3,974 tweets that contain the official hashtag of the festival
(#tiff53) for the period between November 1st
and 13th
3. Film rating and bookmarking data created by the ThessFest
mobile app (available both for iPhone* and Android**).
* https://itunes.apple.com/gr/app/thessfest/id504913309?mt=8
** https://play.google.com/store/apps/details?id=com.mk4droid.FF_pack&hl=el
10 days long event
2-11 November 2012
12. #12
Topic analysis results
• 834 topics (clusters)
• Manual inspection
of topics:
– 53.8% of topic titles
considered
informative
– 98.5% of topics
were found to be
“clean”
Topics in time
Top-10
13. #13
Sentiment analysis results
• Training
– 800K positive & negative tweets for English
– 12K positive & negative tweets for Greek
• Tuning (for threshold)
– Manually annotated dataset from Thessaloniki Documentary Festival
(similar event)
– 325/73/553 in English and 781/216/781 in Greek
• Testing
– 324/33/724 in English and 901/315/1667 in Greek
– Best accuracy (English) ~ 0.75
– Performance in Greek much poorer
compared to English
need for richer training corpus
pos neg neut
14. #14
Aggregation & summarization results (1/2)
#T: number of tweets
Pol: polarity of film tweets
Subj: subjectivity of film tweets
R: average rating
#R: number of ratings
#F: number of times the film was bookmarked
• Films with positive polarity are rated higher.
• Films that are tweeted a lot are also more
likely to be rated.
• Films that are tweet a lot are also more
likely to be added to the users’ bookmarks.
Pearson correlation across film statistics
15. #15
Aggregation & summarization results (2/2)
Most active & influential Twitter
accounts (+sentiment per user)
Most shared photos
(+number of retweets)
16. #16
Summary
• Extract entities of interest from messages
F1 = 0.737 (precision = 0.774, recall = 0.697)
• Detect topics in event related messages
834 topics, 98.5% considered “clean”
• Sentiment analysis per messages, entities & topics
Accuracy: 0.75 for English, 0.62 for Greek
• Aggregation & statistics
Valuable insights and overview information
17. #17
Future Work
• Apply the proposed framework to larger-scale events
of different nature (e.g. music festivals, sports
events).
• Monitoring and processing more OSN sources (e.g.
Facebook, Instagram).
• Refine the proposed methods with the goal of
improving accuracy and robustness over different
datasets.
• Experiment with techniques for automatically
creating visual informative summaries based on the
results of the automatic analysis.
19. #19
References
1. Petrovic S., Osborne M., Lavrenko V. (2010) Streaming first story
detection with application to Twitter. Human Language Technologies:
The 2010 Annual Conference of the North American Chapter of the ACL
(NAACL)
2. A. Go, R. Bhayani, and L. Huang. Twitter sentiment classification using
distant supervision. 2009.