In this paper, we present the methods for event clustering and classi cation de ned by MediaEval 2013. For event clustering, the watershed-based method with external data sources is used. Based on two main observations, the whole metadata is turned into a user-time (UT) image, so that each row of an image contains all records that belong to one user; and the records are sorted by time. For event classi cation, we use supervised machine learning and experiment with Support Vector Machines. We present a composite kernel to jointly learn between text and visual features. The methods prove robustness with F-measure up to 98% in challenge 1, and the composite kernel yields competitive performance across di erent event types in challenge 2.
H.3 [Information Storage and Retrieval]: Information Search and Retrieval
This paper describes the social event detection method that
is specially built to meet challenge 1 and 2 of MediaEval
2013 [
The method proposed for tackling the challenge 1 of SED
2013 is mostly inherited from the method introduced in [
The UT-image (user-time image) is a 2D data structure
where row ith contains all time-ordered data related to user
ith (i.e. time taken of UT-image(i, j) is smaller or equal
to time taken of UT-image(i, j+1)) (for more details please
refer to [
Data: data that need to be clustered Result: set of clusters 1. Translate the original data into UT-image format. 2. For each row ith of UT-image do (#Splitting stage) 2.1 repeat 2.1.1 Split data at column jth if jtime-taken-of-UT-image(i, j) time-taken-of-UT-image(i, j + 1)j <= time threshold. until cannot split anymore; 3. repeat 3.1. For each cluster, create time-taken boundary (e.g. [time-start, time-end]), and a union set of not null (longitude, latitude), tags, titles, and descriptions, respectively. 3.2. For any pair of clusters do MERGING if the following conditions are hold - time-taken boundary intersection does NOT EMPTY or di er not MORE THAN time threshold - distance di erence between two sets of not null (longitude, latitude) is SMALLER than distance threshold - Jaccard index of two sets of tags/title/description is LARGER than tag threshold until cannot merge anymore; 4. End
In order to increase the accuracy of merging stage, \common sense" is taken into account to nd the most \common pattern" in the tags eld (e.g. the most common word users tend to tag for the same event). Here, TF-IDF method is applied on tags of each cluster to extract the most common keywords. These keywords are used as the main clue to merge clusters. The common sense merging procedure is described in algorithm 2.
We used algorithm 1 for run 1, algorithm 1 with di erent parameters for run 2, and both algorithm 1 and 2 for run 3. Data: set of clusters generated by algorithm 1 Result: set of new clusters 1. For each cluster, process TF-IDF on tags set and select the most common keywords to create a \new common sense tags" set. 2. For each row ith of UT-image do (#Splitting stage) 2.1 repeat 2.1. For any two clusters, MERGING if Jaccard index of two \new common sense tags" sets is LARGER then tag threshold. 2.2. Process TF-IDF on \new common sense tags" set, select the most common keywords, and update this set. until cannot merge anymore; 3. End
To tackle the task event class cation in challenge 2, we use supervised machine learning. We experiment with Support Vector Machines, and design a composite kernel to jointly learn between text and visual features.
The data is processed using GATE platform1 for tokenization, POS tagging and basic word features. We used Support Vector Machines to train and test our binary classi er. Here, event classi cation is formulated as a multiclass classi cation problem. The One Vs. Rest strategy is employed by selecting the instance with largest margin as the nal answer. For experimentation, we use 5-fold cross-validation with the svm-light tool2. The feature set for our learning framework is described as follow.
1. wi is text of the title, description, or the tag in each event
2. li is the word wi in lower-case
3. p1i, p2i, p3i, p4i are the four pre xes of wi
4. s1i, s2i, s3i, s4i are the four su xes of wi
5. fi is the part-of-speech of wi
6. gi is the orthographic feature that test whether a word contains
all upper-cased, initial letter upper-cased, all lower-cased.
7. ki is the word form feature that test whether a token is a word,
a number, a symbol, a punctuation mark.
8. oi is the ontological features. We used the ontology and
knowledge base developed in [
Run 1 was done without external resources, i.e., ontological features whereas all the features were used in run 2.
For run 3, the image feature extraction was performed in
a similar manner as in [
In run 5, we used a composite kernel to combine between text and visual features CK = KT + (1 ) KV where is a coe cient, KT and KV is either the kernel applied to text or visual features. We experimented with = 0:5. 4.
The results are reported in tables 1 and 2. In general, the proposed method proves as very competitive whereas there is still room for improvement. With challenge 1, the watershed-based works well with the results being up to 98%. For challenge 2, the classi cation event vs. non-event is acceptable in almost every run, as well as the detection of some classes. In the last run, with the composite kernel to combine between text and visual features, we have 5 classes out of 9 above 55%.
Obviously, we have followed the supervised machine learning for challenge 2, so it could not be learnt e ciently with only 36 positive instances of the class \fashion", it may be better if we used rule-based instead. Moreover, it is not trivial to provide a good detection on the class \other events", which is a rather unde ned class. The combination did the best with class \theater dance". Meanwhile, we also observe that the class \exhibition" has 272 positive instances but could not be learnt with any kind of features and should be studied in more detail.