100 likes | 245 Views
WeKnowIt Emerging, Collective Intelligence for personal, organisational and social use http://www.weknowit.eu Event Detection Processing and Representation Advances, Future Applications, Challenges Yiannis Kompatsiaris CERTH-ITI. Comms. Low- level. Favs. Caption.
E N D
WeKnowItEmerging, Collective Intelligence for personal, organisational and social use http://www.weknowit.eu Event Detection Processing and RepresentationAdvances, Future Applications, Challenges Yiannis Kompatsiaris CERTH-ITI
Comms Low- level Favs Caption Event Processing in User Generated Content / Social Media / Web 2.0 Tags Time Geo Groups Social network User Profile
Event Detection Research approaches • Community Detection (Graph-based) • Image clusters based on finding tag-image communities in social network • Graph-based, fast and scalable community detection approach • Time aware user-tag co-clustering • Co-clustering based • Detects on the same time topics and users relevant to the event • LDA probabilistic approach • generalization of Latent Dirichlet Allocation (LDA) approach • Events are indicated by unusual content or annotation that is localized in space and time
Results and Applications • User-Genrated maps of Points of Interests • Where there is (was) something interesting happeningdemo: www.clusttour.gr • Name events by most important tags
Time-aware user-tag co-clustering User 1 User 2 User 3 fashionweek, fashion, silk, wool Accessories, bags, fashion, hats, Gucci fashion, jeans, NY New York, hat, trousers, fashion, Gucci Cars, football, holidays, horses, sea, turkey, fashion animals, elephants, nature sea, turkey, bags
Research for upcoming Events • Bursts detection in networks of tag co-occurences • Event is an emerging context tag cluster • Detect building-up events by updating tag connectivity strenght from user input stream • Monitor “hot topics” related to specified tags Challenges • System response must by within seconds • Fast updates on large scale graph • Alarm triggered when event reaches threshold • Monitor emerging clusters
Representation – Event Model F • Based on the foundational ontology DOLCE+DnS Ultralight (DUL) - OWL • Representation for time and space, objects and persons • Mereological, causal and correlative relationships between events • Provides flexible means for • event composition • modeling event causality and event correlation • representing different interpretations of the same event. • Available from: • http://west.uni-koblenz.de/eventmodel/
Events Representation - Applications • Monitoring/mergingevent log files • Explore and visualize largesemantically heterogeneousdistributed semantic datasetsin real-time.
Challenges • Granularity of event recognition – trade-off • Few, large, better quality events (e.g. fairs, concerts) • Lots, smaller, noisy events (e.g. birthday parties) • Event naming • Can localize event and display relevant tags, but not always assign simple name (as person would do) • Sparsity of user data • Need large number of geo-localized, timestamped and tagged resources (images) for certain location (e.g.) city and longer time (few years) • Representation • Generating APIs for pattern-based ontologies • Reasoning • Adaptation to domain-specific requirements
Thank you! WeKnowIt http://www.weknowit.eu Yiannis Kompatsiaris http://mklab.iti.gr