Transductive Event Classification through Heterogeneous Networks

Events can be defined as "something that occurs at specific place and time associated with some specific actions". In general, events extracted from news articles and social networks are used to map the information from web to the various phenomena that occur in our physical world. One of the main steps to perform this relationship is the use of machine learning algorithms for event classification, which has received great attention in the web document engineering field in recent years. Traditional machine learning algorithms are based on vector space model representations and supervised classification. However, events are composed of multiple representations such as textual data, temporal information, geographic location and other types of metadata. All these representations are poorly represented together in a vector space model. Moreover, supervised classification requires the labeling of a significant sample of events to construct a training set for learning process, thereby hampering the practical application of event classification. In this paper, we propose a method called TECHN (Transductive Event Classification through Heterogeneous Networks), which considers event metadata as different objects in an heterogeneous network. Besides, the TECHN method has the ability to automatically learn which types of network objects (event metadata) are most efficient in the classification task. In addition, our TECHN method is based on a transductive classification that considers both labeled events and a vast amount of unlabeled events. The experimental results show that TECHN method obtains promising results, especially when we consider different weights of importance for each type of event metadata and a small set of labeled events.