Event Prediction Based on Unsupervised Graph-Based Rank-Fusion Models

This paper introduces an unsupervised graph-based rank aggregation approach for event prediction. The solution is based on the encoding of multiple ranks of a query, defined according to different criteria, into a graph. Later, we embed the generated graph into a feature space, creating fusion vectors. These vectors are then used to train a predictor to determine if an input (even multimodal) object refers to an event or not. Performed experiments in the context of the flooding detection task of the MediaEval 2017 shows that the proposed solution is highly effective for different detection scenarios involving textual, visual, and multimodal features, yielding better detection results than several state-of-the-art methods.