Real-world-events data sifting through ultra-small labeled datasets and graph fusion

Abstract

The information on social media is vital, especially for events such as natural disasters or terrorist attacks, that might cause rapid growth of data sharing through social media networks. However, collecting and processing data of an event is a challenging task and essentially requires a great deal of data cleaning and filtering out what is relevant/irrelevant to the event. Data sifting task endeavors to identifying the related content to the depicted event data. We propose a learning strategy to dynamically learn complementary contributions from different data-driven features through a semi-supervised graph-fusion technique. Our proposed method relies upon minimal training labeled data samples — ultra-small data learning. Learning through a small labeled set is also of particular interest to forensic investigators and medical researchers — concerning massive data labeling and minimizing energy-efficient computing to reduce redundancy and repetitions. We assess the effectiveness of the proposed semi-supervised method on five datasets from real-world events. Compared with prior-art (supervised and semi-supervised ones), experimental results show the proposed method achieves the best classification results and most efficient computational footprint.