Rumor Detection via Recurrent Neural Networks - A Case Study on Adaptivity with Varied Data Compositions.

Rumor detection is a meaningful research problem due to its significance in preventing potential threats to cyber security and social stability. With the recent popularity of recurrent neural networks (RNNs), the application of RNNs in rumor detection has resulted in promising results as RNNs can naturally blend into the task of language processing and sequential data modelling. However, since deep learning models require large data scale for training in order to extract sufficient distinctive patterns, their adaptivity with varied data compositions can become a challenge for real-life application scenarios where rumors are always the minority (outlier) in the streaming data. In this paper, we present a case study to investigate how the ratio of rumors in training data affects the calssification performance of RNN based rumor detection models and successfully address some issues on the model adaptivity.