Discovering temporal, spatial, and contextual anomalous social activities from streaming social media datasets

Abstract

Social media data refers to data collected from social media platforms, including user interactions, posts, comments, likes, etc., while streaming social media data is a specific type of social media data that is continuously and sequentially collected from these platforms. Processing of streaming social media data is crucial for providing quick responses and reacting promptly to issues. Therefore, it requires more specialized solutions compared to traditional social media data processing algorithms. Temporal, spatial, and contextual anomalous social activity of a social media user can be defined as activities of sending a frequent and abnormal number of posts that are contextually similar to other posts in a time frame and at specific locations. However, detecting anomalies in streaming social media datasets has many difficulties due to the size, continuous increase, different characteristics (e.g., temporal, spatial, and contextual) of the data, and shortcomings of existing algorithms to handle streaming social media datasets. In the literature, many studies are conducted on detecting anomalies in social media data or streaming data. Nonetheless, anomaly detection studies on streaming social media datasets are very limited. The focus of this study is the detection of temporal, spatial, and contextual anomalous social activities from streaming social media datasets. In particular, streaming Twitter/X dataset is analyzed in terms of three different aspects, such as temporal, spatial, and contextual anomalous social activity detection. The proposed Temporal, Spatial, and Contextual Anomalous Social Activity Miner (TSCASA-Miner) algorithm is experimentally evaluated with a naive alternative and related studies. Evaluations are conducted using a real-life streaming Twitter/X dataset from New York, USA. The results demonstrate that the proposed TSCASA-Miner algorithm outperforms the other algorithms, with a recall of 0.93, a precision of 0.95, and an accuracy of 0.9323, marking an improvement of 0.018 in accuracy.