Title:Multi-view shaker detection: Insights from a noise-immune influence analysis Perspective

Abstract:Entities whose changes will significantly affect others in a networked system are called shakers. In recent years, some models have been proposed to detect such shaker from evolving entities. However, limited work has focused on shaker detection in very short term, which has many real-world applications. For example, in financial market, it can enable both investors and governors to quickly respond to rapid changes. Under the short-term setting, conventional methods may suffer from limited data sample problems and are sensitive to cynical manipulations, leading to unreliable results. Fortunately, there are multi-attribute evolution records available, which can provide compatible and complementary information. In this paper, we investigate how to learn reliable influence results from the short-term multi-attribute evolution records. We call entities with consistent influence among different views in short term as multi-view shakers and study the new problem of multi-view shaker detection. We identify the challenges as follows: (1) how to jointly detect short-term shakers and model conflicting influence results among different views? (2) how to filter spurious influence relation in each individual view for robust influence inference? In response, a novel solution, called Robust Influence Network from a noise-immune influence analysis perspective is proposed, where the possible outliers are well modelled jointly with multi-view shaker detection task. More specifically, we learn the influence relation from each view and transform influence relation from different views into an intermediate representation. In the meantime, we uncover both the inconsistent and spurious outliers.