PROJECT TITLE :

Deep Learning for Adverse Event Detection from Web Search

ABSTRACT:

The timely identification of product defects, disasters, and major socio-political incidents are just some of the real-world applications that rely heavily on adverse event detection as a crucial component. In the field of medicine, adverse drug events are responsible for a significant number of annual hospitalizations and deaths. Examination of search query logs has become an important detection channel as a result of the fact that users frequently start their information quests and reports with searches conducted online. The difficulty of measuring and analyzing these aspects has prevented their use in previous studies, despite the fact that search context, which includes query intent and heterogeneity in user behaviors, is extremely important for extracting information from search queries. We present DeepSAVE, an innovative deep learning framework that uses user search query logs as its data source to identify potentially harmful events. The context problem that is associated with search-based detection of adverse events is tackled by DeepSAVE with the help of an enriched variational autoencoder that incorporates a novel query embedding and user modeling module. These two modules collaborate to solve the problem. The findings of an evaluation conducted on three very large real-world event datasets demonstrate that DeepSAVE outperforms both the traditional methods of detection currently in use and the competing deep learning auto encoders. According to the results of the ablation analysis, each individual component of DeepSAVE makes a sizable contribution to the overall performance of the system. The findings, taken as a whole, provide conclusive evidence that the proposed architecture can successfully identify potentially harmful occurrences in search query logs.