Scalable Self-taught Deep-Embedded Learning Framework for Drug Abuse Spatial Behaviors Detection

Drug abuse has become an increasingly challenging issue national wide in the United States, while each state has its own legislation regarding such behavior which further stimulates different semantic representations of such behavior over space. To build an accurate and robust classifier to detect such behaviors with spatial variance remains challenging due to the existence of large noise in tweets and limited number of labeled data. Most efforts have utilized humans to label tweets for the base classifier training. The randomness of human labeled data would limit the generalization of base model trained. We propose a deep learning-based centroid-attention framework to consider the spatial variance. We further explore the effect of state-based exemplars on the base model. The performance of the base classifier is thus enhanced.