Stream Analytics for Diverse Applications

点击翻译按钮获取中文摘要

The long-term objective of the proposed program is to develop next-generation tools and algorithms to analyze stream and sensor data collected for diverse applications, such as ocean monitoring, human chronic disease management, and smart cities. In all these applications, the data streams generated are long, time-stamped, and often geo-coded spatially, and typically capture numeric quantities, such as temperature, blood pressure, etc. However, data streams are not restricted to numeric quantities; text streams for disease management and smart cities applications are included and studied in the proposed program. For the next 5 years, we propose four focal areas: (A) aggregate query processing for wireless sensor networks; (B) topic modeling and sentiment extraction for text streams; (C) outlier detection and explanations; and (D) prefix based forecasting. Regarding focal area (A), we will tackle one of the more complex settings for sensor networks when there is a wireless network of sensors tracking moving objects. We will develop novel scalable edge-computing schemes for object de-duplication to provide exact and approximate answers for aggregate “continuous” queries. We also plan to extend those queries with group-by clauses. Text streams is are a special case of data streams. For both chronic disease management and smart cities monitoring, two classes of analytic techniques are valuable, which are topic modeling and sentiment tracking. We will study how topic modeling can be performed effectively with ontologies. We will advance state-of-the-art sentiment analysis to detect the fluctuation of patients' or citizens' affects (such as anger, anxiety or depression) over time, and to conduct discourse coherence analysis to identify possible onsets of mental dysfunction and delirium. Whether a data stream is numeric or text, for many monitoring applications, given long periods of “normalcy” or “background activities”, one of the most interesting analytic tasks is to identify events when “abnormalities” occur. We will study how to make outlier detection scalable for spatio-temporal trajectories, and how to provide possible explanations to detected outliers or groups of outliers. Another important analytic task for long spatio-temporal trajectories is to predict the outcome (such as in the form of a class label) of the trajectory based on the earlier time points, or the prefix, of the trajectory. Accurate prefix based forecasting has numerous applications, including earlier intervention. For area (D), we will develop novel scalable methods for both numeric and text streams. We will also study the tradeoff between the length of the prefix and prediction accuracy. Last but not least, a key component of the proposed program is the tight integration and validation of the new methods within three application domains: a CFI-funded ocean monitoring project , a CIHR-funded tele-monitoring patient cohort, and a smart city initiative.