Skip to content

Cynet Package

Jump to bottom
Alice Saparov edited this page Jun 3, 2020 · 4 revisions

Purpose

The Cynet Package performs analysis on spatio-temporal data. It allows us to study weather events in the US spanning from 2016-2019 using a dataset containing 4 years of recorded weather events (cold, snow, fog, rain, storm) from over 2000 weather stations.

Results

Current average AUC of 0.76 from the 2079 existing models.

Method

  1. Algorithm input is the log of events.
    • An item of event log is the what, where, when of a weather event.
  2. The event log is preprocessed and converted to a time series which contains location and event type.

For example...

 - A time series for location "UofC" and event type “rain” may look like 0 1 0 0 0 1
 - It means no rain at the first time step, rain at the second time step, no rain for the third to fifth time 
   steps, and rain for the last time step.
  1. Cynet generates a directed network for all the time series.
  2. The influence of one time series on another is captured in the edges of the network.

For example...

 - Assume we have time series UofC - “Rain” and O'Hare - “Storm”
 - The edge from O'Hare - “Storm” to UofC - “Rain” is a model showing how storms around O'Hare can be used as           
   predictors for rain around UofC.
 - Let us say that typically in Chicago, wind blows from O'Hare to UofC, then the model will have strong 
   predicting power.
 - We note that the influence may very well be asymmetric.
 - We also infer models for different time lag because some influences may be short-term while others may take 
   some more time to be apparent.
  1. Once the network is generated for all source and target series, it is pruned to remove weak edges.
  2. Each edge serves as a predictor for a weather pattern.
    • Location and time lag are very influential.
  3. The Cynet package integrates all these predictions by producing scalar coefficients from each data source in order to make a final prediction about the weather at the desired location.
  4. Once the network is pruned and scalar coefficients are established for each link in the network, Cynet can make predictions.

Clone this wiki locally