systemml-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Janardhan Pulivarthi <janardhan.pulivar...@gmail.com>
Subject Use case for the SYSTEMML-1437
Date Sat, 15 Apr 2017 18:16:34 GMT
Dear sir / ma'am,

The potential of Factorization machines has yet to be discovered. The
possible uses are traffic prediction, earthquake prediction, movie ratings
and much more... Incorporating the Factorization Machines in *SystemML*
will be very much necessary, for much of the data that we encountering in
many fields is sparse in many respects.


*Description for the use case:🛫 *
The automatic prediction of flight diversion, with no prior knowledge of
flight route requirement. This deviation from the expected flight path
alerts the ground logistics to find the next best alternative, for the
timely delivery of the intended service. The prediction is made by
analyzing the flight data anomalies detected by Factorization Machines (FM)
classification. Often, the datasets associated with the flight data are
sparse, which makes it difficult to analyze the flight path accurately.
But, with FM's this will be easier.

Already, implemented in SVM by some researchers in this article. The doi:
http://doi.org/10.1016/j.dss.2016.05.004

*Diagram: (*Overview of the prediction model for diversion detection.

*)[image: Inline image 1]*
*Features: (*of the aircraft*)*

   1. Distance completed
   2. Distance gained
   3. Velocity deviation
   4. Altitude deviation
   5. Phase

*What to Optimize:*

   1. *Interval length (L):* Interval-length *L* describes the time range
   in which positional updates are gathered, and consequently the amount of
   behavior captured in a single feature vector.
   2. The parameters capture the expected level of noise and the extent to
   which the decision hyperplane is fit to the training data.

Iterations will be made on these three things & we will be trying to have a
sufficient precision as to be needed to safely say that plane will be
diverted: *P* = Precision, *R* = recall,* F*-score
​​

*Dataset inputs:*

   1. a unique flight identifier.
   2. an aircraft identifier.
   3. the flight code.
   4. the timestamp of the event.
   5. the IATA/FAA codes of departure & arrival airports.
   6. the gps coordinates of the aircraft.
   7. the altitude of aircraft.
   8. the speed of the aircraft.

*Datasets:*

   1. http://openflights.org/data.html (open database license)
   2. https://www.flightradar24.com
   3. and there will be more open databases.

*Possible extensions:*

   1. If it can be applied for air traffic, then it will work for Rail🛤,
   Road🛣, Vessel traffic🛳.
   2. Earthquake detection is also possible with the Factorization
   Machines, as the natural event occurrence is subject to variable at any
   given time & place.

Mime
  • Unnamed multipart/related (inline, None, 0 bytes)
View raw message