4 ms·

I'm curious to know what type of machine learning algorithms are being used for the prediction. Anyone have any thoughts?

by physcab 17y ago

dschobel 17y ago
When trying to learn a really complicated concept in absence of a clear domain model the best solution when I was in school a few years back was neural networks. That would be my first instinct if I had to solve this problem.
- bradfordcross 17y ago
  Flightcaster doesn't intend to be a complete black box. To an extent, the more that sophisticated users know, the better - because they can help us by reporting specific details of use cases where we are not doing well so that we can find ways to incorporate those cases into the research and the model. @dschobel - you are right that it is difficult to learn complicated domain models in domains like this with a lot of subtle logic. To deal with this we use a blend of analytical and inductive learning. There is a nice discussion of combining analytical and inductive learning in "Machine Learning" - Mitchell, ch. 11 & 12. As to the specifics of whether we use SVMs, Bayesean Networks, Decision Trees, and so on - we don't feel that it would be beneficial to get into detail. This is not just to keep everything top-secret, but also because all of this is an area of active research and we may completely change the approach at any moment. Most of what gives us in an edge is 1) the ability to deal with the data (preprocessing, joining, etc.) 2) the infrastructure to do fast, distributed learning, and 3) the domain expertise to add richness of logic to the individual classifiers we use. The exact techniques by which we weave everything together are open to constant tweaking. Another important point outside of the specific learning algorithms is that we use rigorous approaches like 100-fold cross-validation and such. But I like to think of it like Popper - so I call it invalidation rather than validation. :-) IMO, the process of bringing scientific rigor to a problem is also more important than the hypothesis under test. A lot of the best stuff comes from a rigorous application of scientific method that yields unexpected results that lead you to a an alternative hypothesis.
zemariamm 17y ago
me too, but I also would love to know which data sources they are using
- mbarr 17y ago
  http://www.flightcaster.com/faq#data_sources http://www.flightcaster.com/faq#data_sources What are FlightCaster's data sources? FlightCaster uses data from: FAA Bureau of Transportation Statistics FAA Air Traffic Control System Command Center FlightStats National Weather Service
- gojomo 17y ago
  They may consider their methods too proprietary to supply details, so I'll speculate: Information on earlier-in-day flights alone would probably be enough (mixed with historical data) to be fairly accurate about later-in-day flight delays -- even those that aren't continuations of the same plane -- because of giant overlaps in delay-causing conditions (weather, airport/mechanical mishaps, crew issues, etc.). Weather forecasts might give another advantage, especially in predicting 'seed' delays that then hint at later delays. If there are any other semi-public feeds related to FAA reporting or air-traffic control -- even if mostly meant for other pilots or General Aviation -- those would be incredibly valuable. If those regional and national maps of planes-in-flight also contain sufficient positioning detail to notice when they're spending a little extra time on runways, or waiting for/at gates, etc. -- another positive early influencer for predictions.
  - physcab 17y ago
    I don't think it would be harmful for them to give a clue about what algorithms they use. There is so much tuning required to get these algorithms to perform the way you'd expect that I think they could still keep their IP locked up even if they gave a general hint. With that said, I'll speculate as well: Perhaps you'd need some type of ideal dataset, one that included departure times/arrivals and distances and weather conditions of flights that came in on time as expected. Then you might start introducing some noisy data, ie. flights that made the same trip but came in late or early with same weather conditions. Then you'd add in the effect of weather conditions and see how flights fared. I'd speculate that you could get away with doing some type of regression analysis, but maybe you'd need to resort to a more complex algorithm for classifying ("On-time", "Early", "Late") based on a series of features ("Distance","Weather","Mechanical", "Time", etc). SVM could pull this off, or perhaps even a naive bayesian classifier. For research purposes I'd probably check out RVM because it might need less information to classify. Not sure if it would be realistic to use it though...this problem is in need of a highly scalable solution.
fizx 17y ago
I'd be somewhat surprised if it isn't Bayesian networks. Oblig. wiki link: http://en.wikipedia.org/wiki/Bayesian_network http://en.wikipedia.org/wiki/Bayesian_network

I'm curious to know what type of machine learning algorithms are being used for the prediction. Anyone have any thoughts?

by physcab 17y ago

I'm curious to know what type of machine learning algorithms are being used for the prediction. Anyone have any thoughts?

dschobel 17y ago

When trying to learn a really complicated concept in absence of a clear domain model the best solution when I was in school a few years back was neural networks. That would be my first instinct if I had to solve this problem.

bradfordcross 17y ago
Flightcaster doesn't intend to be a complete black box. To an extent, the more that sophisticated users know, the better - because they can help us by reporting specific details of use cases where we are not doing well so that we can find ways to incorporate those cases into the research and the model. @dschobel - you are right that it is difficult to learn complicated domain models in domains like this with a lot of subtle logic. To deal with this we use a blend of analytical and inductive learning. There is a nice discussion of combining analytical and inductive learning in "Machine Learning" - Mitchell, ch. 11 & 12. As to the specifics of whether we use SVMs, Bayesean Networks, Decision Trees, and so on - we don't feel that it would be beneficial to get into detail. This is not just to keep everything top-secret, but also because all of this is an area of active research and we may completely change the approach at any moment. Most of what gives us in an edge is 1) the ability to deal with the data (preprocessing, joining, etc.) 2) the infrastructure to do fast, distributed learning, and 3) the domain expertise to add richness of logic to the individual classifiers we use. The exact techniques by which we weave everything together are open to constant tweaking. Another important point outside of the specific learning algorithms is that we use rigorous approaches like 100-fold cross-validation and such. But I like to think of it like Popper - so I call it invalidation rather than validation. :-) IMO, the process of bringing scientific rigor to a problem is also more important than the hypothesis under test. A lot of the best stuff comes from a rigorous application of scientific method that yields unexpected results that lead you to a an alternative hypothesis.

zemariamm 17y ago

me too, but I also would love to know which data sources they are using

mbarr 17y ago
http://www.flightcaster.com/faq#data_sources http://www.flightcaster.com/faq#data_sources What are FlightCaster's data sources? FlightCaster uses data from: FAA Bureau of Transportation Statistics FAA Air Traffic Control System Command Center FlightStats National Weather Service
gojomo 17y ago
They may consider their methods too proprietary to supply details, so I'll speculate: Information on earlier-in-day flights alone would probably be enough (mixed with historical data) to be fairly accurate about later-in-day flight delays -- even those that aren't continuations of the same plane -- because of giant overlaps in delay-causing conditions (weather, airport/mechanical mishaps, crew issues, etc.). Weather forecasts might give another advantage, especially in predicting 'seed' delays that then hint at later delays. If there are any other semi-public feeds related to FAA reporting or air-traffic control -- even if mostly meant for other pilots or General Aviation -- those would be incredibly valuable. If those regional and national maps of planes-in-flight also contain sufficient positioning detail to notice when they're spending a little extra time on runways, or waiting for/at gates, etc. -- another positive early influencer for predictions.
- physcab 17y ago
  I don't think it would be harmful for them to give a clue about what algorithms they use. There is so much tuning required to get these algorithms to perform the way you'd expect that I think they could still keep their IP locked up even if they gave a general hint. With that said, I'll speculate as well: Perhaps you'd need some type of ideal dataset, one that included departure times/arrivals and distances and weather conditions of flights that came in on time as expected. Then you might start introducing some noisy data, ie. flights that made the same trip but came in late or early with same weather conditions. Then you'd add in the effect of weather conditions and see how flights fared. I'd speculate that you could get away with doing some type of regression analysis, but maybe you'd need to resort to a more complex algorithm for classifying ("On-time", "Early", "Late") based on a series of features ("Distance","Weather","Mechanical", "Time", etc). SVM could pull this off, or perhaps even a naive bayesian classifier. For research purposes I'd probably check out RVM because it might need less information to classify. Not sure if it would be realistic to use it though...this problem is in need of a highly scalable solution.

fizx 17y ago

I'd be somewhat surprised if it isn't Bayesian networks. Oblig. wiki link: http://en.wikipedia.org/wiki/Bayesian_network http://en.wikipedia.org/wiki/Bayesian_network