Re:Markov Chains Explained
by thomaspaine on Friday January 09, 2015 @06:01PM
(#8865039)
Markov chains have a lot of nice features you can derive from them, like convergence to a stationary distribution. It's hard to say that it's an inappropriate simplification if it happens to be pragmatic (like most simplifications).
For example PageRank can be modeled using markov chains. You can surely develop more complicated models but markov chains provide a good starting point.
Re:Markov Chains Explained
by dangerlibrary on Friday January 09, 2015 @05:44PM
(#8864970)
The formal definition of a single Markov Chain is one with a fixed transition matrix.
In practice, many software implementations of Markov models update/modify the transition matrix as new information becomes available. Formally, you're no longer working with the same Markov chain, but that doesn't mean that the model isn't useful.
Re:Markov Chains Explained
by learnstats2 on Friday January 09, 2015 @07:04PM
(#8865253)
Right; I don't think this should be described as a Markov Chain variant, since it breaks the fundamental property of a Markov Chain; Any old simulation works in this way.
Re:Markov Chains Explained
by coherentpony on Friday January 09, 2015 @05:40PM
(#8864954)
What you are referring to are non-Markov processes. That is, process with a memory. These can be useful too, but the reason Markovian processes are preferred is because of the independence (lack of memory). Independence is a sufficient condition for many theorems that use Markov processes to prove things about average behaviour.
Re:Markov Chains Explained
by learnstats2 on Friday January 09, 2015 @06:32PM
(#8865160)
Yes, I understand.
But why does a lack of memory help? What useful theorems are there?
Re:Markov Chains Explained
by IndianAstronaut on Friday January 09, 2015 @06:06PM
(#8865059)
Markov chains easily allow you to attach cost functions to them. Not particularly data mining oriented, but good for running simulations and cost-benefit.
What I would prefer to know is why are they more useful than a more nuanced model with non-fixed probabilities, or memory of previous states? - given that these are not really harder to simulate.
It seems to me that Markov Chains are very often used as an inappropriate simplification.
Is there some mathematical advantage to using them?