Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
    Daniel Korzekwa
    @danielkorzekwa
    test message
    Andy Petrella
    @andypetrella
    test response
    I've added my associate who will be very happy to see this ;-)
    Martin Mauch
    @nightscape
    Hah, perfect, I was just checking if bayes-scala has a Gitter channel
    The Sprinkler example still does not work due to danielkorzekwa/bayes-scala#11, right? Maybe that issue should be left open, so that it's more obvious that this is not fixed yet.
    Daniel Korzekwa
    @danielkorzekwa
    reopened, not sure why I closed it
    Francis De Brabandere
    @francisdb
    Hi @danielkorzekwa
    Is any release planned?
    Gowtham Natarajan
    @gowthamnatarajan
    Any example about using it for EM in Multinomial case?
    Daniel Korzekwa
    @danielkorzekwa
    two examples on EM in discrete multinomial bayesian nets: https://github.com/danielkorzekwa/bayes-scala/blob/master/doc/clustergraph/clustergraph.md
    Getting Started - Learning parameters with Expectation Maximisation in Bayesian Networks from incomplete data
    Getting Started - Learning parameters with Expectation Maximisation in Unrolled Dynamic Bayesian Networks from incomplete data
    Gowtham Natarajan
    @gowthamnatarajan
    Thanks. Will look into it. It also supports latent variables I guess?
    Daniel Korzekwa
    @danielkorzekwa
    yes, EM says that!
    the second example show parameters tying
    Gowtham Natarajan
    @gowthamnatarajan
    Any limit on the cardinality of a feature in the graph? In terms of scalability?
    Daniel Korzekwa
    @danielkorzekwa
    well, not much, the biggest I tried was about 6 thousand nodes, max 10 dim each
    with max two parents
    I never really optimized it and moved quite soon to hybrid bayesian nets quite soon
    still a few people used it for some use cases, e.g. biology and medicine and it worked, just give a go
    Gowtham Natarajan
    @gowthamnatarajan
    Thanks, I meant is there any limit on the number of rows of a CPT?
    Daniel Korzekwa
    @danielkorzekwa
    memory/cpu ?, tried max with 100 CPT
    Gowtham Natarajan
    @gowthamnatarajan
    @danielkorzekwa Is this model based on a paper or just an example? https://github.com/danielkorzekwa/bayes-scala/blob/master/doc/dsl/conversionrate/conversion_rate.md
    Gowtham Natarajan
    @gowthamnatarajan
    Is there a different package for hybrid networks? In my case, the latent variable is continuous and its parents are categorical.
    Daniel Korzekwa
    @danielkorzekwa
    For conversion rate example I use Gaussian Process Classification. I didn't follow any paper, which models conversion rate with GPC
    regarding to categorical parents, unfortunately my package does not support it. It is best suited either for discrete models only or for Gaussian Process like models ( Gaussian Process Regression, GPC, Modelling skills in games, large scale GP (e.g. hierarchical, etc.), GP like models are now inside of bayes-scala-gp.
    Hugh Winkler
    @hwinkler
    Hi Daniel, thanks for this great piece of work. I'm just familiar enough with Bayesian networks to have made sense of your ClusterGraph examples. Is it possible to obtain a posterior joint marginal for a subset of the variables? I was looking for something like "loopyBP.marginal(var1.id, var2.id,...)" Is there capability like that already in the program? If not, does it make sense for me to dig in and figure out how to add the capability? Or is there something in the loopy algorithm or cluster graph structure that makes that not possible? (I'm familiar with bucket elimination, not with loopy BP, nor with cluster graphs... but I bought the Koller book to learn about them)
    Hugh Winkler
    @hwinkler
    Minor edit. I may not mean "joint marginal" which Darwiche uses to mean P(A,B, evidence). I mean posterior P(A,B| evidence)
    Daniel Korzekwa
    @danielkorzekwa
    You can add P(A,B) to cluster graph or to factor graph so that you have a join variable. Second option, If A and B are separate variables you can multiply their posteriors in order to get a joint distribution but here you of course make some assumptions about A and B being independent.
    pzhanggithub
    @pzhanggithub
    I have had the problem to compile the project with sbt. I got error message that org.scoverage#sbt-scoverage;1/04: not found and unresolved dependncy: com.codacy#sbt-codacy-coverage; 1.0.3: not fount. Both of them are defined in the plugins.sbt. Does any one have the solution? thanks.
    Francis De Brabandere
    @francisdb
    @pzhanggithub strange, both are available on maven central. You could try newer versions
    kakaner
    @kakaner
    Hallo! It's been a while since the last message. Just curious-- is this channel still active?
    Daniel Korzekwa
    @danielkorzekwa
    kind of, unfortunately not working anymore on bayes-scala project, still happy to help when possible
    Shane Mooney
    @shanem
    Hi Daniel. kakaner and I are working together. We have a question about representing our network as a cluster graph. The edges in our network are A -> B, B -> C, and A -> C, which we represent as two factors, factor 1 over (A, B), and factor 2 over (A, B, C). If we attempt to map each factor to its own cluster, we run into the single variable in supset limitation (with A and B both being in the supset). Is this a sensible cluster graph for our network? Can you suggest any workarounds?
    kakaner
    @kakaner
    (btw @danielkorzekwa just want to thank you for this library! It's the most flexible and functional library we've found.. super grateful :D )
    Daniel Korzekwa
    @danielkorzekwa
    Difficult to advice not knowing what you try to do /infer, but will try.
    Having (A,B) and (A,B,C) clusters makes no sense to me. If you put all variables to a single cluster you don't really need any other clusters as you already have a full joined probability distribution.
    so I can see two approaches (if having a full joined distribution across (A,B,C) is a problem:
    first on cluster graph with clusters (A,B) (B,C) and (A,C)
    second approach almost? equivalent (small diffs during message passing ) on factor graphs with factors (A,B), (B,C) and (A,C)
    kakaner
    @kakaner
    @danielkorzekwa thanks for the feedback. We haven't tried an (A,B,C) cluster yet but were about to try that. It's not really a problem aside from slightly different dimension outcomes. We already tried the two-variable factor approach.
    honestly we don't have great intuition around how 3 two var factors differ from 1 3 var factor so are just kind of learning from experimentation
    ChristienCDK
    @ChristienCDK
    Hi @danielkorzekwa , I'm a student who is doing a research on Bayesian networks to analyze multinomial survey data. Currently I am using R as a programming language. Your tool bayes-scala looks spot on like what I'm looking for. However, I am having difficulties implementing it in R so I was hoping you might give a 101 Introduction on how to use the Scala program in an R environment. Thanks for your hard work and hope to hear from you.
    Daniel Korzekwa
    @danielkorzekwa
    Hello
    though I never used it myself
    or just try using Scala directly
    btw, I'm not working on bayes-scala anymore, I changed companies and I can't support you with more detailed questions, sorry :)