junto inject a seed but node is unlabeled

[castellucc...@gmail.com]

Hi, 

i would like to know if it is possibile (or how i can modify the existing code) of the graph, for injecting a seed to a node still considering it an unlabeled node.

I'm interested in initializing the label distributions of all the nodes of my graph, but i want that some node is changed more than other nodes (the real labeled ones).

Some knows a solution?

[castellucc...@gmail.com]

I'll try to explain better what i mean. I would like to mark some node as effectively labeled and other as unlabeled, but i would like to initialize some label score on both kind of node.

I was thinking about modifying the seed file by including a new column, "true" or "false" indicating if a node is labeled or not.

I modified the GraphInfo.scala file in order to manage this added column in the LabelReader, 

object LabelFileReader {

  def apply (filename: String): List[Label] = {

    (for (line <- io.Source fromFile(filename) getLines) yield {

      val Array(vertex, label, score, gold) = line.trim split("\t")

      new Label(vertex, label, score.toDouble, gold)

    }).toList

  }

}

Then, is it sufficient to set the vertex.isSeedNode=false when label.gold is false in GraphLoader.scala? Or i need to do something more? 

I'm using the MAD algorithm in my experiments, if it can help.

My objective is to inizialize all the nodes in the graph, even the unlabeled with another algorithm, and then let the MAD algorithm refine this labeling by means of other information i put in the graph. If i label all nodes and i use MAD as is, in the objective function it tries to optimize, the first term will make the labeling as close as possible to the seeds, while for some nodes (the test nodes) i would like to be less conservative.

[Partha Pratim Talukdar]

Thanks for the added explanation. Looks like you want to give different emphasis to different seed nodes. If so, then you can control that through the injection probability (p^{inj}_v) per node:

https://github.com/parthatalukdar/junto/blob/0057e5239240548b374faa5f6e90fa6808327a81/src/main/java/upenn/junto/algorithm/mad_sketch/Vertex2.java#L29

You can customize it for each node here:

https://github.com/parthatalukdar/junto/blob/0057e5239240548b374faa5f6e90fa6808327a81/src/main/scala/upenn/junto/graph/Vertex.scala#L157

Essentially you are changing the S matrix in the first term of MAD's objective (Sec 4 of http://talukdar.net/papers/adsorption_ecml09.pdf)

hth,

Partha

[castellucc...@gmail.com]

Thank you for the reply.

I'll try!