Multiple Targets MLN #43
Description
Hello BoostSRL Team!
After reading your paper Learning Markov Logic Networks via Functional Gradient Boosting http://pages.cs.wisc.edu/~tushar/papers/icdm11.pdf i have been trying to learn MLN with multiple targets.
But so far i did not have much success. How do i set up the learning flags and my data properly? Here is what i have tried to do.
Toy-Father
I have manually added female and mother predicates.
For Tree-Based i run following flags:
-mln -trees 10 -l -train datasets\Father-Mother\train\ -target father,mother
the result is 2 models but they appear to be the same as if i learned each target seperatly. I was hoping to see a Joint Model where the 2 models would somehow influence each other.
For Clause-Based i used:
-mln -mlnClause -trees 10 -l -train datasets\Father-Mother\train\ -target father,mother
But the code crashes with following error:
It appears that it is generating wrong examples (mother example for father target)
Toy-Cancer
I was trying to learn a Joint Model for Cancer dataset. In default setting it is meant to learn just the predicate cancer. train_facts contains only friends and smokes predicates and train_pos/train_neg containts exampels for _cancer.
How do i set up to learn all 3 predicates (friends, smokes, cancer) at once?
I tried moving all the facts to train_pos/train_neg or duplicating them in train_facts and train_pos/train_neg (with proper negative examples) but all i was getting were erros like this:
In both Tree-Based and Clause-Based settings.
Even trying to learn 2 predicates out of 3 had similar error messages.
Cora
In your paper you have results for both Tree-Based and Clause-Based approaches on Cora datasets when learning Joint Model for Cora dataset with target predicates SameBib, SameVenue, SameTitle and SameAuthor.
I tried to do Clause-Base learning:
-mln -mlnClause -trees 10 -l -train datasets\Cora\train\ -target sameauthor,samebib,sametitle,samevenue
But after less than a minute i get an error:
I looked at the source code and wiki/documentation and didnt find much on working with multiple targets. Since you got the results in your paper i do really hope that you can help me.
Thank you in advance from D.Ravdin!