Hi,All,
using Ranklib to Click-Through-Rate prediction (or Orders prediction) ranking, such as, there are many items, the ordered label is 1, the no order label is 0, we want rank many items by ranklib.
The LambdaMart is good. But we wish the model is simple and easy to debug ... and using. The Coordinate Ascent model is linear model, Why can't add a ranking model using Logistic regression?
Can add a Logistic regression model for ranking ? because Logistic regression can deal with a little
non linear problem.
How to add a Logistic regression ranking model or change Coordinate Ascent model to a Logistic regression model?
How to change it ?
Best regards,
Star
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
RankLib has a linear regression model for LTR. If you want a logistic one, you'd have to write it yourself.
All the ranking models are subclasses of the Ranker class. You might want to look at LinearRegRank.java to get ideas on how best to proceed. You'd write your own learn() method.
Linear regressions are capable of fitting curves. Could you make use of the linear regression model and simply square or cube or make reciprocal values of some of your features? Doing these things can produce a linear model that fits bends and curves in the data.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thank your reply firstly!
Ihave read the linear regression model in LinearRegRank.java. The linear model using decomposition, factorization or SVD.
Why not using gradient descent?
Thank you very much!
Star
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Since the implementer of RankLib is no longer accessible, I do not know the specifics of why the RankLib algorithms were implemented as they were. Some of the algorithms are documented as to the papers used to guide implementations, but there is no such information for the linear regression ranking.
Other sources state it is a basic least-squares implementation based on Lawson and Hanson, 1995. The code does not look to be basic least-squares to me, but I have not looked at it closely and RankLib has various optimizations that can sometimes hide the flow of the algorithm process.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,All,
using Ranklib to Click-Through-Rate prediction (or Orders prediction) ranking, such as, there are many items, the ordered label is 1, the no order label is 0, we want rank many items by ranklib.
The LambdaMart is good. But we wish the model is simple and easy to debug ... and using. The Coordinate Ascent model is linear model, Why can't add a ranking model using Logistic regression?
Can add a Logistic regression model for ranking ? because Logistic regression can deal with a little
non linear problem.
How to add a Logistic regression ranking model or change Coordinate Ascent model to a Logistic regression model?
How to change it ?
Best regards,
Star
RankLib has a linear regression model for LTR. If you want a logistic one, you'd have to write it yourself.
All the ranking models are subclasses of the Ranker class. You might want to look at LinearRegRank.java to get ideas on how best to proceed. You'd write your own learn() method.
Linear regressions are capable of fitting curves. Could you make use of the linear regression model and simply square or cube or make reciprocal values of some of your features? Doing these things can produce a linear model that fits bends and curves in the data.
Thank your reply firstly!
Ihave read the linear regression model in LinearRegRank.java. The linear model using decomposition, factorization or SVD.
Why not using gradient descent?
Thank you very much!
Star
Since the implementer of RankLib is no longer accessible, I do not know the specifics of why the RankLib algorithms were implemented as they were. Some of the algorithms are documented as to the papers used to guide implementations, but there is no such information for the linear regression ranking.
Other sources state it is a basic least-squares implementation based on Lawson and Hanson, 1995. The code does not look to be basic least-squares to me, but I have not looked at it closely and RankLib has various optimizations that can sometimes hide the flow of the algorithm process.