Learning to Rank
An easy implementation of algorithms of learning to rank. Pairwise (RankNet) and ListWise (ListNet) approach. There implemented also a simple regression of the score with neural network. [Contribution Welcome!]
Requirements
- python 2.7
- tqdm
- matplotlib v1.5.1
- numpy v1.13+
- scipy
- chainer v1.5.1 +
- scikit-learn
- and some basic packages.
RankNet
Pairwise comparison of rank
The original paper was written by Chris Burges et al., "Learning to Rank using Gradient Descent." (available at http://research.microsoft.com/en-us/um/people/cburges/papers/ICML_ranking.pdf)
Usage
Import and initialize
from learning2rank.rank import RankNet
Model = RankNet.RankNet()
Fitting (automatically do training and validation)
Model.fit(X, y)
Here, X
is numpy array with the shape of (num_samples, num_features) and y
is numpy array with the shape of (num_samples, ). y
is the score which you would like to rank based on (e.g., Sales of the products, page view, etc).
Possible options and defaults:
batchsize=100, n_iter=5000, n_units1=512, n_units2=128, tv_ratio=0.95, optimizerAlgorithm="Adam", savefigName="result.pdf", savemodelName="RankNet.model"
n_units1
and n_units2=128
are the number of nodes in hidden layer 1 and 2 in the neural net.
tv_ratio
is the ratio of the data amounts between training and validation.
Predict
Model.predict(X)
ListNet
Listwise comparison of rank
The original paper was written by Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, Hang Li "Learning to Rank: From Pairwise Approach to Listwise Approach." (Available at http://research.microsoft.com/en-us/people/tyliu/listnet.pdf)
NOTICE: The top-k probability is not written. This is listwise approach with neuralnets, comparing two arrays by Jensen-Shannon divergence.
Usage
Import and initialize
from learning2rank.rank import ListNet
Model = ListNet.ListNet()
Fitting (automatically do training and validation)
Model.fit(X, y)
Same as ranknet, X
is numpy array with the shape of (num_samples, num_features) and y
is numpy array with the shape of (num_samples, ). y
is the score which you would like to rank based on (e.g., Sales of the products, page view, etc).
Possible options and defaults:
batchsize=100, n_epoch=200, n_units1=512, n_units2=128, tv_ratio=0.95, optimizerAlgorithm="Adam", savefigName="result.pdf", savemodelName="ListNet.model"
Predict
Model.predict(X)
Regression
Regression the scores with neural network
Usage
Import and initialize
from learning2rank.regression import NN
Model = NN.NN()
Fitting (automatically do training and validation)
Model.fit(X, y)
Possible options and defaults:
batchsize=100, n_iter=5000, n_units1=512, n_units2=128, tv_ratio=0.95, optimizerAlgorithm="Adam", savefigName="result.pdf", savemodelName="RankNet.model"
n_units1
and n_units2=128
are the number of nodes in hidden layer 1 and 2 in the neural net.
tv_ratio
is the ratio of the data amounts between training and validation.
Predict
Model.predict(X)
Author
If you have any troubles or questions, please contact shiba24.
March, 2016