[LTR] RankLib.jar 包介绍

一、介绍

RankLib.jar 是一个学习排名（Learning to rank）算法的库，目前已经实现了如下几种算法：

MART
RankNet
RankBoost
AdaRank
Coordinate Ascent
LambdaMART
ListNet
Random Forests
Linear regression

二、jar 包

Usage: java -jar RankLib.jar <Params>

Params:

  [+] Training (+ tuning and evaluation)

        # 训练数据

        -train <file>           Training data

        # 指定排名算法

        -ranker <type>          Specify which ranking algorithm to use

                                0: MART (gradient boosted regression tree)

                                1: RankNet

                                2: RankBoost

                                3: AdaRank

                                4: Coordinate Ascent

                                6: LambdaMART

                                7: ListNet

                                8: Random Forests

                                9: Linear regression (L2 regularization)

        # 特征描述文件，列出要学习的特征，每行一个特征，默认使用所有特征

        [ -feature <file> ]     Feature description file: list features to be considered by the learner, each on a separate line

                                If not specified, all features will be used.

        #

        [ -metric2t <metric> ]  Metric to optimize on the training data. Supported: MAP, NDCG@k, DCG@k, P@k, RR@k, ERR@k (default=ERR@10)

        [ -gmax <label> ]       Highest judged relevance label. It affects the calculation of ERR (default=4, i.e. 5-point scale {0,1,2,3,4})

        [ -silent ]             Do not print progress messages (which are printed by default)

        # 是否在验证数据集上调整模型

        [ -validate <file> ]    Specify if you want to tune your system on the validation data (default=unspecified)

                                If specified, the final model will be the one that performs best on the validation data

        # 训练-验证数据集的分割比例

        [ -tvs <x \in [0..1]> ] If you don't have separate validation data, use this to set train-validation split to be (x)(1.0-x)

        # 学习模型保存到指定文件

        [ -save <model> ]       Save the model learned (default=not-save)

        # 是否要在数据上测试训练的模型

        [ -test <file> ]        Specify if you want to evaluate the trained model on this data (default=unspecified)

        # 训练-测试数据集的分割比例

        [ -tts <x \in [0..1]> ] Set train-test split to be (x)(1.0-x). -tts will override -tvs

        # 默认与 metric2t 一致

        [ -metric2T <metric> ]  Metric to evaluate on the test data (default to the same as specified for -metric2t)

        # 归一化特征向量，方法包括求和归一化，均值/标准差归一化，最大值/最小值归一化

        [ -norm <method>]       Normalize all feature vectors (default=no-normalization). Method can be:

                                sum: normalize each feature by the sum of all its values

                                zscore: normalize each feature by its mean/standard deviation

                                linear: normalize each feature by its min/max values

        # 在训练数据集上执行交叉验证

        [ -kcv <k> ]            Specify if you want to perform k-fold cross validation using the specified training data (default=NoCV)

                                -tvs can be used to further reserve a portion of the training data in each fold for validation

        # 交叉验证训练库模型的目录

        [ -kcvmd <dir> ]        Directory for models trained via cross-validation (default=not-save)

        [ -kcvmn <model> ]      Name for model learned in each fold. It will be prefix-ed with the fold-number (default=empty)

    [-] RankNet-specific parameters # 特定参数

        # 训练迭代次数

        [ -epoch <T> ]          The number of epochs to train (default=100)

        # 隐含层个数

        [ -layer <layer> ]      The number of hidden layers (default=1)

        # 每层隐含节点个数

        [ -node <node> ]        The number of hidden nodes per layer (default=10)

        # 学习率

        [ -lr <rate> ]          Learning rate (default=0.00005)

    [-] RankBoost-specific parameters # 特定参数

        # 训练迭代次数

        [ -round <T> ]          The number of rounds to train (default=300)

        # 搜索的阈值候选个数

        [ -tc <k> ]             Number of threshold candidates to search. -1 to use all feature values (default=10)

    [-] AdaRank-specific parameters # 特定参数

        # 训练迭代次数

        [ -round <T> ]          The number of rounds to train (default=500)

        #

        [ -noeq ]               Train without enqueuing too-strong features (default=unspecified)

        # 连续两轮学习之间的误差

        [ -tolerance <t> ]      Tolerance between two consecutive rounds of learning (default=0.002)

        # 一个特征可以被连续选择而不改变性能的最大次数

        [ -max <times> ]        The maximum number of times can a feature be consecutively selected without changing performance (default=5)

    [-] Coordinate Ascent-specific parameters # 特定参数

        [ -r <k> ]              The number of random restarts (default=5)

        [ -i <iteration> ]      The number of iterations to search in each dimension (default=25)

        [ -tolerance <t> ]      Performance tolerance between two solutions (default=0.001)

        [ -reg <slack> ]        Regularization parameter (default=no-regularization)

    [-] {MART, LambdaMART}-specific parameters # 特定参数

        # 树的个数

        [ -tree <t> ]           Number of trees (default=1000)

        # 一个叶子的样本个数

        [ -leaf <l> ]           Number of leaves for each tree (default=10)

        # 学习率

        [ -shrinkage <factor> ] Shrinkage, or learning rate (default=0.1)

        # 树分割时的候选特征个数

        [ -tc <k> ]             Number of threshold candidates for tree spliting. -1 to use all feature values (default=256)

        # 一个叶子最少的样本个数

        [ -mls <n> ]            Min leaf support -- minimum #samples each leaf has to contain (default=1)

        [ -estop <e> ]          Stop early when no improvement is observed on validaton data in e consecutive rounds (default=100)

    [-] ListNet-specific parameters

        [ -epoch <T> ]          The number of epochs to train (default=1500)

        [ -lr <rate> ]          Learning rate (default=0.00001)

    [-] Random Forests-specific parameters # 随机森林特定参数

        [ -bag <r> ]            Number of bags (default=300)

        # 子集采样率

        [ -srate <r> ]          Sub-sampling rate (default=1.0)

        # 特征采样率

        [ -frate <r> ]          Feature sampling rate (default=0.3)

        [ -rtype <type> ]       Ranker to bag (default=0, i.e. MART)

        # 树个数

        [ -tree <t> ]           Number of trees in each bag (default=1)

        # 每棵树的叶节点个数

        [ -leaf <l> ]           Number of leaves for each tree (default=100)

        # 学习率

        [ -shrinkage <factor> ] Shrinkage, or learning rate (default=0.1)

        # 树分割时使用的候选特征阈值个数

        [ -tc <k> ]             Number of threshold candidates for tree spliting. -1 to use all feature values (default=256)

        [ -mls <n> ]            Min leaf support -- minimum #samples each leaf has to contain (default=1)

    [-] Linear Regression-specific parameters

        [ -L2 <reg> ]           L2 regularization parameter (default=1.0E-10)

  [+] Testing previously saved models # 测试已保存的模型

        # 加载模型

        -load <model>           The model to load

                                Multiple -load can be used to specify models from multiple folds (in increasing order),

                                  in which case the test/rank data will be partitioned accordingly.

        # 测试数据

        -test <file>            Test data to evaluate the model(s) (specify either this or -rank but not both)

        # 对指定文件中的样本排序，与 -test 不能同时使用

        -rank <file>            Rank the samples in the specified file (specify either this or -test but not both)

        [ -metric2T <metric> ]  Metric to evaluate on the test data (default=ERR@10)

        [ -gmax <label> ]       Highest judged relevance label. It affects the calculation of ERR (default=4, i.e. 5-point scale {0,1,2,3,4})

        [ -score <file>]        Store ranker's score for each object being ranked (has to be used with -rank)

        # 打印单个排名列表上的性能（必须与 -test 一起使用）

        [ -idv <file> ]         Save model performance (in test metric) on individual ranked lists (has to be used with -test)

        # 特征归一化

        [ -norm ]               Normalize feature vectors (similar to -norm for training/tuning)

1. -train <file>

指定训练数据的文件，训练数据格式：

label    qid:$id    $featureid:$featurevalue    $featureid:$featurevalue ... # description

每行代表一个样本，相同查询请求的样本的 qid 相同，label 表示该样本和该查询请求的相关程度，description 描述信息，不参与训练计算。

2、-ranker <type>

指定排名算法

MART（Multiple Additive Regression Tree）多重增量回归树
GBDT（Gradient Boosting Decision Tree）梯度渐进决策树
GBRT（Gradient Boosting Regression Tree）梯度渐进回归树
TreeNet 决策树网络
RankNet
RankBoost
AdaRank
Coordinate Ascent
LambdaMART
ListNet
Random Forests
Linear regression

3、-feature <file>

指定样本的特征定义文件，格式如下：

feature1

feature2

...

# featureK(该特征不参与分析)

4、-metric2t <metric>

指定信息检索中的评价指标，包括：

MAP, NDCG@k, DCG@k, P@k, RR@k, ERR@k

5、Example

java -jar bin/RankLib.jar -train MQ2008/Fold1/train.txt -test MQ2008/Fold1/test.txt -validate MQ2008/Fold1/vali.txt -ranker 6 -metric2t NDCG@10 -metric2T ERR@10 -save mymodel.txt

命令解释 >>>

训练数据：MQ2008/Fold1/train.txt

测试数据：MQ2008/Fold1/test.txt

验证数据：MQ2008/Fold1/vali.txt

排名算法：6，LambdaMART

评估指标：NDCG，取排名前 10 个数据进行计算

测试数据评估指标：ERR，取排名前 10 个数据进行计算

保存模型：mymodel.txt

参数 -validate 是可选的，但可以更好的模型结果，对于 RankNet/MART/LambdaMART 非常重要。
-metric2t 仅应用于 list-wise 算法（AdaRank、Coordinate Ascent 和 LambdaMART）；point-wise 和 Pair-wise 算法（MART、RankNet、RankBoost）是使用自己内部的 RMSE/pair-wise loss 作为评价指标。ListNet 虽然是 list-wise 算法，但是也不用 metric2t 指定评价指标。

6、k-fold cross validation

顺序分区

java -jar bin/RankLib.jar -train MQ2008/Fold1/train.txt -ranker 4 -kcv 5 -kcvmd models/ -kcvmn ca -metric2t NDCG@10 -metric2T ERR@10

按顺序将训练数据拆分5等份，第 i 份数据作为第 i 折叠的测试数据，第 i 折叠的训练数据则是由其他折叠的数据组成。

随机分区

java -cp bin/RankLib.jar ciir.umass.edu.features.FeatureManager -input MQ2008/Fold1/train.txt -output mydata/ -shuffle

将训练数据 train.txt 重新洗牌存储在 mydata/ 目录下 train.txt.shuffled

获取每个折叠中的数据

java -cp bin/RankLib.jar ciir.umass.edu.features.FeatureManager -input MQ2008/Fold1/train.txt.shuffled -output mydata/ -k 5

7、评估已训练的模型

java -jar bin/RankLib.jar -load mymodel.txt -test MQ2008/Fold1/test.txt -metric2T ERR@10

8、模型对比

java -jar bin/RankLib.jar -test MQ2008/Fold1/test.txt -metric2T NDCG@10 -idv output/baseline.ndcg.txt

java -jar bin/RankLib.jar -load ca.model.txt -test MQ2008/Fold1/test.txt -metric2T NDCG@10 -idv output/ca.ndcg.txt

java -jar bin/RankLib.jar -load lm.model.txt -test MQ2008/Fold1/test.txt -metric2T NDCG@10 -idv output/lm.ndcg.txt

输出文件中包含了每条查询的 NDCG@10 指标值，以及所有查询的综合指标，例如：

NDCG@10   170   0.0

NDCG@10   176   0.6722390270733757

NDCG@10   177   0.4772656487866462

NDCG@10   178   0.539003131276382

NDCG@10   185   0.6131471927654585

NDCG@10   189   1.0

NDCG@10   191   0.6309297535714574

NDCG@10   192   1.0

NDCG@10   194   0.2532778777010656

NDCG@10   197   1.0

NDCG@10   200   0.6131471927654585

NDCG@10   204   0.4772656487866462

NDCG@10   207   0.0

NDCG@10   209   0.123151194370365

NDCG@10   221   0.39038004999210174

NDCG@10   all   0.5193204478059303

然后再进行对比：

java -cp RankLib.jar ciir.umass.edu.eval.Analyzer -all output/ -base baseline.ndcg.txt > analysis.txt

对比结果 analysis.txt 如下：

Overall comparison

  ------------------------------------------------------------------------

  System  Performance     Improvement     Win     Loss    p-value

  baseline_ndcg.txt [baseline]    0.093

  LM_ndcg.txt     0.2863  +0.1933 (+207.8%)       9       1       0.03

  CA_ndcg.txt     0.5193  +0.4263 (+458.26%)      12      0       0.0

  Detailed break down

  ------------------------------------------------------------------------

             [ < -100%)  [-100%,-75%)  [-75%,-50%)  [-50%,-25%)  [-25%,0%)  (0%,+25%]  (+25%,+50%]  (+50%,+75%]  (+75%,+100%]  ( > +100%]

  LM_ndcg.txt    0        0           1            0            0         4            2            2            1            0

  CA_ndcg.txt    0             0            0            0            0        1            6            2            3            0

9、利用训练模型重排名

java -jar RankLib.jar -load mymodel.txt -rank myResultLists.txt -score myScoreFile.txt

myScoreFile.txt 文件中只是增加了一列，表示重新计算的排名评分，需要自己另外根据该评分排序获取新的排名顺序。

1   0   -7.528650760650635

1   1   2.9022061824798584

1   2   -0.700125515460968

1   3   2.376657485961914

1   4   -0.29666265845298767

1   5   -2.038628101348877

1   6   -5.267711162567139

1   7   -2.022146463394165

1   8   0.6741248369216919

...

参考

RankLib wiki