Introduction

This open-source project, referred to as PTRanking (Learning to Rank in PyTorch) aims to provide scalable and extendable implementations of typical learning-to-rank methods based on PyTorch. On one hand, this project enables a uniform comparison over several benchmark datasets leading to an in-depth understanding of previous learning-to-rank methods. On the other hand, this project makes it easy to develop and incorporate newly proposed models, so as to expand the territory of techniques on learning-to-rank.

Key Features:

A number of representative learning-to-rank models, including not only the traditional optimization framework via empirical risk minimization but also the adversarial optimization framework
Supports widely used benchmark datasets. Meanwhile, random masking of the ground-truth labels with a specified ratio is also supported
Supports different metrics, such as Precision, MAP, nDCG and nERR
Highly configurable functionalities for fine-tuning hyper-parameters, e.g., grid-search over hyper-parameters of a specific model
Provides easy-to-use APIs for developing a new learning-to-rank model

How-to-Start and Learning more

Demo Scripts

To get a taste of learning-to-rank models without writing any code, you could try the following script. You just need to specify the model name, the dataset id, as well as the directories for input and output.

Jupyter Notebook example on RankNet & LambdaRank

To get familiar with the process of data loading, you could try the following script, namely, get the statistics of a dataset.

Jupyter Notebook example on getting dataset statistics

Develop a new model based on PT-Ranking

PT-Ranking offers deep neural networks as the basis to construct a scoring function based on PyTorch and can thus fully leverage the advantages of PyTorch. NeuralRanker is a class that represents a general learning-to-rank model. A key component of NeuralRanker is the neural scoring function. The configurable hyper-parameters include activation function, number of layers, number of neurons per layer, etc. All specific learning-to-rank models inherit NeuralRanker and mainly differ in the way of computing the training loss. The following figure shows the main step in developing a new learning-to-rank model based on Empirical Risk Minimization, where batch_preds and batch_stds correspond to outputs of the scoring function and ground-truth lables, respectively. We can observe that the main work is to define the surrogate loss function.

Parameter Setting

An easy-to-use parameter setting is necessary for any ML library. PT-Ranking offers a self-contained strategy. In other words, we appeals to particularly designed class objects for setting. For example, DataSetting for data loading, EvalSetting for evaluation setting and ModelParameter for a model's parameter setting.

When incorporating a newly developed model (say ModelA), it is commonly required to develop the subclass ModelAParameter by inheriting ModelParameter and customizing the functions, such as to_para_string(), default_para_dict() and grid_search(). Please refer to LambdaRankParameter as an example.

Thanks to this strategy, on one hand, we can initialize the settings for data-loading, evaluation, and models in a simple way. On the other hand, the parameter setting of a model is self-contained, and easy to customize.

To fully leverage PT-Ranking, one needs to be familiar with PyTorch.

For detailed introduction on learning-to-rank, please refer to the book: Learning to Rank for Information Retrieval.

API Document

Please refer to Learning to Rank in the Wild for details of each component (under construction).

List of implemented models & References

Optimization based on Empirical Risk Minimization

	Model
Pointwise	RankMSE
Pairwise	RankNet
Listwise	ListNet ・ ListMLE ・ RankCosine ・ LambdaRank ・ ApproxNDCG ・ WassRank ・ STListNet ・ LambdaLoss

Adversarial Optimization

	Model
Pointwise	IR_GAN_Point
Pairwise	IR_GAN_Pair
Listwise	IR_GAN_List

Tree-based Model (provided by LightGBM & XGBoost)

	Model
Listwise	LightGBMLambdaMART

References

RankNet: Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. 2005. Learning to rank using gradient descent. In Proceedings of the 22nd ICML. 89–96.
RankSVM: Joachims, Thorsten. Optimizing Search Engines Using Clickthrough Data. Proceedings of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 133–142, 2002.
LambdaRank: Christopher J.C. Burges, Robert Ragno, and Quoc Viet Le. 2006. Learning to Rank with Nonsmooth Cost Functions. In Proceedings of NIPS conference. 193–200.
ListNet: Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to Rank: From Pairwise Approach to Listwise Approach. In Proceedings of the 24th ICML. 129–136.
ListMLE: Fen Xia, Tie-Yan Liu, Jue Wang, Wensheng Zhang, and Hang Li. 2008. Listwise Approach to Learning to Rank: Theory and Algorithm. In Proceedings of the 25th ICML. 1192–1199.
RankCosine: Tao Qin, Xu-Dong Zhang, Ming-Feng Tsai, De-Sheng Wang, Tie-Yan Liu, and Hang Li. 2008. Query-level loss functions for information retrieval. Information Processing and Management 44, 2 (2008), 838–855.
AppoxNDCG: Tao Qin, Tie-Yan Liu, and Hang Li. 2010. A general approximation framework for direct optimization of information retrieval measures. Journal of Information Retrieval 13, 4 (2010), 375–397.
LambdaMART: Q. Wu, C.J.C. Burges, K. Svore and J. Gao. Adapting Boosting for Information Retrieval Measures. Journal of Information Retrieval, 2007. (We note that the implementation is provided by LightGBM)
IRGAN: Wang, Jun and Yu, Lantao and Zhang, Weinan and Gong, Yu and Xu, Yinghui and Wang, Benyou and Zhang, Peng and Zhang, Dell. IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models. Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, 515–524, 2017. (Besides the pointwise and pairiwse adversarial learning-to-rank methods introduced in the paper, we also include the listwise version in PT-Ranking)
LambdaLoss Xuanhui Wang, Cheng Li, Nadav Golbandi, Mike Bendersky and Marc Najork. The LambdaLoss Framework for Ranking Metric Optimization. Proceedings of The 27th ACM International Conference on Information and Knowledge Management (CIKM '18), 1313-1322, 2018.
WassRank: Hai-Tao Yu, Adam Jatowt, Hideo Joho, Joemon Jose, Xiao Yang and Long Chen. WassRank: Listwise Document Ranking Using Optimal Transport Theory. Proceedings of the 12th International Conference on Web Search and Data Mining (WSDM), 24-32, 2019.
Bruch, Sebastian and Han, Shuguang and Bendersky, Michael and Najork, Marc. A Stochastic Treatment of Learning to Rank Scoring Functions. Proceedings of the 13th International Conference on Web Search and Data Mining (WSDM), 61–69, 2020.

Supported Datasets and Formats

The widely used benchmark datasets listed below can be directly used once downloaded

-- MSLR-WEB10K (including MSLR-WEB10K | MSLR-WEB30K)

-- Yahoo! LETOR (including Set1 | Set2)

-- Istella (including Istella-S | Istella | Istella-X)

Some notes on the above datasets

Semi-supervised datasets (MQ2007-semi | MQ2008-semi) have the same format as that for supervised ranking setting. The only difference is that the semi-supervised datasets in this setting contain both judged and undged query-document pairs (in training set but not in validation and testing set)(The relevance label “-1” indicates the query-document pair is not judged) while the datasets for supervised ranking contain only judged query-document pair.
According to Introducing LETOR 4.0 Datasets}, queryLevelNorm version refers to that: conduct query level normalization in the way of using MIN. This data can be directly used for learning. They further provide 5-fold partitions of this version for cross fold validation. Thus there is no need to perform query_level_scale again for {MQ2007_super | MQ2008_super | MQ2007_semi | MQ2008_semi}. But for {MSLRWEB10K | MSLRWEB30K}, the query-level normalization is not conducted yet.
For Yahoo! LETOR, the query-level normalization is already done.
For Istella! LETOR, the query-level normalization is not conducted yet. We note that ISTELLA contains extremely large features, e.g., 1.79769313486e+308, we replace features of this kind with a constant 1000000.

PT-Ranking currently supports to ingest data with the LibSVM formats

LETOR datasets in LibSVM format

<ground-truth label int> qid:<query_id int> <feature_id int>:<feature_value float> ... <feature_id int>:<feature_value float>

For example:

4 qid:105 2:0.4 8:0.7 50:0.5

1 qid:105 5:0.5 30:0.7 32:0.4 48:0.53

0 qid:210 4:0.9 38:0.01 39:0.5 45:0.7

1 qid:210 1:0.2 8:0.9 31:0.93 40:0.6

The above sample dataset includes two queries, the query "105" has 2 documents, the corresponding ground-truth labels are 4 and 1, respectively.

Converting LETOR datasets into LibSVM format with a corresponding group file

This functionality is required when using the implementation of LambdaMART provided in LightGBM.

Test Setting

PyTorch (>=1.3)

Python (3)

Ubuntu 16.04 LTS

Call for Contribution

We are adding more learning-to-rank models all the time. Please submit an issue if there is something you want to have implemented and included. Meanwhile, anyone who are interested in any kinds of contributions and/or collaborations are warmly welcomed.

Citation

If you use PTRanking in your research, please use the following BibTex entry.

@misc{yu2020ptranking,
    title={PT-Ranking: A Benchmarking Platform for Neural Learning-to-Rank},
    author={Hai-Tao Yu},
    year={2020},
    eprint={2008.13368},
    archivePrefix={arXiv},
    primaryClass={cs.IR}
}

Relevant Resources

Name	Language	Deep Learning
PTRanking	Python	PyTorch
TF-Ranking	Python	TensorFlow
MatchZoo	Python	Keras / PyTorch
Shoelace	Python	Chainer
LEROT	Python	x
Rank Lib	Java	x
Propensity SVM^Rank	C	x
QuickRank	C++	x

kiminh / ptranking Goto Github PK

ptranking's Introduction