Running the program sample in this directory requires the use of the PaddlePaddle v0.10.0 version. If your PaddlePaddle installation version is below this requirement, follow the [instructions](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html) in the installation document to update the Paddlepaddle installation version.
Running the program sample in this directory requires the version of the PaddlePaddle is v0.10.0. If the version is below this requirement, following the [instructions](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html) in the document about installation to update your Paddlepaddle's version.
# Learning To Rank
The rank Learning Technology [1] is a machine learning method to build the ranking model,which plays an important role in the computer science scene such as information retrieval, natural language processing and data mining. The primary purpose of sort learning is to order a document that reflects the relevance of any query request for a given set of documents. In this example, using the annotated Corpus training two classical sorting models RankNet[4] and LamdaRank[6],the corresponding sorting model can be generated, and the correlation documents can be sorted by any query request.
Learning to rank[1] is a method to build the ranking model of machine learning,which plays an important role in the computer science scene such as information retrieval, natural language processing and data mining. The primary purpose of learning to rank is to order a document that reflects the relevance of any query request for a given set of documents. In this example, using the annotated Corpus training two classical ranking models RankNet[4] and LamdaRank[6],the corresponding ranking model can be generated, and the correlation documents can be sorted by any query request.
## Background Information
With the rapid growth of Internet, the rank learning has been paid more and more attention, which is one of the common tasks in machine learning. On the one hand, the manual sorting rules can not deal with the large scale of the candidate data, on the other hand can not give the appropriate weight for the candidate data of different channels, so the sort learning is widely used in daily life. Rank learning originated in the field of information retrieval and is still the core module in many information retrieval scenarios, such as search results ranking of the search engine ,candidate data ranking of the recommendation system, online ad sorting, and so on. In this case, we use the document retrieval task to illustrate the ranking learning model.
With the rapid growth of Internet,learning to rank has been paid more and more attention, which is one of the common tasks in machine learning. On the one hand, the manual ranking rules can not deal with the large scale of the candidate data, on the other hand can not give the appropriate weight for the candidate data of different channels, so it is widely used in daily life.Learning to rank originated in the field of information retrieval and is still the core parts of many information retrieval systems,such as the ranking of search results in search engine,ranking of candidate data in the recommendation system,and online advertising, and so on. In this case, we use the document retrieval task to illustrate the learning to rank model.
@@ -13,15 +13,15 @@ Figure.1 the role of ranking model in the typical application search engine of d
Assuming that there is a set of documents $S$, the document retrieval task is based on the relevance of the requests to give the order of the documents. According to the query request, the query engine will score every document according to the query request, and arrange the documents in reverse order according to the grading, and get the query results. In the training model, a query is given and the best ranking and scoring of the corresponding document is given. When predicting, the query request is given, the ranking model generates the document sort. The common ranking learning methods are divided into the following three categories.
- Pointwise Method
- Pointwise approach
The Pointwise method solves the sorting problem by approximating as the regression problem.The input single sample is the **score-document**,the correlation score of each query-Document pair is used as the real number or the sequence number,so the individual query-document pairs are uesd as a sample point (the origin of the word pointwise) to train the ranking model.When predicting,the correlation score of query-document pair is given for the specified input.
- Pairwise Method
In this case,the learning-to-rank problem can be approximated by a regression problem.The input single sample is the **score-document**,the correlation score of each query-Document pair is used as the real number or the sequence number,so the individual query-document pairs are uesd as a sample point (the origin of the word pointwise) to train the ranking model.When predicting,the correlation score of query-document pair is given for the specified input.
- Pairwise approach
The pairwise approach is to solve the ranking problem by approximating as the classification problem,the input single sample is the **label-document pair**.For multiple result documents of one query,any two documents are combined to form document pairs as input samples.Any two documents are combined to form document pairs as the input samples.That is to learn a two classifier, the input is a pair of documents A-B (the origin of Pairwise), according to whether the correlation of A is better than B,the two classifier gives the classification label 1 or 0.After classifying all the document pairs,we can get a set of partial order relations to construct the order relation of the documents.The principle of this kind of the method is to reduce the number of the reverse order document pairs in the order of the given set of document $S$,so as to achieve the goal of optimizing the sorting result.
- Listwise Method
In this case, the learning-to-rank problem is approximated by a classification problem — learning a binary classifier that can tell which document is better in a given pair of documents.The input single sample is the **label-document pair**.For multiple result documents of one query,any two documents are combined to form document pairs as input samples.Any two documents are combined to form document pairs as the input samples.That is to learn a two classifier, the input is a pair of documents A-B (the origin of Pairwise), according to whether the correlation of A is better than B,the two classifier gives the classification label 1 or 0.After classifying all the document pairs,we can get a set of partial order relations to construct the order relation of the documents.The principle of this kind of the method is to reduce the number of the reverse order document pairs in the order of the given pair of documents $S$,so as to achieve the goal of optimizing the sorting result.
- Listwise approach
The Listwise method solve the ranking problem by optimizing the sorting list directly,the single input sample is a **document arranged**. By constructing the appropriate measurement function to measure the difference between the current document sorting and the optimal sorting,then optimizes the measurement function to get the ranking model. It is difficult to optimize because many of the measurement functions have a property of discontinuity.
These algorithms try to directly optimize the value of one of the above evaluation measures, averaged over all queries in the training data.The single input sample is a **document arranged**. By constructing the appropriate measurement function to measure the difference between the current document ranking and the optimal ranking,then optimizes the evaluation measures to get the ranking model. It is difficult to optimize because many of the evaluation measures are not continuous functions with respect to ranking model's parameters.
