Tf-ranking is an open source library for solving large-scale Ranking problems based on the TensorFlow platform. It is based on an underlying technique commonly known as “Rank Learning/Learning to Rank” (LTR). This technique is widely used for things like search and recommendations. Tf-ranking is the first open source deep learning library for LTR problems and has been widely adopted by the industry. Here are the main features of the library and some small examples.
Industry adoption:
- British
- Grubhub
- zhihu
- iQIYI
- .
Sorting learning
Common scenarios are as follows:
TF – Ranking architecture
Tf-ranking Main module
- Loss function: Ranking loss is the core technology of ranking learning
- Index: ranking effect should be based on ranking index as the best evaluation standard
- Pipelining: Pipelining is easy to use and can be used to build and train sequencing learning models
Loss function: can be classified as 3 methods
1) Single document method
- Consider each document separately
- Some examples include: ordered regression, classification, asymptotic gradient regression tree (GBRT)
2) Document correction
- Consider document pairs: Consider whether document A is more relevant than document B
- Some examples include: RankNet, RankSVM, RankBoost
3) Document list method
- Consider ordering the entire list
- Some examples include LambdaMART, ApproxNDCG, List{Net, MLE}
Examples of supported loss functions
Examples of supported sorting indicators
TF – Ranking assembly line
The user only needs to define the green module:
