[ml] PMML of machine learning model -- Overview

The application of machine learning model generally goes through two main processes ：

• 1、 Offline development
• 2、 Online deployment

The offline part is responsible for model training and model export , The online part is responsible for importing models and making predictions .

The above pictures are from ：https://zhuanlan.zhihu.com/p/30378213

One 、PMML Brief introduction

PMML（Predictive Model Markup Language,PMML)） It's a set of bases XML standard , Platform and environment independent model representation language . It is mainly through XML schema Defines and stores the core elements of an algorithm model ：

• The data dictionary ： Describe input data
• Data conversion ： Defines the way of raw data preprocessing , For example, standardization 、 Missing value processing 、 Generation of dummy variables
• Model definition ： Model type and parameters , For example, the split node of the tree model
• Model output ： The output of the model

It can be seen that , By defining PMML All core elements in , It can complete all processes of data mining , That is, when backend developers deploy , Just read the data , call PMML file , Then you can get the output data , Instead of focusing on data transformation 、 Problems such as model parameters , It accelerates the deployment efficiency of the model .

Two 、PMML Model generation and loading related class libraries

PMML The library related to model generation needs to see the offline training library we use . If we use sklearn, Then you can use sklearn2pmml This python Library to generate model files ; If the training model is xgboost, You can also use nyoka In the library xgboost_to_pmml export pmml Model file ;

load PMML The model needs target environment support PMML Loaded Libraries , If it is JAVA, You can use JPMML To load the PMML Model file ; If it is python, Use pypmml To load the model file ;

3、 ... and 、PMML Summary and reflection

1. PMML To meet the cross platform , At the expense of many platform specific optimizations , So many times we use our own algorithm library to save the model API The resulting model file , More than generated PMML The model file is much smaller . meanwhile PMML The loading speed of the file is also much slower than that of the model file in the unique format of the algorithm library .
2. PMML The loaded model is compared with the unique model of algorithm library , There will be a little deviation in the forecast , Of course, this deviation is not big . For example, a certain sample , use sklearn The decision tree model is predicted to be a category 1, But if we make this decision tree into a PMML file , And use JAVA After loading , Continue to predict the sample just now , There is a small probability that the predicted result is not a category 1.
3. For oversized models , For example, large-scale integrated learning model , such as xgboost, Random forests , perhaps tensorflow, Generated PMML It's easy to get a few files G, Even on T（ According to practical experience , The larger the training data set ,pmml The larger the model file is ）, Use this time PMML File load prediction will be very slow , At this point, it is recommended to build a proprietary environment for the model , There's no need to think about cross platform .

Reference resources ：
https://cloud.tencent.com/developer/article/1596754
https://zhuanlan.zhihu.com/p/30378213
https://zhuanlan.zhihu.com/p/458117655