Machine learning difference in the competition and industry application

There is a big difference between machine learning in competitions and industrial applications. Competitions usually focus on the ultimate evaluation indicators, while industrial applications will pay more attention to the stability, interpretability andApplication of domain expert knowledge

Contest

In order to get the ranking of the competition, the evaluation indicators are improved through various methods to the extreme

1. Data quality: the data source remains unchanged, and does not focus on data quality improvement
2. Model application: methods for using new models, complex models, and model fusion
3. Feature Engineering: Using Computationally Expensive Data Augmentation
4. Tune: Do a lot of model tuning
5. Stability: Offline model, low stability requirements
6. Domain expert knowledge: Many competitions even desensitize the original data (such as re-marking field names) to prevent the use of expert knowledge, so domain expert knowledge is less used in the competition

Industrial Applications

We usually pay more attention to the stability of the model and the continuous improvement of data quality under the condition that the application scenario is satisfied

1. Data Quality: Data is constantly changing, so focus on improving data quality
2. Model application: generally use mainstream and relatively simple models, and rarely use complex models and model fusion methods, which are helpful for model interpretability and problem debugging
3. Feature engineering: focus on engineering performance, generally do not use computationally expensive data augmentation
4. Parameter adjustment: After the hyperparameters are fixed, they will not move for a long time (usually adjusted several times a year)
5. Stability: Online real-time model deployment in the production environment requires high stability
6. Domain expert knowledge: will use expert knowledge and theoretical models for modeling