ST⛔P building #AI #ML models by using #data for automation that makes business rules an afterthought.
Most data science 👨🔬 teams understand business rules, and they collaborate with domain experts to make sure they choose the right data, features, algorithms, metrics, and objectives for developing machine learning models. Business rules become filtering criteria on data, features, and algorithms of their choice.
🚩However, this mix approach leads to multiple challenges in developing and operationalizing the models:
🔸Combining business rules in the training process leads to difficulty in the interpretation of #MachineLearning model and most likely overfitting issues
🔸Integration of business rules in the #ML building process requires domain experts' interpretation of the business rules, which will almost always lead to a loss in translation issues!
🔸The life cycle of business rules is vastly different from the model training cycle; therefore, changes in business rules will make model training more difficult as data and features require adjustment, and model re-fitting is needed.
🔸Operationalizing the models becomes challenging because business rules require a different management cycle than machine learning models in production
✨To put it simply, DO NOT Ⓜ️IX THEM.
🤔So what is the solution? Decision-Centric Approach, meaning using decomposition technique to:
🔹Break down the scenario into multiple decision units (i.e., sub-decisions).
🔹Select the right algorithm for each of the sub-decisions. Choose either rule-driven or machine-learning techniques based on the nature of the sub-decisions and their intent.
🔹When the #MachineLearning technique is needed, train the model for just a very specific sub-decision rather than boiling the ocean.
🔹When business rules and constraint optimization are suitable for a sub-decision, there is no need to find data to create a rigid and sub-optimal machine learning model for those sub-decisions.
🔹Additionally, domain experts can directly model the business rules so the rule-driven sub-decisions become the live specification of domain knowledge (no loss in translation issue anymore)
🎯Then, the decision model does its magic and combines all the sub-decisions to create the final decision outcome with a hybrid and smart combination of Business Rules and Machine Learning.
💣Using only data for automation is lazy! Do it properly with a hybrid and composite approach #DecisionIntelligence https://buff.ly/4arEqMr to make sure your models are accurate, can be well maintained, and create consistent and explainable outcomes.
Posted here.
Published March 26th, 2024 at 07:30 am
