Jumping Straight to XGBoost? That’s Why Your Model Isn’t Trusted

If no one understands it, no one will act on it.

Read time: 2.5 minutes

Here's the uncomfortable truth: The reason your Model doesn't work is that the reason it fails is because of the model being too complicated to understand, as opposed to the performance of the model itself.

You created an XGBoost model with 250 variables. The accuracy was superb, so you present your model.

Silence..."How Can You Explain The Output?" You try to explain, but things get especially complicated.

Your audience hesitates because they believe in this model. However, a lack of understanding leads to mistrust in your model's performance.

How to Create Usable Models?

1. Build Simple to Start
Simplicity creates Trust.
• Starting with simple or linear/logistic models first.
• Initial model to set a baseline.

2. Increase Complexity when Required
Performance justifies complexity.
• Compare the results to the interpretation.
• Do not blindly pursue optimization.

3. Prioritize Explanation
Understanding results promotes use.
• Describe the factors that had a significant impact on the outcome.
• Use importance to explain the factor.

4. Decrease Number of Features
Added features do not equal value.
• Naming only impactful features.
• Try to find the right balance of features.

5. Follow with Decision
The Model should lead to some action.
• Establish a connection between model output and an action.
• Be prepared to implement what is learned.

💡Key Takeaway: 

If a model doesn't make sense to anyone, it won't be used.

👉 LIKE this if you’ve ever struggled to debug your own expression.

👉 SUBSCRIBE now to receive valuable advice about constructing models that provide real-world results.

👉 Follow Glenda Carnate for practical guidance on building models that deliver real value.

👉 COMMENT on the most complex model you have created.

👉 SHARE this post with someone who has decided to start their project using just XGBoost.

Reply

or to participate.