My stakeholders keep pushing for "Neural Networks" for every project, but for our current churn prediction, a simple Logistic Regression seems to perform just as well. In the field of Data Science, when is it actually worth the extra compute cost and complexity to go with Deep Learning over a standard Random Forest?
3 answers
Deep Learning is not a silver bullet. It excels primarily with unstructured data like images, audio, and natural language where "feature engineering" is too complex for humans to do manually. For structured, tabular data (like customer churn or credit scoring), traditional ensemble methods like XGBoost or LightGBM often outperform Deep Learning and are much easier to interpret. You should choose your model based on the "Occam's Razor" principle: if two models have similar performance, choose the simpler one. Simple models are faster to train, cheaper to run, and much easier to explain to a regulatory auditor or a business executive.
Since interpretability is so important for churn, have you looked into SHAP or LIME values? Even if you did use a complex model, wouldn't these tools help you explain "why" a customer is predicted to leave?
Data volume is the big decider. If you don't have hundreds of thousands of rows, Deep Learning will almost certainly overfit your data compared to a Random Forest.
Spot on, Gunther. Deep Learning is data-hungry. Without massive datasets, you're just using a very expensive sledgehammer to crack a very small nut.
We actually tried SHAP with a complex Neural Net, but the stakeholders found the explanations too abstract. With the Logistic Regression, we can give them a direct "odds ratio" for each feature—like "a 10% increase in support calls increases churn risk by 15%." That level of directness is invaluable for the Marketing team to build their retention campaigns. While Deep Learning is trendy, the transparency of traditional stats is currently winning us more "trust" with the leadership team, which is the most important metric for us right now.