Little Bits of Artificial Intelligence

1. Want to Continuously Update the Model

• Expected to collect new data regularly
• When wanted the model to refine itself over time, without retraining from scratch
• The system works in a real-time or online learning environment

2. Uncertainty Matters to You

• When not just predictions needed, but also how confident the model is
• The decisions are risk-sensitive (e.g., medicine, finance, hiring)
• When okay with ranges or probability-based outcomes

3. Have Prior Knowledge to Use

• Already have a good idea or assumption about the relationship between variables
• Want to incorporate expert opinion or historical data
• Working in a domain with small data but strong priors

4. The Relationship is Linear (or Almost Linear)

• The relationship between the inputs and output is linear or nearly linear
• Want interpretability of coefficients
• The features are numerical or encoded as numbers

5. Want to Avoid Overfitting in Small Datasets

• The dataset is small or noisy, and regular linear regression overfits
• Want a regularized (controlled) model that doesn’t go extreme on limited data
• Want the model to say “I’m unsure” instead of guessing hard

6. Multivariate Relationships Are Present

• When have multiple factors (e.g., experience, education, city) influencing the output
• Want to model joint effects on the target variable
• When plan to weight or rank feature importance using coefficients

Avoid Bayesian Regression If:

• Need ultra-fast predictions at scale (Bayesian models can be slower)
• Have massive data and no use for uncertainty
• Don’t have priors or don’t care about interpretability