ML project phases and hyperparameters

The lifecycle

Business problem, ML problem framing, data collection and preparation, feature engineering, model training and tuning, evaluation, testing and deployment, monitoring and debugging, retrain with new data. Loop until business goals are met, then keep looping anyway: requirements and data drift.

• Framing: stakeholders define value, budget, KPIs, and whether ML is even appropriate. Data scientists, engineers, and SMEs together.
• Data processing: centralize, preprocess, visualize, engineer features.
• Exploratory data analysis: graph the data. A correlation matrix shows how linked variables are and hints at which features matter.
• Deployment choices: real time, serverless, asynchronous, batch, on-prem.
• Monitoring: catch performance drops early, debug behavior, retrain on schedule or drift.

Hyperparameters

Settings external to the data, fixed before training: they define the model structure and learning process.

• Learning rate: the step size for weight updates. High converges fast but may overshoot; low is precise but slow.
• Batch size: examples per weight update. Small is stabler but slower, large is faster but less stable.
• Epochs: passes over the training set. Too few underfits, too many overfits.
• Regularization: the simplicity vs complexity balance. Increase it to fight overfitting.

Tuning is searching for the best values: grid search, random search, or SageMaker Automatic Model Tuning.

When not to use ML

For deterministic problems where the answer can simply be computed (the probability of drawing a blue card from a known deck), write normal code.

ML or an LLM gives you an approximation of something you could have had exactly.