Organizations need to make predictions, classify objects, or discover patterns — tasks that require mathematical models trained on data.
The process of selecting, training, and evaluating machine learning algorithms to make predictions or discover patterns in data.
- Define the task: regression (predict number), classification (predict category), or clustering (find groups)
- Split data: training set (learn patterns), validation set (tune), test set (evaluate)
- Select algorithms: choose models appropriate for the task and data size
- Train models: fit algorithms to training data by minimizing a loss function
- Evaluate: assess performance using appropriate metrics (RMSE, accuracy, F1)
- Task-dependent: different problems require different model types
- Data-hungry: more data usually improves performance (up to a point)
- Non-convex: many models have multiple local optima during training
- Stochastic: same model with different random seeds can produce different results
- Built from: Feature Engineering — models need engineered features
- Builds into: Supervised Learning — modeling approach for labeled data
- Builds into: Unsupervised Learning — modeling approach for unlabeled data
- Related: Data Science — modeling is the core of data science
- Overfitting: model memorizes training data, fails on new data
- Underfitting: model is too simple to capture underlying patterns
- Data leakage: information from test set inadvertently used during training
- Ignoring model assumptions: linear models assume linear relationships